1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/obdfilter/filter.c
41 * Author: Peter Braam <braam@clusterfs.com>
42 * Author: Andreas Dilger <adilger@clusterfs.com>
46 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
47 * (which need to get journal_lock, may block if journal full).
49 * Invariant: Call filter_start_transno() before any journal ops to avoid the
50 * same deadlock problem. We can (and want) to get rid of the
51 * transno sem in favour of the dir/inode i_mutex to avoid single
52 * threaded operation on the OST.
55 #define DEBUG_SUBSYSTEM S_FILTER
57 #ifndef AUTOCONF_INCLUDED
58 #include <linux/config.h>
60 #include <linux/module.h>
62 #include <linux/dcache.h>
63 #include <linux/init.h>
64 #include <linux/version.h>
65 #include <linux/sched.h>
66 #include <linux/mount.h>
67 #include <linux/buffer_head.h>
69 #include <obd_cksum.h>
70 #include <obd_class.h>
72 #include <lustre_dlm.h>
73 #include <lustre_fsfilt.h>
74 #include <lprocfs_status.h>
75 #include <lustre_log.h>
76 #include <libcfs/list.h>
77 #include <lustre_disk.h>
78 #include <lustre_quota.h>
79 #include <linux/slab.h>
80 #include <lustre_param.h>
81 #include <lustre/ll_fiemap.h>
83 #include "filter_internal.h"
85 static struct lvfs_callback_ops filter_lvfs_ops;
86 cfs_mem_cache_t *ll_fmd_cachep;
88 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
89 void *cb_data, int error)
91 struct obd_export *exp = cb_data;
92 LASSERT(exp->exp_obd == obd);
93 obd_transno_commit_cb(obd, transno, exp, error);
94 class_export_cb_put(exp);
97 int filter_version_get_check(struct obd_export *exp,
98 struct obd_trans_info *oti, struct inode *inode)
102 if (inode == NULL || oti == NULL)
105 curr_version = fsfilt_get_version(exp->exp_obd, inode);
106 if ((__s64)curr_version == -EOPNOTSUPP)
108 /* VBR: version is checked always because costs nothing */
109 if (oti->oti_pre_version != 0 &&
110 oti->oti_pre_version != curr_version) {
111 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112 oti->oti_pre_version, curr_version);
113 cfs_spin_lock(&exp->exp_lock);
114 exp->exp_vbr_failed = 1;
115 cfs_spin_unlock(&exp->exp_lock);
118 oti->oti_pre_version = curr_version;
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124 struct obd_trans_info *oti, int rc, int force_sync)
126 struct obd_device_target *obt = &exp->exp_obd->u.obt;
127 struct tg_export_data *ted = &exp->exp_target_data;
128 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
129 struct lsd_client_data *lcd;
132 int err, log_pri = D_RPCTRACE;
134 /* Propagate error code. */
138 if (!exp->exp_obd->obd_replayable || oti == NULL)
141 cfs_mutex_down(&ted->ted_lcd_lock);
143 /* if the export has already been disconnected, we have no last_rcvd slot,
144 * update server data with latest transno then */
146 cfs_mutex_up(&ted->ted_lcd_lock);
147 CWARN("commit transaction for disconnected client %s: rc %d\n",
148 exp->exp_client_uuid.uuid, rc);
149 err = filter_update_server_data(exp->exp_obd);
153 /* we don't allocate new transnos for replayed requests */
154 cfs_spin_lock(&obt->obt_lut->lut_translock);
155 if (oti->oti_transno == 0) {
156 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
157 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
158 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
160 last_rcvd = oti->oti_transno;
161 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
162 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
163 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
164 CERROR("Trying to overwrite bigger transno, on-disk: "
165 LPU64", new: "LPU64"\n",
166 le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
167 cfs_spin_lock(&exp->exp_lock);
168 exp->exp_vbr_failed = 1;
169 cfs_spin_unlock(&exp->exp_lock);
170 cfs_spin_unlock(&obt->obt_lut->lut_translock);
171 cfs_mutex_up(&ted->ted_lcd_lock);
175 oti->oti_transno = last_rcvd;
177 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
178 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
179 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
180 cfs_spin_unlock(&obt->obt_lut->lut_translock);
183 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
185 off = ted->ted_lr_off;
187 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
188 ted->ted_lr_idx, ted->ted_lr_off);
191 class_export_cb_get(exp); /* released when the cb is called */
193 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
199 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
200 lcd, sizeof(*lcd), &off,
201 force_sync | exp->exp_need_sync);
203 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
211 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
212 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
213 cfs_mutex_up(&ted->ted_lcd_lock);
217 void f_dput(struct dentry *dentry)
219 /* Can't go inside filter_ddelete because it can block */
220 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
221 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
222 LASSERT(atomic_read(&dentry->d_count) > 0);
227 static void init_brw_stats(struct brw_stats *brw_stats)
230 for (i = 0; i < BRW_LAST; i++)
231 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
234 static int lprocfs_init_rw_stats(struct obd_device *obd,
235 struct lprocfs_stats **stats)
239 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
240 LPROC_FILTER_LAST - 1;
241 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
245 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
246 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
247 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
248 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
249 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
254 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
255 plus the procfs overhead :( */
256 static int filter_export_stats_init(struct obd_device *obd,
257 struct obd_export *exp,
263 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
264 /* Self-export gets no proc entry */
267 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
269 /* Mask error for already created
277 struct nid_stat *tmp = exp->exp_nid_stats;
278 LASSERT(tmp != NULL);
280 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
281 if (tmp->nid_brw_stats == NULL)
282 GOTO(clean, rc = -ENOMEM);
284 init_brw_stats(tmp->nid_brw_stats);
285 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
286 0644, &filter_per_nid_stats_fops,
289 CWARN("Error adding the brw_stats file\n");
291 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
295 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
299 rc = lprocfs_nid_ldlm_stats_init(tmp);
309 /* Add client data to the FILTER. We use a bitmap to locate a free space
310 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
311 * Otherwise, we have just read the data from the last_rcvd file and
312 * we know its offset. */
313 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
316 struct obd_device_target *obt = &obd->u.obt;
317 struct tg_export_data *ted = &exp->exp_target_data;
318 struct lr_server_data *lsd = class_server_data(obd);
319 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
320 int new_client = (cl_idx == -1);
324 LASSERT(bitmap != NULL);
325 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
328 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
331 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
332 * there's no need for extra complication here
335 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
337 if (cl_idx >= LR_MAX_CLIENTS) {
338 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
342 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
348 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
349 CERROR("FILTER client %d: bit already set in bitmap!\n",
355 ted->ted_lr_idx = cl_idx;
356 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
357 cl_idx * le16_to_cpu(lsd->lsd_client_size);
358 cfs_init_mutex(&ted->ted_lcd_lock);
359 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
361 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
362 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
365 struct lvfs_run_ctxt saved;
366 loff_t off = ted->ted_lr_off;
370 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
371 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
373 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
376 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
377 /* Transaction needed to fix bug 1403 */
378 handle = fsfilt_start(obd,
379 obt->obt_rcvd_filp->f_dentry->d_inode,
380 FSFILT_OP_SETATTR, NULL);
381 if (IS_ERR(handle)) {
382 rc = PTR_ERR(handle);
383 CERROR("unable to start transaction: rc %d\n", rc);
385 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
386 exp->exp_last_request_time = cfs_time_current_sec();
387 rc = fsfilt_add_journal_cb(obd, 0, handle,
388 target_client_add_cb,
389 class_export_cb_get(exp));
391 cfs_spin_lock(&exp->exp_lock);
392 exp->exp_need_sync = 1;
393 cfs_spin_unlock(&exp->exp_lock);
395 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
397 sizeof(*ted->ted_lcd),
398 &off, rc /* sync if no cb */);
400 obt->obt_rcvd_filp->f_dentry->d_inode,
403 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
406 CERROR("error writing %s client idx %u: rc %d\n",
407 LAST_RCVD, ted->ted_lr_idx, rc);
414 static int filter_client_del(struct obd_export *exp)
416 struct tg_export_data *ted = &exp->exp_target_data;
417 struct obd_device_target *obt = &exp->exp_obd->u.obt;
418 struct lvfs_run_ctxt saved;
423 if (ted->ted_lcd == NULL)
426 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
427 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
430 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
432 off = ted->ted_lr_off;
434 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
435 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
437 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
438 * we leak a client slot that will be cleaned on the next recovery. */
440 CERROR("%s: client idx %d has med_off %lld\n",
441 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
442 GOTO(free, rc = -EINVAL);
445 /* Clear the bit _after_ zeroing out the client so we don't
446 race with filter_client_add and zero out new clients.*/
447 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
448 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
453 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454 /* Make sure the server's last_transno is up to date.
455 * This should be done before zeroing client slot so last_transno will
456 * be in server data or in client data in case of failure */
457 filter_update_server_data(exp->exp_obd);
459 cfs_mutex_down(&ted->ted_lcd_lock);
460 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
461 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
463 sizeof(*ted->ted_lcd), &off, 0);
464 cfs_mutex_up(&ted->ted_lcd_lock);
465 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
467 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
468 "zero out client %s at idx %u/%llu in %s, rc %d\n",
469 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
476 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
477 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
478 struct filter_mod_data *fmd)
480 LASSERT_SPIN_LOCKED(&fed->fed_lock);
481 if (--fmd->fmd_refcount == 0) {
482 /* XXX when we have persistent reservations and the handle
483 * is stored herein we need to drop it here. */
484 fed->fed_mod_count--;
485 cfs_list_del(&fmd->fmd_list);
486 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
490 /* drop fmd reference, free it if last ref */
491 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
493 struct filter_export_data *fed;
498 fed = &exp->exp_filter_data;
499 cfs_spin_lock(&fed->fed_lock);
500 filter_fmd_put_nolock(fed, fmd); /* caller reference */
501 cfs_spin_unlock(&fed->fed_lock);
504 /* expire entries from the end of the list if there are too many
505 * or they are too old */
506 static void filter_fmd_expire_nolock(struct filter_obd *filter,
507 struct filter_export_data *fed,
508 struct filter_mod_data *keep)
510 struct filter_mod_data *fmd, *tmp;
512 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
516 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
517 fed->fed_mod_count < filter->fo_fmd_max_num)
520 cfs_list_del_init(&fmd->fmd_list);
521 filter_fmd_put_nolock(fed, fmd); /* list reference */
525 void filter_fmd_expire(struct obd_export *exp)
527 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
528 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
529 &exp->exp_filter_data, NULL);
530 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
533 /* find specified objid, group in export fmd list.
534 * caller must hold fed_lock and take fmd reference itself */
535 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
536 struct filter_export_data *fed,
537 obd_id objid, obd_seq group)
539 struct filter_mod_data *found = NULL, *fmd;
541 LASSERT_SPIN_LOCKED(&fed->fed_lock);
543 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
544 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
546 cfs_list_del(&fmd->fmd_list);
547 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
548 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
553 filter_fmd_expire_nolock(filter, fed, found);
558 /* Find fmd based on objid and group, or return NULL if not found. */
559 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
560 obd_id objid, obd_seq group)
562 struct filter_mod_data *fmd;
564 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
565 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
566 &exp->exp_filter_data, objid, group);
568 fmd->fmd_refcount++; /* caller reference */
569 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
574 /* Find fmd based on objid and group, or create a new one if none is found.
575 * It is possible for this function to return NULL under memory pressure,
576 * or if objid = 0 is passed (which will only cause old entries to expire).
577 * Currently this is not fatal because any fmd state is transient and
578 * may also be freed when it gets sufficiently old. */
579 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
580 obd_id objid, obd_seq group)
582 struct filter_export_data *fed = &exp->exp_filter_data;
583 struct filter_mod_data *found = NULL, *fmd_new = NULL;
585 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
587 cfs_spin_lock(&fed->fed_lock);
588 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
591 cfs_list_add_tail(&fmd_new->fmd_list,
593 fmd_new->fmd_id = objid;
594 fmd_new->fmd_gr = group;
595 fmd_new->fmd_refcount++; /* list reference */
597 fed->fed_mod_count++;
599 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
603 found->fmd_refcount++; /* caller reference */
604 found->fmd_expire = jiffies +
605 exp->exp_obd->u.filter.fo_fmd_max_age;
608 cfs_spin_unlock(&fed->fed_lock);
614 /* drop fmd list reference so it will disappear when last reference is put.
