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 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 cfs_spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 cfs_spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct obd_device_target *obt = &exp->exp_obd->u.obt;
124 struct tg_export_data *ted = &exp->exp_target_data;
125 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126 struct lsd_client_data *lcd;
129 int err, log_pri = D_RPCTRACE;
131 /* Propagate error code. */
135 if (!exp->exp_obd->obd_replayable || oti == NULL)
138 cfs_mutex_down(&ted->ted_lcd_lock);
140 /* if the export has already been disconnected, we have no last_rcvd slot,
141 * update server data with latest transno then */
143 cfs_mutex_up(&ted->ted_lcd_lock);
144 CWARN("commit transaction for disconnected client %s: rc %d\n",
145 exp->exp_client_uuid.uuid, rc);
146 err = filter_update_server_data(exp->exp_obd);
150 /* we don't allocate new transnos for replayed requests */
151 cfs_spin_lock(&obt->obt_lut->lut_translock);
152 if (oti->oti_transno == 0) {
153 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156 last_rcvd = oti->oti_transno;
157 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160 oti->oti_transno = last_rcvd;
162 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166 cfs_spin_unlock(&obt->obt_lut->lut_translock);
169 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
171 off = ted->ted_lr_off;
173 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174 ted->ted_lr_idx, ted->ted_lr_off);
177 class_export_cb_get(exp); /* released when the cb is called */
179 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
185 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186 lcd, sizeof(*lcd), &off,
187 force_sync | exp->exp_need_sync);
189 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
197 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199 cfs_mutex_up(&ted->ted_lcd_lock);
203 void f_dput(struct dentry *dentry)
205 /* Can't go inside filter_ddelete because it can block */
206 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208 LASSERT(atomic_read(&dentry->d_count) > 0);
213 static void init_brw_stats(struct brw_stats *brw_stats)
216 for (i = 0; i < BRW_LAST; i++)
217 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221 struct lprocfs_stats **stats)
225 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226 LPROC_FILTER_LAST - 1;
227 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
231 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241 plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243 struct obd_export *exp,
249 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250 /* Self-export gets no proc entry */
253 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
255 /* Mask error for already created
263 struct nid_stat *tmp = exp->exp_nid_stats;
264 LASSERT(tmp != NULL);
266 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267 if (tmp->nid_brw_stats == NULL)
268 GOTO(clean, rc = -ENOMEM);
270 init_brw_stats(tmp->nid_brw_stats);
271 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272 0644, &filter_per_nid_stats_fops,
275 CWARN("Error adding the brw_stats file\n");
277 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
285 rc = lprocfs_nid_ldlm_stats_init(tmp);
295 /* Add client data to the FILTER. We use a bitmap to locate a free space
296 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
297 * Otherwise, we have just read the data from the last_rcvd file and
298 * we know its offset. */
299 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
302 struct obd_device_target *obt = &obd->u.obt;
303 struct tg_export_data *ted = &exp->exp_target_data;
304 struct lr_server_data *lsd = class_server_data(obd);
305 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
306 int new_client = (cl_idx == -1);
310 LASSERT(bitmap != NULL);
311 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
314 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
317 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
318 * there's no need for extra complication here
321 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
323 if (cl_idx >= LR_MAX_CLIENTS) {
324 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
328 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
329 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
334 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
335 CERROR("FILTER client %d: bit already set in bitmap!\n",
341 ted->ted_lr_idx = cl_idx;
342 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
343 cl_idx * le16_to_cpu(lsd->lsd_client_size);
344 cfs_init_mutex(&ted->ted_lcd_lock);
345 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
347 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
348 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
351 struct lvfs_run_ctxt saved;
352 loff_t off = ted->ted_lr_off;
356 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
357 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
359 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
362 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
363 /* Transaction needed to fix bug 1403 */
364 handle = fsfilt_start(obd,
365 obt->obt_rcvd_filp->f_dentry->d_inode,
366 FSFILT_OP_SETATTR, NULL);
367 if (IS_ERR(handle)) {
368 rc = PTR_ERR(handle);
369 CERROR("unable to start transaction: rc %d\n", rc);
371 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
372 exp->exp_last_request_time = cfs_time_current_sec();
373 rc = fsfilt_add_journal_cb(obd, 0, handle,
374 target_client_add_cb,
375 class_export_cb_get(exp));
377 cfs_spin_lock(&exp->exp_lock);
378 exp->exp_need_sync = 1;
379 cfs_spin_unlock(&exp->exp_lock);
381 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
383 sizeof(*ted->ted_lcd),
384 &off, rc /* sync if no cb */);
386 obt->obt_rcvd_filp->f_dentry->d_inode,
389 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
392 CERROR("error writing %s client idx %u: rc %d\n",
393 LAST_RCVD, ted->ted_lr_idx, rc);
400 static int filter_client_del(struct obd_export *exp)
402 struct tg_export_data *ted = &exp->exp_target_data;
403 struct obd_device_target *obt = &exp->exp_obd->u.obt;
404 struct lvfs_run_ctxt saved;
409 if (ted->ted_lcd == NULL)
412 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
413 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
416 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
418 off = ted->ted_lr_off;
420 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
421 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
423 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
424 * we leak a client slot that will be cleaned on the next recovery. */
426 CERROR("%s: client idx %d has med_off %lld\n",
427 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
428 GOTO(free, rc = -EINVAL);
431 /* Clear the bit _after_ zeroing out the client so we don't
432 race with filter_client_add and zero out new clients.*/
433 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
434 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
439 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
440 /* Make sure the server's last_transno is up to date.
441 * This should be done before zeroing client slot so last_transno will
442 * be in server data or in client data in case of failure */
443 filter_update_server_data(exp->exp_obd);
445 cfs_mutex_down(&ted->ted_lcd_lock);
446 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
447 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
449 sizeof(*ted->ted_lcd), &off, 0);
450 cfs_mutex_up(&ted->ted_lcd_lock);
451 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
453 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
454 "zero out client %s at idx %u/%llu in %s, rc %d\n",
455 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
462 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
463 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
464 struct filter_mod_data *fmd)
466 LASSERT_SPIN_LOCKED(&fed->fed_lock);
467 if (--fmd->fmd_refcount == 0) {
468 /* XXX when we have persistent reservations and the handle
469 * is stored herein we need to drop it here. */
470 fed->fed_mod_count--;
471 cfs_list_del(&fmd->fmd_list);
472 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
476 /* drop fmd reference, free it if last ref */
477 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
479 struct filter_export_data *fed;
484 fed = &exp->exp_filter_data;
485 cfs_spin_lock(&fed->fed_lock);
486 filter_fmd_put_nolock(fed, fmd); /* caller reference */
487 cfs_spin_unlock(&fed->fed_lock);
490 /* expire entries from the end of the list if there are too many
491 * or they are too old */
492 static void filter_fmd_expire_nolock(struct filter_obd *filter,
493 struct filter_export_data *fed,
494 struct filter_mod_data *keep)
496 struct filter_mod_data *fmd, *tmp;
498 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
502 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
503 fed->fed_mod_count < filter->fo_fmd_max_num)
506 cfs_list_del_init(&fmd->fmd_list);
507 filter_fmd_put_nolock(fed, fmd); /* list reference */
511 void filter_fmd_expire(struct obd_export *exp)
513 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
514 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
515 &exp->exp_filter_data, NULL);
516 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
519 /* find specified objid, group in export fmd list.
520 * caller must hold fed_lock and take fmd reference itself */
521 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
522 struct filter_export_data *fed,
523 obd_id objid, obd_seq group)
525 struct filter_mod_data *found = NULL, *fmd;
527 LASSERT_SPIN_LOCKED(&fed->fed_lock);
529 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
530 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
532 cfs_list_del(&fmd->fmd_list);
533 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
534 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
539 filter_fmd_expire_nolock(filter, fed, found);
544 /* Find fmd based on objid and group, or return NULL if not found. */
545 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
546 obd_id objid, obd_seq group)
548 struct filter_mod_data *fmd;
550 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
551 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
552 &exp->exp_filter_data, objid, group);
554 fmd->fmd_refcount++; /* caller reference */
555 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
560 /* Find fmd based on objid and group, or create a new one if none is found.
561 * It is possible for this function to return NULL under memory pressure,
562 * or if objid = 0 is passed (which will only cause old entries to expire).
563 * Currently this is not fatal because any fmd state is transient and
564 * may also be freed when it gets sufficiently old. */
565 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
566 obd_id objid, obd_seq group)
568 struct filter_export_data *fed = &exp->exp_filter_data;
569 struct filter_mod_data *found = NULL, *fmd_new = NULL;
571 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
573 cfs_spin_lock(&fed->fed_lock);
574 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
577 cfs_list_add_tail(&fmd_new->fmd_list,
579 fmd_new->fmd_id = objid;
580 fmd_new->fmd_gr = group;
581 fmd_new->fmd_refcount++; /* list reference */
583 fed->fed_mod_count++;
585 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
589 found->fmd_refcount++; /* caller reference */
590 found->fmd_expire = jiffies +
591 exp->exp_obd->u.filter.fo_fmd_max_age;
594 cfs_spin_unlock(&fed->fed_lock);
600 /* drop fmd list reference so it will disappear when last reference is put.
601 * This isn't so critical because it would in fact only affect the one client
602 * that is doing the unlink and at worst we have an stale entry referencing
603 * an object that should never be used again. */
604 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
606 struct filter_mod_data *found = NULL;
608 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
609 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
611 cfs_list_del_init(&found->fmd_list);
612 filter_fmd_put_nolock(&exp->exp_filter_data, found);
614 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
617 #define filter_fmd_drop(exp, objid, group)
620 /* remove all entries from fmd list */
621 static void filter_fmd_cleanup(struct obd_export *exp)
623 struct filter_export_data *fed = &exp->exp_filter_data;
624 struct filter_mod_data *fmd = NULL, *tmp;
626 cfs_spin_lock(&fed->fed_lock);
627 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
628 cfs_list_del_init(&fmd->fmd_list);
629 filter_fmd_put_nolock(fed, fmd);
631 cfs_spin_unlock(&fed->fed_lock);
634 static int filter_init_export(struct obd_export *exp)
637 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
638 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
640 cfs_spin_lock(&exp->exp_lock);
641 exp->exp_connecting = 1;
642 cfs_spin_unlock(&exp->exp_lock);
643 rc = lut_client_alloc(exp);
645 rc = ldlm_init_export(exp);
650 static int filter_free_server_data(struct obd_device_target *obt)
652 lut_fini(NULL, obt->obt_lut);
653 OBD_FREE_PTR(obt->obt_lut);
657 /* assumes caller is already in kernel ctxt */
658 int filter_update_server_data(struct obd_device *obd)
660 struct file *filp = obd->u.obt.obt_rcvd_filp;
661 struct lr_server_data *lsd = class_server_data(obd);
666 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
667 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
668 le64_to_cpu(lsd->lsd_last_transno));
669 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
670 le64_to_cpu(lsd->lsd_mount_count));
672 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
674 CERROR("error writing lr_server_data: rc = %d\n", rc);
679 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
682 struct filter_obd *filter = &obd->u.filter;
688 if (filter->fo_last_objid_files[group] == NULL) {
689 CERROR("Object seq "LPU64" not fully setup; not updating "
690 "last_objid\n", group);
694 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
695 obd->obd_name, filter->fo_last_objids[group], group);
697 tmp = cpu_to_le64(filter->fo_last_objids[group]);
698 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
699 &tmp, sizeof(tmp), &off, force_sync);
701 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
705 extern int ost_handle(struct ptlrpc_request *req);
706 /* assumes caller has already in kernel ctxt */
707 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
709 struct filter_obd *filter = &obd->u.filter;
710 struct lr_server_data *lsd;
711 struct lsd_client_data *lcd = NULL;
712 struct inode *inode = filp->f_dentry->d_inode;
713 unsigned long last_rcvd_size = i_size_read(inode);
714 struct lu_target *lut;
721 /* ensure padding in the struct is the correct size */
722 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
723 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
724 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
725 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
727 /* allocate and initialize lu_target */
731 rc = lut_init(NULL, lut, obd, NULL);
734 lsd = class_server_data(obd);
735 if (last_rcvd_size == 0) {
736 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
738 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
739 lsd->lsd_last_transno = 0;
740 mount_count = lsd->lsd_mount_count = 0;
741 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
742 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
743 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
744 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
745 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
746 /* OBD_COMPAT_OST is set in filter_connect_internal when the
747 * MDS first connects and assigns the OST index number. */
748 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
751 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
753 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
757 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
758 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
759 "using the wrong disk %s. Were the "
760 "/dev/ assignments rearranged?\n",
761 obd->obd_uuid.uuid, lsd->lsd_uuid);
762 GOTO(err_lut, rc = -EINVAL);
764 mount_count = le64_to_cpu(lsd->lsd_mount_count);
765 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
767 /* Assume old last_rcvd format unless I_C_LR is set */
768 if (!(lsd->lsd_feature_incompat &
769 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
770 lsd->lsd_last_transno = lsd->lsd_compat14;
772 /* OBD_COMPAT_OST is set in filter_connect_internal when the
773 * MDS first connects and assigns the OST index number. */
774 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
778 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
779 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
780 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
781 ~FILTER_INCOMPAT_SUPP);
782 GOTO(err_lut, rc = -EINVAL);
784 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
785 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
786 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
787 ~FILTER_ROCOMPAT_SUPP);
788 /* Do something like remount filesystem read-only */
789 GOTO(err_lut, rc = -EINVAL);
792 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
794 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
795 obd->obd_name, start_epoch);
796 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
797 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
798 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
799 obd->obd_name, mount_count + 1);
800 CDEBUG(D_INODE, "%s: server data size: %u\n",
801 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
802 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
803 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
804 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
805 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
806 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
807 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
808 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
809 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
810 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
811 le16_to_cpu(lsd->lsd_client_size));
813 if (!obd->obd_replayable) {
814 CWARN("%s: recovery support OFF\n", obd->obd_name);
820 GOTO(err_client, rc = -ENOMEM);
822 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
823 off < last_rcvd_size; cl_idx++) {
825 struct obd_export *exp;
826 struct filter_export_data *fed;
828 /* Don't assume off is incremented properly by
829 * fsfilt_read_record(), in case sizeof(*lcd)
830 * isn't the same as lsd->lsd_client_size. */
831 off = le32_to_cpu(lsd->lsd_client_start) +
832 cl_idx * le16_to_cpu(lsd->lsd_client_size);
833 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
835 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
836 LAST_RCVD, cl_idx, off, rc);
837 break; /* read error shouldn't cause startup to fail */
840 if (lcd->lcd_uuid[0] == '\0') {
841 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
846 check_lcd(obd->obd_name, cl_idx, lcd);
848 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
850 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
851 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
852 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
854 /* These exports are cleaned up by filter_disconnect(), so they
855 * need to be set up like real exports as filter_connect() does.
