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 2008 Sun Microsystems, Inc. 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 filter_obd *filter = &exp->exp_obd->u.filter;
124 struct filter_export_data *fed = &exp->exp_filter_data;
125 struct lsd_client_data *lcd;
128 int err, log_pri = D_RPCTRACE;
130 /* Propagate error code. */
134 if (!exp->exp_obd->obd_replayable || oti == NULL)
137 cfs_mutex_down(&fed->fed_lcd_lock);
139 /* if the export has already been disconnected, we have no last_rcvd slot,
140 * update server data with latest transno then */
142 cfs_mutex_up(&fed->fed_lcd_lock);
143 CWARN("commit transaction for disconnected client %s: rc %d\n",
144 exp->exp_client_uuid.uuid, rc);
145 err = filter_update_server_data(exp->exp_obd,
146 filter->fo_rcvd_filp,
151 /* we don't allocate new transnos for replayed requests */
152 cfs_spin_lock(&filter->fo_translock);
153 if (oti->oti_transno == 0) {
154 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
155 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
157 last_rcvd = oti->oti_transno;
158 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
159 filter->fo_fsd->lsd_last_transno =
160 cpu_to_le64(last_rcvd);
162 oti->oti_transno = last_rcvd;
164 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
165 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
166 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
167 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
168 cfs_spin_unlock(&filter->fo_translock);
171 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
173 off = fed->fed_lr_off;
175 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
176 fed->fed_lr_idx, fed->fed_lr_off);
179 class_export_cb_get(exp); /* released when the cb is called */
181 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
187 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
188 lcd, sizeof(*lcd), &off,
189 force_sync | exp->exp_need_sync);
191 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
199 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
200 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
201 cfs_mutex_up(&fed->fed_lcd_lock);
205 void f_dput(struct dentry *dentry)
207 /* Can't go inside filter_ddelete because it can block */
208 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
209 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
210 LASSERT(atomic_read(&dentry->d_count) > 0);
215 static void init_brw_stats(struct brw_stats *brw_stats)
218 for (i = 0; i < BRW_LAST; i++)
219 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
222 static int lprocfs_init_rw_stats(struct obd_device *obd,
223 struct lprocfs_stats **stats)
227 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
228 LPROC_FILTER_LAST - 1;
229 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
233 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
234 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
236 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
237 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
242 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
243 plus the procfs overhead :( */
244 static int filter_export_stats_init(struct obd_device *obd,
245 struct obd_export *exp,
251 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
252 /* Self-export gets no proc entry */
255 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
257 /* Mask error for already created
265 struct nid_stat *tmp = exp->exp_nid_stats;
266 LASSERT(tmp != NULL);
268 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
269 if (tmp->nid_brw_stats == NULL)
272 init_brw_stats(tmp->nid_brw_stats);
273 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
274 0644, &filter_per_nid_stats_fops,
277 CWARN("Error adding the brw_stats file\n");
279 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
283 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
287 /* Always add in ldlm_stats */
288 tmp->nid_ldlm_stats =
289 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
290 LPROCFS_STATS_FLAG_NOPERCPU);
291 if (tmp->nid_ldlm_stats == NULL)
294 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
295 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
296 tmp->nid_ldlm_stats);
304 /* Add client data to the FILTER. We use a bitmap to locate a free space
305 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
306 * Otherwise, we have just read the data from the last_rcvd file and
307 * we know its offset. */
308 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
311 struct filter_obd *filter = &obd->u.filter;
312 struct filter_export_data *fed = &exp->exp_filter_data;
313 unsigned long *bitmap = filter->fo_last_rcvd_slots;
314 int new_client = (cl_idx == -1);
318 LASSERT(bitmap != NULL);
319 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
322 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
325 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
326 * there's no need for extra complication here
329 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
331 if (cl_idx >= LR_MAX_CLIENTS) {
332 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
337 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
342 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343 CERROR("FILTER client %d: bit already set in bitmap!\n",
349 fed->fed_lr_idx = cl_idx;
350 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
351 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
352 cfs_init_mutex(&fed->fed_lcd_lock);
353 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
355 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
356 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
359 struct lvfs_run_ctxt saved;
360 loff_t off = fed->fed_lr_off;
364 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
365 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
367 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
368 /* Transaction needed to fix bug 1403 */
369 handle = fsfilt_start(obd,
370 filter->fo_rcvd_filp->f_dentry->d_inode,
371 FSFILT_OP_SETATTR, NULL);
372 if (IS_ERR(handle)) {
373 rc = PTR_ERR(handle);
374 CERROR("unable to start transaction: rc %d\n", rc);
376 fed->fed_lcd->lcd_last_epoch =
377 filter->fo_fsd->lsd_start_epoch;
378 exp->exp_last_request_time = cfs_time_current_sec();
379 rc = fsfilt_add_journal_cb(obd, 0, handle,
380 target_client_add_cb,
381 class_export_cb_get(exp));
383 cfs_spin_lock(&exp->exp_lock);
384 exp->exp_need_sync = 1;
385 cfs_spin_unlock(&exp->exp_lock);
387 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
389 sizeof(*fed->fed_lcd),
390 &off, rc /* sync if no cb */);
392 filter->fo_rcvd_filp->f_dentry->d_inode,
395 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398 CERROR("error writing %s client idx %u: rc %d\n",
399 LAST_RCVD, fed->fed_lr_idx, rc);
406 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
408 static int filter_client_free(struct obd_export *exp)
410 struct filter_export_data *fed = &exp->exp_filter_data;
411 struct filter_obd *filter = &exp->exp_obd->u.filter;
412 struct obd_device *obd = exp->exp_obd;
413 struct lvfs_run_ctxt saved;
414 struct lsd_client_data *lcd = fed->fed_lcd;
422 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
423 if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
426 LASSERT(filter->fo_last_rcvd_slots != NULL);
428 off = fed->fed_lr_off;
430 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
431 fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
433 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
434 * we leak a client slot that will be cleaned on the next recovery. */
436 CERROR("%s: client idx %d has med_off %lld\n",
437 obd->obd_name, fed->fed_lr_idx, off);
438 GOTO(free, rc = -EINVAL);
441 /* Clear the bit _after_ zeroing out the client so we don't
442 race with filter_client_add and zero out new clients.*/
443 if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
444 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
449 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
450 /* Make sure the server's last_transno is up to date.
451 * This should be done before zeroing client slot so last_transno will
452 * be in server data or in client data in case of failure */
453 filter_update_server_data(obd, filter->fo_rcvd_filp, filter->fo_fsd);
455 cfs_mutex_down(&fed->fed_lcd_lock);
456 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
457 sizeof(zero_lcd), &off, 0);
459 cfs_mutex_up(&fed->fed_lcd_lock);
460 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
462 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
463 "zero out client %s at idx %u/%llu in %s, rc %d\n",
464 lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
467 if (!cfs_test_and_clear_bit(fed->fed_lr_idx,
468 filter->fo_last_rcvd_slots)) {
469 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
476 cfs_mutex_down(&fed->fed_lcd_lock);
478 cfs_mutex_up(&fed->fed_lcd_lock);
484 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
485 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
486 struct filter_mod_data *fmd)
488 LASSERT_SPIN_LOCKED(&fed->fed_lock);
489 if (--fmd->fmd_refcount == 0) {
490 /* XXX when we have persistent reservations and the handle
491 * is stored herein we need to drop it here. */
492 fed->fed_mod_count--;
493 cfs_list_del(&fmd->fmd_list);
494 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
498 /* drop fmd reference, free it if last ref */
499 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
501 struct filter_export_data *fed;
506 fed = &exp->exp_filter_data;
507 cfs_spin_lock(&fed->fed_lock);
508 filter_fmd_put_nolock(fed, fmd); /* caller reference */
509 cfs_spin_unlock(&fed->fed_lock);
512 /* expire entries from the end of the list if there are too many
513 * or they are too old */
514 static void filter_fmd_expire_nolock(struct filter_obd *filter,
515 struct filter_export_data *fed,
516 struct filter_mod_data *keep)
518 struct filter_mod_data *fmd, *tmp;
520 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
524 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
525 fed->fed_mod_count < filter->fo_fmd_max_num)
528 cfs_list_del_init(&fmd->fmd_list);
529 filter_fmd_put_nolock(fed, fmd); /* list reference */
533 void filter_fmd_expire(struct obd_export *exp)
535 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
536 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
537 &exp->exp_filter_data, NULL);
538 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
541 /* find specified objid, group in export fmd list.
542 * caller must hold fed_lock and take fmd reference itself */
543 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
544 struct filter_export_data *fed,
545 obd_id objid, obd_gr group)
547 struct filter_mod_data *found = NULL, *fmd;
549 LASSERT_SPIN_LOCKED(&fed->fed_lock);
551 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
552 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
554 cfs_list_del(&fmd->fmd_list);
555 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
556 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
561 filter_fmd_expire_nolock(filter, fed, found);
566 /* Find fmd based on objid and group, or return NULL if not found. */
567 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
568 obd_id objid, obd_gr group)
570 struct filter_mod_data *fmd;
572 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
573 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
574 &exp->exp_filter_data, objid, group);
576 fmd->fmd_refcount++; /* caller reference */
577 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
582 /* Find fmd based on objid and group, or create a new one if none is found.
583 * It is possible for this function to return NULL under memory pressure,
584 * or if objid = 0 is passed (which will only cause old entries to expire).