615 * This isn't so critical because it would in fact only affect the one client
616 * that is doing the unlink and at worst we have an stale entry referencing
617 * an object that should never be used again. */
618 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
620 struct filter_mod_data *found = NULL;
622 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
623 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
625 cfs_list_del_init(&found->fmd_list);
626 filter_fmd_put_nolock(&exp->exp_filter_data, found);
628 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
631 #define filter_fmd_drop(exp, objid, group)
634 /* remove all entries from fmd list */
635 static void filter_fmd_cleanup(struct obd_export *exp)
637 struct filter_export_data *fed = &exp->exp_filter_data;
638 struct filter_mod_data *fmd = NULL, *tmp;
640 cfs_spin_lock(&fed->fed_lock);
641 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
642 cfs_list_del_init(&fmd->fmd_list);
643 filter_fmd_put_nolock(fed, fmd);
645 cfs_spin_unlock(&fed->fed_lock);
648 static int filter_init_export(struct obd_export *exp)
653 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
654 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
656 cfs_spin_lock(&exp->exp_lock);
657 exp->exp_connecting = 1;
658 cfs_spin_unlock(&exp->exp_lock);
660 /* self-export doesn't need client data and ldlm initialization */
661 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
662 &exp->exp_client_uuid)))
665 rc = lut_client_alloc(exp);
667 rc = ldlm_init_export(exp);
669 CERROR("%s: Can't initialize export: rc %d\n",
670 exp->exp_obd->obd_name, rc);
675 static int filter_free_server_data(struct obd_device_target *obt)
677 lut_fini(NULL, obt->obt_lut);
678 OBD_FREE_PTR(obt->obt_lut);
682 /* assumes caller is already in kernel ctxt */
683 int filter_update_server_data(struct obd_device *obd)
685 struct file *filp = obd->u.obt.obt_rcvd_filp;
686 struct lr_server_data *lsd = class_server_data(obd);
691 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
692 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
693 le64_to_cpu(lsd->lsd_last_transno));
694 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
695 le64_to_cpu(lsd->lsd_mount_count));
697 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
699 CERROR("error writing lr_server_data: rc = %d\n", rc);
704 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
707 struct filter_obd *filter = &obd->u.filter;
713 if (filter->fo_last_objid_files[group] == NULL) {
714 CERROR("Object seq "LPU64" not fully setup; not updating "
715 "last_objid\n", group);
719 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
720 obd->obd_name, filter->fo_last_objids[group], group);
722 tmp = cpu_to_le64(filter->fo_last_objids[group]);
723 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
724 &tmp, sizeof(tmp), &off, force_sync);
726 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
730 extern int ost_handle(struct ptlrpc_request *req);
731 /* assumes caller has already in kernel ctxt */
732 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
734 struct filter_obd *filter = &obd->u.filter;
735 struct lr_server_data *lsd;
736 struct lsd_client_data *lcd = NULL;
737 struct inode *inode = filp->f_dentry->d_inode;
738 unsigned long last_rcvd_size = i_size_read(inode);
739 struct lu_target *lut;
746 /* ensure padding in the struct is the correct size */
747 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
748 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
749 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
750 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
752 /* allocate and initialize lu_target */
756 rc = lut_init(NULL, lut, obd, NULL);
759 lsd = class_server_data(obd);
760 if (last_rcvd_size == 0) {
761 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
763 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
764 lsd->lsd_last_transno = 0;
765 mount_count = lsd->lsd_mount_count = 0;
766 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
767 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
768 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
769 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
770 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
771 /* OBD_COMPAT_OST is set in filter_connect_internal when the
772 * MDS first connects and assigns the OST index number. */
773 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
776 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
778 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
782 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
783 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
784 "using the wrong disk %s. Were the "
785 "/dev/ assignments rearranged?\n",
786 obd->obd_uuid.uuid, lsd->lsd_uuid);
787 GOTO(err_lut, rc = -EINVAL);
789 mount_count = le64_to_cpu(lsd->lsd_mount_count);
790 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
792 /* Assume old last_rcvd format unless I_C_LR is set */
793 if (!(lsd->lsd_feature_incompat &
794 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
795 lsd->lsd_last_transno = lsd->lsd_compat14;
797 /* OBD_COMPAT_OST is set in filter_connect_internal when the
798 * MDS first connects and assigns the OST index number. */
799 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
803 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
804 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
805 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
806 ~FILTER_INCOMPAT_SUPP);
807 GOTO(err_lut, rc = -EINVAL);
809 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
810 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
811 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
812 ~FILTER_ROCOMPAT_SUPP);
813 /* Do something like remount filesystem read-only */
814 GOTO(err_lut, rc = -EINVAL);
817 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
819 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
820 obd->obd_name, start_epoch);
821 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
822 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
823 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
824 obd->obd_name, mount_count + 1);
825 CDEBUG(D_INODE, "%s: server data size: %u\n",
826 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
827 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
828 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
829 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
830 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
831 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
832 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
833 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
834 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
835 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
836 le16_to_cpu(lsd->lsd_client_size));
838 if (!obd->obd_replayable) {
839 CWARN("%s: recovery support OFF\n", obd->obd_name);
845 GOTO(err_client, rc = -ENOMEM);
847 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
848 off < last_rcvd_size; cl_idx++) {
850 struct obd_export *exp;
851 struct filter_export_data *fed;
853 /* Don't assume off is incremented properly by
854 * fsfilt_read_record(), in case sizeof(*lcd)
855 * isn't the same as lsd->lsd_client_size. */
856 off = le32_to_cpu(lsd->lsd_client_start) +
857 cl_idx * le16_to_cpu(lsd->lsd_client_size);
858 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
860 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
861 LAST_RCVD, cl_idx, off, rc);
862 break; /* read error shouldn't cause startup to fail */
865 if (lcd->lcd_uuid[0] == '\0') {
866 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
871 check_lcd(obd->obd_name, cl_idx, lcd);
873 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
875 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
876 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
877 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
879 /* These exports are cleaned up by filter_disconnect(), so they
880 * need to be set up like real exports as filter_connect() does.
882 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
884 if (PTR_ERR(exp) == -EALREADY) {
885 /* export already exists, zero out this one */
886 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
890 GOTO(err_client, rc = PTR_ERR(exp));
893 fed = &exp->exp_filter_data;
894 *fed->fed_ted.ted_lcd = *lcd;
895 fed->fed_group = 0; /* will be assigned at connect */
896 filter_export_stats_init(obd, exp, NULL);
897 rc = filter_client_add(obd, exp, cl_idx);
898 /* can't fail for existing client */
899 LASSERTF(rc == 0, "rc = %d\n", rc);
901 /* VBR: set export last committed */
902 exp->exp_last_committed = last_rcvd;
903 cfs_spin_lock(&exp->exp_lock);
904 exp->exp_connecting = 0;
905 exp->exp_in_recovery = 0;
906 cfs_spin_unlock(&exp->exp_lock);
907 obd->obd_max_recoverable_clients++;
908 class_export_put(exp);
910 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
911 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
915 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
917 obd->u.obt.obt_mount_count = mount_count + 1;
918 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
919 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
921 /* save it, so mount count and last_transno is current */
922 rc = filter_update_server_data(obd);
924 GOTO(err_client, rc);
929 class_disconnect_exports(obd);
931 filter_free_server_data(&obd->u.obt);
935 static int filter_cleanup_groups(struct obd_device *obd)
937 struct filter_obd *filter = &obd->u.filter;
939 struct dentry *dentry;
943 if (filter->fo_dentry_O_groups != NULL) {
944 for (i = 0; i < filter->fo_group_count; i++) {
945 dentry = filter->fo_dentry_O_groups[i];
949 OBD_FREE(filter->fo_dentry_O_groups,
950 filter->fo_group_count *
951 sizeof(*filter->fo_dentry_O_groups));
952 filter->fo_dentry_O_groups = NULL;
954 if (filter->fo_last_objid_files != NULL) {
955 for (i = 0; i < filter->fo_group_count; i++) {
956 filp = filter->fo_last_objid_files[i];
960 OBD_FREE(filter->fo_last_objid_files,
961 filter->fo_group_count *
962 sizeof(*filter->fo_last_objid_files));
963 filter->fo_last_objid_files = NULL;
965 if (filter->fo_dentry_O_sub != NULL) {
966 for (i = 0; i < filter->fo_group_count; i++) {
967 for (j = 0; j < filter->fo_subdir_count; j++) {
968 dentry = filter->fo_dentry_O_sub[i].dentry[j];
973 OBD_FREE(filter->fo_dentry_O_sub,
974 filter->fo_group_count *
975 sizeof(*filter->fo_dentry_O_sub));
976 filter->fo_dentry_O_sub = NULL;
978 if (filter->fo_last_objids != NULL) {
979 OBD_FREE(filter->fo_last_objids,
980 filter->fo_group_count *
981 sizeof(*filter->fo_last_objids));
982 filter->fo_last_objids = NULL;
984 if (filter->fo_dentry_O != NULL) {
985 f_dput(filter->fo_dentry_O);
986 filter->fo_dentry_O = NULL;
991 static int filter_update_last_group(struct obd_device *obd, int group)
993 struct filter_obd *filter = &obd->u.filter;
994 struct file *filp = NULL;
995 int last_group = 0, rc;
999 if (group <= filter->fo_committed_group)
1002 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1006 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1010 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1012 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1016 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1017 obd->obd_name, last_group, group);
1021 /* must be sync: bXXXX */
1022 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1024 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1028 filter->fo_committed_group = group;
1031 filp_close(filp, 0);
1035 static int filter_read_group_internal(struct obd_device *obd, int group,
1038 struct filter_obd *filter = &obd->u.filter;
1039 __u64 *new_objids = NULL;
1040 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1041 struct dentry **new_groups = NULL;
1042 struct file **new_files = NULL;
1043 struct dentry *dentry;
1045 int old_count = filter->fo_group_count, rc, stage = 0, i;
1049 int len = group + 1;
1051 snprintf(name, 24, "%d", group);
1055 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1057 if (IS_ERR(dentry)) {
1058 CERROR("Cannot lookup expected object group %d: %ld\n",
1059 group, PTR_ERR(dentry));
1060 RETURN(PTR_ERR(dentry));
1063 dentry = simple_mkdir(filter->fo_dentry_O,
1064 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1065 if (IS_ERR(dentry)) {
1066 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1068 RETURN(PTR_ERR(dentry));
1073 snprintf(name, 24, "O/%d/LAST_ID", group);
1075 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1077 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1078 GOTO(cleanup, rc = PTR_ERR(filp));
1082 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1084 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1088 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1089 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1090 if (tmp_subdirs == NULL)
1091 GOTO(cleanup, rc = -ENOMEM);
1094 for (i = 0; i < filter->fo_subdir_count; i++) {
1096 snprintf(dir, sizeof(dir), "d%u", i);
1098 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1099 obd->u.obt.obt_vfsmnt,
1101 if (IS_ERR(tmp_subdirs->dentry[i])) {
1102 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1103 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1108 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1109 tmp_subdirs->dentry[i]);
1113 /* 'group' is an index; we need an array of length 'group + 1' */
1114 if (group + 1 > old_count) {
1115 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1116 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1117 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1118 OBD_ALLOC(new_files, len * sizeof(*new_files));
1120 if (new_objids == NULL || new_subdirs == NULL ||
1121 new_groups == NULL || new_files == NULL)
1122 GOTO(cleanup, rc = -ENOMEM);
1125 memcpy(new_objids, filter->fo_last_objids,
1126 old_count * sizeof(*new_objids));
1127 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1128 old_count * sizeof(*new_subdirs));
1129 memcpy(new_groups, filter->fo_dentry_O_groups,
1130 old_count * sizeof(*new_groups));
1131 memcpy(new_files, filter->fo_last_objid_files,
1132 old_count * sizeof(*new_files));
1134 OBD_FREE(filter->fo_last_objids,
1135 old_count * sizeof(*new_objids));
1136 OBD_FREE(filter->fo_dentry_O_sub,
1137 old_count * sizeof(*new_subdirs));
1138 OBD_FREE(filter->fo_dentry_O_groups,
1139 old_count * sizeof(*new_groups));
1140 OBD_FREE(filter->fo_last_objid_files,
1141 old_count * sizeof(*new_files));
1143 filter->fo_last_objids = new_objids;
1144 filter->fo_dentry_O_sub = new_subdirs;
1145 filter->fo_dentry_O_groups = new_groups;
1146 filter->fo_last_objid_files = new_files;
1147 filter->fo_group_count = len;
1150 filter->fo_dentry_O_groups[group] = dentry;
1151 filter->fo_last_objid_files[group] = filp;
1152 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1153 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1154 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1157 filter_update_last_group(obd, group);
1159 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1160 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1161 rc = filter_update_last_objid(obd, group, 1);
1165 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1166 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1167 obd->obd_name, group, last_objid);
1172 if (new_objids != NULL)
1173 OBD_FREE(new_objids, len * sizeof(*new_objids));
1174 if (new_subdirs != NULL)
1175 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1176 if (new_groups != NULL)
1177 OBD_FREE(new_groups, len * sizeof(*new_groups));
1178 if (new_files != NULL)
1179 OBD_FREE(new_files, len * sizeof(*new_files));
1181 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1182 for (i = 0; i < filter->fo_subdir_count; i++) {
1183 if (tmp_subdirs->dentry[i] != NULL)
1184 dput(tmp_subdirs->dentry[i]);
1186 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1189 filp_close(filp, 0);
1196 static int filter_read_groups(struct obd_device *obd, int last_group,
1199 struct filter_obd *filter = &obd->u.filter;
1200 int old_count, group, rc = 0;
1202 cfs_down(&filter->fo_init_lock);
1203 old_count = filter->fo_group_count;
1204 for (group = old_count; group <= last_group; group++) {
1205 rc = filter_read_group_internal(obd, group, create);
1209 cfs_up(&filter->fo_init_lock);
1213 /* FIXME: object groups */
1214 static int filter_prep_groups(struct obd_device *obd)
1216 struct filter_obd *filter = &obd->u.filter;
1217 struct dentry *O_dentry;
1219 int last_group, rc = 0, cleanup_phase = 0;