857 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
859 if (PTR_ERR(exp) == -EALREADY) {
860 /* export already exists, zero out this one */
861 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
865 GOTO(err_client, rc = PTR_ERR(exp));
868 fed = &exp->exp_filter_data;
869 *fed->fed_ted.ted_lcd = *lcd;
870 fed->fed_group = 0; /* will be assigned at connect */
871 filter_export_stats_init(obd, exp, NULL);
872 rc = filter_client_add(obd, exp, cl_idx);
873 /* can't fail for existing client */
874 LASSERTF(rc == 0, "rc = %d\n", rc);
876 /* VBR: set export last committed */
877 exp->exp_last_committed = last_rcvd;
878 cfs_spin_lock(&exp->exp_lock);
879 exp->exp_connecting = 0;
880 exp->exp_in_recovery = 0;
881 cfs_spin_unlock(&exp->exp_lock);
882 obd->obd_max_recoverable_clients++;
883 class_export_put(exp);
885 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
886 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
890 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
892 obd->u.obt.obt_mount_count = mount_count + 1;
893 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
895 /* save it, so mount count and last_transno is current */
896 rc = filter_update_server_data(obd);
898 GOTO(err_client, rc);
903 class_disconnect_exports(obd);
905 filter_free_server_data(&obd->u.obt);
909 static int filter_cleanup_groups(struct obd_device *obd)
911 struct filter_obd *filter = &obd->u.filter;
913 struct dentry *dentry;
917 if (filter->fo_dentry_O_groups != NULL) {
918 for (i = 0; i < filter->fo_group_count; i++) {
919 dentry = filter->fo_dentry_O_groups[i];
923 OBD_FREE(filter->fo_dentry_O_groups,
924 filter->fo_group_count *
925 sizeof(*filter->fo_dentry_O_groups));
926 filter->fo_dentry_O_groups = NULL;
928 if (filter->fo_last_objid_files != NULL) {
929 for (i = 0; i < filter->fo_group_count; i++) {
930 filp = filter->fo_last_objid_files[i];
934 OBD_FREE(filter->fo_last_objid_files,
935 filter->fo_group_count *
936 sizeof(*filter->fo_last_objid_files));
937 filter->fo_last_objid_files = NULL;
939 if (filter->fo_dentry_O_sub != NULL) {
940 for (i = 0; i < filter->fo_group_count; i++) {
941 for (j = 0; j < filter->fo_subdir_count; j++) {
942 dentry = filter->fo_dentry_O_sub[i].dentry[j];
947 OBD_FREE(filter->fo_dentry_O_sub,
948 filter->fo_group_count *
949 sizeof(*filter->fo_dentry_O_sub));
950 filter->fo_dentry_O_sub = NULL;
952 if (filter->fo_last_objids != NULL) {
953 OBD_FREE(filter->fo_last_objids,
954 filter->fo_group_count *
955 sizeof(*filter->fo_last_objids));
956 filter->fo_last_objids = NULL;
958 if (filter->fo_dentry_O != NULL) {
959 f_dput(filter->fo_dentry_O);
960 filter->fo_dentry_O = NULL;
965 static int filter_update_last_group(struct obd_device *obd, int group)
967 struct filter_obd *filter = &obd->u.filter;
968 struct file *filp = NULL;
969 int last_group = 0, rc;
973 if (group <= filter->fo_committed_group)
976 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
980 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
984 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
986 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
990 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
991 obd->obd_name, last_group, group);
995 /* must be sync: bXXXX */
996 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
998 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1002 filter->fo_committed_group = group;
1005 filp_close(filp, 0);
1009 static int filter_read_group_internal(struct obd_device *obd, int group,
1012 struct filter_obd *filter = &obd->u.filter;
1013 __u64 *new_objids = NULL;
1014 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1015 struct dentry **new_groups = NULL;
1016 struct file **new_files = NULL;
1017 struct dentry *dentry;
1019 int old_count = filter->fo_group_count, rc, stage = 0, i;
1023 int len = group + 1;
1025 snprintf(name, 24, "%d", group);
1029 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1031 if (IS_ERR(dentry)) {
1032 CERROR("Cannot lookup expected object group %d: %ld\n",
1033 group, PTR_ERR(dentry));
1034 RETURN(PTR_ERR(dentry));
1037 dentry = simple_mkdir(filter->fo_dentry_O,
1038 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1039 if (IS_ERR(dentry)) {
1040 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1042 RETURN(PTR_ERR(dentry));
1047 snprintf(name, 24, "O/%d/LAST_ID", group);
1049 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1051 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1052 GOTO(cleanup, rc = PTR_ERR(filp));
1056 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1058 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1062 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1063 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1064 if (tmp_subdirs == NULL)
1065 GOTO(cleanup, rc = -ENOMEM);
1068 for (i = 0; i < filter->fo_subdir_count; i++) {
1070 snprintf(dir, sizeof(dir), "d%u", i);
1072 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1073 obd->u.obt.obt_vfsmnt,
1075 if (IS_ERR(tmp_subdirs->dentry[i])) {
1076 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1077 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1082 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1083 tmp_subdirs->dentry[i]);
1087 /* 'group' is an index; we need an array of length 'group + 1' */
1088 if (group + 1 > old_count) {
1089 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1090 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1091 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1092 OBD_ALLOC(new_files, len * sizeof(*new_files));
1094 if (new_objids == NULL || new_subdirs == NULL ||
1095 new_groups == NULL || new_files == NULL)
1096 GOTO(cleanup, rc = -ENOMEM);
1099 memcpy(new_objids, filter->fo_last_objids,
1100 old_count * sizeof(*new_objids));
1101 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1102 old_count * sizeof(*new_subdirs));
1103 memcpy(new_groups, filter->fo_dentry_O_groups,
1104 old_count * sizeof(*new_groups));
1105 memcpy(new_files, filter->fo_last_objid_files,
1106 old_count * sizeof(*new_files));
1108 OBD_FREE(filter->fo_last_objids,
1109 old_count * sizeof(*new_objids));
1110 OBD_FREE(filter->fo_dentry_O_sub,
1111 old_count * sizeof(*new_subdirs));
1112 OBD_FREE(filter->fo_dentry_O_groups,
1113 old_count * sizeof(*new_groups));
1114 OBD_FREE(filter->fo_last_objid_files,
1115 old_count * sizeof(*new_files));
1117 filter->fo_last_objids = new_objids;
1118 filter->fo_dentry_O_sub = new_subdirs;
1119 filter->fo_dentry_O_groups = new_groups;
1120 filter->fo_last_objid_files = new_files;
1121 filter->fo_group_count = len;
1124 filter->fo_dentry_O_groups[group] = dentry;
1125 filter->fo_last_objid_files[group] = filp;
1126 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1127 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1128 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1131 filter_update_last_group(obd, group);
1133 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1134 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1135 rc = filter_update_last_objid(obd, group, 1);
1139 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1140 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1141 obd->obd_name, group, last_objid);
1146 if (new_objids != NULL)
1147 OBD_FREE(new_objids, len * sizeof(*new_objids));
1148 if (new_subdirs != NULL)
1149 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1150 if (new_groups != NULL)
1151 OBD_FREE(new_groups, len * sizeof(*new_groups));
1152 if (new_files != NULL)
1153 OBD_FREE(new_files, len * sizeof(*new_files));
1155 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1156 for (i = 0; i < filter->fo_subdir_count; i++) {
1157 if (tmp_subdirs->dentry[i] != NULL)
1158 dput(tmp_subdirs->dentry[i]);
1160 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1163 filp_close(filp, 0);
1170 static int filter_read_groups(struct obd_device *obd, int last_group,
1173 struct filter_obd *filter = &obd->u.filter;
1174 int old_count, group, rc = 0;
1176 cfs_down(&filter->fo_init_lock);
1177 old_count = filter->fo_group_count;
1178 for (group = old_count; group <= last_group; group++) {
1179 rc = filter_read_group_internal(obd, group, create);
1183 cfs_up(&filter->fo_init_lock);
1187 /* FIXME: object groups */
1188 static int filter_prep_groups(struct obd_device *obd)
1190 struct filter_obd *filter = &obd->u.filter;
1191 struct dentry *O_dentry;
1193 int last_group, rc = 0, cleanup_phase = 0;
1197 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1199 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1200 if (IS_ERR(O_dentry)) {
1201 rc = PTR_ERR(O_dentry);
1202 CERROR("cannot open/create O: rc = %d\n", rc);
1205 filter->fo_dentry_O = O_dentry;
1206 cleanup_phase = 1; /* O_dentry */
1208 /* we have to initialize all groups before first connections from
1209 * clients because they may send create/destroy for any group -bzzz */
1210 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1212 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1213 GOTO(cleanup, rc = PTR_ERR(filp));
1215 cleanup_phase = 2; /* filp */
1217 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1219 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1224 last_group = FID_SEQ_OST_MDT0;
1226 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1227 FID_SEQ_OST_MDT0, last_group);
1228 filter->fo_committed_group = last_group;
1229 rc = filter_read_groups(obd, last_group, 1);
1233 filp_close(filp, 0);
1237 switch (cleanup_phase) {
1239 filp_close(filp, 0);
1241 filter_cleanup_groups(obd);
1242 f_dput(filter->fo_dentry_O);
1243 filter->fo_dentry_O = NULL;
1251 /* setup the object store with correct subdirectories */
1252 static int filter_prep(struct obd_device *obd)
1254 struct lvfs_run_ctxt saved;
1255 struct filter_obd *filter = &obd->u.filter;
1257 struct inode *inode;
1261 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1262 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1263 if (!file || IS_ERR(file)) {
1265 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1269 obd->u.obt.obt_rcvd_filp = file;
1270 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1271 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1272 file->f_dentry->d_inode->i_mode);
1273 GOTO(err_filp, rc = -ENOENT);
1276 inode = file->f_dentry->d_parent->d_inode;
1277 /* We use i_op->unlink directly in filter_vfs_unlink() */
1278 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1279 CERROR("%s: filesystem does not support create/unlink ops\n",
1281 GOTO(err_filp, rc = -EOPNOTSUPP);
1284 rc = filter_init_server_data(obd, file);
1286 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1289 LASSERT(obd->u.obt.obt_lut);
1290 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1292 /* open and create health check io file*/
1293 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1296 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1298 GOTO(err_server_data, rc);
1300 filter->fo_obt.obt_health_check_filp = file;
1301 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1302 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1303 file->f_dentry->d_inode->i_mode);
1304 GOTO(err_health_check, rc = -ENOENT);
1306 rc = lvfs_check_io_health(obd, file);
1308 GOTO(err_health_check, rc);
1310 rc = filter_prep_groups(obd);
1312 GOTO(err_health_check, rc);
1314 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1319 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1320 CERROR("can't close %s after error\n", HEALTH_CHECK);
1321 filter->fo_obt.obt_health_check_filp = NULL;
1323 target_recovery_fini(obd);
1324 filter_free_server_data(&obd->u.obt);
1326 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1327 CERROR("can't close %s after error\n", LAST_RCVD);
1328 obd->u.obt.obt_rcvd_filp = NULL;
1332 /* cleanup the filter: write last used object id to status file */
1333 static void filter_post(struct obd_device *obd)
1335 struct lvfs_run_ctxt saved;
1336 struct filter_obd *filter = &obd->u.filter;
1339 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1340 * best to start a transaction with h_sync, because we removed this
1343 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1344 rc = filter_update_server_data(obd);
1346 CERROR("error writing server data: rc = %d\n", rc);
1348 for (i = 0; i < filter->fo_group_count; i++) {
1349 rc = filter_update_last_objid(obd, i,
1350 (i == filter->fo_group_count - 1));
1352 CERROR("error writing group %d lastobjid: rc = %d\n",
1356 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1357 obd->u.obt.obt_rcvd_filp = NULL;
1359 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1361 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1362 filter->fo_obt.obt_health_check_filp = NULL;
1364 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1366 filter_cleanup_groups(obd);
1367 filter_free_server_data(&obd->u.obt);
1368 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1370 filter_free_capa_keys(filter);
1371 cleanup_capa_hash(filter->fo_capa_hash);
1374 static void filter_set_last_id(struct filter_obd *filter,
1375 obd_id id, obd_seq group)
1377 LASSERT(group <= filter->fo_group_count);
1379 cfs_spin_lock(&filter->fo_objidlock);
1380 filter->fo_last_objids[group] = id;
1381 cfs_spin_unlock(&filter->fo_objidlock);
1384 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1387 LASSERT(group <= filter->fo_group_count);
1388 LASSERT(filter->fo_last_objids != NULL);
1390 /* FIXME: object groups */
1391 cfs_spin_lock(&filter->fo_objidlock);
1392 id = filter->fo_last_objids[group];
1393 cfs_spin_unlock(&filter->fo_objidlock);
1397 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1399 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1403 /* We never dget the object parent, so DON'T dput it either */
1404 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1406 struct filter_obd *filter = &obd->u.filter;
1407 struct filter_subdirs *subdirs;
1409 if (group >= filter->fo_group_count) /* FIXME: object groups */
1410 return ERR_PTR(-EBADF);
1412 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1413 return filter->fo_dentry_O_groups[group];
1415 subdirs = &filter->fo_dentry_O_sub[group];
1416 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1419 /* We never dget the object parent, so DON'T dput it either */
1420 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1423 unsigned long now = jiffies;
1424 struct dentry *dparent = filter_parent(obd, group, objid);
1427 if (IS_ERR(dparent))
1429 if (dparent == NULL)
1430 return ERR_PTR(-ENOENT);
1432 rc = filter_lock_dentry(obd, dparent);
1433 fsfilt_check_slow(obd, now, "parent lock");
1434 return rc ? ERR_PTR(rc) : dparent;
1437 /* We never dget the object parent, so DON'T dput it either */
1438 static void filter_parent_unlock(struct dentry *dparent)
1440 UNLOCK_INODE_MUTEX(dparent->d_inode);
1443 /* How to get files, dentries, inodes from object id's.