585 * Currently this is not fatal because any fmd state is transient and
586 * may also be freed when it gets sufficiently old. */
587 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
588 obd_id objid, obd_gr group)
590 struct filter_export_data *fed = &exp->exp_filter_data;
591 struct filter_mod_data *found = NULL, *fmd_new = NULL;
593 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
595 cfs_spin_lock(&fed->fed_lock);
596 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
599 cfs_list_add_tail(&fmd_new->fmd_list,
601 fmd_new->fmd_id = objid;
602 fmd_new->fmd_gr = group;
603 fmd_new->fmd_refcount++; /* list reference */
605 fed->fed_mod_count++;
607 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
611 found->fmd_refcount++; /* caller reference */
612 found->fmd_expire = jiffies +
613 exp->exp_obd->u.filter.fo_fmd_max_age;
616 cfs_spin_unlock(&fed->fed_lock);
622 /* drop fmd list reference so it will disappear when last reference is put.
623 * This isn't so critical because it would in fact only affect the one client
624 * that is doing the unlink and at worst we have an stale entry referencing
625 * an object that should never be used again. */
626 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
628 struct filter_mod_data *found = NULL;
630 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
631 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
633 cfs_list_del_init(&found->fmd_list);
634 filter_fmd_put_nolock(&exp->exp_filter_data, found);
636 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
639 #define filter_fmd_drop(exp, objid, group)
642 /* remove all entries from fmd list */
643 static void filter_fmd_cleanup(struct obd_export *exp)
645 struct filter_export_data *fed = &exp->exp_filter_data;
646 struct filter_mod_data *fmd = NULL, *tmp;
648 cfs_spin_lock(&fed->fed_lock);
649 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
650 cfs_list_del_init(&fmd->fmd_list);
651 filter_fmd_put_nolock(fed, fmd);
653 cfs_spin_unlock(&fed->fed_lock);
656 static int filter_init_export(struct obd_export *exp)
658 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
659 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
661 cfs_spin_lock(&exp->exp_lock);
662 exp->exp_connecting = 1;
663 cfs_spin_unlock(&exp->exp_lock);
665 return ldlm_init_export(exp);
668 static int filter_free_server_data(struct filter_obd *filter)
670 OBD_FREE_PTR(filter->fo_fsd);
671 filter->fo_fsd = NULL;
672 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
673 filter->fo_last_rcvd_slots = NULL;
677 /* assumes caller is already in kernel ctxt */
678 int filter_update_server_data(struct obd_device *obd, struct file *filp,
679 struct lr_server_data *fsd)
685 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
686 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
687 le64_to_cpu(fsd->lsd_last_transno));
688 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
689 le64_to_cpu(fsd->lsd_mount_count));
691 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
693 CERROR("error writing lr_server_data: rc = %d\n", rc);
698 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
701 struct filter_obd *filter = &obd->u.filter;
707 if (filter->fo_last_objid_files[group] == NULL) {
708 CERROR("Object group "LPU64" not fully setup; not updating "
709 "last_objid\n", group);
713 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
714 obd->obd_name, group, filter->fo_last_objids[group]);
716 tmp = cpu_to_le64(filter->fo_last_objids[group]);
717 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
718 &tmp, sizeof(tmp), &off, force_sync);
720 CERROR("error writing group "LPU64" last objid: rc = %d\n",
724 extern int ost_handle(struct ptlrpc_request *req);
725 /* assumes caller has already in kernel ctxt */
726 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
728 struct filter_obd *filter = &obd->u.filter;
729 struct lr_server_data *fsd;
730 struct lsd_client_data *lcd = NULL;
731 struct inode *inode = filp->f_dentry->d_inode;
732 unsigned long last_rcvd_size = i_size_read(inode);
739 /* ensure padding in the struct is the correct size */
740 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
741 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
742 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
743 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
745 OBD_ALLOC(fsd, sizeof(*fsd));
748 filter->fo_fsd = fsd;
750 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
751 if (filter->fo_last_rcvd_slots == NULL) {
752 OBD_FREE(fsd, sizeof(*fsd));
756 if (last_rcvd_size == 0) {
757 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
759 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
760 fsd->lsd_last_transno = 0;
761 mount_count = fsd->lsd_mount_count = 0;
762 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
763 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
764 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
765 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
766 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
767 /* OBD_COMPAT_OST is set in filter_connect_internal when the
768 * MDS first connects and assigns the OST index number. */
769 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
772 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
774 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
778 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
779 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
780 "using the wrong disk %s. Were the "
781 "/dev/ assignments rearranged?\n",
782 obd->obd_uuid.uuid, fsd->lsd_uuid);
783 GOTO(err_fsd, rc = -EINVAL);
785 mount_count = le64_to_cpu(fsd->lsd_mount_count);
786 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
788 /* Assume old last_rcvd format unless I_C_LR is set */
789 if (!(fsd->lsd_feature_incompat &
790 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
791 fsd->lsd_last_transno = fsd->lsd_compat14;
793 /* OBD_COMPAT_OST is set in filter_connect_internal when the
794 * MDS first connects and assigns the OST index number. */
795 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
799 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
800 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
801 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
802 ~FILTER_INCOMPAT_SUPP);
803 GOTO(err_fsd, rc = -EINVAL);
805 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
806 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
807 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
808 ~FILTER_ROCOMPAT_SUPP);
809 /* Do something like remount filesystem read-only */
810 GOTO(err_fsd, rc = -EINVAL);
813 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
815 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
816 obd->obd_name, start_epoch);
817 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
818 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
819 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
820 obd->obd_name, mount_count + 1);
821 CDEBUG(D_INODE, "%s: server data size: %u\n",
822 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
823 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
824 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
825 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
826 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
827 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
828 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
829 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
830 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
831 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
832 le16_to_cpu(fsd->lsd_client_size));
834 if (!obd->obd_replayable) {
835 CWARN("%s: recovery support OFF\n", obd->obd_name);
839 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
840 off < last_rcvd_size; cl_idx++) {
842 struct obd_export *exp;
843 struct filter_export_data *fed;
848 GOTO(err_client, rc = -ENOMEM);
851 /* Don't assume off is incremented properly by
852 * fsfilt_read_record(), in case sizeof(*lcd)
853 * isn't the same as fsd->lsd_client_size. */
854 off = le32_to_cpu(fsd->lsd_client_start) +
855 cl_idx * le16_to_cpu(fsd->lsd_client_size);
856 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
858 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
859 LAST_RCVD, cl_idx, off, rc);
860 break; /* read error shouldn't cause startup to fail */
863 if (lcd->lcd_uuid[0] == '\0') {
864 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
869 check_lcd(obd->obd_name, cl_idx, lcd);
871 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
873 /* These exports are cleaned up by filter_disconnect(), so they
874 * need to be set up like real exports as filter_connect() does.