1223 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1225 CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1226 if (IS_ERR(O_dentry)) {
1227 rc = PTR_ERR(O_dentry);
1228 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1231 filter->fo_dentry_O = O_dentry;
1232 cleanup_phase = 1; /* O_dentry */
1234 /* we have to initialize all groups before first connections from
1235 * clients because they may send create/destroy for any group -bzzz */
1236 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1238 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1239 obd->obd_name, PTR_ERR(filp));
1240 GOTO(cleanup, rc = PTR_ERR(filp));
1242 cleanup_phase = 2; /* filp */
1244 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1246 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1252 last_group = FID_SEQ_OST_MDT0;
1254 CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1255 FID_SEQ_OST_MDT0, last_group);
1256 filter->fo_committed_group = last_group;
1257 rc = filter_read_groups(obd, last_group, 1);
1261 filp_close(filp, 0);
1265 switch (cleanup_phase) {
1267 filp_close(filp, 0);
1269 filter_cleanup_groups(obd);
1270 f_dput(filter->fo_dentry_O);
1271 filter->fo_dentry_O = NULL;
1279 /* setup the object store with correct subdirectories */
1280 static int filter_prep(struct obd_device *obd)
1282 struct lvfs_run_ctxt saved;
1283 struct filter_obd *filter = &obd->u.filter;
1285 struct inode *inode;
1289 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1290 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1291 if (!file || IS_ERR(file)) {
1293 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1297 obd->u.obt.obt_rcvd_filp = file;
1298 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1299 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1300 file->f_dentry->d_inode->i_mode);
1301 GOTO(err_filp, rc = -ENOENT);
1304 inode = file->f_dentry->d_parent->d_inode;
1305 /* We use i_op->unlink directly in filter_vfs_unlink() */
1306 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1307 CERROR("%s: filesystem does not support create/unlink ops\n",
1309 GOTO(err_filp, rc = -EOPNOTSUPP);
1312 rc = filter_init_server_data(obd, file);
1314 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1317 LASSERT(obd->u.obt.obt_lut);
1318 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1320 /* open and create health check io file*/
1321 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1324 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1326 GOTO(err_server_data, rc);
1328 filter->fo_obt.obt_health_check_filp = file;
1329 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1330 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1331 file->f_dentry->d_inode->i_mode);
1332 GOTO(err_health_check, rc = -ENOENT);
1334 rc = lvfs_check_io_health(obd, file);
1336 GOTO(err_health_check, rc);
1338 rc = filter_prep_groups(obd);
1340 GOTO(err_health_check, rc);
1342 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1347 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1348 CERROR("can't close %s after error\n", HEALTH_CHECK);
1349 filter->fo_obt.obt_health_check_filp = NULL;
1351 target_recovery_fini(obd);
1352 filter_free_server_data(&obd->u.obt);
1354 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1355 CERROR("can't close %s after error\n", LAST_RCVD);
1356 obd->u.obt.obt_rcvd_filp = NULL;
1360 /* cleanup the filter: write last used object id to status file */
1361 static void filter_post(struct obd_device *obd)
1363 struct lvfs_run_ctxt saved;
1364 struct filter_obd *filter = &obd->u.filter;
1367 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1368 * best to start a transaction with h_sync, because we removed this
1371 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1372 rc = filter_update_server_data(obd);
1374 CERROR("error writing server data: rc = %d\n", rc);
1376 for (i = 0; i < filter->fo_group_count; i++) {
1377 rc = filter_update_last_objid(obd, i,
1378 (i == filter->fo_group_count - 1));
1380 CERROR("error writing group %d lastobjid: rc = %d\n",
1384 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1385 obd->u.obt.obt_rcvd_filp = NULL;
1387 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1389 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1390 filter->fo_obt.obt_health_check_filp = NULL;
1392 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1394 filter_cleanup_groups(obd);
1395 filter_free_server_data(&obd->u.obt);
1396 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1398 filter_free_capa_keys(filter);
1399 cleanup_capa_hash(filter->fo_capa_hash);
1402 static void filter_set_last_id(struct filter_obd *filter,
1403 obd_id id, obd_seq group)
1405 LASSERT(group <= filter->fo_group_count);
1407 cfs_spin_lock(&filter->fo_objidlock);
1408 filter->fo_last_objids[group] = id;
1409 cfs_spin_unlock(&filter->fo_objidlock);
1412 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1415 LASSERT(group <= filter->fo_group_count);
1416 LASSERT(filter->fo_last_objids != NULL);
1418 /* FIXME: object groups */
1419 cfs_spin_lock(&filter->fo_objidlock);
1420 id = filter->fo_last_objids[group];
1421 cfs_spin_unlock(&filter->fo_objidlock);
1425 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1427 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1431 /* We never dget the object parent, so DON'T dput it either */
1432 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1434 struct filter_obd *filter = &obd->u.filter;
1435 struct filter_subdirs *subdirs;
1437 if (group >= filter->fo_group_count) /* FIXME: object groups */
1438 return ERR_PTR(-EBADF);
1440 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1441 return filter->fo_dentry_O_groups[group];
1443 subdirs = &filter->fo_dentry_O_sub[group];
1444 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1447 /* We never dget the object parent, so DON'T dput it either */
1448 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1451 unsigned long now = jiffies;
1452 struct dentry *dparent = filter_parent(obd, group, objid);
1455 if (IS_ERR(dparent))
1457 if (dparent == NULL)
1458 return ERR_PTR(-ENOENT);
1460 rc = filter_lock_dentry(obd, dparent);
1461 fsfilt_check_slow(obd, now, "parent lock");
1462 return rc ? ERR_PTR(rc) : dparent;
1465 /* We never dget the object parent, so DON'T dput it either */
1466 static void filter_parent_unlock(struct dentry *dparent)
1468 UNLOCK_INODE_MUTEX(dparent->d_inode);
1471 /* How to get files, dentries, inodes from object id's.
1473 * If dir_dentry is passed, the caller has already locked the parent
1474 * appropriately for this operation (normally a write lock). If
1475 * dir_dentry is NULL, we do a read lock while we do the lookup to
1476 * avoid races with create/destroy and such changing the directory
1477 * internal to the filesystem code. */
1478 struct dentry *filter_fid2dentry(struct obd_device *obd,
1479 struct dentry *dir_dentry,
1480 obd_seq group, obd_id id)
1482 struct dentry *dparent = dir_dentry;
1483 struct dentry *dchild;
1488 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1489 obd->u.filter.fo_destroys_in_progress == 0) {
1490 /* don't fail lookups for orphan recovery, it causes
1491 * later LBUGs when objects still exist during precreate */
1492 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1493 RETURN(ERR_PTR(-ENOENT));
1496 CERROR("fatal: invalid object id 0\n");
1497 RETURN(ERR_PTR(-ESTALE));
1500 len = sprintf(name, LPU64, id);
1501 if (dir_dentry == NULL) {
1502 dparent = filter_parent_lock(obd, group, id);
1503 if (IS_ERR(dparent)) {
1504 CERROR("%s: error getting object "POSTID
1505 " parent: rc %ld\n", obd->obd_name,
1506 id, group, PTR_ERR(dparent));
1510 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1511 dparent->d_name.len, dparent->d_name.name, name);
1512 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1513 dchild = lookup_one_len(name, dparent, len);
1514 if (dir_dentry == NULL)
1515 filter_parent_unlock(dparent);
1516 if (IS_ERR(dchild)) {
1517 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1518 obd->obd_name, id, group, PTR_ERR(dchild));
1522 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1523 CERROR("%s: got bad object "LPU64" inode %lu\n",
1524 obd->obd_name, id, dchild->d_inode->i_ino);
1526 RETURN(ERR_PTR(-ENOENT));
1529 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1530 name, dchild, atomic_read(&dchild->d_count));
1532 LASSERT(atomic_read(&dchild->d_count) > 0);
1537 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1538 obd_id group, struct lustre_handle *lockh)
1540 int flags = LDLM_AST_DISCARD_DATA, rc;
1541 struct ldlm_res_id res_id;
1542 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1545 osc_build_res_name(objid, group, &res_id);
1546 /* Tell the clients that the object is gone now and that they should
1547 * throw away any cached pages. */
1548 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1549 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1550 ldlm_completion_ast, NULL, NULL, 0, NULL,
1557 static void filter_fini_destroy(struct obd_device *obd,
1558 struct lustre_handle *lockh)
1560 if (lustre_handle_is_used(lockh))
1561 ldlm_lock_decref(lockh, LCK_PW);
1564 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1565 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1566 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1567 * reverse. Caller must take i_sem before starting the transaction and we
1568 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1569 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1570 struct vfsmount *mnt)
1575 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1576 * itself we already hold dir->i_mutex for child create/unlink ops */
1577 LASSERT(dentry->d_inode != NULL);
1578 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1579 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1583 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1584 GOTO(out, rc = -ENOENT);
1586 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1591 GOTO(out, rc = -EPERM);
1593 /* check_sticky() */
1594 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1595 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1596 IS_IMMUTABLE(dentry->d_inode))
1597 GOTO(out, rc = -EPERM);
1599 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1600 * that was done because of journal_start (which is already done
1601 * here) or some other ordering issue. */
1602 ll_vfs_dq_init(dir);
1604 rc = ll_security_inode_unlink(dir, dentry, mnt);
1608 rc = dir->i_op->unlink(dir, dentry);
1610 /* need to drop i_mutex before we lose inode reference */
1611 UNLOCK_INODE_MUTEX(dentry->d_inode);
1618 /* Caller must hold LCK_PW on parent and push us into kernel context.
1619 * Caller must hold child i_mutex, we drop it always.
1620 * Caller is also required to ensure that dchild->d_inode exists. */
1621 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1622 obd_seq group, struct dentry *dparent,
1623 struct dentry *dchild)
1625 struct inode *inode = dchild->d_inode;
1628 /* There should be 2 references to the inode:
1629 * 1) taken by filter_prepare_destroy
1630 * 2) taken by filter_destroy */
1631 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1632 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1633 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1634 (unsigned long)inode->i_nlink,
1635 atomic_read(&inode->i_count));
1638 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1640 CERROR("error unlinking objid %.*s: rc %d\n",
1641 dchild->d_name.len, dchild->d_name.name, rc);
1645 struct filter_intent_args {
1646 struct ldlm_lock **victim;
1651 static enum interval_iter filter_intent_cb(struct interval_node *n,
1654 struct ldlm_interval *node = (struct ldlm_interval *)n;
1655 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1656 __u64 size = arg->size;
1657 struct ldlm_lock **v = arg->victim;
1658 struct ldlm_lock *lck;
1660 /* If the interval is lower than the current file size,
1662 if (interval_high(n) <= size)
1663 return INTERVAL_ITER_STOP;
1665 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1666 /* Don't send glimpse ASTs to liblustre clients.
1667 * They aren't listening for them, and they do
1668 * entirely synchronous I/O anyways. */
1669 if (lck->l_export == NULL ||
1670 lck->l_export->exp_libclient == 1)
1673 if (*arg->liblustre)
1674 *arg->liblustre = 0;
1677 *v = LDLM_LOCK_GET(lck);
1678 } else if ((*v)->l_policy_data.l_extent.start <
1679 lck->l_policy_data.l_extent.start) {
1680 LDLM_LOCK_RELEASE(*v);
1681 *v = LDLM_LOCK_GET(lck);
1684 /* the same policy group - every lock has the
1685 * same extent, so needn't do it any more */
1689 return INTERVAL_ITER_CONT;
1692 static int filter_intent_policy(struct ldlm_namespace *ns,
1693 struct ldlm_lock **lockp, void *req_cookie,
1694 ldlm_mode_t mode, int flags, void *data)
1696 CFS_LIST_HEAD(rpc_list);
1697 struct ptlrpc_request *req = req_cookie;
1698 struct ldlm_lock *lock = *lockp, *l = NULL;
1699 struct ldlm_resource *res = lock->l_resource;
1700 ldlm_processing_policy policy;
1701 struct ost_lvb *res_lvb, *reply_lvb;
1702 struct ldlm_reply *rep;
1704 int idx, rc, tmpflags = 0, only_liblustre = 1;
1705 struct ldlm_interval_tree *tree;
1706 struct filter_intent_args arg;
1707 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1708 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1709 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1712 policy = ldlm_get_processing_policy(res);
1713 LASSERT(policy != NULL);
1714 LASSERT(req != NULL);
1716 rc = lustre_pack_reply(req, 3, repsize, NULL);
1718 RETURN(req->rq_status = rc);
1720 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1721 LASSERT(rep != NULL);
1723 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1724 sizeof(*reply_lvb));
1725 LASSERT(reply_lvb != NULL);
1727 //fixup_handle_for_resent_req(req, lock, &lockh);
1729 /* Call the extent policy function to see if our request can be
1730 * granted, or is blocked.
1731 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1732 * lock, and should not be granted if the lock will be blocked.
1735 LASSERT(ns == ldlm_res_to_ns(res));
1737 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1738 check_res_locked(res);
1740 /* FIXME: we should change the policy function slightly, to not make
1741 * this list at all, since we just turn around and free it */
1742 while (!cfs_list_empty(&rpc_list)) {
1743 struct ldlm_lock *wlock =
1744 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1746 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1747 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1748 lock->l_flags &= ~LDLM_FL_CP_REQD;
1749 cfs_list_del_init(&wlock->l_cp_ast);
1750 LDLM_LOCK_RELEASE(wlock);
1753 /* The lock met with no resistance; we're finished. */
1754 if (rc == LDLM_ITER_CONTINUE) {
1755 /* do not grant locks to the liblustre clients: they cannot
1756 * handle ASTs robustly. We need to do this while still
1757 * holding lr_lock to avoid the lock remaining on the res_link
1758 * list (and potentially being added to l_pending_list by an
1759 * AST) when we are going to drop this lock ASAP. */
1760 if (lock->l_export->exp_libclient ||
1761 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1762 ldlm_resource_unlink_lock(lock);
1763 err = ELDLM_LOCK_ABORTED;
1765 err = ELDLM_LOCK_REPLACED;
1771 /* Do not grant any lock, but instead send GL callbacks. The extent
1772 * policy nicely created a list of all PW locks for us. We will choose
1773 * the highest of those which are larger than the size in the LVB, if
1774 * any, and perform a glimpse callback. */
1775 res_lvb = res->lr_lvb_data;
1776 LASSERT(res_lvb != NULL);
1777 *reply_lvb = *res_lvb;
1780 * lr_lock guarantees that no new locks are granted, and,
1781 * therefore, that res->lr_lvb_data cannot increase beyond the
1782 * end of already granted lock. As a result, it is safe to
1783 * check against "stale" reply_lvb->lvb_size value without
1786 arg.size = reply_lvb->lvb_size;
1788 arg.liblustre = &only_liblustre;
1789 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1790 tree = &res->lr_itree[idx];
1791 if (tree->lit_mode == LCK_PR)
1794 interval_iterate_reverse(tree->lit_root,
1795 filter_intent_cb, &arg);
1799 /* There were no PW locks beyond the size in the LVB; finished. */
1801 if (only_liblustre) {
1802 /* If we discovered a liblustre client with a PW lock,
1803 * however, the LVB may be out of date! The LVB is
1804 * updated only on glimpse (which we don't do for
1805 * liblustre clients) and cancel (which the client
1806 * obviously has not yet done). So if it has written
1807 * data but kept the lock, the LVB is stale and needs
1808 * to be updated from disk.