1445 * If dir_dentry is passed, the caller has already locked the parent
1446 * appropriately for this operation (normally a write lock). If
1447 * dir_dentry is NULL, we do a read lock while we do the lookup to
1448 * avoid races with create/destroy and such changing the directory
1449 * internal to the filesystem code. */
1450 struct dentry *filter_fid2dentry(struct obd_device *obd,
1451 struct dentry *dir_dentry,
1452 obd_seq group, obd_id id)
1454 struct dentry *dparent = dir_dentry;
1455 struct dentry *dchild;
1460 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1461 obd->u.filter.fo_destroys_in_progress == 0) {
1462 /* don't fail lookups for orphan recovery, it causes
1463 * later LBUGs when objects still exist during precreate */
1464 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1465 RETURN(ERR_PTR(-ENOENT));
1468 CERROR("fatal: invalid object id 0\n");
1469 RETURN(ERR_PTR(-ESTALE));
1472 len = sprintf(name, LPU64, id);
1473 if (dir_dentry == NULL) {
1474 dparent = filter_parent_lock(obd, group, id);
1475 if (IS_ERR(dparent)) {
1476 CERROR("%s: error getting object "POSTID
1477 " parent: rc %ld\n", obd->obd_name,
1478 id, group, PTR_ERR(dparent));
1482 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1483 dparent->d_name.len, dparent->d_name.name, name);
1484 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1485 dchild = lookup_one_len(name, dparent, len);
1486 if (dir_dentry == NULL)
1487 filter_parent_unlock(dparent);
1488 if (IS_ERR(dchild)) {
1489 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1490 obd->obd_name, id, group, PTR_ERR(dchild));
1494 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1495 CERROR("%s: got bad object "LPU64" inode %lu\n",
1496 obd->obd_name, id, dchild->d_inode->i_ino);
1498 RETURN(ERR_PTR(-ENOENT));
1501 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1502 name, dchild, atomic_read(&dchild->d_count));
1504 LASSERT(atomic_read(&dchild->d_count) > 0);
1509 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1510 obd_id group, struct lustre_handle *lockh)
1512 int flags = LDLM_AST_DISCARD_DATA, rc;
1513 struct ldlm_res_id res_id;
1514 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1517 osc_build_res_name(objid, group, &res_id);
1518 /* Tell the clients that the object is gone now and that they should
1519 * throw away any cached pages. */
1520 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1521 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1522 ldlm_completion_ast, NULL, NULL, 0, NULL,
1529 static void filter_fini_destroy(struct obd_device *obd,
1530 struct lustre_handle *lockh)
1532 if (lustre_handle_is_used(lockh))
1533 ldlm_lock_decref(lockh, LCK_PW);
1536 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1537 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1538 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1539 * reverse. Caller must take i_sem before starting the transaction and we
1540 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1541 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1542 struct vfsmount *mnt)
1547 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1548 * itself we already hold dir->i_mutex for child create/unlink ops */
1549 LASSERT(dentry->d_inode != NULL);
1550 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1551 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1555 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1556 GOTO(out, rc = -ENOENT);
1558 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1563 GOTO(out, rc = -EPERM);
1565 /* check_sticky() */
1566 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1567 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1568 IS_IMMUTABLE(dentry->d_inode))
1569 GOTO(out, rc = -EPERM);
1571 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1572 * that was done because of journal_start (which is already done
1573 * here) or some other ordering issue. */
1574 ll_vfs_dq_init(dir);
1576 rc = ll_security_inode_unlink(dir, dentry, mnt);
1580 rc = dir->i_op->unlink(dir, dentry);
1582 /* need to drop i_mutex before we lose inode reference */
1583 UNLOCK_INODE_MUTEX(dentry->d_inode);
1590 /* Caller must hold LCK_PW on parent and push us into kernel context.
1591 * Caller must hold child i_mutex, we drop it always.
1592 * Caller is also required to ensure that dchild->d_inode exists. */
1593 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1594 obd_seq group, struct dentry *dparent,
1595 struct dentry *dchild)
1597 struct inode *inode = dchild->d_inode;
1600 /* There should be 2 references to the inode:
1601 * 1) taken by filter_prepare_destroy
1602 * 2) taken by filter_destroy */
1603 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1604 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1605 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1606 (unsigned long)inode->i_nlink,
1607 atomic_read(&inode->i_count));
1610 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1612 CERROR("error unlinking objid %.*s: rc %d\n",
1613 dchild->d_name.len, dchild->d_name.name, rc);
1617 struct filter_intent_args {
1618 struct ldlm_lock **victim;
1623 static enum interval_iter filter_intent_cb(struct interval_node *n,
1626 struct ldlm_interval *node = (struct ldlm_interval *)n;
1627 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1628 __u64 size = arg->size;
1629 struct ldlm_lock **v = arg->victim;
1630 struct ldlm_lock *lck;
1632 /* If the interval is lower than the current file size,
1634 if (interval_high(n) <= size)
1635 return INTERVAL_ITER_STOP;
1637 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1638 /* Don't send glimpse ASTs to liblustre clients.
1639 * They aren't listening for them, and they do
1640 * entirely synchronous I/O anyways. */
1641 if (lck->l_export == NULL ||
1642 lck->l_export->exp_libclient == 1)
1645 if (*arg->liblustre)
1646 *arg->liblustre = 0;
1649 *v = LDLM_LOCK_GET(lck);
1650 } else if ((*v)->l_policy_data.l_extent.start <
1651 lck->l_policy_data.l_extent.start) {
1652 LDLM_LOCK_RELEASE(*v);
1653 *v = LDLM_LOCK_GET(lck);
1656 /* the same policy group - every lock has the
1657 * same extent, so needn't do it any more */
1661 return INTERVAL_ITER_CONT;
1664 static int filter_intent_policy(struct ldlm_namespace *ns,
1665 struct ldlm_lock **lockp, void *req_cookie,
1666 ldlm_mode_t mode, int flags, void *data)
1668 CFS_LIST_HEAD(rpc_list);
1669 struct ptlrpc_request *req = req_cookie;
1670 struct ldlm_lock *lock = *lockp, *l = NULL;
1671 struct ldlm_resource *res = lock->l_resource;
1672 ldlm_processing_policy policy;
1673 struct ost_lvb *res_lvb, *reply_lvb;
1674 struct ldlm_reply *rep;
1676 int idx, rc, tmpflags = 0, only_liblustre = 1;
1677 struct ldlm_interval_tree *tree;
1678 struct filter_intent_args arg;
1679 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1680 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1681 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1684 policy = ldlm_get_processing_policy(res);
1685 LASSERT(policy != NULL);
1686 LASSERT(req != NULL);
1688 rc = lustre_pack_reply(req, 3, repsize, NULL);
1690 RETURN(req->rq_status = rc);
1692 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1693 LASSERT(rep != NULL);
1695 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1696 sizeof(*reply_lvb));
1697 LASSERT(reply_lvb != NULL);
1699 //fixup_handle_for_resent_req(req, lock, &lockh);
1701 /* Call the extent policy function to see if our request can be
1702 * granted, or is blocked.
1703 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1704 * lock, and should not be granted if the lock will be blocked.
1707 LASSERT(ns == ldlm_res_to_ns(res));
1709 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1710 check_res_locked(res);
1712 /* FIXME: we should change the policy function slightly, to not make
1713 * this list at all, since we just turn around and free it */
1714 while (!cfs_list_empty(&rpc_list)) {
1715 struct ldlm_lock *wlock =
1716 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1718 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1719 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1720 lock->l_flags &= ~LDLM_FL_CP_REQD;
1721 cfs_list_del_init(&wlock->l_cp_ast);
1722 LDLM_LOCK_RELEASE(wlock);
1725 /* The lock met with no resistance; we're finished. */
1726 if (rc == LDLM_ITER_CONTINUE) {
1727 /* do not grant locks to the liblustre clients: they cannot
1728 * handle ASTs robustly. We need to do this while still
1729 * holding lr_lock to avoid the lock remaining on the res_link
1730 * list (and potentially being added to l_pending_list by an
1731 * AST) when we are going to drop this lock ASAP. */
1732 if (lock->l_export->exp_libclient ||
1733 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1734 ldlm_resource_unlink_lock(lock);
1735 err = ELDLM_LOCK_ABORTED;
1737 err = ELDLM_LOCK_REPLACED;
1743 /* Do not grant any lock, but instead send GL callbacks. The extent
1744 * policy nicely created a list of all PW locks for us. We will choose
1745 * the highest of those which are larger than the size in the LVB, if
1746 * any, and perform a glimpse callback. */
1747 res_lvb = res->lr_lvb_data;
1748 LASSERT(res_lvb != NULL);
1749 *reply_lvb = *res_lvb;
1752 * lr_lock guarantees that no new locks are granted, and,
1753 * therefore, that res->lr_lvb_data cannot increase beyond the
1754 * end of already granted lock. As a result, it is safe to
1755 * check against "stale" reply_lvb->lvb_size value without
1758 arg.size = reply_lvb->lvb_size;
1760 arg.liblustre = &only_liblustre;
1761 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1762 tree = &res->lr_itree[idx];
1763 if (tree->lit_mode == LCK_PR)
1766 interval_iterate_reverse(tree->lit_root,
1767 filter_intent_cb, &arg);
1771 /* There were no PW locks beyond the size in the LVB; finished. */
1773 if (only_liblustre) {
1774 /* If we discovered a liblustre client with a PW lock,
1775 * however, the LVB may be out of date! The LVB is
1776 * updated only on glimpse (which we don't do for
1777 * liblustre clients) and cancel (which the client
1778 * obviously has not yet done). So if it has written
1779 * data but kept the lock, the LVB is stale and needs
1780 * to be updated from disk.
1782 * Of course, this will all disappear when we switch to
1783 * taking liblustre locks on the OST. */
1784 ldlm_res_lvbo_update(res, NULL, 1);
1786 RETURN(ELDLM_LOCK_ABORTED);
1790 * This check is for lock taken in filter_prepare_destroy() that does
1791 * not have l_glimpse_ast set. So the logic is: if there is a lock
1792 * with no l_glimpse_ast set, this object is being destroyed already.
1794 * Hence, if you are grabbing DLM locks on the server, always set
1795 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1797 if (l->l_glimpse_ast == NULL) {
1798 /* We are racing with unlink(); just return -ENOENT */
1799 rep->lock_policy_res1 = -ENOENT;
1803 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1804 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1807 *reply_lvb = *res_lvb;
1811 LDLM_LOCK_RELEASE(l);
1813 RETURN(ELDLM_LOCK_ABORTED);
1817 * per-obd_device iobuf pool.
1819 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1820 * allocations in write-path has to be minimized (see bug 5137).
1822 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1823 * OST threads (see ost_thread_{init,done}()).
1825 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1826 * at the OST layer there are only (potentially) multiple obd_device of type
1827 * unknown at the time of OST thread creation.
1829 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1830 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1831 * it's very own iobuf.
1835 * filter_kiobuf_pool_init()
1837 * filter_kiobuf_pool_done()
1839 * filter_iobuf_get()
1841 * operate on this array. They are "generic" in a sense that they don't depend
1842 * on actual type of iobuf's (the latter depending on Linux kernel version).