876 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
878 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
879 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
880 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
882 if (PTR_ERR(exp) == -EALREADY) {
883 /* export already exists, zero out this one */
884 CERROR("Zeroing out duplicate export due to "
886 lcd->lcd_uuid[0] = '\0';
888 GOTO(err_client, rc = PTR_ERR(exp));
891 fed = &exp->exp_filter_data;
893 fed->fed_group = 0; /* will be assigned at connect */
894 filter_export_stats_init(obd, exp, NULL);
895 rc = filter_client_add(obd, exp, cl_idx);
896 /* can't fail for existing client */
897 LASSERTF(rc == 0, "rc = %d\n", rc);
899 /* VBR: set export last committed */
900 exp->exp_last_committed = last_rcvd;
901 cfs_spin_lock(&exp->exp_lock);
902 exp->exp_connecting = 0;
903 exp->exp_in_recovery = 0;
904 cfs_spin_unlock(&exp->exp_lock);
905 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
906 obd->obd_max_recoverable_clients++;
907 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
909 class_export_put(exp);
912 /* Need to check last_rcvd even for duplicated exports. */
913 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
916 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
917 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
923 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
925 filter->fo_mount_count = mount_count + 1;
926 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
928 /* save it, so mount count and last_transno is current */
929 rc = filter_update_server_data(obd, filp, filter->fo_fsd);
931 GOTO(err_client, rc);
936 class_disconnect_exports(obd);
938 filter_free_server_data(filter);
942 static int filter_cleanup_groups(struct obd_device *obd)
944 struct filter_obd *filter = &obd->u.filter;
946 struct dentry *dentry;
950 if (filter->fo_dentry_O_groups != NULL) {
951 for (i = 0; i < filter->fo_group_count; i++) {
952 dentry = filter->fo_dentry_O_groups[i];
956 OBD_FREE(filter->fo_dentry_O_groups,
957 filter->fo_group_count *
958 sizeof(*filter->fo_dentry_O_groups));
959 filter->fo_dentry_O_groups = NULL;
961 if (filter->fo_last_objid_files != NULL) {
962 for (i = 0; i < filter->fo_group_count; i++) {
963 filp = filter->fo_last_objid_files[i];
967 OBD_FREE(filter->fo_last_objid_files,
968 filter->fo_group_count *
969 sizeof(*filter->fo_last_objid_files));
970 filter->fo_last_objid_files = NULL;
972 if (filter->fo_dentry_O_sub != NULL) {
973 for (i = 0; i < filter->fo_group_count; i++) {
974 for (j = 0; j < filter->fo_subdir_count; j++) {
975 dentry = filter->fo_dentry_O_sub[i].dentry[j];
980 OBD_FREE(filter->fo_dentry_O_sub,
981 filter->fo_group_count *
982 sizeof(*filter->fo_dentry_O_sub));
983 filter->fo_dentry_O_sub = NULL;
985 if (filter->fo_last_objids != NULL) {
986 OBD_FREE(filter->fo_last_objids,
987 filter->fo_group_count *
988 sizeof(*filter->fo_last_objids));
989 filter->fo_last_objids = NULL;
991 if (filter->fo_dentry_O != NULL) {
992 f_dput(filter->fo_dentry_O);
993 filter->fo_dentry_O = NULL;
998 static int filter_update_last_group(struct obd_device *obd, int group)
1000 struct filter_obd *filter = &obd->u.filter;
1001 struct file *filp = NULL;
1002 int last_group = 0, rc;
1006 if (group <= filter->fo_committed_group)
1009 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1013 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1017 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1019 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1023 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1024 obd->obd_name, last_group, group);
1028 /* must be sync: bXXXX */
1029 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1031 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1035 filter->fo_committed_group = group;
1038 filp_close(filp, 0);
1042 static int filter_read_group_internal(struct obd_device *obd, int group,
1045 struct filter_obd *filter = &obd->u.filter;
1046 __u64 *new_objids = NULL;
1047 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1048 struct dentry **new_groups = NULL;
1049 struct file **new_files = NULL;
1050 struct dentry *dentry;
1052 int old_count = filter->fo_group_count, rc, stage = 0, i;
1056 int len = group + 1;
1058 snprintf(name, 24, "%d", group);
1062 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1064 if (IS_ERR(dentry)) {
1065 CERROR("Cannot lookup expected object group %d: %ld\n",
1066 group, PTR_ERR(dentry));
1067 RETURN(PTR_ERR(dentry));
1070 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1072 if (IS_ERR(dentry)) {
1073 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1075 RETURN(PTR_ERR(dentry));
1080 snprintf(name, 24, "O/%d/LAST_ID", group);
1082 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1084 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1085 GOTO(cleanup, rc = PTR_ERR(filp));
1089 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1091 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1095 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1096 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1097 if (tmp_subdirs == NULL)
1098 GOTO(cleanup, rc = -ENOMEM);
1101 for (i = 0; i < filter->fo_subdir_count; i++) {
1103 snprintf(dir, sizeof(dir), "d%u", i);
1105 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1108 if (IS_ERR(tmp_subdirs->dentry[i])) {
1109 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1110 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1115 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1116 tmp_subdirs->dentry[i]);
1120 /* 'group' is an index; we need an array of length 'group + 1' */
1121 if (group + 1 > old_count) {
1122 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1123 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1124 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1125 OBD_ALLOC(new_files, len * sizeof(*new_files));
1127 if (new_objids == NULL || new_subdirs == NULL ||
1128 new_groups == NULL || new_files == NULL)
1129 GOTO(cleanup, rc = -ENOMEM);
1132 memcpy(new_objids, filter->fo_last_objids,
1133 old_count * sizeof(*new_objids));
1134 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1135 old_count * sizeof(*new_subdirs));
1136 memcpy(new_groups, filter->fo_dentry_O_groups,
1137 old_count * sizeof(*new_groups));
1138 memcpy(new_files, filter->fo_last_objid_files,
1139 old_count * sizeof(*new_files));
1141 OBD_FREE(filter->fo_last_objids,
1142 old_count * sizeof(*new_objids));
1143 OBD_FREE(filter->fo_dentry_O_sub,
1144 old_count * sizeof(*new_subdirs));
1145 OBD_FREE(filter->fo_dentry_O_groups,
1146 old_count * sizeof(*new_groups));
1147 OBD_FREE(filter->fo_last_objid_files,
1148 old_count * sizeof(*new_files));
1150 filter->fo_last_objids = new_objids;
1151 filter->fo_dentry_O_sub = new_subdirs;
1152 filter->fo_dentry_O_groups = new_groups;
1153 filter->fo_last_objid_files = new_files;
1154 filter->fo_group_count = len;
1157 filter->fo_dentry_O_groups[group] = dentry;
1158 filter->fo_last_objid_files[group] = filp;
1159 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1160 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1161 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1164 filter_update_last_group(obd, group);
1166 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1167 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1168 rc = filter_update_last_objid(obd, group, 1);
1172 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1173 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1174 obd->obd_name, group, last_objid);
1179 if (new_objids != NULL)
1180 OBD_FREE(new_objids, len * sizeof(*new_objids));
1181 if (new_subdirs != NULL)
1182 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1183 if (new_groups != NULL)
1184 OBD_FREE(new_groups, len * sizeof(*new_groups));
1185 if (new_files != NULL)
1186 OBD_FREE(new_files, len * sizeof(*new_files));
1188 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1189 for (i = 0; i < filter->fo_subdir_count; i++) {
1190 if (tmp_subdirs->dentry[i] != NULL)
1191 dput(tmp_subdirs->dentry[i]);
1193 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1196 filp_close(filp, 0);
1203 static int filter_read_groups(struct obd_device *obd, int last_group,
1206 struct filter_obd *filter = &obd->u.filter;
1207 int old_count, group, rc = 0;
1209 cfs_down(&filter->fo_init_lock);
1210 old_count = filter->fo_group_count;
1211 for (group = old_count; group <= last_group; group++) {
1212 rc = filter_read_group_internal(obd, group, create);
1216 cfs_up(&filter->fo_init_lock);
1220 /* FIXME: object groups */
1221 static int filter_prep_groups(struct obd_device *obd)
1223 struct filter_obd *filter = &obd->u.filter;
1224 struct dentry *O_dentry;
1226 int last_group, rc = 0, cleanup_phase = 0;
1230 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1232 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1233 if (IS_ERR(O_dentry)) {
1234 rc = PTR_ERR(O_dentry);
1235 CERROR("cannot open/create O: rc = %d\n", rc);
1238 filter->fo_dentry_O = O_dentry;
1239 cleanup_phase = 1; /* O_dentry */
1241 /* we have to initialize all groups before first connections from
1242 * clients because they may send create/destroy for any group -bzzz */
1243 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1245 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1246 GOTO(cleanup, rc = PTR_ERR(filp));
1248 cleanup_phase = 2; /* filp */
1250 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1252 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1257 last_group = FILTER_GROUP_MDS0;
1259 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1260 FILTER_GROUP_MDS0, last_group);
1261 filter->fo_committed_group = last_group;
1262 rc = filter_read_groups(obd, last_group, 1);
1266 filp_close(filp, 0);
1270 switch (cleanup_phase) {
1272 filp_close(filp, 0);
1274 filter_cleanup_groups(obd);
1275 f_dput(filter->fo_dentry_O);
1276 filter->fo_dentry_O = NULL;
1284 /* setup the object store with correct subdirectories */
1285 static int filter_prep(struct obd_device *obd)
1287 struct lvfs_run_ctxt saved;
1288 struct filter_obd *filter = &obd->u.filter;
1290 struct inode *inode;
1294 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1295 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1296 if (!file || IS_ERR(file)) {
1298 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1302 filter->fo_rcvd_filp = file;
1303 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1304 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1305 file->f_dentry->d_inode->i_mode);
1306 GOTO(err_filp, rc = -ENOENT);
1309 inode = file->f_dentry->d_parent->d_inode;
1310 /* We use i_op->unlink directly in filter_vfs_unlink() */
1311 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1312 CERROR("%s: filesystem does not support create/unlink ops\n",
1314 GOTO(err_filp, rc = -EOPNOTSUPP);
1317 /** lu_target has very limited use in filter now */
1318 lut_init(NULL, &filter->fo_lut, obd, NULL);
1320 rc = filter_init_server_data(obd, file);
1322 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1326 target_recovery_init(&filter->fo_lut, ost_handle);
1328 /* open and create health check io file*/
1329 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1332 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1334 GOTO(err_server_data, rc);
1336 filter->fo_obt.obt_health_check_filp = file;
1337 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1338 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1339 file->f_dentry->d_inode->i_mode);
1340 GOTO(err_health_check, rc = -ENOENT);
1342 rc = lvfs_check_io_health(obd, file);
1344 GOTO(err_health_check, rc);
1346 rc = filter_prep_groups(obd);
1348 GOTO(err_health_check, rc);
1350 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1355 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1356 CERROR("can't close %s after error\n", HEALTH_CHECK);
1357 filter->fo_obt.obt_health_check_filp = NULL;
1359 target_recovery_fini(obd);
1360 filter_free_server_data(filter);
1362 if (filp_close(filter->fo_rcvd_filp, 0))
1363 CERROR("can't close %s after error\n", LAST_RCVD);
1364 filter->fo_rcvd_filp = NULL;
1368 /* cleanup the filter: write last used object id to status file */
1369 static void filter_post(struct obd_device *obd)
1371 struct lvfs_run_ctxt saved;
1372 struct filter_obd *filter = &obd->u.filter;
1375 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1376 * best to start a transaction with h_sync, because we removed this
1379 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1380 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1383 CERROR("error writing server data: rc = %d\n", rc);
1385 for (i = 0; i < filter->fo_group_count; i++) {
1386 rc = filter_update_last_objid(obd, i,
1387 (i == filter->fo_group_count - 1));
1389 CERROR("error writing group %d lastobjid: rc = %d\n",
1393 rc = filp_close(filter->fo_rcvd_filp, 0);
1394 filter->fo_rcvd_filp = NULL;
1396 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1398 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1399 filter->fo_obt.obt_health_check_filp = NULL;
1401 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1403 filter_cleanup_groups(obd);
1404 filter_free_server_data(filter);
1405 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1407 filter_free_capa_keys(filter);
1408 cleanup_capa_hash(filter->fo_capa_hash);
1411 static void filter_set_last_id(struct filter_obd *filter,
1412 obd_id id, obd_gr group)
1414 LASSERT(filter->fo_fsd != NULL);
1415 LASSERT(group <= filter->fo_group_count);
1417 cfs_spin_lock(&filter->fo_objidlock);
1418 filter->fo_last_objids[group] = id;
1419 cfs_spin_unlock(&filter->fo_objidlock);
1422 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1425 LASSERT(filter->fo_fsd != NULL);
1426 LASSERT(group <= filter->fo_group_count);
1427 LASSERT(filter->fo_last_objids != NULL);
1429 /* FIXME: object groups */
1430 cfs_spin_lock(&filter->fo_objidlock);
1431 id = filter->fo_last_objids[group];
1432 cfs_spin_unlock(&filter->fo_objidlock);
1436 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1438 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1442 /* We never dget the object parent, so DON'T dput it either */
1443 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1445 struct filter_obd *filter = &obd->u.filter;
1446 struct filter_subdirs *subdirs;
1447 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1449 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1450 return filter->fo_dentry_O_groups[group];
1452 subdirs = &filter->fo_dentry_O_sub[group];
1453 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1456 /* We never dget the object parent, so DON'T dput it either */
1457 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1460 unsigned long now = jiffies;
1461 struct dentry *dparent = filter_parent(obd, group, objid);
1464 if (IS_ERR(dparent))
1466 if (dparent == NULL)
1467 return ERR_PTR(-ENOENT);
1469 rc = filter_lock_dentry(obd, dparent);
1470 fsfilt_check_slow(obd, now, "parent lock");
1471 return rc ? ERR_PTR(rc) : dparent;
1474 /* We never dget the object parent, so DON'T dput it either */
1475 static void filter_parent_unlock(struct dentry *dparent)
1477 UNLOCK_INODE_MUTEX(dparent->d_inode);
1480 /* How to get files, dentries, inodes from object id's.