1810 * Of course, this will all disappear when we switch to
1811 * taking liblustre locks on the OST. */
1812 ldlm_res_lvbo_update(res, NULL, 1);
1814 RETURN(ELDLM_LOCK_ABORTED);
1818 * This check is for lock taken in filter_prepare_destroy() that does
1819 * not have l_glimpse_ast set. So the logic is: if there is a lock
1820 * with no l_glimpse_ast set, this object is being destroyed already.
1822 * Hence, if you are grabbing DLM locks on the server, always set
1823 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1825 if (l->l_glimpse_ast == NULL) {
1826 /* We are racing with unlink(); just return -ENOENT */
1827 rep->lock_policy_res1 = -ENOENT;
1831 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1832 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1835 *reply_lvb = *res_lvb;
1839 LDLM_LOCK_RELEASE(l);
1841 RETURN(ELDLM_LOCK_ABORTED);
1845 * per-obd_device iobuf pool.
1847 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1848 * allocations in write-path has to be minimized (see bug 5137).
1850 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1851 * OST threads (see ost_thread_{init,done}()).
1853 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1854 * at the OST layer there are only (potentially) multiple obd_device of type
1855 * unknown at the time of OST thread creation.
1857 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1858 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1859 * it's very own iobuf.
1863 * filter_kiobuf_pool_init()
1865 * filter_kiobuf_pool_done()
1867 * filter_iobuf_get()
1869 * operate on this array. They are "generic" in a sense that they don't depend
1870 * on actual type of iobuf's (the latter depending on Linux kernel version).
1874 * destroy pool created by filter_iobuf_pool_init
1876 static void filter_iobuf_pool_done(struct filter_obd *filter)
1878 struct filter_iobuf **pool;
1883 pool = filter->fo_iobuf_pool;
1885 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1886 if (pool[i] != NULL)
1887 filter_free_iobuf(pool[i]);
1889 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1890 filter->fo_iobuf_pool = NULL;
1895 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1897 struct filter_obd *filter = &obd->u.filter;
1898 struct sptlrpc_rule_set tmp_rset;
1901 sptlrpc_rule_set_init(&tmp_rset);
1902 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1904 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1909 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1911 cfs_write_lock(&filter->fo_sptlrpc_lock);
1912 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1913 filter->fo_sptlrpc_rset = tmp_rset;
1914 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1920 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1922 static int filter_iobuf_pool_init(struct filter_obd *filter)
1929 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1931 if (filter->fo_iobuf_pool == NULL)
1934 filter->fo_iobuf_count = OSS_THREADS_MAX;
1939 /* Return iobuf allocated for @thread_id. We don't know in advance how
1940 * many threads there will be so we allocate a large empty array and only
1941 * fill in those slots that are actually in use.
1942 * If we haven't allocated a pool entry for this thread before, do so now. */
1943 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1945 int thread_id = (oti && oti->oti_thread) ?
1946 oti->oti_thread->t_id : -1;
1947 struct filter_iobuf *pool = NULL;
1948 struct filter_iobuf **pool_place = NULL;
1950 if (thread_id >= 0) {
1951 LASSERT(thread_id < filter->fo_iobuf_count);
1952 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1955 if (unlikely(pool == NULL)) {
1956 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1957 PTLRPC_MAX_BRW_PAGES);
1958 if (pool_place != NULL)
1965 /* mount the file system (secretly). lustre_cfg parameters are:
1968 * 3 = flags: failover=f, failout=n
1971 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1974 struct filter_obd *filter = &obd->u.filter;
1975 struct vfsmount *mnt;
1976 struct lustre_mount_info *lmi;
1977 struct obd_uuid uuid;
1981 struct request_queue *q;
1985 if (lcfg->lcfg_bufcount < 3 ||
1986 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1987 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1990 lmi = server_get_mount(obd->obd_name);
1992 /* We already mounted in lustre_fill_super.
1993 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1994 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1996 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1998 /* old path - used by lctl */
1999 CERROR("Using old MDS mount method\n");
2000 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2001 MS_NOATIME|MS_NODIRATIME,
2002 lustre_cfg_string(lcfg, 1), option);
2005 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2006 lustre_cfg_string(lcfg, 1), rc);
2010 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2012 if (IS_ERR(obd->obd_fsops))
2013 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2015 rc = filter_iobuf_pool_init(filter);
2019 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2020 CERROR("%s: Underlying device is marked as read-only. "
2021 "Setup failed\n", obd->obd_name);
2022 GOTO(err_ops, rc = -EROFS);
2025 /* failover is the default */
2026 obd->obd_replayable = 1;
2028 /* disable connection until configuration finishes */
2029 obd->obd_no_conn = 1;
2031 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2032 str = lustre_cfg_string(lcfg, 3);
2033 if (strchr(str, 'n')) {
2034 CWARN("%s: recovery disabled\n", obd->obd_name);
2035 obd->obd_replayable = 0;
2039 obd->u.obt.obt_magic = OBT_MAGIC;
2040 obd->u.obt.obt_vfsmnt = mnt;
2041 obd->u.obt.obt_sb = mnt->mnt_sb;
2042 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2043 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2045 fsfilt_setup(obd, obd->u.obt.obt_sb);
2047 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2048 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2049 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2050 obd->obd_lvfs_ctxt.fs = get_ds();
2051 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2053 cfs_init_mutex(&filter->fo_init_lock);
2054 filter->fo_committed_group = 0;
2055 filter->fo_destroys_in_progress = 0;
2056 for (i = 0; i < 32; i++)
2057 cfs_sema_init(&filter->fo_create_locks[i], 1);
2059 cfs_spin_lock_init(&filter->fo_objidlock);
2060 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2061 cfs_sema_init(&filter->fo_alloc_lock, 1);
2062 init_brw_stats(&filter->fo_filter_stats);
2063 cfs_spin_lock_init(&filter->fo_flags_lock);
2064 filter->fo_read_cache = 1; /* enable read-only cache by default */
2065 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2066 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2067 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2068 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2069 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2070 filter_slc_set(filter); /* initialize sync on lock cancel */
2072 rc = filter_prep(obd);
2076 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2077 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2079 filter->fo_fl_oss_capa = 1;
2081 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2082 filter->fo_capa_hash = init_capa_hash();
2083 if (filter->fo_capa_hash == NULL)
2084 GOTO(err_post, rc = -ENOMEM);
2086 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2087 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2088 LDLM_NAMESPACE_SERVER,
2089 LDLM_NAMESPACE_GREEDY,
2091 if (obd->obd_namespace == NULL)
2092 GOTO(err_post, rc = -ENOMEM);
2093 obd->obd_namespace->ns_lvbp = obd;
2094 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2095 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2097 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2098 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2100 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2102 CERROR("failed to setup llogging subsystems\n");
2106 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2107 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2108 /* do this after llog being initialized */
2109 filter_adapt_sptlrpc_conf(obd, 1);
2111 rc = lquota_setup(filter_quota_interface_ref, obd);
2115 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2116 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2117 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2118 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2119 "for larger I/O requests: max_sectors = %u "
2120 "could be up to max_hw_sectors=%u\n",
2121 obd->obd_name, mnt->mnt_sb->s_id,
2122 queue_max_sectors(q), queue_max_hw_sectors(q));
2124 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2125 if (uuid_ptr != NULL) {
2126 class_uuid_unparse(uuid_ptr, &uuid);
2132 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2133 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2134 obd->obd_name, label ?: str, lmi ? "on " : "",
2135 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2136 obd->obd_replayable ? "enabled" : "disabled");
2143 fsfilt_put_ops(obd->obd_fsops);
2144 filter_iobuf_pool_done(filter);
2146 server_put_mount(obd->obd_name, mnt);
2147 obd->u.obt.obt_sb = 0;
2151 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2153 struct lprocfs_static_vars lvars;
2154 cfs_proc_dir_entry_t *entry;
2160 CLASSERT(offsetof(struct obd_device, u.obt) ==
2161 offsetof(struct obd_device, u.filter.fo_obt));
2163 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2166 /* lprocfs must be setup before the filter so state can be safely added
2167 * to /proc incrementally as the filter is setup */
2168 lprocfs_filter_init_vars(&lvars);
2169 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2171 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2176 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2178 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2180 GOTO(obd_cleanup, rc);
2183 /* Init obdfilter private stats here */
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2185 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2187 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2188 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2189 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2190 "get_page", "usec");
2191 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2192 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2194 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2196 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2197 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2198 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2200 rc = lproc_filter_attach_seqstat(obd);
2202 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2203 GOTO(free_obd_stats, rc);
2206 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2207 if (IS_ERR(entry)) {
2208 rc = PTR_ERR(entry);
2209 CERROR("%s: error %d setting up lprocfs for %s\n",
2210 obd->obd_name, rc, "exports");
2211 GOTO(free_obd_stats, rc);
2213 obd->obd_proc_exports_entry = entry;
2215 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2216 lprocfs_nid_stats_clear_read,
2217 lprocfs_nid_stats_clear_write, obd, NULL);
2218 if (IS_ERR(entry)) {
2219 rc = PTR_ERR(entry);
2220 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2222 GOTO(free_obd_stats, rc);
2225 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2226 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2228 GOTO(remove_entry_clear, rc = -ENOMEM);
2229 addr = (unsigned long)cfs_page_address(page);
2230 clear_page((void *)addr);
2231 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2232 LUSTRE_CFG_BUFLEN(lcfg, 4));
2233 rc = filter_common_setup(obd, lcfg, (void *)addr);
2234 OBD_PAGE_FREE(page);
2236 CERROR("%s: filter_common_setup failed: %d.\n",
2238 GOTO(remove_entry_clear, rc);
2244 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2246 lprocfs_free_obd_stats(obd);
2248 lprocfs_obd_cleanup(obd);
2252 static struct llog_operations filter_mds_ost_repl_logops;
2254 static struct llog_operations filter_size_orig_logops = {
2255 .lop_setup = llog_obd_origin_setup,
2256 .lop_cleanup = llog_obd_origin_cleanup,
2257 .lop_add = llog_obd_origin_add
2260 static int filter_olg_fini(struct obd_llog_group *olg)
2262 struct llog_ctxt *ctxt;
2263 int rc = 0, rc2 = 0;
2266 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2268 rc = llog_cleanup(ctxt);
2270 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2272 rc2 = llog_cleanup(ctxt);
2277 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2279 rc2 = llog_cleanup(ctxt);
2288 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2289 struct obd_device *tgt)
2294 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2295 &filter_mds_ost_repl_logops);
2299 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2300 &filter_size_orig_logops);
2306 filter_olg_fini(olg);
2311 * Init the default olg, which is embeded in the obd_device, for filter.
2314 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2315 struct obd_device *tgt)
2317 struct filter_obd *filter = &obd->u.filter;
2318 struct llog_ctxt *ctxt;
2322 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2323 if (!filter->fo_lcm)
2326 filter_mds_ost_repl_logops = llog_client_ops;
2327 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2328 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2329 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2331 rc = filter_olg_init(obd, olg, tgt);
2333 GOTO(cleanup_lcm, rc);
2335 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2338 GOTO(cleanup_olg, rc);
2340 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2342 CERROR("Can't get ctxt for %p:%x\n", olg,
2343 LLOG_MDS_OST_REPL_CTXT);
2344 GOTO(cleanup_olg, rc = -ENODEV);
2346 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2347 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2348 llog_ctxt_put(ctxt);
2352 filter_olg_fini(olg);
2354 llog_recov_thread_fini(filter->fo_lcm, 1);
2355 filter->fo_lcm = NULL;
2360 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2361 struct obd_device *tgt, int *index)
2363 struct filter_obd *filter = &obd->u.filter;
2364 struct llog_ctxt *ctxt;
2368 LASSERT(olg != NULL);
2369 if (olg == &obd->obd_olg)
2370 return filter_default_olg_init(obd, olg, tgt);
2372 LASSERT(filter->fo_lcm != NULL);
2373 rc = filter_olg_init(obd, olg, tgt);
2376 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2378 CERROR("Can't get ctxt for %p:%x\n", olg,
2379 LLOG_MDS_OST_REPL_CTXT);
2380 filter_olg_fini(olg);
2383 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2384 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2385 llog_ctxt_put(ctxt);
2389 static int filter_llog_finish(struct obd_device *obd, int count)
2391 struct filter_obd *filter = &obd->u.filter;
2392 struct llog_ctxt *ctxt;
2395 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2398 * Make sure that no cached llcds left in recov_thread.
2399 * We actually do sync in disconnect time, but disconnect
2400 * may not come being marked rq_no_resend = 1.
2402 llog_sync(ctxt, NULL);
2405 * Balance class_import_get() in llog_receptor_accept().
2406 * This is safe to do, as llog is already synchronized
2407 * and its import may go.