1846 * destroy pool created by filter_iobuf_pool_init
1848 static void filter_iobuf_pool_done(struct filter_obd *filter)
1850 struct filter_iobuf **pool;
1855 pool = filter->fo_iobuf_pool;
1857 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1858 if (pool[i] != NULL)
1859 filter_free_iobuf(pool[i]);
1861 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1862 filter->fo_iobuf_pool = NULL;
1867 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1869 struct filter_obd *filter = &obd->u.filter;
1870 struct sptlrpc_rule_set tmp_rset;
1873 sptlrpc_rule_set_init(&tmp_rset);
1874 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1876 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1881 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1883 cfs_write_lock(&filter->fo_sptlrpc_lock);
1884 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1885 filter->fo_sptlrpc_rset = tmp_rset;
1886 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1892 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1894 static int filter_iobuf_pool_init(struct filter_obd *filter)
1901 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1903 if (filter->fo_iobuf_pool == NULL)
1906 filter->fo_iobuf_count = OSS_THREADS_MAX;
1911 /* Return iobuf allocated for @thread_id. We don't know in advance how
1912 * many threads there will be so we allocate a large empty array and only
1913 * fill in those slots that are actually in use.
1914 * If we haven't allocated a pool entry for this thread before, do so now. */
1915 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1917 int thread_id = (oti && oti->oti_thread) ?
1918 oti->oti_thread->t_id : -1;
1919 struct filter_iobuf *pool = NULL;
1920 struct filter_iobuf **pool_place = NULL;
1922 if (thread_id >= 0) {
1923 LASSERT(thread_id < filter->fo_iobuf_count);
1924 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1927 if (unlikely(pool == NULL)) {
1928 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1929 PTLRPC_MAX_BRW_PAGES);
1930 if (pool_place != NULL)
1937 /* mount the file system (secretly). lustre_cfg parameters are:
1940 * 3 = flags: failover=f, failout=n
1943 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1946 struct filter_obd *filter = &obd->u.filter;
1947 struct vfsmount *mnt;
1948 struct lustre_mount_info *lmi;
1949 struct obd_uuid uuid;
1953 struct request_queue *q;
1957 if (lcfg->lcfg_bufcount < 3 ||
1958 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1959 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1962 lmi = server_get_mount(obd->obd_name);
1964 /* We already mounted in lustre_fill_super.
1965 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1966 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1968 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1970 /* gets recovery timeouts from mount data */
1971 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1972 obd->obd_recovery_timeout =
1973 lsi->lsi_lmd->lmd_recovery_time_soft;
1974 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1975 obd->obd_recovery_time_hard =
1976 lsi->lsi_lmd->lmd_recovery_time_hard;
1978 /* old path - used by lctl */
1979 CERROR("Using old MDS mount method\n");
1980 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1981 MS_NOATIME|MS_NODIRATIME,
1982 lustre_cfg_string(lcfg, 1), option);
1985 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1986 lustre_cfg_string(lcfg, 1), rc);
1990 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1992 if (IS_ERR(obd->obd_fsops))
1993 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1995 rc = filter_iobuf_pool_init(filter);
1999 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2000 CERROR("%s: Underlying device is marked as read-only. "
2001 "Setup failed\n", obd->obd_name);
2002 GOTO(err_ops, rc = -EROFS);
2005 /* failover is the default */
2006 obd->obd_replayable = 1;
2008 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2009 str = lustre_cfg_string(lcfg, 3);
2010 if (strchr(str, 'n')) {
2011 CWARN("%s: recovery disabled\n", obd->obd_name);
2012 obd->obd_replayable = 0;
2016 obd->u.obt.obt_vfsmnt = mnt;
2017 obd->u.obt.obt_sb = mnt->mnt_sb;
2018 obd->u.obt.obt_magic = OBT_MAGIC;
2019 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2020 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2022 fsfilt_setup(obd, obd->u.obt.obt_sb);
2024 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2025 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2026 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2027 obd->obd_lvfs_ctxt.fs = get_ds();
2028 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2030 cfs_init_mutex(&filter->fo_init_lock);
2031 filter->fo_committed_group = 0;
2032 filter->fo_destroys_in_progress = 0;
2033 for (i = 0; i < 32; i++)
2034 cfs_sema_init(&filter->fo_create_locks[i], 1);
2036 cfs_spin_lock_init(&filter->fo_objidlock);
2037 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2038 cfs_sema_init(&filter->fo_alloc_lock, 1);
2039 init_brw_stats(&filter->fo_filter_stats);
2040 cfs_spin_lock_init(&filter->fo_flags_lock);
2041 filter->fo_read_cache = 1; /* enable read-only cache by default */
2042 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2043 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2044 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2045 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2046 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2047 filter_slc_set(filter); /* initialize sync on lock cancel */
2049 rc = filter_prep(obd);
2053 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2054 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2056 filter->fo_fl_oss_capa = 1;
2058 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2059 filter->fo_capa_hash = init_capa_hash();
2060 if (filter->fo_capa_hash == NULL)
2061 GOTO(err_post, rc = -ENOMEM);
2063 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2064 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2065 LDLM_NAMESPACE_SERVER,
2066 LDLM_NAMESPACE_GREEDY,
2068 if (obd->obd_namespace == NULL)
2069 GOTO(err_post, rc = -ENOMEM);
2070 obd->obd_namespace->ns_lvbp = obd;
2071 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2072 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2074 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2075 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2077 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2079 CERROR("failed to setup llogging subsystems\n");
2083 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2084 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2085 /* do this after llog being initialized */
2086 filter_adapt_sptlrpc_conf(obd, 1);
2088 rc = lquota_setup(filter_quota_interface_ref, obd);
2092 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2093 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2094 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2095 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2096 "for larger I/O requests: max_sectors = %u "
2097 "could be up to max_hw_sectors=%u\n",
2098 obd->obd_name, mnt->mnt_sb->s_id,
2099 queue_max_sectors(q), queue_max_hw_sectors(q));
2101 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2102 if (uuid_ptr != NULL) {
2103 class_uuid_unparse(uuid_ptr, &uuid);
2109 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2110 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2111 obd->obd_name, label ?: str, lmi ? "on " : "",
2112 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2113 obd->obd_replayable ? "enabled" : "disabled");
2115 if (obd->obd_recovering)
2116 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2117 "or until %d client%s reconnect%s\n",
2119 obd->obd_recovery_timeout / 60,
2120 obd->obd_recovery_timeout % 60,
2121 obd->obd_max_recoverable_clients,
2122 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2123 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2131 fsfilt_put_ops(obd->obd_fsops);
2132 filter_iobuf_pool_done(filter);
2134 server_put_mount(obd->obd_name, mnt);
2135 obd->u.obt.obt_sb = 0;
2139 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2141 struct lprocfs_static_vars lvars;
2146 CLASSERT(offsetof(struct obd_device, u.obt) ==
2147 offsetof(struct obd_device, u.filter.fo_obt));
2149 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2152 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2153 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2156 addr = (unsigned long)cfs_page_address(page);
2157 clear_page((void *)addr);
2159 /* lprocfs must be setup before the filter so state can be safely added
2160 * to /proc incrementally as the filter is setup */
2161 lprocfs_filter_init_vars(&lvars);
2162 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2163 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2164 /* Init obdfilter private stats here */
2165 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2166 LPROCFS_CNTR_AVGMINMAX,
2167 "read_bytes", "bytes");
2168 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2169 LPROCFS_CNTR_AVGMINMAX,
2170 "write_bytes", "bytes");
2171 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2172 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2173 "get_page", "usec");
2174 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2175 LPROCFS_CNTR_AVGMINMAX,
2176 "get_page_failures", "num");
2177 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2178 LPROCFS_CNTR_AVGMINMAX,
2179 "cache_access", "pages");
2180 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2181 LPROCFS_CNTR_AVGMINMAX,
2182 "cache_hit", "pages");
2183 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2184 LPROCFS_CNTR_AVGMINMAX,
2185 "cache_miss", "pages");
2187 lproc_filter_attach_seqstat(obd);
2188 obd->obd_proc_exports_entry = lprocfs_register("exports",
2189 obd->obd_proc_entry,
2191 if (IS_ERR(obd->obd_proc_exports_entry)) {
2192 rc = PTR_ERR(obd->obd_proc_exports_entry);
2193 CERROR("error %d setting up lprocfs for %s\n",
2195 obd->obd_proc_exports_entry = NULL;
2198 if (obd->obd_proc_exports_entry)
2199 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2200 lprocfs_nid_stats_clear_read,
2201 lprocfs_nid_stats_clear_write, obd, NULL);
2203 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2204 LUSTRE_CFG_BUFLEN(lcfg, 4));
2205 rc = filter_common_setup(obd, lcfg, (void *)addr);
2206 OBD_PAGE_FREE(page);
2209 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2210 lprocfs_free_per_client_stats(obd);
2211 lprocfs_free_obd_stats(obd);
2212 lprocfs_obd_cleanup(obd);
2218 static struct llog_operations filter_mds_ost_repl_logops;
2220 static struct llog_operations filter_size_orig_logops = {
2221 .lop_setup = llog_obd_origin_setup,
2222 .lop_cleanup = llog_obd_origin_cleanup,
2223 .lop_add = llog_obd_origin_add
2226 static int filter_olg_fini(struct obd_llog_group *olg)
2228 struct llog_ctxt *ctxt;
2229 int rc = 0, rc2 = 0;
2232 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2234 rc = llog_cleanup(ctxt);
2236 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2238 rc2 = llog_cleanup(ctxt);
2243 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2245 rc2 = llog_cleanup(ctxt);
2254 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2255 struct obd_device *tgt)
2260 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2261 &filter_mds_ost_repl_logops);
2265 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2266 &filter_size_orig_logops);
2272 filter_olg_fini(olg);
2277 * Init the default olg, which is embeded in the obd_device, for filter.
2280 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2281 struct obd_device *tgt)
2283 struct filter_obd *filter = &obd->u.filter;
2284 struct llog_ctxt *ctxt;
2288 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2289 if (!filter->fo_lcm)
2292 filter_mds_ost_repl_logops = llog_client_ops;
2293 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2294 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2295 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2297 rc = filter_olg_init(obd, olg, tgt);
2299 GOTO(cleanup_lcm, rc);
2301 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2304 GOTO(cleanup_olg, rc);
2306 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2308 CERROR("Can't get ctxt for %p:%x\n", olg,
2309 LLOG_MDS_OST_REPL_CTXT);
2310 GOTO(cleanup_olg, rc = -ENODEV);
2312 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2313 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2314 llog_ctxt_put(ctxt);
2318 filter_olg_fini(olg);
2320 llog_recov_thread_fini(filter->fo_lcm, 1);
2321 filter->fo_lcm = NULL;
2326 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2327 struct obd_device *tgt, int *index)
2329 struct filter_obd *filter = &obd->u.filter;
2330 struct llog_ctxt *ctxt;
2334 LASSERT(olg != NULL);
2335 if (olg == &obd->obd_olg)
2336 return filter_default_olg_init(obd, olg, tgt);
2338 LASSERT(filter->fo_lcm != NULL);
2339 rc = filter_olg_init(obd, olg, tgt);
2342 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2344 CERROR("Can't get ctxt for %p:%x\n", olg,
2345 LLOG_MDS_OST_REPL_CTXT);
2346 filter_olg_fini(olg);
2349 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2350 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2351 llog_ctxt_put(ctxt);
2355 static int filter_llog_finish(struct obd_device *obd, int count)
2357 struct filter_obd *filter = &obd->u.filter;
2358 struct llog_ctxt *ctxt;
2361 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2364 * Make sure that no cached llcds left in recov_thread.
2365 * We actually do sync in disconnect time, but disconnect
2366 * may not come being marked rq_no_resend = 1.
2368 llog_sync(ctxt, NULL);
2371 * Balance class_import_get() in llog_receptor_accept().
2372 * This is safe to do, as llog is already synchronized
2373 * and its import may go.
2375 cfs_mutex_down(&ctxt->loc_sem);
2376 if (ctxt->loc_imp) {
2377 class_import_put(ctxt->loc_imp);
2378 ctxt->loc_imp = NULL;
2380 cfs_mutex_up(&ctxt->loc_sem);
2381 llog_ctxt_put(ctxt);
2384 if (filter->fo_lcm) {
2385 cfs_mutex_down(&ctxt->loc_sem);
2386 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2387 filter->fo_lcm = NULL;
2388 cfs_mutex_up(&ctxt->loc_sem);
2390 RETURN(filter_olg_fini(&obd->obd_olg));
2394 * Find the group llog according to group index in the llog group list.