1482 * If dir_dentry is passed, the caller has already locked the parent
1483 * appropriately for this operation (normally a write lock). If
1484 * dir_dentry is NULL, we do a read lock while we do the lookup to
1485 * avoid races with create/destroy and such changing the directory
1486 * internal to the filesystem code. */
1487 struct dentry *filter_fid2dentry(struct obd_device *obd,
1488 struct dentry *dir_dentry,
1489 obd_gr group, obd_id id)
1491 struct dentry *dparent = dir_dentry;
1492 struct dentry *dchild;
1497 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1498 obd->u.filter.fo_destroys_in_progress == 0) {
1499 /* don't fail lookups for orphan recovery, it causes
1500 * later LBUGs when objects still exist during precreate */
1501 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1502 RETURN(ERR_PTR(-ENOENT));
1505 CERROR("fatal: invalid object id 0\n");
1506 RETURN(ERR_PTR(-ESTALE));
1509 len = sprintf(name, LPU64, id);
1510 if (dir_dentry == NULL) {
1511 dparent = filter_parent_lock(obd, group, id);
1512 if (IS_ERR(dparent)) {
1513 CERROR("%s: error getting object "LPU64":"LPU64
1514 " parent: rc %ld\n", obd->obd_name,
1515 id, group, PTR_ERR(dparent));
1519 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1520 dparent->d_name.len, dparent->d_name.name, name);
1521 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1522 if (dir_dentry == NULL)
1523 filter_parent_unlock(dparent);
1524 if (IS_ERR(dchild)) {
1525 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1530 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1531 CERROR("%s: got bad object "LPU64" inode %lu\n",
1532 obd->obd_name, id, dchild->d_inode->i_ino);
1534 RETURN(ERR_PTR(-ENOENT));
1537 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1538 name, dchild, atomic_read(&dchild->d_count));
1540 LASSERT(atomic_read(&dchild->d_count) > 0);
1545 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1546 obd_id group, struct lustre_handle *lockh)
1548 int flags = LDLM_AST_DISCARD_DATA, rc;
1549 struct ldlm_res_id res_id;
1550 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1553 osc_build_res_name(objid, group, &res_id);
1554 /* Tell the clients that the object is gone now and that they should
1555 * throw away any cached pages. */
1556 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1557 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1558 ldlm_completion_ast, NULL, NULL, 0, NULL,
1565 static void filter_fini_destroy(struct obd_device *obd,
1566 struct lustre_handle *lockh)
1568 if (lustre_handle_is_used(lockh))
1569 ldlm_lock_decref(lockh, LCK_PW);
1572 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1573 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1574 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1575 * reverse. Caller must take i_sem before starting the transaction and we
1576 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1577 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1578 struct vfsmount *mnt)
1583 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1584 * itself we already hold dir->i_mutex for child create/unlink ops */
1585 LASSERT(dentry->d_inode != NULL);
1586 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1587 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1591 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1592 GOTO(out, rc = -ENOENT);
1594 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1599 GOTO(out, rc = -EPERM);
1601 /* check_sticky() */
1602 if ((dentry->d_inode->i_uid != current->fsuid &&
1603 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1604 IS_IMMUTABLE(dentry->d_inode))
1605 GOTO(out, rc = -EPERM);
1607 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1608 * that was done because of journal_start (which is already done
1609 * here) or some other ordering issue. */
1612 rc = ll_security_inode_unlink(dir, dentry, mnt);
1616 rc = dir->i_op->unlink(dir, dentry);
1618 /* need to drop i_mutex before we lose inode reference */
1619 UNLOCK_INODE_MUTEX(dentry->d_inode);
1626 /* Caller must hold LCK_PW on parent and push us into kernel context.
1627 * Caller must hold child i_mutex, we drop it always.
1628 * Caller is also required to ensure that dchild->d_inode exists. */
1629 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1630 obd_gr group, struct dentry *dparent,
1631 struct dentry *dchild)
1633 struct inode *inode = dchild->d_inode;
1634 struct filter_obd *filter = &obd->u.filter;
1637 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1638 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1639 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1640 (unsigned long)inode->i_nlink,
1641 atomic_read(&inode->i_count));
1644 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1646 CERROR("error unlinking objid %.*s: rc %d\n",
1647 dchild->d_name.len, dchild->d_name.name, rc);
1651 struct filter_intent_args {
1652 struct ldlm_lock **victim;
1657 static enum interval_iter filter_intent_cb(struct interval_node *n,
1660 struct ldlm_interval *node = (struct ldlm_interval *)n;
1661 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1662 __u64 size = arg->size;
1663 struct ldlm_lock **v = arg->victim;
1664 struct ldlm_lock *lck;
1666 /* If the interval is lower than the current file size,
1668 if (interval_high(n) <= size)
1669 return INTERVAL_ITER_STOP;
1671 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1672 /* Don't send glimpse ASTs to liblustre clients.
1673 * They aren't listening for them, and they do
1674 * entirely synchronous I/O anyways. */
1675 if (lck->l_export == NULL ||
1676 lck->l_export->exp_libclient == 1)
1679 if (*arg->liblustre)
1680 *arg->liblustre = 0;
1683 *v = LDLM_LOCK_GET(lck);
1684 } else if ((*v)->l_policy_data.l_extent.start <
1685 lck->l_policy_data.l_extent.start) {
1686 LDLM_LOCK_RELEASE(*v);
1687 *v = LDLM_LOCK_GET(lck);
1690 /* the same policy group - every lock has the
1691 * same extent, so needn't do it any more */
1695 return INTERVAL_ITER_CONT;
1698 static int filter_intent_policy(struct ldlm_namespace *ns,
1699 struct ldlm_lock **lockp, void *req_cookie,
1700 ldlm_mode_t mode, int flags, void *data)
1702 CFS_LIST_HEAD(rpc_list);
1703 struct ptlrpc_request *req = req_cookie;
1704 struct ldlm_lock *lock = *lockp, *l = NULL;
1705 struct ldlm_resource *res = lock->l_resource;
1706 ldlm_processing_policy policy;
1707 struct ost_lvb *res_lvb, *reply_lvb;
1708 struct ldlm_reply *rep;
1710 int idx, rc, tmpflags = 0, only_liblustre = 1;
1711 struct ldlm_interval_tree *tree;
1712 struct filter_intent_args arg;
1713 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1714 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1715 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1718 policy = ldlm_get_processing_policy(res);
1719 LASSERT(policy != NULL);
1720 LASSERT(req != NULL);
1722 rc = lustre_pack_reply(req, 3, repsize, NULL);
1724 RETURN(req->rq_status = rc);
1726 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1727 LASSERT(rep != NULL);
1729 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1730 sizeof(*reply_lvb));
1731 LASSERT(reply_lvb != NULL);
1733 //fixup_handle_for_resent_req(req, lock, &lockh);
1735 /* Call the extent policy function to see if our request can be
1736 * granted, or is blocked.
1737 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1738 * lock, and should not be granted if the lock will be blocked.