2409 cfs_mutex_down(&ctxt->loc_sem);
2410 if (ctxt->loc_imp) {
2411 class_import_put(ctxt->loc_imp);
2412 ctxt->loc_imp = NULL;
2414 cfs_mutex_up(&ctxt->loc_sem);
2415 llog_ctxt_put(ctxt);
2418 if (filter->fo_lcm) {
2419 cfs_mutex_down(&ctxt->loc_sem);
2420 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2421 filter->fo_lcm = NULL;
2422 cfs_mutex_up(&ctxt->loc_sem);
2424 RETURN(filter_olg_fini(&obd->obd_olg));
2428 * Find the group llog according to group index in the llog group list.
2430 static struct obd_llog_group *
2431 filter_find_olg_internal(struct filter_obd *filter, int group)
2433 struct obd_llog_group *olg;
2435 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2436 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2437 if (olg->olg_seq == group)
2444 * Find the group llog according to group index on the filter
2446 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2448 struct obd_llog_group *olg = NULL;
2449 struct filter_obd *filter;
2451 filter = &obd->u.filter;
2453 if (group == FID_SEQ_LLOG)
2454 RETURN(&obd->obd_olg);
2456 cfs_spin_lock(&filter->fo_llog_list_lock);
2457 olg = filter_find_olg_internal(filter, group);
2458 cfs_spin_unlock(&filter->fo_llog_list_lock);
2463 * Find the llog_group of the filter according to the group. If it can not
2464 * find, create the llog_group, which only happens when mds is being synced
2467 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2469 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2470 struct filter_obd *filter;
2473 filter = &obd->u.filter;
2475 if (group == FID_SEQ_LLOG)
2476 RETURN(&obd->obd_olg);
2478 OBD_ALLOC_PTR(olg_new);
2479 if (olg_new == NULL)
2480 RETURN(ERR_PTR(-ENOMEM));
2482 cfs_spin_lock(&filter->fo_llog_list_lock);
2483 olg = filter_find_olg_internal(filter, group);
2485 if (olg->olg_initializing) {
2486 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2488 GOTO(out_unlock, olg);
2491 /* set as the newly allocated one */
2496 llog_group_init(olg, group);
2497 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2498 olg->olg_initializing = 1;
2499 cfs_spin_unlock(&filter->fo_llog_list_lock);
2501 rc = obd_llog_init(obd, olg, obd, NULL);
2503 cfs_spin_lock(&filter->fo_llog_list_lock);
2504 cfs_list_del(&olg->olg_list);
2505 cfs_spin_unlock(&filter->fo_llog_list_lock);
2507 GOTO(out, olg = ERR_PTR(-ENOMEM));
2509 cfs_spin_lock(&filter->fo_llog_list_lock);
2510 olg->olg_initializing = 0;
2511 cfs_spin_unlock(&filter->fo_llog_list_lock);
2512 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2513 obd->obd_name, group, olg);
2518 cfs_spin_unlock(&filter->fo_llog_list_lock);
2520 OBD_FREE_PTR(olg_new);
2524 static int filter_llog_connect(struct obd_export *exp,
2525 struct llogd_conn_body *body)
2527 struct obd_device *obd = exp->exp_obd;
2528 struct llog_ctxt *ctxt;
2529 struct obd_llog_group *olg;
2533 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2534 obd->obd_name, body->lgdc_logid.lgl_oid,
2535 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2537 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2539 CERROR(" %s: can not find olg of group %d\n",
2540 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2543 llog_group_set_export(olg, exp);
2545 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2546 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2547 body->lgdc_ctxt_idx);
2549 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2550 obd->obd_name, body->lgdc_logid.lgl_oid,
2551 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2553 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2554 obd->u.filter.fo_mds_ost_sync = 1;
2555 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2556 rc = llog_connect(ctxt, &body->lgdc_logid,
2557 &body->lgdc_gen, NULL);
2558 llog_ctxt_put(ctxt);
2560 CERROR("failed to connect rc %d idx %d\n", rc,
2561 body->lgdc_ctxt_idx);
2566 static int filter_llog_preclean(struct obd_device *obd)
2568 struct obd_llog_group *olg, *tmp;
2569 struct filter_obd *filter;
2570 cfs_list_t remove_list;
2574 rc = obd_llog_finish(obd, 0);
2576 CERROR("failed to cleanup llogging subsystem\n");
2578 filter = &obd->u.filter;
2579 CFS_INIT_LIST_HEAD(&remove_list);
2581 cfs_spin_lock(&filter->fo_llog_list_lock);
2582 while (!cfs_list_empty(&filter->fo_llog_list)) {
2583 olg = cfs_list_entry(filter->fo_llog_list.next,
2584 struct obd_llog_group, olg_list);
2585 cfs_list_del(&olg->olg_list);
2586 cfs_list_add(&olg->olg_list, &remove_list);
2588 cfs_spin_unlock(&filter->fo_llog_list_lock);
2590 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2591 cfs_list_del_init(&olg->olg_list);
2592 rc = filter_olg_fini(olg);
2594 CERROR("failed to cleanup llogging subsystem for %u\n",
2602 static int filter_precleanup(struct obd_device *obd,
2603 enum obd_cleanup_stage stage)
2609 case OBD_CLEANUP_EARLY:
2611 case OBD_CLEANUP_EXPORTS:
2612 /* Stop recovery before namespace cleanup. */
2613 target_recovery_fini(obd);
2614 rc = filter_llog_preclean(obd);
2620 static int filter_cleanup(struct obd_device *obd)
2622 struct filter_obd *filter = &obd->u.filter;
2626 LCONSOLE_WARN("%s: shutting down for failover; client state "
2627 "will be preserved.\n", obd->obd_name);
2629 obd_exports_barrier(obd);
2630 obd_zombie_barrier();
2632 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2633 lprocfs_free_per_client_stats(obd);
2634 lprocfs_free_obd_stats(obd);
2635 lprocfs_obd_cleanup(obd);
2636 lquota_cleanup(filter_quota_interface_ref, obd);
2638 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2639 obd->obd_namespace = NULL;
2641 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2643 if (obd->u.obt.obt_sb == NULL)
2648 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2649 shrink_dcache_sb(obd->u.obt.obt_sb);
2651 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2652 obd->u.obt.obt_sb = NULL;
2654 fsfilt_put_ops(obd->obd_fsops);
2656 filter_iobuf_pool_done(filter);
2658 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2663 static int filter_connect_internal(struct obd_export *exp,
2664 struct obd_connect_data *data,
2667 struct filter_export_data *fed = &exp->exp_filter_data;
2672 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2673 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2674 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2675 data->ocd_connect_flags, data->ocd_version,
2676 data->ocd_grant, data->ocd_index);
2678 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2679 CWARN("!!! This export (nid %s) used object group %d "
2680 "earlier; now it's trying to use group %d! This could "
2681 "be a bug in the MDS. Please report to "
2682 "http://bugs.whamcloud.com/\n",
2683 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2686 fed->fed_group = data->ocd_group;
2688 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2689 exp->exp_connect_flags = data->ocd_connect_flags;
2690 data->ocd_version = LUSTRE_VERSION_CODE;
2692 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2693 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2694 CWARN("%s: Received MDS connection for group %u\n",
2695 exp->exp_obd->obd_name, data->ocd_group);
2696 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
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 data->ocd_cksum_types &= cksum_types_supported();
2772 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2773 if (unlikely(data->ocd_cksum_types == 0))
2774 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2776 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2777 "%x\n", exp->exp_obd->obd_name,
2778 obd_export_nid2str(exp), cksum_types,
2779 data->ocd_cksum_types);
2781 /* This client does not support OBD_CONNECT_CKSUM
2782 * fall back to CRC32 */
2783 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2784 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2785 exp->exp_obd->obd_name,
2786 obd_export_nid2str(exp));
2789 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2790 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2795 static int filter_reconnect(const struct lu_env *env,
2796 struct obd_export *exp, struct obd_device *obd,
2797 struct obd_uuid *cluuid,
2798 struct obd_connect_data *data,
2804 if (exp == NULL || obd == NULL || cluuid == NULL)
2807 rc = filter_connect_internal(exp, data, 1);
2809 filter_export_stats_init(obd, exp, localdata);
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;
2825 if (exp == NULL || obd == NULL || cluuid == NULL)
2828 rc = class_connect(&conn, obd, cluuid);
2831 lexp = class_conn2export(&conn);
2832 LASSERT(lexp != NULL);
2834 rc = filter_connect_internal(lexp, data, 0);
2838 filter_export_stats_init(obd, lexp, localdata);
2839 if (obd->obd_replayable) {
2840 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2842 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2843 rc = filter_client_add(obd, lexp, -1);
2848 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2849 rc = filter_read_groups(obd, data->ocd_group, 1);
2850 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2852 CERROR("can't read group %u\n", data->ocd_group);
2860 class_disconnect(lexp);
2869 /* Do extra sanity checks for grant accounting. We do this at connect,
2870 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2871 * always get rid of it or turn it off when we know accounting is good. */
2872 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2874 struct filter_export_data *fed;
2875 struct obd_export *exp;
2876 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2877 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2878 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2880 if (cfs_list_empty(&obd->obd_exports))
2883 /* We don't want to do this for large machines that do lots of
2884 mounts or unmounts. It burns... */
2885 if (obd->obd_num_exports > 100)
2888 cfs_spin_lock(&obd->obd_osfs_lock);
2889 cfs_spin_lock(&obd->obd_dev_lock);
2890 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2892 fed = &exp->exp_filter_data;
2893 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2896 if (maxsize > 0) { /* we may not have done a statfs yet */
2897 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2898 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2899 exp->exp_client_uuid.uuid, exp,
2900 fed->fed_grant, fed->fed_pending, maxsize);
2901 LASSERTF(fed->fed_dirty <= maxsize,
2902 "%s: cli %s/%p %ld > "LPU64"\n", func,
2903 exp->exp_client_uuid.uuid, exp,
2904 fed->fed_dirty, maxsize);
2907 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2908 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2909 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2911 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2912 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2913 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2914 tot_granted += fed->fed_grant + fed->fed_pending;
2915 tot_pending += fed->fed_pending;
2916 tot_dirty += fed->fed_dirty;
2918 fo_tot_granted = obd->u.filter.fo_tot_granted;
2919 fo_tot_pending = obd->u.filter.fo_tot_pending;
2920 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2921 cfs_spin_unlock(&obd->obd_dev_lock);
2922 cfs_spin_unlock(&obd->obd_osfs_lock);
2924 /* Do these assertions outside the spinlocks so we don't kill system */
2925 if (tot_granted != fo_tot_granted)
2926 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2927 func, tot_granted, fo_tot_granted);
2928 if (tot_pending != fo_tot_pending)
2929 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2930 func, tot_pending, fo_tot_pending);
2931 if (tot_dirty != fo_tot_dirty)
2932 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2933 func, tot_dirty, fo_tot_dirty);
2934 if (tot_pending > tot_granted)
2935 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2936 func, tot_pending, tot_granted);
2937 if (tot_granted > maxsize)
2938 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2939 func, tot_granted, maxsize);
2940 if (tot_dirty > maxsize)
2941 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2942 func, tot_dirty, maxsize);
2945 /* Remove this client from the grant accounting totals. We also remove
2946 * the export from the obd device under the osfs and dev locks to ensure
2947 * that the filter_grant_sanity_check() calculations are always valid.