2396 static struct obd_llog_group *
2397 filter_find_olg_internal(struct filter_obd *filter, int group)
2399 struct obd_llog_group *olg;
2401 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2402 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2403 if (olg->olg_seq == group)
2410 * Find the group llog according to group index on the filter
2412 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2414 struct obd_llog_group *olg = NULL;
2415 struct filter_obd *filter;
2417 filter = &obd->u.filter;
2419 if (group == FID_SEQ_LLOG)
2420 RETURN(&obd->obd_olg);
2422 cfs_spin_lock(&filter->fo_llog_list_lock);
2423 olg = filter_find_olg_internal(filter, group);
2424 cfs_spin_unlock(&filter->fo_llog_list_lock);
2429 * Find the llog_group of the filter according to the group. If it can not
2430 * find, create the llog_group, which only happens when mds is being synced
2433 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2435 struct obd_llog_group *olg = NULL;
2436 struct filter_obd *filter;
2439 filter = &obd->u.filter;
2441 if (group == FID_SEQ_LLOG)
2442 RETURN(&obd->obd_olg);
2444 cfs_spin_lock(&filter->fo_llog_list_lock);
2445 olg = filter_find_olg_internal(filter, group);
2447 if (olg->olg_initializing) {
2448 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2450 GOTO(out_unlock, olg);
2455 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2457 llog_group_init(olg, group);
2458 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2459 olg->olg_initializing = 1;
2460 cfs_spin_unlock(&filter->fo_llog_list_lock);
2462 rc = obd_llog_init(obd, olg, obd, NULL);
2464 cfs_spin_lock(&filter->fo_llog_list_lock);
2465 cfs_list_del(&olg->olg_list);
2466 cfs_spin_unlock(&filter->fo_llog_list_lock);
2468 GOTO(out, olg = ERR_PTR(-ENOMEM));
2470 cfs_spin_lock(&filter->fo_llog_list_lock);
2471 olg->olg_initializing = 0;
2472 cfs_spin_unlock(&filter->fo_llog_list_lock);
2473 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2474 obd->obd_name, group, olg);
2479 cfs_spin_unlock(&filter->fo_llog_list_lock);
2483 static int filter_llog_connect(struct obd_export *exp,
2484 struct llogd_conn_body *body)
2486 struct obd_device *obd = exp->exp_obd;
2487 struct llog_ctxt *ctxt;
2488 struct obd_llog_group *olg;
2492 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2493 obd->obd_name, body->lgdc_logid.lgl_oid,
2494 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2496 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2498 CERROR(" %s: can not find olg of group %d\n",
2499 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2502 llog_group_set_export(olg, exp);
2504 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2505 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2506 body->lgdc_ctxt_idx);
2508 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2509 obd->obd_name, body->lgdc_logid.lgl_oid,
2510 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2512 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2513 obd->u.filter.fo_mds_ost_sync = 1;
2514 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2515 rc = llog_connect(ctxt, &body->lgdc_logid,
2516 &body->lgdc_gen, NULL);
2517 llog_ctxt_put(ctxt);
2519 CERROR("failed to connect rc %d idx %d\n", rc,
2520 body->lgdc_ctxt_idx);
2525 static int filter_llog_preclean(struct obd_device *obd)
2527 struct obd_llog_group *olg, *tmp;
2528 struct filter_obd *filter;
2529 cfs_list_t remove_list;
2533 rc = obd_llog_finish(obd, 0);
2535 CERROR("failed to cleanup llogging subsystem\n");
2537 filter = &obd->u.filter;
2538 CFS_INIT_LIST_HEAD(&remove_list);
2540 cfs_spin_lock(&filter->fo_llog_list_lock);
2541 while (!cfs_list_empty(&filter->fo_llog_list)) {
2542 olg = cfs_list_entry(filter->fo_llog_list.next,
2543 struct obd_llog_group, olg_list);
2544 cfs_list_del(&olg->olg_list);
2545 cfs_list_add(&olg->olg_list, &remove_list);
2547 cfs_spin_unlock(&filter->fo_llog_list_lock);
2549 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2550 cfs_list_del_init(&olg->olg_list);
2551 rc = filter_olg_fini(olg);
2553 CERROR("failed to cleanup llogging subsystem for %u\n",
2561 static int filter_precleanup(struct obd_device *obd,
2562 enum obd_cleanup_stage stage)
2568 case OBD_CLEANUP_EARLY:
2570 case OBD_CLEANUP_EXPORTS:
2571 /* Stop recovery before namespace cleanup. */
2572 target_recovery_fini(obd);
2573 rc = filter_llog_preclean(obd);
2579 static int filter_cleanup(struct obd_device *obd)
2581 struct filter_obd *filter = &obd->u.filter;
2585 LCONSOLE_WARN("%s: shutting down for failover; client state "
2586 "will be preserved.\n", obd->obd_name);
2588 obd_exports_barrier(obd);
2589 obd_zombie_barrier();
2591 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2592 lprocfs_free_per_client_stats(obd);
2593 lprocfs_free_obd_stats(obd);
2594 lprocfs_obd_cleanup(obd);
2595 lquota_cleanup(filter_quota_interface_ref, obd);
2597 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2598 obd->obd_namespace = NULL;
2600 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2602 if (obd->u.obt.obt_sb == NULL)
2607 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2608 shrink_dcache_sb(obd->u.obt.obt_sb);
2610 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2611 obd->u.obt.obt_sb = NULL;
2613 fsfilt_put_ops(obd->obd_fsops);
2615 filter_iobuf_pool_done(filter);
2617 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2622 static int filter_connect_internal(struct obd_export *exp,
2623 struct obd_connect_data *data,
2626 struct filter_export_data *fed = &exp->exp_filter_data;
2631 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2632 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2633 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2634 data->ocd_connect_flags, data->ocd_version,
2635 data->ocd_grant, data->ocd_index);
2637 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2638 CWARN("!!! This export (nid %s) used object group %d "
2639 "earlier; now it's trying to use group %d! This could "
2640 "be a bug in the MDS. Please report to "
2641 "http://bugzilla.lustre.org/\n",
2642 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2645 fed->fed_group = data->ocd_group;
2647 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2648 exp->exp_connect_flags = data->ocd_connect_flags;
2649 data->ocd_version = LUSTRE_VERSION_CODE;
2651 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2652 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2653 data->ocd_connect_flags & OBD_CONNECT_MDS))
2656 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2657 struct filter_obd *filter = &exp->exp_obd->u.filter;
2658 obd_size left, want;
2660 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2661 left = filter_grant_space_left(exp);
2662 want = data->ocd_grant;
2663 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2664 data->ocd_grant = fed->fed_grant;
2665 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2667 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2668 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2669 exp->exp_client_uuid.uuid, exp,
2670 data->ocd_grant, want, left);
2672 filter->fo_tot_granted_clients ++;
2675 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2676 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2677 int index = le32_to_cpu(lsd->lsd_ost_index);
2679 if (!(lsd->lsd_feature_compat &
2680 cpu_to_le32(OBD_COMPAT_OST))) {
2681 /* this will only happen on the first connect */
2682 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2683 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2684 /* sync is not needed here as filter_client_add will
2685 * set exp_need_sync flag */
2686 filter_update_server_data(exp->exp_obd);
2687 } else if (index != data->ocd_index) {
2688 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2689 " %u doesn't match actual OST index"
2690 " %u in last_rcvd file, bad "
2692 obd_export_nid2str(exp), index,
2696 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2697 * FIXME: We don't strictly need the COMPAT flag for that,
2698 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2699 * FIXME: We needed it for the index, as index 0 is valid. */
2702 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2703 data->ocd_brw_size = 65536;
2704 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2705 data->ocd_brw_size = min(data->ocd_brw_size,
2706 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2707 if (data->ocd_brw_size == 0) {
2708 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2709 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2710 "ocd_brw_size is unexpectedly zero, "
2711 "network data corruption?"
2712 "Refusing connection of this client\n",
2713 exp->exp_obd->obd_name,
2714 exp->exp_client_uuid.uuid,
2715 exp, data->ocd_connect_flags, data->ocd_version,
2716 data->ocd_grant, data->ocd_index);
2721 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2722 __u32 cksum_types = data->ocd_cksum_types;
2724 /* The client set in ocd_cksum_types the checksum types it
2725 * supports. We have to mask off the algorithms that we don't
2727 if (cksum_types & OBD_CKSUM_ALL)
2728 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2730 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2732 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2733 "%x\n", exp->exp_obd->obd_name,
2734 obd_export_nid2str(exp), cksum_types,
2735 data->ocd_cksum_types);
2737 /* This client does not support OBD_CONNECT_CKSUM
2738 * fall back to CRC32 */
2739 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2740 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2741 exp->exp_obd->obd_name,
2742 obd_export_nid2str(exp));
2748 static int filter_reconnect(const struct lu_env *env,
2749 struct obd_export *exp, struct obd_device *obd,
2750 struct obd_uuid *cluuid,
2751 struct obd_connect_data *data,
2757 if (exp == NULL || obd == NULL || cluuid == NULL)
2760 rc = filter_connect_internal(exp, data, 1);
2762 filter_export_stats_init(obd, exp, localdata);
2767 /* nearly identical to mds_connect */
2768 static int filter_connect(const struct lu_env *env,
2769 struct obd_export **exp, struct obd_device *obd,
2770 struct obd_uuid *cluuid,
2771 struct obd_connect_data *data, void *localdata)
2773 struct lvfs_run_ctxt saved;
2774 struct lustre_handle conn = { 0 };
2775 struct obd_export *lexp;
2780 if (exp == NULL || obd == NULL || cluuid == NULL)
2783 rc = class_connect(&conn, obd, cluuid);
2786 lexp = class_conn2export(&conn);
2787 LASSERT(lexp != NULL);
2789 rc = filter_connect_internal(lexp, data, 0);
2793 filter_export_stats_init(obd, lexp, localdata);
2794 if (obd->obd_replayable) {
2795 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2797 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2798 rc = filter_client_add(obd, lexp, -1);
2803 group = data->ocd_group;
2805 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2806 obd->obd_name, lexp->exp_handle.h_cookie, group);
2808 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2809 rc = filter_read_groups(obd, group, 1);
2810 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2812 CERROR("can't read group %u\n", group);
2820 class_disconnect(lexp);
2829 /* Do extra sanity checks for grant accounting. We do this at connect,
2830 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2831 * always get rid of it or turn it off when we know accounting is good. */
2832 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2834 struct filter_export_data *fed;
2835 struct obd_export *exp;
2836 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2837 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2838 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2840 if (cfs_list_empty(&obd->obd_exports))
2843 /* We don't want to do this for large machines that do lots of
2844 mounts or unmounts. It burns... */
2845 if (obd->obd_num_exports > 100)
2848 cfs_spin_lock(&obd->obd_osfs_lock);
2849 cfs_spin_lock(&obd->obd_dev_lock);
2850 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2852 fed = &exp->exp_filter_data;
2853 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2856 if (maxsize > 0) { /* we may not have done a statfs yet */
2857 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2858 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2859 exp->exp_client_uuid.uuid, exp,
2860 fed->fed_grant, fed->fed_pending, maxsize);
2861 LASSERTF(fed->fed_dirty <= maxsize,
2862 "%s: cli %s/%p %ld > "LPU64"\n", func,
2863 exp->exp_client_uuid.uuid, exp,
2864 fed->fed_dirty, maxsize);
2867 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2868 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2869 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2871 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2872 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2873 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2874 tot_granted += fed->fed_grant + fed->fed_pending;
2875 tot_pending += fed->fed_pending;
2876 tot_dirty += fed->fed_dirty;
2878 fo_tot_granted = obd->u.filter.fo_tot_granted;
2879 fo_tot_pending = obd->u.filter.fo_tot_pending;
2880 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2881 cfs_spin_unlock(&obd->obd_dev_lock);
2882 cfs_spin_unlock(&obd->obd_osfs_lock);
2884 /* Do these assertions outside the spinlocks so we don't kill system */
2885 if (tot_granted != fo_tot_granted)
2886 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2887 func, tot_granted, fo_tot_granted);
2888 if (tot_pending != fo_tot_pending)
2889 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2890 func, tot_pending, fo_tot_pending);
2891 if (tot_dirty != fo_tot_dirty)
2892 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2893 func, tot_dirty, fo_tot_dirty);
2894 if (tot_pending > tot_granted)
2895 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2896 func, tot_pending, tot_granted);
2897 if (tot_granted > maxsize)
2898 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2899 func, tot_granted, maxsize);
2900 if (tot_dirty > maxsize)
2901 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2902 func, tot_dirty, maxsize);
2905 /* Remove this client from the grant accounting totals. We also remove
2906 * the export from the obd device under the osfs and dev locks to ensure
2907 * that the filter_grant_sanity_check() calculations are always valid.