1741 LASSERT(ns == res->lr_namespace);
1743 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1744 check_res_locked(res);
1746 /* FIXME: we should change the policy function slightly, to not make
1747 * this list at all, since we just turn around and free it */
1748 while (!cfs_list_empty(&rpc_list)) {
1749 struct ldlm_lock *wlock =
1750 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1752 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1753 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1754 lock->l_flags &= ~LDLM_FL_CP_REQD;
1755 cfs_list_del_init(&wlock->l_cp_ast);
1756 LDLM_LOCK_RELEASE(wlock);
1759 /* The lock met with no resistance; we're finished. */
1760 if (rc == LDLM_ITER_CONTINUE) {
1761 /* do not grant locks to the liblustre clients: they cannot
1762 * handle ASTs robustly. We need to do this while still
1763 * holding ns_lock to avoid the lock remaining on the res_link
1764 * list (and potentially being added to l_pending_list by an
1765 * AST) when we are going to drop this lock ASAP. */
1766 if (lock->l_export->exp_libclient ||
1767 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1768 ldlm_resource_unlink_lock(lock);
1769 err = ELDLM_LOCK_ABORTED;
1771 err = ELDLM_LOCK_REPLACED;
1777 /* Do not grant any lock, but instead send GL callbacks. The extent
1778 * policy nicely created a list of all PW locks for us. We will choose
1779 * the highest of those which are larger than the size in the LVB, if
1780 * any, and perform a glimpse callback. */
1781 res_lvb = res->lr_lvb_data;
1782 LASSERT(res_lvb != NULL);
1783 *reply_lvb = *res_lvb;
1786 * ->ns_lock guarantees that no new locks are granted, and,
1787 * therefore, that res->lr_lvb_data cannot increase beyond the
1788 * end of already granted lock. As a result, it is safe to
1789 * check against "stale" reply_lvb->lvb_size value without
1792 arg.size = reply_lvb->lvb_size;
1794 arg.liblustre = &only_liblustre;
1795 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1796 tree = &res->lr_itree[idx];
1797 if (tree->lit_mode == LCK_PR)
1800 interval_iterate_reverse(tree->lit_root,
1801 filter_intent_cb, &arg);
1805 /* There were no PW locks beyond the size in the LVB; finished. */
1807 if (only_liblustre) {
1808 /* If we discovered a liblustre client with a PW lock,
1809 * however, the LVB may be out of date! The LVB is
1810 * updated only on glimpse (which we don't do for
1811 * liblustre clients) and cancel (which the client
1812 * obviously has not yet done). So if it has written
1813 * data but kept the lock, the LVB is stale and needs
1814 * to be updated from disk.
1816 * Of course, this will all disappear when we switch to
1817 * taking liblustre locks on the OST. */
1818 ldlm_res_lvbo_update(res, NULL, 1);
1820 RETURN(ELDLM_LOCK_ABORTED);
1824 * This check is for lock taken in filter_prepare_destroy() that does
1825 * not have l_glimpse_ast set. So the logic is: if there is a lock
1826 * with no l_glimpse_ast set, this object is being destroyed already.
1828 * Hence, if you are grabbing DLM locks on the server, always set
1829 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1831 if (l->l_glimpse_ast == NULL) {
1832 /* We are racing with unlink(); just return -ENOENT */
1833 rep->lock_policy_res1 = -ENOENT;
1837 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1838 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1839 /* Update the LVB from disk if the AST failed (this is a legal race) */
1841 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1842 * sending ast is not handled. This can result in lost client writes.
1845 ldlm_res_lvbo_update(res, NULL, 1);
1848 *reply_lvb = *res_lvb;
1852 LDLM_LOCK_RELEASE(l);
1854 RETURN(ELDLM_LOCK_ABORTED);
1858 * per-obd_device iobuf pool.
1860 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1861 * allocations in write-path has to be minimized (see bug 5137).
1863 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1864 * OST threads (see ost_thread_{init,done}()).
1866 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1867 * at the OST layer there are only (potentially) multiple obd_device of type
1868 * unknown at the time of OST thread creation.
1870 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1871 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1872 * it's very own iobuf.
1876 * filter_kiobuf_pool_init()
1878 * filter_kiobuf_pool_done()
1880 * filter_iobuf_get()
1882 * operate on this array. They are "generic" in a sense that they don't depend
1883 * on actual type of iobuf's (the latter depending on Linux kernel version).
1887 * destroy pool created by filter_iobuf_pool_init
1889 static void filter_iobuf_pool_done(struct filter_obd *filter)
1891 struct filter_iobuf **pool;
1896 pool = filter->fo_iobuf_pool;
1898 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1899 if (pool[i] != NULL)
1900 filter_free_iobuf(pool[i]);
1902 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1903 filter->fo_iobuf_pool = NULL;
1908 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1910 struct filter_obd *filter = &obd->u.filter;
1911 struct sptlrpc_rule_set tmp_rset;
1914 sptlrpc_rule_set_init(&tmp_rset);
1915 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1917 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1922 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1924 cfs_write_lock(&filter->fo_sptlrpc_lock);
1925 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1926 filter->fo_sptlrpc_rset = tmp_rset;
1927 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1933 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1935 static int filter_iobuf_pool_init(struct filter_obd *filter)
1942 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1944 if (filter->fo_iobuf_pool == NULL)
1947 filter->fo_iobuf_count = OSS_THREADS_MAX;
1952 /* Return iobuf allocated for @thread_id. We don't know in advance how
1953 * many threads there will be so we allocate a large empty array and only
1954 * fill in those slots that are actually in use.
1955 * If we haven't allocated a pool entry for this thread before, do so now. */
1956 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1958 int thread_id = (oti && oti->oti_thread) ?
1959 oti->oti_thread->t_id : -1;
1960 struct filter_iobuf *pool = NULL;
1961 struct filter_iobuf **pool_place = NULL;
1963 if (thread_id >= 0) {
1964 LASSERT(thread_id < filter->fo_iobuf_count);
1965 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1968 if (unlikely(pool == NULL)) {
1969 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1970 PTLRPC_MAX_BRW_PAGES);
1971 if (pool_place != NULL)
1978 /* mount the file system (secretly). lustre_cfg parameters are:
1981 * 3 = flags: failover=f, failout=n
1984 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1987 struct filter_obd *filter = &obd->u.filter;
1988 struct vfsmount *mnt;
1989 struct lustre_mount_info *lmi;
1990 struct obd_uuid uuid;
1998 if (lcfg->lcfg_bufcount < 3 ||
1999 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2000 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2003 lmi = server_get_mount(obd->obd_name);
2005 /* We already mounted in lustre_fill_super.
2006 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2007 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2009 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2011 /* old path - used by lctl */
2012 CERROR("Using old MDS mount method\n");
2013 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2014 MS_NOATIME|MS_NODIRATIME,
2015 lustre_cfg_string(lcfg, 1), option);
2018 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2019 lustre_cfg_string(lcfg, 1), rc);
2023 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2025 if (IS_ERR(obd->obd_fsops))
2026 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2028 rc = filter_iobuf_pool_init(filter);
2032 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2033 CERROR("%s: Underlying device is marked as read-only. "
2034 "Setup failed\n", obd->obd_name);
2035 GOTO(err_ops, rc = -EROFS);
2038 /* failover is the default */
2039 obd->obd_replayable = 1;
2041 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2042 str = lustre_cfg_string(lcfg, 3);
2043 if (strchr(str, 'n')) {
2044 CWARN("%s: recovery disabled\n", obd->obd_name);
2045 obd->obd_replayable = 0;
2049 filter->fo_vfsmnt = mnt;
2050 obd->u.obt.obt_sb = mnt->mnt_sb;
2051 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2052 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2054 fsfilt_setup(obd, obd->u.obt.obt_sb);
2056 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2057 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2058 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2059 obd->obd_lvfs_ctxt.fs = get_ds();
2060 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2062 cfs_init_mutex(&filter->fo_init_lock);
2063 filter->fo_committed_group = 0;
2064 filter->fo_destroys_in_progress = 0;
2065 for (i = 0; i < 32; i++)
2066 cfs_sema_init(&filter->fo_create_locks[i], 1);
2068 cfs_spin_lock_init(&filter->fo_translock);
2069 cfs_spin_lock_init(&filter->fo_objidlock);
2070 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2071 cfs_sema_init(&filter->fo_alloc_lock, 1);
2072 init_brw_stats(&filter->fo_filter_stats);
2073 filter->fo_read_cache = 1; /* enable read-only cache by default */
2074 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2075 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2076 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2077 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2079 rc = filter_prep(obd);
2083 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2084 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2086 filter->fo_fl_oss_capa = 1;
2088 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2089 filter->fo_capa_hash = init_capa_hash();
2090 if (filter->fo_capa_hash == NULL)
2091 GOTO(err_post, rc = -ENOMEM);
2093 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2094 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2095 LDLM_NAMESPACE_GREEDY);
2096 if (obd->obd_namespace == NULL)
2097 GOTO(err_post, rc = -ENOMEM);
2098 obd->obd_namespace->ns_lvbp = obd;
2099 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2100 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2102 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2103 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2105 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2107 CERROR("failed to setup llogging subsystems\n");
2111 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2112 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2113 /* do this after llog being initialized */
2114 filter_adapt_sptlrpc_conf(obd, 1);
2116 rc = lquota_setup(filter_quota_interface_ref, obd);
2120 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2121 if (q->max_sectors < q->max_hw_sectors &&