2948 * The client should do something similar when it invalidates its import. */
2949 static void filter_grant_discard(struct obd_export *exp)
2951 struct obd_device *obd = exp->exp_obd;
2952 struct filter_obd *filter = &obd->u.filter;
2953 struct filter_export_data *fed = &exp->exp_filter_data;
2955 cfs_spin_lock(&obd->obd_osfs_lock);
2956 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2957 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2958 obd->obd_name, filter->fo_tot_granted,
2959 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2960 filter->fo_tot_granted -= fed->fed_grant;
2961 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2962 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2963 obd->obd_name, filter->fo_tot_pending,
2964 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2965 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2966 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2967 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2968 obd->obd_name, filter->fo_tot_dirty,
2969 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2970 filter->fo_tot_dirty -= fed->fed_dirty;
2974 cfs_spin_unlock(&obd->obd_osfs_lock);
2977 static int filter_destroy_export(struct obd_export *exp)
2979 struct filter_export_data *fed = &exp->exp_filter_data;
2982 if (fed->fed_pending)
2983 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2984 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2985 exp, fed->fed_pending);
2987 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2989 target_destroy_export(exp);
2991 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
2992 &exp->exp_client_uuid)))
2995 ldlm_destroy_export(exp);
2996 lut_client_free(exp);
2998 if (!exp->exp_obd->obd_replayable)
2999 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3001 filter_grant_discard(exp);
3002 filter_fmd_cleanup(exp);
3004 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3005 struct filter_obd *filter = &exp->exp_obd->u.filter;
3006 if (filter->fo_tot_granted_clients > 0)
3007 filter->fo_tot_granted_clients --;
3010 if (!(exp->exp_flags & OBD_OPT_FORCE))
3011 filter_grant_sanity_check(exp->exp_obd, __func__);
3016 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3018 struct obd_llog_group *olg_min, *olg;
3019 struct filter_obd *filter;
3020 int worked = -1, group;
3021 struct llog_ctxt *ctxt;
3024 filter = &obd->u.filter;
3026 /* we can't sync log holding spinlock. also, we do not want to get
3027 * into livelock. so we do following: loop over MDS's exports in
3028 * group order and skip already synced llogs -bzzz */
3030 /* look for group with min. number, but > worked */
3033 cfs_spin_lock(&filter->fo_llog_list_lock);
3034 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3035 if (olg->olg_seq <= worked) {
3036 /* this group is already synced */
3039 if (group < olg->olg_seq) {
3040 /* we have group with smaller number to sync */
3043 /* store current minimal group */
3045 group = olg->olg_seq;
3047 cfs_spin_unlock(&filter->fo_llog_list_lock);
3049 if (olg_min == NULL)
3052 worked = olg_min->olg_seq;
3053 if (olg_min->olg_exp &&
3054 (dexp == olg_min->olg_exp || dexp == NULL)) {
3056 ctxt = llog_group_get_ctxt(olg_min,
3057 LLOG_MDS_OST_REPL_CTXT);
3059 err = llog_sync(ctxt, olg_min->olg_exp);
3060 llog_ctxt_put(ctxt);
3062 CERROR("error flushing logs to MDS: "
3067 } while (olg_min != NULL);
3070 /* Also incredibly similar to mds_disconnect */
3071 static int filter_disconnect(struct obd_export *exp)
3073 struct obd_device *obd = exp->exp_obd;
3078 class_export_get(exp);
3080 if (!(exp->exp_flags & OBD_OPT_FORCE))
3081 filter_grant_sanity_check(obd, __func__);
3082 filter_grant_discard(exp);
3084 /* Flush any remaining cancel messages out to the target */
3085 filter_sync_llogs(obd, exp);
3087 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3089 rc = server_disconnect_export(exp);
3091 /* Do not erase record for recoverable client. */
3092 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3093 filter_client_del(exp);
3095 fsfilt_sync(obd, obd->u.obt.obt_sb);
3097 class_export_put(exp);
3101 /* reverse import is changed, sync all cancels */
3102 static void filter_revimp_update(struct obd_export *exp)
3107 class_export_get(exp);
3109 /* flush any remaining cancel messages out to the target */
3110 filter_sync_llogs(exp->exp_obd, exp);
3111 class_export_put(exp);
3115 static int filter_ping(struct obd_export *exp)
3117 filter_fmd_expire(exp);
3121 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3122 const char *what, int quiet)
3124 struct dentry *dchild = NULL;
3126 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3128 if (IS_ERR(dchild)) {
3129 CERROR("%s error looking up object: "POSTID"\n",
3130 what, ostid->oi_id, ostid->oi_seq);
3134 if (dchild->d_inode == NULL) {
3136 CERROR("%s: %s on non-existent object: "POSTID" \n",
3137 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3139 RETURN(ERR_PTR(-ENOENT));
3145 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3147 struct dentry *dentry = NULL;
3148 struct obd_device *obd;
3152 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3153 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3157 obd = class_exp2obd(exp);
3159 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3163 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3165 RETURN(PTR_ERR(dentry));
3167 /* Limit the valid bits in the return data to what we actually use */
3168 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3169 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3175 /* this should be enabled/disabled in condition to enabled/disabled large
3176 * inodes (fast EAs) in backing store FS. */
3177 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3178 void *handle, struct obdo *oa)
3180 struct obd_device *obd = exp->exp_obd;
3184 if (oa->o_valid & OBD_MD_FLFID) {
3185 struct filter_fid ff;
3187 if (!(oa->o_valid & OBD_MD_FLGROUP))
3189 /* packing fid and converting it to LE for storing into EA.
3190 * Here ->o_stripe_idx should be filled by LOV and rest of
3191 * fields - by client. */
3192 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3193 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3194 /* XXX: we are ignoring o_parent_ver here, since this should
3195 * be the same for all objects in this fileset. */
3196 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3197 ff.ff_objid = cpu_to_le64(oa->o_id);
3198 ff.ff_seq = cpu_to_le64(oa->o_seq);
3200 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3201 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3204 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3206 CERROR("store fid in object failed! rc: %d\n", rc);
3208 CDEBUG(D_HA, "OSS object without fid info!\n");
3214 /* this is called from filter_truncate() until we have filter_punch() */
3215 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3216 struct obdo *oa, struct obd_trans_info *oti)
3218 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3219 struct llog_cookie *fcc = NULL;
3220 struct filter_obd *filter;
3221 int rc, err, sync = 0;
3222 loff_t old_size = 0;
3223 unsigned int ia_valid;
3224 struct inode *inode;
3225 struct page *page = NULL;
3230 LASSERT(dentry != NULL);
3231 LASSERT(!IS_ERR(dentry));
3233 inode = dentry->d_inode;
3234 LASSERT(inode != NULL);
3236 filter = &exp->exp_obd->u.filter;
3237 iattr_from_obdo(&iattr, oa, oa->o_valid);
3238 ia_valid = iattr.ia_valid;
3240 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3241 OBD_ALLOC(fcc, sizeof(*fcc));
3243 *fcc = oa->o_lcookie;
3245 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3246 unsigned long now = jiffies;
3247 ll_vfs_dq_init(inode);
3248 /* Filter truncates and writes are serialized by
3249 * i_alloc_sem, see the comment in
3250 * filter_preprw_write.*/
3251 if (ia_valid & ATTR_SIZE)
3252 down_write(&inode->i_alloc_sem);
3253 LOCK_INODE_MUTEX(inode);
3254 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3255 old_size = i_size_read(inode);
3258 /* VBR: version recovery check */
3259 rc = filter_version_get_check(exp, oti, inode);
3261 GOTO(out_unlock, rc);
3263 /* Let's pin the last page so that ldiskfs_truncate
3264 * should not start GFP_FS allocation. */
3265 if (ia_valid & ATTR_SIZE) {
3266 page = grab_cache_page(inode->i_mapping,
3267 iattr.ia_size >> PAGE_CACHE_SHIFT);
3269 GOTO(out_unlock, rc = -ENOMEM);
3274 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3275 * then we will accept the UID+GID sent by the client during write for
3276 * initializing the ownership of this inode. We only allow this to
3277 * happen once so clear these bits in setattr. In 2.6 kernels it is
3278 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3279 * the flags that are actually being set. */
3280 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3281 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3282 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3284 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3285 if (!(ia_valid & ATTR_MODE)) {
3286 iattr.ia_mode = inode->i_mode;
3287 iattr.ia_valid |= ATTR_MODE;
3289 iattr.ia_mode &= ~S_ISUID;
3291 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3292 if (!(iattr.ia_valid & ATTR_MODE)) {
3293 iattr.ia_mode = inode->i_mode;
3294 iattr.ia_valid |= ATTR_MODE;
3296 iattr.ia_mode &= ~S_ISGID;
3299 orig_ids[USRQUOTA] = inode->i_uid;
3300 orig_ids[GRPQUOTA] = inode->i_gid;
3301 handle = fsfilt_start_log(exp->exp_obd, inode,
3302 FSFILT_OP_SETATTR, oti, 1);
3304 GOTO(out_unlock, rc = PTR_ERR(handle));
3306 /* update inode EA only once when inode is suid bit marked. As
3307 * on 2.6.x UID and GID may be set separately, we check here
3308 * only one of them to avoid double setting. */
3309 if (inode->i_mode & S_ISUID)
3310 filter_update_fidea(exp, inode, handle, oa);
3312 handle = fsfilt_start(exp->exp_obd, inode,
3313 FSFILT_OP_SETATTR, oti);
3315 GOTO(out_unlock, rc = PTR_ERR(handle));
3317 if (oa->o_valid & OBD_MD_FLFLAGS) {
3318 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3319 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3321 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3323 /* set cancel cookie callback function */
3324 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3325 filter_cancel_cookies_cb,
3329 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3330 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3332 /* The truncate might have used up our transaction credits. Make sure
3333 * we have two left for the last_rcvd and VBR inode version updates. */
3334 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3336 /* Update inode version only if data has changed => size has changed */
3337 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3341 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3345 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3347 CERROR("error on commit, err = %d\n", err);
3354 /* For a partial-page truncate flush the page to disk immediately
3355 * to avoid data corruption during direct disk write. b=17397 */
3356 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3357 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3358 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3368 page_cache_release(page);
3370 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3371 UNLOCK_INODE_MUTEX(inode);
3372 if (ia_valid & ATTR_SIZE)
3373 up_write(&inode->i_alloc_sem);
3375 OBD_FREE(fcc, sizeof(*fcc));
3377 /* trigger quota release */
3378 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3379 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3380 int rc2 = lquota_adjust(filter_quota_interface_ref,
3381 exp->exp_obd, cur_ids,
3382 orig_ids, rc, FSFILT_OP_SETATTR);
3383 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3384 "filter adjust qunit. (rc:%d)\n", rc2);
3389 /* this is called from filter_truncate() until we have filter_punch() */
3390 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3391 struct obd_trans_info *oti)
3393 struct obdo *oa = oinfo->oi_oa;
3394 struct lustre_capa *capa = oinfo_capa(oinfo);
3395 struct ldlm_res_id res_id;
3396 struct filter_mod_data *fmd;
3397 struct lvfs_run_ctxt saved;
3398 struct filter_obd *filter;
3399 struct ldlm_resource *res;
3400 struct dentry *dentry;
3401 __u64 opc = CAPA_OPC_META_WRITE;
3405 if (oinfo->oi_flags & OBD_FL_PUNCH)
3406 opc |= CAPA_OPC_OSS_TRUNC;
3408 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3412 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3413 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3418 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3419 /* This would be very bad - accidentally truncating a file when
3420 * changing the time or similar - bug 12203. */
3421 if (oa->o_valid & OBD_MD_FLSIZE &&
3422 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3423 static char mdsinum[48];
3425 if (oa->o_valid & OBD_MD_FLFID)
3426 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3427 oa->o_parent_seq, oa->o_parent_oid,
3432 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3433 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3434 oa->o_id, oa->o_seq, mdsinum);
3438 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3440 RETURN(PTR_ERR(dentry));
3442 filter = &exp->exp_obd->u.filter;
3443 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3446 * We need to be atomic against a concurrent write
3447 * (which takes the semaphore for reading). fmd_mactime_xid
3448 * checks will have no effect if a write request with lower
3449 * xid starts just before a setattr and finishes later than
3450 * the setattr (see bug 21489, comment 27).
3453 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3454 unsigned long now = jiffies;
3455 down_write(&dentry->d_inode->i_alloc_sem);
3456 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3457 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3458 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3459 fmd->fmd_mactime_xid = oti->oti_xid;
3460 filter_fmd_put(exp, fmd);
3461 up_write(&dentry->d_inode->i_alloc_sem);
3464 /* setting objects attributes (including owner/group) */
3465 rc = filter_setattr_internal(exp, dentry, oa, oti);
3467 GOTO(out_unlock, rc);
3469 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3470 &res_id, LDLM_EXTENT, 0);
3473 LDLM_RESOURCE_ADDREF(res);
3474 rc = ldlm_res_lvbo_update(res, NULL, 0);
3475 LDLM_RESOURCE_DELREF(res);
3476 ldlm_resource_putref(res);
3479 oa->o_valid = OBD_MD_FLID;
3481 /* Quota release need uid/gid info */
3482 obdo_from_inode(oa, dentry->d_inode, NULL,
3483 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3488 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3492 /* XXX identical to osc_unpackmd */
3493 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3494 struct lov_mds_md *lmm, int lmm_bytes)
3500 if (lmm_bytes < sizeof (*lmm)) {
3501 CERROR("lov_mds_md too small: %d, need %d\n",
3502 lmm_bytes, (int)sizeof(*lmm));
3505 /* XXX LOV_MAGIC etc check? */
3507 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3508 CERROR("lov_mds_md: zero lmm_object_id\n");
3513 lsm_size = lov_stripe_md_size(1);
3517 if (*lsmp != NULL && lmm == NULL) {
3518 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3519 OBD_FREE(*lsmp, lsm_size);
3524 if (*lsmp == NULL) {
3525 OBD_ALLOC(*lsmp, lsm_size);
3529 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3530 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3531 OBD_FREE(*lsmp, lsm_size);
3534 loi_init((*lsmp)->lsm_oinfo[0]);
3538 /* XXX zero *lsmp? */
3539 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3540 LASSERT((*lsmp)->lsm_object_id);
3543 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3548 /* caller must hold fo_create_locks[oa->o_seq] */
3549 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3550 struct filter_obd *filter)
3552 struct obdo doa = { 0 }; /* XXX obdo on stack */
3558 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3560 memset(&doa, 0, sizeof(doa));
3562 doa.o_valid |= OBD_MD_FLGROUP;
3563 doa.o_seq = oa->o_seq;
3564 doa.o_mode = S_IFREG;
3566 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3567 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3568 exp->exp_obd->obd_name, doa.o_seq);
3572 last = filter_last_id(filter, doa.o_seq);
3574 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3576 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3577 exp->exp_obd->obd_name, oa->o_id + 1, last,
3578 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3580 for (id = last; id > oa->o_id; id--) {
3582 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3583 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3584 CEMERG("error destroying precreate objid "LPU64": %d\n",
3587 /* update last_id on disk periodically so that if we restart
3588 * we don't need to re-scan all of the just-deleted objects. */
3589 if ((id & 511) == 0 && !skip_orphan) {
3590 filter_set_last_id(filter, id - 1, doa.o_seq);
3591 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3595 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3596 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3599 filter_set_last_id(filter, id, doa.o_seq);
3600 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3603 * We have destroyed orphan objects, but don't want to reuse
3604 * them. Therefore we don't reset last_id to the last created
3605 * objects. Instead, we report back to the MDS the object id
3606 * of the last orphan, so that the MDS can restart allocating
3607 * objects from this id + 1 and thus skip the whole orphan
3613 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3618 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3619 obd_seq group, int *num);
3620 /* returns a negative error or a nonnegative number of files to create */
3621 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3622 obd_seq group, struct obd_trans_info *oti)
3624 struct obd_device *obd = exp->exp_obd;
3625 struct filter_obd *filter = &obd->u.filter;
3629 /* delete orphans request */
3630 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3631 obd_id last = filter_last_id(filter, group);
3633 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3634 CERROR("%s: dropping old orphan cleanup request\n",
3638 /* This causes inflight precreates to abort and drop lock */
3639 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3640 cfs_down(&filter->fo_create_locks[group]);
3641 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3642 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3643 exp->exp_obd->obd_name, group);
3644 cfs_up(&filter->fo_create_locks[group]);
3647 diff = oa->o_id - last;
3648 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3651 if (-diff > OST_MAX_PRECREATE) {
3652 CERROR("%s: ignoring bogus orphan destroy request: "
3653 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3655 /* FIXME: should reset precreate_next_id on MDS */
3656 GOTO(out, rc = -EINVAL);
3659 rc = filter_destroy_precreated(exp, oa, filter);
3661 CERROR("%s: unable to write lastobjid, but "
3662 "orphans were deleted\n", obd->obd_name);
3665 /* XXX: Used by MDS for the first time! */
3666 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3669 cfs_down(&filter->fo_create_locks[group]);
3670 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3671 CERROR("%s: dropping old precreate request\n",
3675 /* only precreate if group == 0 and o_id is specfied */
3676 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3679 diff = oa->o_id - filter_last_id(filter, group);
3680 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3681 filter_last_id(filter, group), diff);
3683 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3684 oa->o_id, filter_last_id(filter, group), diff);
3688 oa->o_id = filter_last_id(&obd->u.filter, group);
3689 rc = filter_precreate(obd, oa, group, &diff);
3690 oa->o_id = filter_last_id(&obd->u.filter, group);
3692 oa->o_valid |= (OBD_MD_FLID | OBD_MD_FLGROUP);
3695 /* else diff == 0 */
3698 cfs_up(&filter->fo_create_locks[group]);
3702 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3703 __u64 max_age, __u32 flags)
3705 struct filter_obd *filter = &obd->u.filter;
3706 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3710 /* at least try to account for cached pages. its still racey and
3711 * might be under-reporting if clients haven't announced their
3712 * caches with brw recently */
3713 cfs_spin_lock(&obd->obd_osfs_lock);
3714 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3715 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3716 cfs_spin_unlock(&obd->obd_osfs_lock);
3718 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3719 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3720 filter->fo_tot_dirty, filter->fo_tot_granted,
3721 filter->fo_tot_pending,
3722 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3724 filter_grant_sanity_check(obd, __func__);
3726 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3727 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3728 osfs->os_bsize - 1) >> blockbits));
3730 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3731 struct lr_server_data *lsd = class_server_data(obd);
3732 int index = le32_to_cpu(lsd->lsd_ost_index);
3734 if (cfs_fail_val == -1 || index == cfs_fail_val)
3735 osfs->os_bfree = osfs->os_bavail = 2;
3736 else if (cfs_fail_loc & OBD_FAIL_ONCE)
3737 cfs_fail_loc &= ~OBD_FAILED; /* reset flag */
3740 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3741 * stop creating files on MDS if OST is not good shape to create
3745 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3746 osfs->os_state = OS_STATE_READONLY;
3748 if (filter->fo_raid_degraded)
3749 osfs->os_state |= OS_STATE_DEGRADED;
3753 static int filter_use_existing_obj(struct obd_device *obd,
3754 struct dentry *dchild, void **handle,
3757 struct inode *inode = dchild->d_inode;
3761 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3764 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3765 if (IS_ERR(*handle))
3766 return PTR_ERR(*handle);
3768 iattr.ia_valid = ATTR_MODE;
3769 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3770 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3777 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3780 __u64 os_ffree = -1;
3782 cfs_spin_lock(&obd->obd_osfs_lock);
3783 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3785 os_ffree = obd->obd_osfs.os_ffree;
3786 cfs_spin_unlock(&obd->obd_osfs_lock);
3791 /* We rely on the fact that only one thread will be creating files in a given
3792 * group at a time, which is why we don't need an atomic filter_get_new_id.