2908 * The client should do something similar when it invalidates its import. */
2909 static void filter_grant_discard(struct obd_export *exp)
2911 struct obd_device *obd = exp->exp_obd;
2912 struct filter_obd *filter = &obd->u.filter;
2913 struct filter_export_data *fed = &exp->exp_filter_data;
2915 cfs_spin_lock(&obd->obd_osfs_lock);
2916 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2917 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2918 obd->obd_name, filter->fo_tot_granted,
2919 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2920 filter->fo_tot_granted -= fed->fed_grant;
2921 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2922 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2923 obd->obd_name, filter->fo_tot_pending,
2924 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2925 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2926 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2927 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2928 obd->obd_name, filter->fo_tot_dirty,
2929 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2930 filter->fo_tot_dirty -= fed->fed_dirty;
2934 cfs_spin_unlock(&obd->obd_osfs_lock);
2937 static int filter_destroy_export(struct obd_export *exp)
2939 struct filter_export_data *fed = &exp->exp_filter_data;
2942 if (fed->fed_pending)
2943 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2944 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2945 exp, fed->fed_pending);
2947 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2949 target_destroy_export(exp);
2950 ldlm_destroy_export(exp);
2951 lut_client_free(exp);
2953 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2956 if (!exp->exp_obd->obd_replayable)
2957 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2959 filter_grant_discard(exp);
2960 filter_fmd_cleanup(exp);
2962 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2963 struct filter_obd *filter = &exp->exp_obd->u.filter;
2964 if (filter->fo_tot_granted_clients > 0)
2965 filter->fo_tot_granted_clients --;
2968 if (!(exp->exp_flags & OBD_OPT_FORCE))
2969 filter_grant_sanity_check(exp->exp_obd, __func__);
2974 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2976 struct obd_llog_group *olg_min, *olg;
2977 struct filter_obd *filter;
2978 int worked = -1, group;
2979 struct llog_ctxt *ctxt;
2982 filter = &obd->u.filter;
2984 /* we can't sync log holding spinlock. also, we do not want to get
2985 * into livelock. so we do following: loop over MDS's exports in
2986 * group order and skip already synced llogs -bzzz */
2988 /* look for group with min. number, but > worked */
2991 cfs_spin_lock(&filter->fo_llog_list_lock);
2992 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2993 if (olg->olg_seq <= worked) {
2994 /* this group is already synced */
2997 if (group < olg->olg_seq) {
2998 /* we have group with smaller number to sync */
3001 /* store current minimal group */
3003 group = olg->olg_seq;
3005 cfs_spin_unlock(&filter->fo_llog_list_lock);
3007 if (olg_min == NULL)
3010 worked = olg_min->olg_seq;
3011 if (olg_min->olg_exp &&
3012 (dexp == olg_min->olg_exp || dexp == NULL)) {
3014 ctxt = llog_group_get_ctxt(olg_min,
3015 LLOG_MDS_OST_REPL_CTXT);
3017 err = llog_sync(ctxt, olg_min->olg_exp);
3018 llog_ctxt_put(ctxt);
3020 CERROR("error flushing logs to MDS: "
3025 } while (olg_min != NULL);
3028 /* Also incredibly similar to mds_disconnect */
3029 static int filter_disconnect(struct obd_export *exp)
3031 struct obd_device *obd = exp->exp_obd;
3036 class_export_get(exp);
3038 if (!(exp->exp_flags & OBD_OPT_FORCE))
3039 filter_grant_sanity_check(obd, __func__);
3040 filter_grant_discard(exp);
3042 /* Flush any remaining cancel messages out to the target */
3043 filter_sync_llogs(obd, exp);
3045 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3047 rc = server_disconnect_export(exp);
3049 /* Do not erase record for recoverable client. */
3050 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3051 filter_client_del(exp);
3053 fsfilt_sync(obd, obd->u.obt.obt_sb);
3055 class_export_put(exp);
3059 /* reverse import is changed, sync all cancels */
3060 static void filter_revimp_update(struct obd_export *exp)
3065 class_export_get(exp);
3067 /* flush any remaining cancel messages out to the target */
3068 filter_sync_llogs(exp->exp_obd, exp);
3069 class_export_put(exp);
3073 static int filter_ping(struct obd_export *exp)
3075 filter_fmd_expire(exp);
3079 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3080 const char *what, int quiet)
3082 struct dentry *dchild = NULL;
3084 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3086 if (IS_ERR(dchild)) {
3087 CERROR("%s error looking up object: "POSTID"\n",
3088 what, ostid->oi_id, ostid->oi_seq);
3092 if (dchild->d_inode == NULL) {
3094 CERROR("%s: %s on non-existent object: "POSTID" \n",
3095 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3097 RETURN(ERR_PTR(-ENOENT));
3103 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3105 struct dentry *dentry = NULL;
3106 struct obd_device *obd;
3110 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3111 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3115 obd = class_exp2obd(exp);
3117 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3121 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3123 RETURN(PTR_ERR(dentry));
3125 /* Limit the valid bits in the return data to what we actually use */
3126 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3127 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3133 /* this should be enabled/disabled in condition to enabled/disabled large
3134 * inodes (fast EAs) in backing store FS. */
3135 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3136 void *handle, struct obdo *oa)
3138 struct obd_device *obd = exp->exp_obd;
3142 if (oa->o_valid & OBD_MD_FLFID) {
3143 struct filter_fid ff;
3145 if (!(oa->o_valid & OBD_MD_FLGROUP))
3147 /* packing fid and converting it to LE for storing into EA.
3148 * Here ->o_stripe_idx should be filled by LOV and rest of
3149 * fields - by client. */
3150 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3151 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3152 /* XXX: we are ignoring o_parent_ver here, since this should
3153 * be the same for all objects in this fileset. */
3154 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3155 ff.ff_objid = cpu_to_le64(oa->o_id);
3156 ff.ff_seq = cpu_to_le64(oa->o_seq);
3158 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3159 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3162 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3164 CERROR("store fid in object failed! rc: %d\n", rc);
3166 CDEBUG(D_HA, "OSS object without fid info!\n");
3172 /* this is called from filter_truncate() until we have filter_punch() */
3173 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3174 struct obdo *oa, struct obd_trans_info *oti)
3176 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3177 struct llog_cookie *fcc = NULL;
3178 struct filter_obd *filter;
3179 int rc, err, sync = 0;
3180 loff_t old_size = 0;
3181 unsigned int ia_valid;
3182 struct inode *inode;
3183 struct page *page = NULL;
3188 LASSERT(dentry != NULL);
3189 LASSERT(!IS_ERR(dentry));
3191 inode = dentry->d_inode;
3192 LASSERT(inode != NULL);
3194 filter = &exp->exp_obd->u.filter;
3195 iattr_from_obdo(&iattr, oa, oa->o_valid);
3196 ia_valid = iattr.ia_valid;
3198 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3199 OBD_ALLOC(fcc, sizeof(*fcc));
3201 *fcc = oa->o_lcookie;
3203 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3204 unsigned long now = jiffies;
3205 ll_vfs_dq_init(inode);
3206 /* Filter truncates and writes are serialized by
3207 * i_alloc_sem, see the comment in
3208 * filter_preprw_write.*/
3209 if (ia_valid & ATTR_SIZE)
3210 down_write(&inode->i_alloc_sem);
3211 LOCK_INODE_MUTEX(inode);
3212 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3213 old_size = i_size_read(inode);
3216 /* VBR: version recovery check */
3217 rc = filter_version_get_check(exp, oti, inode);
3219 GOTO(out_unlock, rc);
3221 /* Let's pin the last page so that ldiskfs_truncate
3222 * should not start GFP_FS allocation. */
3223 if (ia_valid & ATTR_SIZE) {
3224 page = grab_cache_page(inode->i_mapping,
3225 iattr.ia_size >> PAGE_CACHE_SHIFT);
3227 GOTO(out_unlock, rc = -ENOMEM);
3232 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3233 * then we will accept the UID+GID sent by the client during write for
3234 * initializing the ownership of this inode. We only allow this to
3235 * happen once so clear these bits in setattr. In 2.6 kernels it is
3236 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3237 * the flags that are actually being set. */
3238 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3239 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3240 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3242 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3243 if (!(ia_valid & ATTR_MODE)) {
3244 iattr.ia_mode = inode->i_mode;
3245 iattr.ia_valid |= ATTR_MODE;
3247 iattr.ia_mode &= ~S_ISUID;
3249 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3250 if (!(iattr.ia_valid & ATTR_MODE)) {
3251 iattr.ia_mode = inode->i_mode;
3252 iattr.ia_valid |= ATTR_MODE;
3254 iattr.ia_mode &= ~S_ISGID;
3257 orig_ids[USRQUOTA] = inode->i_uid;
3258 orig_ids[GRPQUOTA] = inode->i_gid;
3259 handle = fsfilt_start_log(exp->exp_obd, inode,
3260 FSFILT_OP_SETATTR, oti, 1);
3262 GOTO(out_unlock, rc = PTR_ERR(handle));
3264 /* update inode EA only once when inode is suid bit marked. As
3265 * on 2.6.x UID and GID may be set separately, we check here
3266 * only one of them to avoid double setting. */
3267 if (inode->i_mode & S_ISUID)
3268 filter_update_fidea(exp, inode, handle, oa);
3270 handle = fsfilt_start(exp->exp_obd, inode,
3271 FSFILT_OP_SETATTR, oti);
3273 GOTO(out_unlock, rc = PTR_ERR(handle));
3275 if (oa->o_valid & OBD_MD_FLFLAGS) {
3276 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3277 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3279 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3281 /* set cancel cookie callback function */
3282 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3283 filter_cancel_cookies_cb,
3287 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3288 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3290 /* The truncate might have used up our transaction credits. Make sure
3291 * we have two left for the last_rcvd and VBR inode version updates. */
3292 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3294 /* Update inode version only if data has changed => size has changed */
3295 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3299 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3303 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3305 CERROR("error on commit, err = %d\n", err);
3312 /* For a partial-page truncate flush the page to disk immediately
3313 * to avoid data corruption during direct disk write. b=17397 */
3314 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3315 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3316 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3326 page_cache_release(page);
3328 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3329 UNLOCK_INODE_MUTEX(inode);
3330 if (ia_valid & ATTR_SIZE)
3331 up_write(&inode->i_alloc_sem);
3333 OBD_FREE(fcc, sizeof(*fcc));
3335 /* trigger quota release */
3336 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3337 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3338 int rc2 = lquota_adjust(filter_quota_interface_ref,
3339 exp->exp_obd, cur_ids,
3340 orig_ids, rc, FSFILT_OP_SETATTR);
3341 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3342 "filter adjust qunit. (rc:%d)\n", rc2);
3347 /* this is called from filter_truncate() until we have filter_punch() */
3348 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3349 struct obd_trans_info *oti)
3351 struct obdo *oa = oinfo->oi_oa;
3352 struct lustre_capa *capa = oinfo_capa(oinfo);
3353 struct ldlm_res_id res_id;
3354 struct filter_mod_data *fmd;
3355 struct lvfs_run_ctxt saved;
3356 struct filter_obd *filter;
3357 struct ldlm_resource *res;
3358 struct dentry *dentry;
3359 __u64 opc = CAPA_OPC_META_WRITE;
3363 if (oa->o_valid & OBD_FL_TRUNC)
3364 opc |= CAPA_OPC_OSS_TRUNC;
3366 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3370 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3371 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3376 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3377 /* This would be very bad - accidentally truncating a file when
3378 * changing the time or similar - bug 12203. */
3379 if (oa->o_valid & OBD_MD_FLSIZE &&
3380 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3381 static char mdsinum[48];
3383 if (oa->o_valid & OBD_MD_FLFID)
3384 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3385 oa->o_parent_seq, oa->o_parent_oid,
3390 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3391 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3392 oa->o_id, oa->o_seq, mdsinum);
3396 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3398 RETURN(PTR_ERR(dentry));
3400 filter = &exp->exp_obd->u.filter;
3401 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3404 * We need to be atomic against a concurrent write
3405 * (which takes the semaphore for reading). fmd_mactime_xid
3406 * checks will have no effect if a write request with lower
3407 * xid starts just before a setattr and finishes later than
3408 * the setattr (see bug 21489, comment 27).