2122 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2123 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2124 "for larger I/O requests: max_sectors = %u "
2125 "could be up to max_hw_sectors=%u\n",
2126 obd->obd_name, mnt->mnt_sb->s_id,
2127 q->max_sectors, q->max_hw_sectors);
2129 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2130 if (uuid_ptr != NULL) {
2131 class_uuid_unparse(uuid_ptr, &uuid);
2137 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2139 if (obd->obd_recovering) {
2140 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2141 "recovery for at least %d:%.02d, or until %d "
2142 "client%s reconnect%s.\n",
2143 obd->obd_name, lustre_cfg_string(lcfg, 1),
2144 label ?: "", label ? "/" : "", str,
2145 obd->obd_recovery_timeout / 60,
2146 obd->obd_recovery_timeout % 60,
2147 obd->obd_max_recoverable_clients,
2148 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2149 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2151 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2152 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2153 label ?: "", label ? "/" : "", str,
2154 obd->obd_replayable ? "enabled" : "disabled");
2162 fsfilt_put_ops(obd->obd_fsops);
2163 filter_iobuf_pool_done(filter);
2165 server_put_mount(obd->obd_name, mnt);
2166 obd->u.obt.obt_sb = 0;
2170 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2172 struct lprocfs_static_vars lvars;
2177 CLASSERT(offsetof(struct obd_device, u.obt) ==
2178 offsetof(struct obd_device, u.filter.fo_obt));
2180 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2183 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2184 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2187 addr = (unsigned long)cfs_page_address(page);
2188 clear_page((void *)addr);
2190 /* lprocfs must be setup before the filter so state can be safely added
2191 * to /proc incrementally as the filter is setup */
2192 lprocfs_filter_init_vars(&lvars);
2193 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2194 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2195 /* Init obdfilter private stats here */
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2197 LPROCFS_CNTR_AVGMINMAX,
2198 "read_bytes", "bytes");
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2200 LPROCFS_CNTR_AVGMINMAX,
2201 "write_bytes", "bytes");
2202 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2203 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2204 "get_page", "usec");
2205 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2206 LPROCFS_CNTR_AVGMINMAX,
2207 "get_page_failures", "num");
2208 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2209 LPROCFS_CNTR_AVGMINMAX,
2210 "cache_access", "pages");
2211 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2212 LPROCFS_CNTR_AVGMINMAX,
2213 "cache_hit", "pages");
2214 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2215 LPROCFS_CNTR_AVGMINMAX,
2216 "cache_miss", "pages");
2218 lproc_filter_attach_seqstat(obd);
2219 obd->obd_proc_exports_entry = lprocfs_register("exports",
2220 obd->obd_proc_entry,
2222 if (IS_ERR(obd->obd_proc_exports_entry)) {
2223 rc = PTR_ERR(obd->obd_proc_exports_entry);
2224 CERROR("error %d setting up lprocfs for %s\n",
2226 obd->obd_proc_exports_entry = NULL;
2229 if (obd->obd_proc_exports_entry)
2230 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2231 lprocfs_nid_stats_clear_read,
2232 lprocfs_nid_stats_clear_write, obd, NULL);
2234 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2235 LUSTRE_CFG_BUFLEN(lcfg, 4));
2236 rc = filter_common_setup(obd, lcfg, (void *)addr);
2237 OBD_PAGE_FREE(page);
2240 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2241 lprocfs_free_per_client_stats(obd);
2242 lprocfs_free_obd_stats(obd);
2243 lprocfs_obd_cleanup(obd);
2249 static struct llog_operations filter_mds_ost_repl_logops;
2251 static struct llog_operations filter_size_orig_logops = {
2252 .lop_setup = llog_obd_origin_setup,
2253 .lop_cleanup = llog_obd_origin_cleanup,
2254 .lop_add = llog_obd_origin_add
2257 static int filter_olg_fini(struct obd_llog_group *olg)
2259 struct llog_ctxt *ctxt;
2260 int rc = 0, rc2 = 0;
2263 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2265 rc = llog_cleanup(ctxt);
2267 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2269 rc2 = llog_cleanup(ctxt);
2274 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2276 rc2 = llog_cleanup(ctxt);
2285 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2286 struct obd_device *tgt)
2291 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2292 &filter_mds_ost_repl_logops);
2296 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2297 &filter_size_orig_logops);
2303 filter_olg_fini(olg);
2308 * Init the default olg, which is embeded in the obd_device, for filter.
2311 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2312 struct obd_device *tgt)
2314 struct filter_obd *filter = &obd->u.filter;
2315 struct llog_ctxt *ctxt;
2319 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2320 if (!filter->fo_lcm)
2323 filter_mds_ost_repl_logops = llog_client_ops;
2324 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2325 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2326 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2328 rc = filter_olg_init(obd, olg, tgt);
2330 GOTO(cleanup_lcm, rc);
2332 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2335 GOTO(cleanup_olg, rc);
2337 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2339 CERROR("Can't get ctxt for %p:%x\n", olg,
2340 LLOG_MDS_OST_REPL_CTXT);
2341 GOTO(cleanup_olg, rc = -ENODEV);
2343 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2344 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2345 llog_ctxt_put(ctxt);
2349 filter_olg_fini(olg);
2351 llog_recov_thread_fini(filter->fo_lcm, 1);
2352 filter->fo_lcm = NULL;
2357 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2358 struct obd_device *tgt, int *index)
2360 struct filter_obd *filter = &obd->u.filter;
2361 struct llog_ctxt *ctxt;
2365 LASSERT(olg != NULL);
2366 if (olg == &obd->obd_olg)
2367 return filter_default_olg_init(obd, olg, tgt);
2369 LASSERT(filter->fo_lcm != NULL);
2370 rc = filter_olg_init(obd, olg, tgt);
2373 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2375 CERROR("Can't get ctxt for %p:%x\n", olg,
2376 LLOG_MDS_OST_REPL_CTXT);
2377 filter_olg_fini(olg);
2380 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2381 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2382 llog_ctxt_put(ctxt);
2386 static int filter_llog_finish(struct obd_device *obd, int count)
2388 struct filter_obd *filter = &obd->u.filter;
2389 struct llog_ctxt *ctxt;
2392 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2395 * Make sure that no cached llcds left in recov_thread.
2396 * We actually do sync in disconnect time, but disconnect
2397 * may not come being marked rq_no_resend = 1.
2399 llog_sync(ctxt, NULL);
2402 * Balance class_import_get() in llog_receptor_accept().
2403 * This is safe to do, as llog is already synchronized
2404 * and its import may go.
2406 cfs_mutex_down(&ctxt->loc_sem);
2407 if (ctxt->loc_imp) {
2408 class_import_put(ctxt->loc_imp);
2409 ctxt->loc_imp = NULL;
2411 cfs_mutex_up(&ctxt->loc_sem);
2412 llog_ctxt_put(ctxt);
2415 if (filter->fo_lcm) {
2416 cfs_mutex_down(&ctxt->loc_sem);
2417 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2418 filter->fo_lcm = NULL;
2419 cfs_mutex_up(&ctxt->loc_sem);
2421 RETURN(filter_olg_fini(&obd->obd_olg));
2425 * Find the group llog according to group index in the llog group list.
2427 static struct obd_llog_group *
2428 filter_find_olg_internal(struct filter_obd *filter, int group)
2430 struct obd_llog_group *olg;
2432 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2433 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2434 if (olg->olg_group == group)
2441 * Find the group llog according to group index on the filter
2443 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2445 struct obd_llog_group *olg = NULL;
2446 struct filter_obd *filter;
2448 filter = &obd->u.filter;
2450 if (group == FILTER_GROUP_LLOG)
2451 RETURN(&obd->obd_olg);
2453 cfs_spin_lock(&filter->fo_llog_list_lock);
2454 olg = filter_find_olg_internal(filter, group);
2455 cfs_spin_unlock(&filter->fo_llog_list_lock);
2460 * Find the llog_group of the filter according to the group. If it can not
2461 * find, create the llog_group, which only happens when mds is being synced
2464 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2466 struct obd_llog_group *olg = NULL;
2467 struct filter_obd *filter;
2470 filter = &obd->u.filter;
2472 if (group == FILTER_GROUP_LLOG)
2473 RETURN(&obd->obd_olg);
2475 cfs_spin_lock(&filter->fo_llog_list_lock);
2476 olg = filter_find_olg_internal(filter, group);
2478 if (olg->olg_initializing) {
2479 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2481 GOTO(out_unlock, olg);
2486 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2488 llog_group_init(olg, group);
2489 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2490 olg->olg_initializing = 1;
2491 cfs_spin_unlock(&filter->fo_llog_list_lock);
2493 rc = obd_llog_init(obd, olg, obd, NULL);
2495 cfs_spin_lock(&filter->fo_llog_list_lock);
2496 cfs_list_del(&olg->olg_list);
2497 cfs_spin_unlock(&filter->fo_llog_list_lock);
2499 GOTO(out, olg = ERR_PTR(-ENOMEM));
2501 cfs_spin_lock(&filter->fo_llog_list_lock);
2502 olg->olg_initializing = 0;
2503 cfs_spin_unlock(&filter->fo_llog_list_lock);
2504 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2505 obd->obd_name, group, olg);
2510 cfs_spin_unlock(&filter->fo_llog_list_lock);
2514 static int filter_llog_connect(struct obd_export *exp,
2515 struct llogd_conn_body *body)
2517 struct obd_device *obd = exp->exp_obd;
2518 struct llog_ctxt *ctxt;
2519 struct obd_llog_group *olg;
2523 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2524 obd->obd_name, body->lgdc_logid.lgl_oid,
2525 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2527 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2529 CERROR(" %s: can not find olg of group %d\n",
2530 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2533 llog_group_set_export(olg, exp);
2535 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2536 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2537 body->lgdc_ctxt_idx);
2539 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2540 obd->obd_name, body->lgdc_logid.lgl_oid,
2541 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2543 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2544 obd->u.filter.fo_mds_ost_sync = 1;
2545 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2546 rc = llog_connect(ctxt, &body->lgdc_logid,
2547 &body->lgdc_gen, NULL);
2548 llog_ctxt_put(ctxt);
2550 CERROR("failed to connect rc %d idx %d\n", rc,
2551 body->lgdc_ctxt_idx);
2556 static int filter_llog_preclean(struct obd_device *obd)