3793 * Even if we had that atomic function, the following race would exist:
3795 * thread 1: gets id x from filter_next_id
3796 * thread 2: gets id (x + 1) from filter_next_id
3797 * thread 2: creates object (x + 1)
3798 * thread 1: tries to create object x, gets -ENOSPC
3800 * Caller must hold fo_create_locks[group]
3802 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3803 obd_seq group, int *num)
3805 struct dentry *dchild = NULL, *dparent = NULL;
3806 struct filter_obd *filter;
3807 struct obd_statfs *osfs;
3809 int err = 0, rc = 0, recreate_obj = 0, i;
3810 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3813 void *handle = NULL;
3816 filter = &obd->u.filter;
3818 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3820 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3822 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3823 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3826 OBD_ALLOC(osfs, sizeof(*osfs));
3829 rc = filter_statfs(obd, osfs,
3830 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3832 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3833 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3834 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3835 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3837 if (oa->o_valid & OBD_MD_FLFLAGS)
3838 oa->o_flags |= OBD_FL_NOSPC_BLK;
3840 oa->o_valid |= OBD_MD_FLFLAGS;
3841 oa->o_flags = OBD_FL_NOSPC_BLK;
3846 OBD_FREE(osfs, sizeof(*osfs));
3851 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3852 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3854 for (i = 0; i < *num && err == 0; i++) {
3855 int cleanup_phase = 0;
3857 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3858 CWARN("%s: create aborted by destroy\n",
3867 last_id = filter_last_id(filter, group);
3868 if (next_id > last_id) {
3869 CERROR("Error: Trying to recreate obj greater"
3870 "than last id "LPD64" > "LPD64"\n",
3872 GOTO(cleanup, rc = -EINVAL);
3875 next_id = filter_last_id(filter, group) + 1;
3877 /* Don't create objects beyond the valid range for this SEQ */
3878 if (unlikely(fid_seq_is_mdt0(group) &&
3879 next_id >= IDIF_MAX_OID)) {
3880 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3881 obd->obd_name, next_id, group);
3882 GOTO(cleanup, rc = -ENOSPC);
3883 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3884 next_id >= OBIF_MAX_OID)) {
3885 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3886 obd->obd_name, next_id, group);
3887 GOTO(cleanup, rc = -ENOSPC);
3890 dparent = filter_parent_lock(obd, group, next_id);
3891 if (IS_ERR(dparent))
3892 GOTO(cleanup, rc = PTR_ERR(dparent));
3893 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3895 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3897 GOTO(cleanup, rc = PTR_ERR(dchild));
3898 cleanup_phase = 2; /* f_dput(dchild) */
3900 if (dchild->d_inode != NULL) {
3901 /* This would only happen if lastobjid was bad on disk*/
3902 /* Could also happen if recreating missing obj but it
3903 * already exists. */
3905 CERROR("%s: recreating existing object %.*s?\n",
3906 obd->obd_name, dchild->d_name.len,
3907 dchild->d_name.name);
3909 /* Use these existing objects if they are
3911 if (dchild->d_inode->i_size == 0) {
3912 rc = filter_use_existing_obj(obd,dchild,
3913 &handle, &cleanup_phase);
3920 CERROR("%s: Serious error: objid %.*s already "
3921 "exists; is this filesystem corrupt?\n",
3922 obd->obd_name, dchild->d_name.len,
3923 dchild->d_name.name);
3926 GOTO(cleanup, rc = -EEXIST);
3929 handle = fsfilt_start_log(obd, dparent->d_inode,
3930 FSFILT_OP_CREATE, NULL, 1);
3932 GOTO(cleanup, rc = PTR_ERR(handle));
3935 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3936 ",od->o_id="LPU64")\n", obd->obd_name, group,
3939 /* We mark object SUID+SGID to flag it for accepting UID+GID
3940 * from client on first write. Currently the permission bits
3941 * on the OST are never used, so this is OK. */
3942 rc = ll_vfs_create(dparent->d_inode, dchild,
3943 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3945 CERROR("create failed rc = %d\n", rc);
3946 if (rc == -ENOSPC) {
3947 os_ffree = filter_calc_free_inodes(obd);
3951 if (obd->obd_osfs.os_bavail <
3952 (obd->obd_osfs.os_blocks >> 10)) {
3953 if (oa->o_valid & OBD_MD_FLFLAGS)
3954 oa->o_flags |= OBD_FL_NOSPC_BLK;
3956 oa->o_valid |= OBD_MD_FLFLAGS;
3957 oa->o_flags = OBD_FL_NOSPC_BLK;
3960 CERROR("%s: free inode "LPU64"\n",
3961 obd->obd_name, os_ffree);
3967 if (dchild->d_inode)
3968 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3969 dchild->d_inode->i_ino);
3972 /* Set a/c/m time to a insane large negative value at creation
3973 * time so that any timestamp arriving from the client will
3974 * always be newer and update the inode.
3975 * See LU-221 for details */
3976 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
3977 LTIME_S(iattr.ia_atime) = INT_MIN + 24 * 3600;
3978 LTIME_S(iattr.ia_mtime) = INT_MIN + 24 * 3600;
3979 LTIME_S(iattr.ia_ctime) = INT_MIN + 24 * 3600;
3980 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
3982 CERROR("unable to initialize a/c/m time of newly"
3985 if (!recreate_obj) {
3986 filter_set_last_id(filter, next_id, group);
3987 err = filter_update_last_objid(obd, group, 0);
3989 CERROR("unable to write lastobjid "
3990 "but file created\n");
3994 switch(cleanup_phase) {
3996 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3998 CERROR("error on commit, err = %d\n", err);
4005 filter_parent_unlock(dparent);
4012 if (cfs_time_after(jiffies, enough_time)) {
4015 "%s: precreate slow - want %d got %d \n",
4016 obd->obd_name, *num, i);
4023 "%s: created %d objects for group "POSTID" rc %d\n",
4024 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4029 int filter_create(struct obd_export *exp, struct obdo *oa,
4030 struct lov_stripe_md **ea, struct obd_trans_info *oti)
4032 struct obd_device *obd = exp->exp_obd;
4033 struct filter_export_data *fed;
4034 struct filter_obd *filter;
4035 struct lvfs_run_ctxt saved;
4036 struct lov_stripe_md *lsm = NULL;
4040 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4041 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4043 fed = &exp->exp_filter_data;
4044 filter = &obd->u.filter;
4046 /* 1.8 client doesn't carry the ocd_group with connect request,
4047 * so the fed_group will always be zero for 1.8 client. */
4048 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4049 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4050 oa->o_seq != FID_SEQ_LLOG &&
4051 oa->o_seq != FID_SEQ_ECHO) {
4052 CERROR("The request from older client has invalid"
4053 " group "LPU64"!\n", oa->o_seq);
4056 } else if (fed->fed_group != oa->o_seq) {
4057 CERROR("%s: this export (nid %s) used object group %d "
4058 "earlier; now it's trying to use group "LPU64"!"
4059 " This could be a bug in the MDS. Please report to "
4060 "http://bugzilla.lustre.org/\n", obd->obd_name,
4061 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4068 rc = obd_alloc_memmd(exp, &lsm);
4075 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4077 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4078 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4079 if (!obd->obd_recovering ||
4080 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4081 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4082 oa->o_id, filter_last_id(filter, oa->o_seq));
4086 cfs_down(&filter->fo_create_locks[oa->o_seq]);
4087 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4088 cfs_up(&filter->fo_create_locks[oa->o_seq]);
4091 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4094 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4095 if (rc && ea != NULL && *ea != lsm) {
4096 obd_free_memmd(exp, &lsm);
4097 } else if (rc == 0 && ea != NULL) {
4098 /* XXX LOV STACKING: the lsm that is passed to us from
4099 * LOV does not have valid lsm_oinfo data structs, so
4100 * don't go touching that. This needs to be fixed in a
4102 lsm->lsm_object_id = oa->o_id;
4109 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4110 struct lov_stripe_md *md, struct obd_trans_info *oti,
4111 struct obd_export *md_exp, void *capa)
4113 unsigned int qcids[MAXQUOTAS] = {0, 0};
4114 struct obd_device *obd;
4115 struct filter_obd *filter;
4116 struct dentry *dchild = NULL, *dparent = NULL;
4117 struct lustre_handle lockh = { 0 };
4118 struct lvfs_run_ctxt saved;
4119 void *handle = NULL;
4120 struct llog_cookie *fcc = NULL;
4121 int rc, rc2, cleanup_phase = 0, sync = 0;
4126 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4127 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4132 filter = &obd->u.filter;
4134 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4137 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4138 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4140 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4142 GOTO(cleanup, rc = PTR_ERR(dchild));
4145 if (dchild->d_inode == NULL) {
4146 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4147 oa->o_id, oa->o_seq);
4148 /* If object already gone, cancel cookie right now */
4149 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4150 struct llog_ctxt *ctxt;
4151 struct obd_llog_group *olg;
4153 olg = filter_find_olg(obd, oa->o_seq);
4155 CERROR(" %s: can not find olg of group %d\n",
4156 obd->obd_name, (int)oa->o_seq);
4157 GOTO(cleanup, rc = PTR_ERR(olg));
4159 fcc = &oa->o_lcookie;
4160 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4161 llog_cancel(ctxt, NULL, 1, fcc, 0);
4162 llog_ctxt_put(ctxt);
4163 fcc = NULL; /* we didn't allocate fcc, don't free it */
4165 GOTO(cleanup, rc = -ENOENT);
4168 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4172 /* Our MDC connection is established by the MDS to us */
4173 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4174 OBD_ALLOC(fcc, sizeof(*fcc));
4176 *fcc = oa->o_lcookie;
4178 ll_vfs_dq_init(dchild->d_inode);
4180 /* we're gonna truncate it first in order to avoid possible deadlock:
4182 * open trasaction open transaction
4183 * down(i_zombie) down(i_zombie)
4184 * restart transaction
4185 * (see BUG 4180) -bzzz
4187 * take i_alloc_sem too to prevent other threads from writing to the
4188 * file while we are truncating it. This can cause lock ordering issue
4189 * between page lock, i_mutex & starting new journal handle.