3411 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3412 unsigned long now = jiffies;
3413 down_write(&dentry->d_inode->i_alloc_sem);
3414 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3415 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3416 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3417 fmd->fmd_mactime_xid = oti->oti_xid;
3418 filter_fmd_put(exp, fmd);
3419 up_write(&dentry->d_inode->i_alloc_sem);
3422 /* setting objects attributes (including owner/group) */
3423 rc = filter_setattr_internal(exp, dentry, oa, oti);
3425 GOTO(out_unlock, rc);
3427 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3428 &res_id, LDLM_EXTENT, 0);
3431 LDLM_RESOURCE_ADDREF(res);
3432 rc = ldlm_res_lvbo_update(res, NULL, 0);
3433 LDLM_RESOURCE_DELREF(res);
3434 ldlm_resource_putref(res);
3437 oa->o_valid = OBD_MD_FLID;
3439 /* Quota release need uid/gid info */
3440 obdo_from_inode(oa, dentry->d_inode, NULL,
3441 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3446 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3450 /* XXX identical to osc_unpackmd */
3451 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3452 struct lov_mds_md *lmm, int lmm_bytes)
3458 if (lmm_bytes < sizeof (*lmm)) {
3459 CERROR("lov_mds_md too small: %d, need %d\n",
3460 lmm_bytes, (int)sizeof(*lmm));
3463 /* XXX LOV_MAGIC etc check? */
3465 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3466 CERROR("lov_mds_md: zero lmm_object_id\n");
3471 lsm_size = lov_stripe_md_size(1);
3475 if (*lsmp != NULL && lmm == NULL) {
3476 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3477 OBD_FREE(*lsmp, lsm_size);
3482 if (*lsmp == NULL) {
3483 OBD_ALLOC(*lsmp, lsm_size);
3487 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3488 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3489 OBD_FREE(*lsmp, lsm_size);
3492 loi_init((*lsmp)->lsm_oinfo[0]);
3496 /* XXX zero *lsmp? */
3497 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3498 LASSERT((*lsmp)->lsm_object_id);
3501 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3506 /* caller must hold fo_create_locks[oa->o_seq] */
3507 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3508 struct filter_obd *filter)
3510 struct obdo doa = { 0 }; /* XXX obdo on stack */
3516 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3518 memset(&doa, 0, sizeof(doa));
3520 doa.o_valid |= OBD_MD_FLGROUP;
3521 doa.o_seq = oa->o_seq;
3522 doa.o_mode = S_IFREG;
3524 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3525 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3526 exp->exp_obd->obd_name, doa.o_seq);
3530 last = filter_last_id(filter, doa.o_seq);
3532 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3534 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3535 exp->exp_obd->obd_name, oa->o_id + 1, last,
3536 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3538 for (id = last; id > oa->o_id; id--) {
3540 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3541 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3542 CEMERG("error destroying precreate objid "LPU64": %d\n",
3545 /* update last_id on disk periodically so that if we restart
3546 * we don't need to re-scan all of the just-deleted objects. */
3547 if ((id & 511) == 0 && !skip_orphan) {
3548 filter_set_last_id(filter, id - 1, doa.o_seq);
3549 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3553 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3554 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3557 filter_set_last_id(filter, id, doa.o_seq);
3558 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3561 * We have destroyed orphan objects, but don't want to reuse
3562 * them. Therefore we don't reset last_id to the last created
3563 * objects. Instead, we report back to the MDS the object id
3564 * of the last orphan, so that the MDS can restart allocating
3565 * objects from this id + 1 and thus skip the whole orphan
3571 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3576 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3577 obd_seq group, int *num);
3578 /* returns a negative error or a nonnegative number of files to create */
3579 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3580 obd_seq group, struct obd_trans_info *oti)
3582 struct obd_device *obd = exp->exp_obd;
3583 struct filter_obd *filter = &obd->u.filter;
3587 /* delete orphans request */
3588 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3589 obd_id last = filter_last_id(filter, group);
3591 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3592 CERROR("%s: dropping old orphan cleanup request\n",
3596 /* This causes inflight precreates to abort and drop lock */
3597 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3598 cfs_down(&filter->fo_create_locks[group]);
3599 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3600 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3601 exp->exp_obd->obd_name, group);
3602 cfs_up(&filter->fo_create_locks[group]);
3605 diff = oa->o_id - last;
3606 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3609 if (-diff > OST_MAX_PRECREATE) {
3610 CERROR("%s: ignoring bogus orphan destroy request: "
3611 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3613 /* FIXME: should reset precreate_next_id on MDS */
3614 GOTO(out, rc = -EINVAL);
3617 rc = filter_destroy_precreated(exp, oa, filter);
3619 CERROR("%s: unable to write lastobjid, but "
3620 "orphans were deleted\n", obd->obd_name);
3623 /* XXX: Used by MDS for the first time! */
3624 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3627 cfs_down(&filter->fo_create_locks[group]);
3628 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3629 CERROR("%s: dropping old precreate request\n",
3633 /* only precreate if group == 0 and o_id is specfied */
3634 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3637 diff = oa->o_id - filter_last_id(filter, group);
3638 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3639 filter_last_id(filter, group), diff);
3641 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3642 oa->o_id, filter_last_id(filter, group), diff);
3646 oa->o_id = filter_last_id(&obd->u.filter, group);
3647 rc = filter_precreate(obd, oa, group, &diff);
3648 oa->o_id = filter_last_id(&obd->u.filter, group);
3650 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3653 /* else diff == 0 */
3656 cfs_up(&filter->fo_create_locks[group]);
3660 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3661 __u64 max_age, __u32 flags)
3663 struct filter_obd *filter = &obd->u.filter;
3664 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3668 /* at least try to account for cached pages. its still racey and
3669 * might be under-reporting if clients haven't announced their
3670 * caches with brw recently */
3671 cfs_spin_lock(&obd->obd_osfs_lock);
3672 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3673 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3674 cfs_spin_unlock(&obd->obd_osfs_lock);
3676 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3677 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3678 filter->fo_tot_dirty, filter->fo_tot_granted,
3679 filter->fo_tot_pending,
3680 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3682 filter_grant_sanity_check(obd, __func__);
3684 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3685 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3686 osfs->os_bsize - 1) >> blockbits));
3688 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3689 struct lr_server_data *lsd = class_server_data(obd);
3690 int index = le32_to_cpu(lsd->lsd_ost_index);
3692 if (obd_fail_val == -1 ||
3693 index == obd_fail_val)
3694 osfs->os_bfree = osfs->os_bavail = 2;
3695 else if (obd_fail_loc & OBD_FAIL_ONCE)
3696 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3699 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3700 * stop creating files on MDS if OST is not good shape to create
3704 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3705 osfs->os_state = OS_STATE_READONLY;
3707 if (filter->fo_raid_degraded)
3708 osfs->os_state |= OS_STATE_DEGRADED;
3712 static int filter_use_existing_obj(struct obd_device *obd,
3713 struct dentry *dchild, void **handle,
3716 struct inode *inode = dchild->d_inode;
3720 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3723 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3724 if (IS_ERR(*handle))
3725 return PTR_ERR(*handle);
3727 iattr.ia_valid = ATTR_MODE;
3728 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3729 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3736 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3739 __u64 os_ffree = -1;
3741 cfs_spin_lock(&obd->obd_osfs_lock);
3742 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3744 os_ffree = obd->obd_osfs.os_ffree;
3745 cfs_spin_unlock(&obd->obd_osfs_lock);
3750 /* We rely on the fact that only one thread will be creating files in a given
3751 * group at a time, which is why we don't need an atomic filter_get_new_id.
3752 * Even if we had that atomic function, the following race would exist:
3754 * thread 1: gets id x from filter_next_id
3755 * thread 2: gets id (x + 1) from filter_next_id
3756 * thread 2: creates object (x + 1)
3757 * thread 1: tries to create object x, gets -ENOSPC
3759 * Caller must hold fo_create_locks[group]
3761 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3762 obd_seq group, int *num)
3764 struct dentry *dchild = NULL, *dparent = NULL;
3765 struct filter_obd *filter;
3766 struct obd_statfs *osfs;
3767 int err = 0, rc = 0, recreate_obj = 0, i;
3768 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3771 void *handle = NULL;
3774 filter = &obd->u.filter;
3776 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3778 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3780 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3781 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3784 OBD_ALLOC(osfs, sizeof(*osfs));
3787 rc = filter_statfs(obd, osfs,
3788 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3790 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3791 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3792 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3793 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3797 OBD_FREE(osfs, sizeof(*osfs));
3802 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3803 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3805 for (i = 0; i < *num && err == 0; i++) {
3806 int cleanup_phase = 0;
3808 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3809 CWARN("%s: create aborted by destroy\n",
3818 last_id = filter_last_id(filter, group);
3819 if (next_id > last_id) {
3820 CERROR("Error: Trying to recreate obj greater"
3821 "than last id "LPD64" > "LPD64"\n",
3823 GOTO(cleanup, rc = -EINVAL);
3826 next_id = filter_last_id(filter, group) + 1;
3828 /* Don't create objects beyond the valid range for this SEQ */
3829 if (unlikely(fid_seq_is_mdt0(group) &&
3830 next_id >= IDIF_MAX_OID)) {
3831 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3832 obd->obd_name, next_id, group);
3833 GOTO(cleanup, rc = -ENOSPC);
3834 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3835 next_id >= OBIF_MAX_OID)) {
3836 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3837 obd->obd_name, next_id, group);
3838 GOTO(cleanup, rc = -ENOSPC);
3841 dparent = filter_parent_lock(obd, group, next_id);
3842 if (IS_ERR(dparent))
3843 GOTO(cleanup, rc = PTR_ERR(dparent));
3844 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3846 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3848 GOTO(cleanup, rc = PTR_ERR(dchild));
3849 cleanup_phase = 2; /* f_dput(dchild) */
3851 if (dchild->d_inode != NULL) {
3852 /* This would only happen if lastobjid was bad on disk*/
3853 /* Could also happen if recreating missing obj but it
3854 * already exists. */
3856 CERROR("%s: recreating existing object %.*s?\n",
3857 obd->obd_name, dchild->d_name.len,
3858 dchild->d_name.name);
3860 /* Use these existing objects if they are
3862 if (dchild->d_inode->i_size == 0) {
3863 rc = filter_use_existing_obj(obd,dchild,
3864 &handle, &cleanup_phase);
3871 CERROR("%s: Serious error: objid %.*s already "
3872 "exists; is this filesystem corrupt?\n",
3873 obd->obd_name, dchild->d_name.len,
3874 dchild->d_name.name);
3877 GOTO(cleanup, rc = -EEXIST);
3880 handle = fsfilt_start_log(obd, dparent->d_inode,
3881 FSFILT_OP_CREATE, NULL, 1);
3883 GOTO(cleanup, rc = PTR_ERR(handle));
3886 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3887 ",od->o_id="LPU64")\n", obd->obd_name, group,
3890 /* We mark object SUID+SGID to flag it for accepting UID+GID
3891 * from client on first write. Currently the permission bits
3892 * on the OST are never used, so this is OK. */
3893 rc = ll_vfs_create(dparent->d_inode, dchild,
3894 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3896 CERROR("create failed rc = %d\n", rc);
3897 if (rc == -ENOSPC) {
3898 os_ffree = filter_calc_free_inodes(obd);
3900 CERROR("%s: free inode "LPU64"\n",
3901 obd->obd_name, os_ffree);
3906 if (dchild->d_inode)
3907 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3908 dchild->d_inode->i_ino);
3911 if (!recreate_obj) {
3912 filter_set_last_id(filter, next_id, group);
3913 err = filter_update_last_objid(obd, group, 0);
3915 CERROR("unable to write lastobjid "
3916 "but file created\n");
3920 switch(cleanup_phase) {
3922 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3924 CERROR("error on commit, err = %d\n", err);
3931 filter_parent_unlock(dparent);
3938 if (cfs_time_after(jiffies, enough_time)) {
3941 "%s: precreate slow - want %d got %d \n",
3942 obd->obd_name, *num, i);
3949 "%s: created %d objects for group "POSTID" rc %d\n",
3950 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
3955 int filter_create(struct obd_export *exp, struct obdo *oa,
3956 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3958 struct obd_device *obd = exp->exp_obd;
3959 struct filter_export_data *fed;
3960 struct filter_obd *filter;
3961 struct lvfs_run_ctxt saved;
3962 struct lov_stripe_md *lsm = NULL;
3966 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
3967 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
3969 fed = &exp->exp_filter_data;
3970 filter = &obd->u.filter;
3972 if (fed->fed_group != oa->o_seq) {
3973 CERROR("%s: this export (nid %s) used object group %d "
3974 "earlier; now it's trying to use group "LPU64"!"
3975 " This could be a bug in the MDS. Please report to "
3976 "http://bugzilla.lustre.org/\n", obd->obd_name,
3977 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
3984 rc = obd_alloc_memmd(exp, &lsm);
3991 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3993 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3994 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3995 if (!obd->obd_recovering ||
3996 oa->o_id > filter_last_id(filter, oa->o_seq)) {
3997 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3998 oa->o_id, filter_last_id(filter, oa->o_seq));
4002 cfs_down(&filter->fo_create_locks[oa->o_seq]);
4003 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4004 cfs_up(&filter->fo_create_locks[oa->o_seq]);
4007 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4010 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4011 if (rc && ea != NULL && *ea != lsm) {
4012 obd_free_memmd(exp, &lsm);
4013 } else if (rc == 0 && ea != NULL) {
4014 /* XXX LOV STACKING: the lsm that is passed to us from
4015 * LOV does not have valid lsm_oinfo data structs, so
4016 * don't go touching that. This needs to be fixed in a
4018 lsm->lsm_object_id = oa->o_id;
4025 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4026 struct lov_stripe_md *md, struct obd_trans_info *oti,
4027 struct obd_export *md_exp, void *capa)
4029 unsigned int qcids[MAXQUOTAS] = {0, 0};
4030 struct obd_device *obd;
4031 struct filter_obd *filter;
4032 struct dentry *dchild = NULL, *dparent = NULL;
4033 struct lustre_handle lockh = { 0 };
4034 struct lvfs_run_ctxt saved;
4035 void *handle = NULL;
4036 struct llog_cookie *fcc = NULL;
4037 int rc, rc2, cleanup_phase = 0, sync = 0;
4042 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4043 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4048 filter = &obd->u.filter;
4050 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4053 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4054 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4056 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4058 GOTO(cleanup, rc = PTR_ERR(dchild));
4061 if (dchild->d_inode == NULL) {
4062 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4063 oa->o_id, oa->o_seq);
4064 /* If object already gone, cancel cookie right now */
4065 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4066 struct llog_ctxt *ctxt;
4067 struct obd_llog_group *olg;
4069 olg = filter_find_olg(obd, oa->o_seq);
4071 CERROR(" %s: can not find olg of group %d\n",
4072 obd->obd_name, (int)oa->o_seq);
4073 GOTO(cleanup, rc = PTR_ERR(olg));
4075 fcc = &oa->o_lcookie;
4076 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4077 llog_cancel(ctxt, NULL, 1, fcc, 0);
4078 llog_ctxt_put(ctxt);
4079 fcc = NULL; /* we didn't allocate fcc, don't free it */
4081 GOTO(cleanup, rc = -ENOENT);
4084 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4088 /* Our MDC connection is established by the MDS to us */
4089 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4090 OBD_ALLOC(fcc, sizeof(*fcc));
4092 *fcc = oa->o_lcookie;
4094 ll_vfs_dq_init(dchild->d_inode);
4096 /* we're gonna truncate it first in order to avoid possible deadlock:
4098 * open trasaction open transaction
4099 * down(i_zombie) down(i_zombie)
4100 * restart transaction
4101 * (see BUG 4180) -bzzz
4103 * take i_alloc_sem too to prevent other threads from writing to the
4104 * file while we are truncating it. This can cause lock ordering issue
4105 * between page lock, i_mutex & starting new journal handle.