2558 struct obd_llog_group *olg, *tmp;
2559 struct filter_obd *filter;
2560 cfs_list_t remove_list;
2564 rc = obd_llog_finish(obd, 0);
2566 CERROR("failed to cleanup llogging subsystem\n");
2568 filter = &obd->u.filter;
2569 CFS_INIT_LIST_HEAD(&remove_list);
2571 cfs_spin_lock(&filter->fo_llog_list_lock);
2572 while (!cfs_list_empty(&filter->fo_llog_list)) {
2573 olg = cfs_list_entry(filter->fo_llog_list.next,
2574 struct obd_llog_group, olg_list);
2575 cfs_list_del(&olg->olg_list);
2576 cfs_list_add(&olg->olg_list, &remove_list);
2578 cfs_spin_unlock(&filter->fo_llog_list_lock);
2580 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2581 cfs_list_del_init(&olg->olg_list);
2582 rc = filter_olg_fini(olg);
2584 CERROR("failed to cleanup llogging subsystem for %u\n",
2592 static int filter_precleanup(struct obd_device *obd,
2593 enum obd_cleanup_stage stage)
2599 case OBD_CLEANUP_EARLY:
2601 case OBD_CLEANUP_EXPORTS:
2602 /* Stop recovery before namespace cleanup. */
2603 target_recovery_fini(obd);
2604 rc = filter_llog_preclean(obd);
2610 static int filter_cleanup(struct obd_device *obd)
2612 struct filter_obd *filter = &obd->u.filter;
2616 LCONSOLE_WARN("%s: shutting down for failover; client state "
2617 "will be preserved.\n", obd->obd_name);
2619 obd_exports_barrier(obd);
2620 obd_zombie_barrier();
2622 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2623 lprocfs_free_per_client_stats(obd);
2624 lprocfs_free_obd_stats(obd);
2625 lprocfs_obd_cleanup(obd);
2626 lquota_cleanup(filter_quota_interface_ref, obd);
2628 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2629 obd->obd_namespace = NULL;
2631 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2633 if (obd->u.obt.obt_sb == NULL)
2638 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2639 shrink_dcache_sb(obd->u.obt.obt_sb);
2641 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2642 obd->u.obt.obt_sb = NULL;
2644 fsfilt_put_ops(obd->obd_fsops);
2646 filter_iobuf_pool_done(filter);
2648 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2653 static int filter_connect_internal(struct obd_export *exp,
2654 struct obd_connect_data *data,
2657 struct filter_export_data *fed = &exp->exp_filter_data;
2662 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2663 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2664 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2665 data->ocd_connect_flags, data->ocd_version,
2666 data->ocd_grant, data->ocd_index);
2668 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2669 CWARN("!!! This export (nid %s) used object group %d "
2670 "earlier; now it's trying to use group %d! This could "
2671 "be a bug in the MDS. Please report to "
2672 "http://bugzilla.lustre.org/\n",
2673 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2676 fed->fed_group = data->ocd_group;
2678 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2679 exp->exp_connect_flags = data->ocd_connect_flags;
2680 data->ocd_version = LUSTRE_VERSION_CODE;
2682 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2683 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2684 data->ocd_connect_flags & OBD_CONNECT_MDS))
2687 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2688 struct filter_obd *filter = &exp->exp_obd->u.filter;
2689 obd_size left, want;
2691 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2692 left = filter_grant_space_left(exp);
2693 want = data->ocd_grant;
2694 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2695 data->ocd_grant = fed->fed_grant;
2696 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2698 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2699 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2700 exp->exp_client_uuid.uuid, exp,
2701 data->ocd_grant, want, left);
2703 filter->fo_tot_granted_clients ++;
2706 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2707 struct filter_obd *filter = &exp->exp_obd->u.filter;
2708 struct lr_server_data *lsd = filter->fo_fsd;
2709 int index = le32_to_cpu(lsd->lsd_ost_index);
2711 if (!(lsd->lsd_feature_compat &
2712 cpu_to_le32(OBD_COMPAT_OST))) {
2713 /* this will only happen on the first connect */
2714 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2715 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2716 /* sync is not needed here as filter_client_add will
2717 * set exp_need_sync flag */
2718 filter_update_server_data(exp->exp_obd,
2719 filter->fo_rcvd_filp, lsd);
2720 } else if (index != data->ocd_index) {
2721 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2722 " %u doesn't match actual OST index"
2723 " %u in last_rcvd file, bad "
2725 obd_export_nid2str(exp), index,
2729 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2730 * FIXME: We don't strictly need the COMPAT flag for that,
2731 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2732 * FIXME: We needed it for the index, as index 0 is valid. */
2735 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2736 data->ocd_brw_size = 65536;
2737 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2738 data->ocd_brw_size = min(data->ocd_brw_size,
2739 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2740 LASSERT(data->ocd_brw_size);
2743 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2744 __u32 cksum_types = data->ocd_cksum_types;
2746 /* The client set in ocd_cksum_types the checksum types it
2747 * supports. We have to mask off the algorithms that we don't
2749 if (cksum_types & OBD_CKSUM_ALL)
2750 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2752 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2754 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2755 "%x\n", exp->exp_obd->obd_name,
2756 obd_export_nid2str(exp), cksum_types,
2757 data->ocd_cksum_types);
2759 /* This client does not support OBD_CONNECT_CKSUM
2760 * fall back to CRC32 */
2761 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2762 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2763 exp->exp_obd->obd_name,
2764 obd_export_nid2str(exp));
2770 static int filter_reconnect(const struct lu_env *env,
2771 struct obd_export *exp, struct obd_device *obd,
2772 struct obd_uuid *cluuid,
2773 struct obd_connect_data *data,
2779 if (exp == NULL || obd == NULL || cluuid == NULL)
2782 rc = filter_connect_internal(exp, data, 1);
2784 filter_export_stats_init(obd, exp, localdata);
2789 /* nearly identical to mds_connect */
2790 static int filter_connect(const struct lu_env *env,
2791 struct obd_export **exp, struct obd_device *obd,
2792 struct obd_uuid *cluuid,
2793 struct obd_connect_data *data, void *localdata)
2795 struct lvfs_run_ctxt saved;
2796 struct lustre_handle conn = { 0 };
2797 struct obd_export *lexp;
2798 struct filter_export_data *fed;
2799 struct lsd_client_data *lcd = NULL;
2804 if (exp == NULL || obd == NULL || cluuid == NULL)
2807 rc = class_connect(&conn, obd, cluuid);
2810 lexp = class_conn2export(&conn);
2811 LASSERT(lexp != NULL);
2813 fed = &lexp->exp_filter_data;
2815 rc = filter_connect_internal(lexp, data, 0);
2819 filter_export_stats_init(obd, lexp, localdata);
2820 if (obd->obd_replayable) {
2821 OBD_ALLOC(lcd, sizeof(*lcd));
2823 CERROR("filter: out of memory for client data\n");
2824 GOTO(cleanup, rc = -ENOMEM);
2827 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2829 rc = filter_client_add(obd, lexp, -1);
2834 group = data->ocd_group;
2836 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2837 obd->obd_name, lexp->exp_handle.h_cookie, group);
2839 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2840 rc = filter_read_groups(obd, group, 1);
2841 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2843 CERROR("can't read group %u\n", group);
2853 fed->fed_lcd = NULL;
2855 class_disconnect(lexp);
2864 /* Do extra sanity checks for grant accounting. We do this at connect,
2865 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2866 * always get rid of it or turn it off when we know accounting is good. */
2867 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2869 struct filter_export_data *fed;
2870 struct obd_export *exp;
2871 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2872 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2873 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2875 if (cfs_list_empty(&obd->obd_exports))
2878 /* We don't want to do this for large machines that do lots of
2879 mounts or unmounts. It burns... */
2880 if (obd->obd_num_exports > 100)
2883 cfs_spin_lock(&obd->obd_osfs_lock);
2884 cfs_spin_lock(&obd->obd_dev_lock);
2885 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2887 fed = &exp->exp_filter_data;
2888 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2891 if (maxsize > 0) { /* we may not have done a statfs yet */
2892 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2893 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2894 exp->exp_client_uuid.uuid, exp,
2895 fed->fed_grant, fed->fed_pending, maxsize);
2896 LASSERTF(fed->fed_dirty <= maxsize,
2897 "%s: cli %s/%p %ld > "LPU64"\n", func,
2898 exp->exp_client_uuid.uuid, exp,
2899 fed->fed_dirty, maxsize);
2902 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2903 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2904 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2906 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2907 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2908 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2909 tot_granted += fed->fed_grant + fed->fed_pending;
2910 tot_pending += fed->fed_pending;
2911 tot_dirty += fed->fed_dirty;
2913 fo_tot_granted = obd->u.filter.fo_tot_granted;
2914 fo_tot_pending = obd->u.filter.fo_tot_pending;
2915 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2916 cfs_spin_unlock(&obd->obd_dev_lock);
2917 cfs_spin_unlock(&obd->obd_osfs_lock);
2919 /* Do these assertions outside the spinlocks so we don't kill system */
2920 if (tot_granted != fo_tot_granted)
2921 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2922 func, tot_granted, fo_tot_granted);
2923 if (tot_pending != fo_tot_pending)
2924 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2925 func, tot_pending, fo_tot_pending);
2926 if (tot_dirty != fo_tot_dirty)
2927 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2928 func, tot_dirty, fo_tot_dirty);
2929 if (tot_pending > tot_granted)
2930 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2931 func, tot_pending, tot_granted);
2932 if (tot_granted > maxsize)
2933 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2934 func, tot_granted, maxsize);
2935 if (tot_dirty > maxsize)
2936 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2937 func, tot_dirty, maxsize);
2940 /* Remove this client from the grant accounting totals. We also remove
2941 * the export from the obd device under the osfs and dev locks to ensure
2942 * that the filter_grant_sanity_check() calculations are always valid.