4190 * (see bug 20321) -johann
4193 down_write(&dchild->d_inode->i_alloc_sem);
4194 LOCK_INODE_MUTEX(dchild->d_inode);
4195 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4197 /* VBR: version recovery check */
4198 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4200 UNLOCK_INODE_MUTEX(dchild->d_inode);
4201 up_write(&dchild->d_inode->i_alloc_sem);
4205 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4207 if (IS_ERR(handle)) {
4208 UNLOCK_INODE_MUTEX(dchild->d_inode);
4209 up_write(&dchild->d_inode->i_alloc_sem);
4210 GOTO(cleanup, rc = PTR_ERR(handle));
4213 iattr.ia_valid = ATTR_SIZE;
4215 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4216 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4217 UNLOCK_INODE_MUTEX(dchild->d_inode);
4218 up_write(&dchild->d_inode->i_alloc_sem);
4222 GOTO(cleanup, rc = rc2);
4224 /* We don't actually need to lock the parent until we are unlinking
4225 * here, and not while truncating above. That avoids holding the
4226 * parent lock for a long time during truncate, which can block other
4227 * threads from doing anything to objects in that directory. bug 7171 */
4228 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4229 if (IS_ERR(dparent))
4230 GOTO(cleanup, rc = PTR_ERR(dparent));
4231 cleanup_phase = 3; /* filter_parent_unlock */
4233 LOCK_INODE_MUTEX(dchild->d_inode);
4234 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4235 if (IS_ERR(handle)) {
4236 UNLOCK_INODE_MUTEX(dchild->d_inode);
4237 GOTO(cleanup, rc = PTR_ERR(handle));
4239 cleanup_phase = 4; /* fsfilt_commit */
4241 /* Quota release need uid/gid of inode */
4242 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4244 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4246 /* this drops dchild->d_inode->i_mutex unconditionally */
4247 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4251 switch(cleanup_phase) {
4254 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4255 oti->oti_handle : handle,
4256 filter_cancel_cookies_cb,
4258 /* If add_journal_cb failed, then filter_finish_transno
4259 * will commit the handle and we will do a sync
4260 * on commit. then we call callback directly to free
4263 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4265 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4268 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4270 CERROR("error on commit, err = %d\n", rc2);
4277 filter_parent_unlock(dparent);
4279 filter_fini_destroy(obd, &lockh);
4283 OBD_FREE(fcc, sizeof(*fcc));
4285 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4288 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4292 /* trigger quota release */
4293 qcids[USRQUOTA] = oa->o_uid;
4294 qcids[GRPQUOTA] = oa->o_gid;
4295 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4298 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4302 /* NB start and end are used for punch, but not truncate */
4303 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4304 struct obd_trans_info *oti,
4305 struct ptlrpc_request_set *rqset)
4310 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4311 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4312 oinfo->oi_policy.l_extent.end);
4316 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4317 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4318 oinfo->oi_policy.l_extent.start);
4320 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4321 rc = filter_setattr(exp, oinfo, oti);
4325 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4326 obd_off start, obd_off end,
4327 struct ptlrpc_request_set *set)
4329 struct lvfs_run_ctxt saved;
4330 struct obd_device_target *obt;
4331 struct dentry *dentry;
4335 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4336 (struct lustre_capa *)oinfo->oi_capa,
4337 CAPA_OPC_OSS_WRITE);
4341 obt = &exp->exp_obd->u.obt;
4343 /* An objid of zero is taken to mean "sync whole filesystem" */
4344 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4345 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4346 /* Flush any remaining cancel messages out to the target */
4347 filter_sync_llogs(exp->exp_obd, exp);
4351 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4353 RETURN(PTR_ERR(dentry));
4355 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4357 LOCK_INODE_MUTEX(dentry->d_inode);
4359 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4361 /* just any file to grab fsync method - "file" arg unused */
4362 struct file *file = obt->obt_rcvd_filp;
4364 if (file->f_op && file->f_op->fsync)
4365 rc = file->f_op->fsync(NULL, dentry, 1);
4367 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4371 UNLOCK_INODE_MUTEX(dentry->d_inode);
4373 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4374 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4375 FILTER_VALID_FLAGS);
4377 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4383 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4384 void *key, __u32 *vallen, void *val,
4385 struct lov_stripe_md *lsm)
4387 struct obd_device *obd;
4390 obd = class_exp2obd(exp);
4392 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4396 if (KEY_IS(KEY_BLOCKSIZE)) {
4397 __u32 *blocksize = val;
4399 if (*vallen < sizeof(*blocksize))
4401 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4403 *vallen = sizeof(*blocksize);
4407 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4408 __u32 *blocksize_bits = val;
4409 if (blocksize_bits) {
4410 if (*vallen < sizeof(*blocksize_bits))
4412 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4414 *vallen = sizeof(*blocksize_bits);
4418 if (KEY_IS(KEY_LAST_ID)) {
4419 obd_id *last_id = val;
4420 /* FIXME: object groups */
4422 if (*vallen < sizeof(*last_id))
4424 *last_id = filter_last_id(&obd->u.filter,
4425 exp->exp_filter_data.fed_group);
4427 *vallen = sizeof(*last_id);
4431 if (KEY_IS(KEY_FIEMAP)) {
4432 struct ll_fiemap_info_key *fm_key = key;
4433 struct dentry *dentry;
4434 struct ll_user_fiemap *fiemap = val;
4435 struct lvfs_run_ctxt saved;
4438 if (fiemap == NULL) {
4439 *vallen = fiemap_count_to_size(
4440 fm_key->fiemap.fm_extent_count);
4444 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4447 RETURN(PTR_ERR(dentry));
4449 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4450 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4451 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4453 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4459 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4460 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4461 *vallen = sizeof(__u32);
4465 CDEBUG(D_IOCTL, "invalid key\n");
4469 static inline int filter_setup_llog_group(struct obd_export *exp,
4470 struct obd_device *obd,
4473 struct obd_llog_group *olg;
4474 struct llog_ctxt *ctxt;
4477 olg = filter_find_create_olg(obd, group);
4479 RETURN(PTR_ERR(olg));
4481 llog_group_set_export(olg, exp);
4483 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4484 LASSERTF(ctxt != NULL, "ctxt is null\n");
4486 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4487 llog_ctxt_put(ctxt);
4491 static int filter_set_grant_shrink(struct obd_export *exp,
4492 struct ost_body *body)
4494 /* handle shrink grant */
4495 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4496 filter_grant_incoming(exp, &body->oa);
4497 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4503 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4505 struct obd_device *obd;
4511 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4515 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4516 obd_export_nid2str(exp));
4517 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4519 /* setup llog imports */
4521 group = (int)(*(__u32 *)val);
4523 group = 0; /* default value */
4525 LASSERT_SEQ_IS_MDT(group);
4526 rc = filter_setup_llog_group(exp, obd, group);
4530 if (group == FID_SEQ_OST_MDT0) {
4531 /* setup llog group 1 for interop */
4532 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4535 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4540 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4541 void *key, __u32 vallen, void *val,
4542 struct ptlrpc_request_set *set)
4544 struct obd_device *obd;
4549 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4553 if (KEY_IS(KEY_CAPA_KEY)) {
4555 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4557 CERROR("filter update capability key failed: %d\n", rc);
4561 if (KEY_IS(KEY_REVIMP_UPD)) {
4562 filter_revimp_update(exp);
4563 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4567 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4568 filter_adapt_sptlrpc_conf(obd, 0);
4572 if (KEY_IS(KEY_MDS_CONN))
4573 RETURN(filter_set_mds_conn(exp, val));
4575 if (KEY_IS(KEY_GRANT_SHRINK))
4576 RETURN(filter_set_grant_shrink(exp, val));
4581 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4582 int len, void *karg, void *uarg)
4584 struct obd_device *obd = exp->exp_obd;
4585 struct obd_ioctl_data *data = karg;
4589 case OBD_IOC_ABORT_RECOVERY: {
4590 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4591 target_stop_recovery_thread(obd);
4595 case OBD_IOC_SYNC: {
4596 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4597 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4601 case OBD_IOC_SET_READONLY: {
4603 struct super_block *sb = obd->u.obt.obt_sb;
4604 struct inode *inode = sb->s_root->d_inode;
4605 BDEVNAME_DECLARE_STORAGE(tmp);
4606 CERROR("*** setting device %s read-only ***\n",
4607 ll_bdevname(sb, tmp));
4609 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4610 if (!IS_ERR(handle))
4611 rc = fsfilt_commit(obd, inode, handle, 1);
4613 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4614 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4616 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4620 case OBD_IOC_CATLOGLIST: {
4621 rc = llog_catalog_list(obd, 1, data);
4625 case OBD_IOC_LLOG_CANCEL:
4626 case OBD_IOC_LLOG_REMOVE:
4627 case OBD_IOC_LLOG_INFO:
4628 case OBD_IOC_LLOG_PRINT: {
4629 /* FIXME to be finished */
4630 RETURN(-EOPNOTSUPP);
4632 struct llog_ctxt *ctxt = NULL;
4634 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4635 rc = llog_ioctl(ctxt, cmd, data);
4636 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4649 static int filter_health_check(struct obd_device *obd)
4651 #ifdef USE_HEALTH_CHECK_WRITE
4652 struct filter_obd *filter = &obd->u.filter;
4657 * health_check to return 0 on healthy
4658 * and 1 on unhealthy.
4660 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4663 #ifdef USE_HEALTH_CHECK_WRITE
4664 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4665 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4670 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4673 return filter_fid2dentry(data, NULL, gr, id);
4676 static int filter_process_config(struct obd_device *obd, obd_count len,
4679 struct lustre_cfg *lcfg = buf;
4680 struct lprocfs_static_vars lvars;
4683 switch (lcfg->lcfg_command) {
4685 lprocfs_filter_init_vars(&lvars);
4687 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4697 static int filter_notify(struct obd_device *obd,
4698 struct obd_device *unused,
4699 enum obd_notify_event ev, void *data)
4702 case OBD_NOTIFY_CONFIG:
4703 LASSERT(obd->obd_no_conn);
4704 cfs_spin_lock(&obd->obd_dev_lock);
4705 obd->obd_no_conn = 0;
4706 cfs_spin_unlock(&obd->obd_dev_lock);
4709 CDEBUG(D_INFO, "%s: Unhandled notification %#x\n",
4715 static struct lvfs_callback_ops filter_lvfs_ops = {
4716 l_fid2dentry: filter_lvfs_fid2dentry,
4719 static struct obd_ops filter_obd_ops = {
4720 .o_owner = THIS_MODULE,
4721 .o_get_info = filter_get_info,
4722 .o_set_info_async = filter_set_info_async,
4723 .o_setup = filter_setup,
4724 .o_precleanup = filter_precleanup,
4725 .o_cleanup = filter_cleanup,
4726 .o_connect = filter_connect,
4727 .o_reconnect = filter_reconnect,
4728 .o_disconnect = filter_disconnect,
4729 .o_ping = filter_ping,
4730 .o_init_export = filter_init_export,
4731 .o_destroy_export = filter_destroy_export,
4732 .o_statfs = filter_statfs,
4733 .o_getattr = filter_getattr,
4734 .o_unpackmd = filter_unpackmd,
4735 .o_create = filter_create,
4736 .o_setattr = filter_setattr,
4737 .o_destroy = filter_destroy,
4738 .o_brw = filter_brw,
4739 .o_punch = filter_truncate,
4740 .o_sync = filter_sync,
4741 .o_preprw = filter_preprw,
4742 .o_commitrw = filter_commitrw,
4743 .o_llog_init = filter_llog_init,
4744 .o_llog_connect = filter_llog_connect,
4745 .o_llog_finish = filter_llog_finish,
4746 .o_iocontrol = filter_iocontrol,
4747 .o_health_check = filter_health_check,
4748 .o_process_config = filter_process_config,
4749 .o_notify = filter_notify,
4752 quota_interface_t *filter_quota_interface_ref;
4753 extern quota_interface_t filter_quota_interface;
4755 static int __init obdfilter_init(void)
4757 struct lprocfs_static_vars lvars;
4760 /** sanity check for group<->mdsno conversion */
4761 for (i = 0; i < MAX_MDT_COUNT; i++)
4762 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4764 lprocfs_filter_init_vars(&lvars);
4766 cfs_request_module("%s", "lquota");
4767 OBD_ALLOC(obdfilter_created_scratchpad,
4768 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4769 sizeof(*obdfilter_created_scratchpad));
4770 if (obdfilter_created_scratchpad == NULL)
4773 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4774 sizeof(struct filter_mod_data),
4777 GOTO(out, rc = -ENOMEM);
4779 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4780 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4782 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4783 LUSTRE_OST_NAME, NULL);
4787 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4788 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4789 ll_fmd_cachep = NULL;
4791 if (filter_quota_interface_ref)
4792 PORTAL_SYMBOL_PUT(filter_quota_interface);
4794 OBD_FREE(obdfilter_created_scratchpad,
4795 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4796 sizeof(*obdfilter_created_scratchpad));
4802 static void __exit obdfilter_exit(void)
4804 if (filter_quota_interface_ref)
4805 PORTAL_SYMBOL_PUT(filter_quota_interface);
4807 if (ll_fmd_cachep) {
4808 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4809 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4810 ll_fmd_cachep = NULL;
4813 class_unregister_type(LUSTRE_OST_NAME);
4814 OBD_FREE(obdfilter_created_scratchpad,
4815 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4816 sizeof(*obdfilter_created_scratchpad));
4819 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4820 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4821 MODULE_LICENSE("GPL");
4823 module_init(obdfilter_init);
4824 module_exit(obdfilter_exit);