4106 * (see bug 20321) -johann
4109 down_write(&dchild->d_inode->i_alloc_sem);
4110 LOCK_INODE_MUTEX(dchild->d_inode);
4111 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4113 /* VBR: version recovery check */
4114 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4116 UNLOCK_INODE_MUTEX(dchild->d_inode);
4117 up_write(&dchild->d_inode->i_alloc_sem);
4121 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4123 if (IS_ERR(handle)) {
4124 UNLOCK_INODE_MUTEX(dchild->d_inode);
4125 up_write(&dchild->d_inode->i_alloc_sem);
4126 GOTO(cleanup, rc = PTR_ERR(handle));
4129 iattr.ia_valid = ATTR_SIZE;
4131 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4132 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4133 UNLOCK_INODE_MUTEX(dchild->d_inode);
4134 up_write(&dchild->d_inode->i_alloc_sem);
4138 GOTO(cleanup, rc = rc2);
4140 /* We don't actually need to lock the parent until we are unlinking
4141 * here, and not while truncating above. That avoids holding the
4142 * parent lock for a long time during truncate, which can block other
4143 * threads from doing anything to objects in that directory. bug 7171 */
4144 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4145 if (IS_ERR(dparent))
4146 GOTO(cleanup, rc = PTR_ERR(dparent));
4147 cleanup_phase = 3; /* filter_parent_unlock */
4149 LOCK_INODE_MUTEX(dchild->d_inode);
4150 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4151 if (IS_ERR(handle)) {
4152 UNLOCK_INODE_MUTEX(dchild->d_inode);
4153 GOTO(cleanup, rc = PTR_ERR(handle));
4155 cleanup_phase = 4; /* fsfilt_commit */
4157 /* Quota release need uid/gid of inode */
4158 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4160 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4162 /* this drops dchild->d_inode->i_mutex unconditionally */
4163 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4167 switch(cleanup_phase) {
4170 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4171 oti->oti_handle : handle,
4172 filter_cancel_cookies_cb,
4174 /* If add_journal_cb failed, then filter_finish_transno
4175 * will commit the handle and we will do a sync
4176 * on commit. then we call callback directly to free
4179 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4181 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4184 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4186 CERROR("error on commit, err = %d\n", rc2);
4193 filter_parent_unlock(dparent);
4195 filter_fini_destroy(obd, &lockh);
4199 OBD_FREE(fcc, sizeof(*fcc));
4201 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4204 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4208 /* trigger quota release */
4209 qcids[USRQUOTA] = oa->o_uid;
4210 qcids[GRPQUOTA] = oa->o_gid;
4211 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4214 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4218 /* NB start and end are used for punch, but not truncate */
4219 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4220 struct obd_trans_info *oti,
4221 struct ptlrpc_request_set *rqset)
4226 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4227 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4228 oinfo->oi_policy.l_extent.end);
4232 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4233 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4234 oinfo->oi_policy.l_extent.start);
4236 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4237 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4238 rc = filter_setattr(exp, oinfo, oti);
4239 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4243 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4244 obd_off start, obd_off end,
4245 struct ptlrpc_request_set *set)
4247 struct lvfs_run_ctxt saved;
4248 struct obd_device_target *obt;
4249 struct dentry *dentry;
4253 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4254 (struct lustre_capa *)oinfo->oi_capa,
4255 CAPA_OPC_OSS_WRITE);
4259 obt = &exp->exp_obd->u.obt;
4261 /* An objid of zero is taken to mean "sync whole filesystem" */
4262 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4263 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4264 /* Flush any remaining cancel messages out to the target */
4265 filter_sync_llogs(exp->exp_obd, exp);
4269 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4271 RETURN(PTR_ERR(dentry));
4273 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4275 LOCK_INODE_MUTEX(dentry->d_inode);
4277 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4279 /* just any file to grab fsync method - "file" arg unused */
4280 struct file *file = obt->obt_rcvd_filp;
4282 if (file->f_op && file->f_op->fsync)
4283 rc = file->f_op->fsync(NULL, dentry, 1);
4285 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4289 UNLOCK_INODE_MUTEX(dentry->d_inode);
4291 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4292 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4293 FILTER_VALID_FLAGS);
4295 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4301 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4302 void *key, __u32 *vallen, void *val,
4303 struct lov_stripe_md *lsm)
4305 struct obd_device *obd;
4308 obd = class_exp2obd(exp);
4310 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4314 if (KEY_IS(KEY_BLOCKSIZE)) {
4315 __u32 *blocksize = val;
4317 if (*vallen < sizeof(*blocksize))
4319 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4321 *vallen = sizeof(*blocksize);
4325 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4326 __u32 *blocksize_bits = val;
4327 if (blocksize_bits) {
4328 if (*vallen < sizeof(*blocksize_bits))
4330 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4332 *vallen = sizeof(*blocksize_bits);
4336 if (KEY_IS(KEY_LAST_ID)) {
4337 obd_id *last_id = val;
4338 /* FIXME: object groups */
4340 if (*vallen < sizeof(*last_id))
4342 *last_id = filter_last_id(&obd->u.filter,
4343 exp->exp_filter_data.fed_group);
4345 *vallen = sizeof(*last_id);
4349 if (KEY_IS(KEY_FIEMAP)) {
4350 struct ll_fiemap_info_key *fm_key = key;
4351 struct dentry *dentry;
4352 struct ll_user_fiemap *fiemap = val;
4353 struct lvfs_run_ctxt saved;
4356 if (fiemap == NULL) {
4357 *vallen = fiemap_count_to_size(
4358 fm_key->fiemap.fm_extent_count);
4362 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4365 RETURN(PTR_ERR(dentry));
4367 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4368 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4369 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4371 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4377 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4378 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4379 *vallen = sizeof(__u32);
4383 CDEBUG(D_IOCTL, "invalid key\n");
4387 static inline int filter_setup_llog_group(struct obd_export *exp,
4388 struct obd_device *obd,
4391 struct obd_llog_group *olg;
4392 struct llog_ctxt *ctxt;
4395 olg = filter_find_create_olg(obd, group);
4397 RETURN(PTR_ERR(olg));
4399 llog_group_set_export(olg, exp);
4401 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4402 LASSERTF(ctxt != NULL, "ctxt is null\n");
4404 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4405 llog_ctxt_put(ctxt);
4409 static int filter_set_grant_shrink(struct obd_export *exp,
4410 struct ost_body *body)
4412 /* handle shrink grant */
4413 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4414 filter_grant_incoming(exp, &body->oa);
4415 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4421 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4423 struct obd_device *obd;
4429 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4433 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4434 obd_export_nid2str(exp));
4435 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4437 /* setup llog imports */
4439 group = (int)(*(__u32 *)val);
4441 group = 0; /* default value */
4443 LASSERT_SEQ_IS_MDT(group);
4444 rc = filter_setup_llog_group(exp, obd, group);
4448 if (group == FID_SEQ_OST_MDT0) {
4449 /* setup llog group 1 for interop */
4450 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4453 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4458 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4459 void *key, __u32 vallen, void *val,
4460 struct ptlrpc_request_set *set)
4462 struct obd_device *obd;
4467 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4471 if (KEY_IS(KEY_CAPA_KEY)) {
4473 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4475 CERROR("filter update capability key failed: %d\n", rc);
4479 if (KEY_IS(KEY_REVIMP_UPD)) {
4480 filter_revimp_update(exp);
4481 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4485 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4486 filter_adapt_sptlrpc_conf(obd, 0);
4490 if (KEY_IS(KEY_MDS_CONN))
4491 RETURN(filter_set_mds_conn(exp, val));
4493 if (KEY_IS(KEY_GRANT_SHRINK))
4494 RETURN(filter_set_grant_shrink(exp, val));
4499 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4500 int len, void *karg, void *uarg)
4502 struct obd_device *obd = exp->exp_obd;
4503 struct obd_ioctl_data *data = karg;
4507 case OBD_IOC_ABORT_RECOVERY: {
4508 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4509 target_stop_recovery_thread(obd);
4513 case OBD_IOC_SYNC: {
4514 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4515 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4519 case OBD_IOC_SET_READONLY: {
4521 struct super_block *sb = obd->u.obt.obt_sb;
4522 struct inode *inode = sb->s_root->d_inode;
4523 BDEVNAME_DECLARE_STORAGE(tmp);
4524 CERROR("*** setting device %s read-only ***\n",
4525 ll_bdevname(sb, tmp));
4527 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4528 if (!IS_ERR(handle))
4529 rc = fsfilt_commit(obd, inode, handle, 1);
4531 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4532 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4534 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4538 case OBD_IOC_CATLOGLIST: {
4539 rc = llog_catalog_list(obd, 1, data);
4543 case OBD_IOC_LLOG_CANCEL:
4544 case OBD_IOC_LLOG_REMOVE:
4545 case OBD_IOC_LLOG_INFO:
4546 case OBD_IOC_LLOG_PRINT: {
4547 /* FIXME to be finished */
4548 RETURN(-EOPNOTSUPP);
4550 struct llog_ctxt *ctxt = NULL;
4552 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4553 rc = llog_ioctl(ctxt, cmd, data);
4554 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4567 static int filter_health_check(struct obd_device *obd)
4569 #ifdef USE_HEALTH_CHECK_WRITE
4570 struct filter_obd *filter = &obd->u.filter;
4575 * health_check to return 0 on healthy
4576 * and 1 on unhealthy.
4578 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4581 #ifdef USE_HEALTH_CHECK_WRITE
4582 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4583 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4588 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4591 return filter_fid2dentry(data, NULL, gr, id);
4594 static int filter_process_config(struct obd_device *obd, obd_count len,
4597 struct lustre_cfg *lcfg = buf;
4598 struct lprocfs_static_vars lvars;
4601 switch (lcfg->lcfg_command) {
4603 lprocfs_filter_init_vars(&lvars);
4605 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4615 static struct lvfs_callback_ops filter_lvfs_ops = {
4616 l_fid2dentry: filter_lvfs_fid2dentry,
4619 static struct obd_ops filter_obd_ops = {
4620 .o_owner = THIS_MODULE,
4621 .o_get_info = filter_get_info,
4622 .o_set_info_async = filter_set_info_async,
4623 .o_setup = filter_setup,
4624 .o_precleanup = filter_precleanup,
4625 .o_cleanup = filter_cleanup,
4626 .o_connect = filter_connect,
4627 .o_reconnect = filter_reconnect,
4628 .o_disconnect = filter_disconnect,
4629 .o_ping = filter_ping,
4630 .o_init_export = filter_init_export,
4631 .o_destroy_export = filter_destroy_export,
4632 .o_statfs = filter_statfs,
4633 .o_getattr = filter_getattr,
4634 .o_unpackmd = filter_unpackmd,
4635 .o_create = filter_create,
4636 .o_setattr = filter_setattr,
4637 .o_destroy = filter_destroy,
4638 .o_brw = filter_brw,
4639 .o_punch = filter_truncate,
4640 .o_sync = filter_sync,
4641 .o_preprw = filter_preprw,
4642 .o_commitrw = filter_commitrw,
4643 .o_llog_init = filter_llog_init,
4644 .o_llog_connect = filter_llog_connect,
4645 .o_llog_finish = filter_llog_finish,
4646 .o_iocontrol = filter_iocontrol,
4647 .o_health_check = filter_health_check,
4648 .o_process_config = filter_process_config,
4651 quota_interface_t *filter_quota_interface_ref;
4652 extern quota_interface_t filter_quota_interface;
4654 static int __init obdfilter_init(void)
4656 struct lprocfs_static_vars lvars;
4659 /** sanity check for group<->mdsno conversion */
4660 for (i = 0; i < MAX_MDT_COUNT; i++)
4661 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4663 lprocfs_filter_init_vars(&lvars);
4665 cfs_request_module("%s", "lquota");
4666 OBD_ALLOC(obdfilter_created_scratchpad,
4667 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4668 sizeof(*obdfilter_created_scratchpad));
4669 if (obdfilter_created_scratchpad == NULL)
4672 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4673 sizeof(struct filter_mod_data),
4676 GOTO(out, rc = -ENOMEM);
4678 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4679 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4681 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4682 LUSTRE_OST_NAME, NULL);
4686 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4687 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4688 ll_fmd_cachep = NULL;
4690 if (filter_quota_interface_ref)
4691 PORTAL_SYMBOL_PUT(filter_quota_interface);
4693 OBD_FREE(obdfilter_created_scratchpad,
4694 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4695 sizeof(*obdfilter_created_scratchpad));
4701 static void __exit obdfilter_exit(void)
4703 if (filter_quota_interface_ref)
4704 PORTAL_SYMBOL_PUT(filter_quota_interface);
4706 if (ll_fmd_cachep) {
4707 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4708 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4709 ll_fmd_cachep = NULL;
4712 class_unregister_type(LUSTRE_OST_NAME);
4713 OBD_FREE(obdfilter_created_scratchpad,
4714 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4715 sizeof(*obdfilter_created_scratchpad));
4718 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4719 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4720 MODULE_LICENSE("GPL");
4722 module_init(obdfilter_init);
4723 module_exit(obdfilter_exit);