2943 * The client should do something similar when it invalidates its import. */
2944 static void filter_grant_discard(struct obd_export *exp)
2946 struct obd_device *obd = exp->exp_obd;
2947 struct filter_obd *filter = &obd->u.filter;
2948 struct filter_export_data *fed = &exp->exp_filter_data;
2950 cfs_spin_lock(&obd->obd_osfs_lock);
2951 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2952 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2953 obd->obd_name, filter->fo_tot_granted,
2954 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2955 filter->fo_tot_granted -= fed->fed_grant;
2956 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2957 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2958 obd->obd_name, filter->fo_tot_pending,
2959 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2960 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2961 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2962 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2963 obd->obd_name, filter->fo_tot_dirty,
2964 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2965 filter->fo_tot_dirty -= fed->fed_dirty;
2969 cfs_spin_unlock(&obd->obd_osfs_lock);
2972 static int filter_destroy_export(struct obd_export *exp)
2976 if (exp->exp_filter_data.fed_pending)
2977 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2978 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2979 exp, exp->exp_filter_data.fed_pending);
2981 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2983 target_destroy_export(exp);
2984 ldlm_destroy_export(exp);
2986 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2990 if (!exp->exp_obd->obd_replayable)
2991 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2993 filter_grant_discard(exp);
2994 filter_fmd_cleanup(exp);
2996 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2997 struct filter_obd *filter = &exp->exp_obd->u.filter;
2998 if (filter->fo_tot_granted_clients > 0)
2999 filter->fo_tot_granted_clients --;
3002 if (!(exp->exp_flags & OBD_OPT_FORCE))
3003 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3008 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3010 struct obd_llog_group *olg_min, *olg;
3011 struct filter_obd *filter;
3012 int worked = -1, group;
3013 struct llog_ctxt *ctxt;
3016 filter = &obd->u.filter;
3018 /* we can't sync log holding spinlock. also, we do not want to get
3019 * into livelock. so we do following: loop over MDS's exports in
3020 * group order and skip already synced llogs -bzzz */
3022 /* look for group with min. number, but > worked */
3025 cfs_spin_lock(&filter->fo_llog_list_lock);
3026 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3027 if (olg->olg_group <= worked) {
3028 /* this group is already synced */
3031 if (group < olg->olg_group) {
3032 /* we have group with smaller number to sync */
3035 /* store current minimal group */
3037 group = olg->olg_group;
3039 cfs_spin_unlock(&filter->fo_llog_list_lock);
3041 if (olg_min == NULL)
3044 worked = olg_min->olg_group;
3045 if (olg_min->olg_exp &&
3046 (dexp == olg_min->olg_exp || dexp == NULL)) {
3048 ctxt = llog_group_get_ctxt(olg_min,
3049 LLOG_MDS_OST_REPL_CTXT);
3051 err = llog_sync(ctxt, olg_min->olg_exp);
3052 llog_ctxt_put(ctxt);
3054 CERROR("error flushing logs to MDS: "
3059 } while (olg_min != NULL);
3062 /* Also incredibly similar to mds_disconnect */
3063 static int filter_disconnect(struct obd_export *exp)
3065 struct obd_device *obd = exp->exp_obd;
3070 class_export_get(exp);
3072 if (!(exp->exp_flags & OBD_OPT_FORCE))
3073 filter_grant_sanity_check(obd, __FUNCTION__);
3074 filter_grant_discard(exp);
3076 /* Flush any remaining cancel messages out to the target */
3077 filter_sync_llogs(obd, exp);
3079 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3081 rc = server_disconnect_export(exp);
3083 if (exp->exp_obd->obd_replayable)
3084 filter_client_free(exp);
3086 fsfilt_sync(obd, obd->u.obt.obt_sb);
3088 class_export_put(exp);
3092 /* reverse import is changed, sync all cancels */
3093 static void filter_revimp_update(struct obd_export *exp)
3098 class_export_get(exp);
3100 /* flush any remaining cancel messages out to the target */
3101 filter_sync_llogs(exp->exp_obd, exp);
3102 class_export_put(exp);
3106 static int filter_ping(struct obd_export *exp)
3108 filter_fmd_expire(exp);
3112 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3113 const char *what, int quiet)
3115 struct dentry *dchild = NULL;
3118 if (oa->o_valid & OBD_MD_FLGROUP)
3121 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3123 if (IS_ERR(dchild)) {
3124 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3125 what, group, oa->o_id);
3129 if (dchild->d_inode == NULL) {
3131 CERROR("%s: %s on non-existent object: "LPU64"\n",
3132 obd->obd_name, what, oa->o_id);
3134 RETURN(ERR_PTR(-ENOENT));
3140 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3142 struct dentry *dentry = NULL;
3143 struct obd_device *obd;
3147 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3148 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3149 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3153 obd = class_exp2obd(exp);
3155 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3159 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3161 RETURN(PTR_ERR(dentry));
3163 /* Limit the valid bits in the return data to what we actually use */
3164 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3165 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3171 /* this should be enabled/disabled in condition to enabled/disabled large
3172 * inodes (fast EAs) in backing store FS. */
3173 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3174 void *handle, struct obdo *oa)
3176 struct obd_device *obd = exp->exp_obd;
3180 if (oa->o_valid & OBD_MD_FLFID) {
3181 struct filter_fid ff;
3183 if (!(oa->o_valid & OBD_MD_FLGROUP))
3185 /* packing fid and converting it to LE for storing into EA.
3186 * Here ->o_stripe_idx should be filled by LOV and rest of
3187 * fields - by client. */
3188 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3189 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3190 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3191 ff.ff_objid = cpu_to_le64(oa->o_id);
3192 ff.ff_group = cpu_to_le64(oa->o_gr);
3194 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3195 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3196 oa->o_generation, oa->o_id, oa->o_gr);
3198 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3200 CERROR("store fid in object failed! rc: %d\n", rc);
3202 CDEBUG(D_HA, "OSS object without fid info!\n");
3208 /* this is called from filter_truncate() until we have filter_punch() */
3209 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3210 struct obdo *oa, struct obd_trans_info *oti)
3212 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3213 struct llog_cookie *fcc = NULL;
3214 struct filter_obd *filter;
3215 int rc, err, sync = 0;
3216 loff_t old_size = 0;
3217 unsigned int ia_valid;
3218 struct inode *inode;
3219 struct page *page = NULL;
3224 LASSERT(dentry != NULL);
3225 LASSERT(!IS_ERR(dentry));
3227 inode = dentry->d_inode;
3228 LASSERT(inode != NULL);
3230 filter = &exp->exp_obd->u.filter;
3231 iattr_from_obdo(&iattr, oa, oa->o_valid);
3232 ia_valid = iattr.ia_valid;
3234 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3235 OBD_ALLOC(fcc, sizeof(*fcc));
3237 *fcc = oa->o_lcookie;
3239 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3241 /* Filter truncates and writes are serialized by
3242 * i_alloc_sem, see the comment in
3243 * filter_preprw_write.*/
3244 if (ia_valid & ATTR_SIZE)
3245 down_write(&inode->i_alloc_sem);
3246 LOCK_INODE_MUTEX(inode);
3247 old_size = i_size_read(inode);
3250 /* VBR: version recovery check */
3251 rc = filter_version_get_check(exp, oti, inode);
3253 GOTO(out_unlock, rc);
3255 /* Let's pin the last page so that ldiskfs_truncate
3256 * should not start GFP_FS allocation. */
3257 if (ia_valid & ATTR_SIZE) {
3258 page = grab_cache_page(inode->i_mapping,
3259 iattr.ia_size >> PAGE_CACHE_SHIFT);
3261 GOTO(out_unlock, rc = -ENOMEM);
3266 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3267 * then we will accept the UID+GID sent by the client during write for
3268 * initializing the ownership of this inode. We only allow this to
3269 * happen once so clear these bits in setattr. In 2.6 kernels it is
3270 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3271 * the flags that are actually being set. */
3272 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3273 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3274 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3276 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3277 if (!(ia_valid & ATTR_MODE)) {
3278 iattr.ia_mode = inode->i_mode;
3279 iattr.ia_valid |= ATTR_MODE;
3281 iattr.ia_mode &= ~S_ISUID;
3283 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3284 if (!(iattr.ia_valid & ATTR_MODE)) {
3285 iattr.ia_mode = inode->i_mode;
3286 iattr.ia_valid |= ATTR_MODE;
3288 iattr.ia_mode &= ~S_ISGID;
3291 orig_ids[USRQUOTA] = inode->i_uid;
3292 orig_ids[GRPQUOTA] = inode->i_gid;
3293 handle = fsfilt_start_log(exp->exp_obd, inode,
3294 FSFILT_OP_SETATTR, oti, 1);
3296 GOTO(out_unlock, rc = PTR_ERR(handle));
3298 /* update inode EA only once when inode is suid bit marked. As
3299 * on 2.6.x UID and GID may be set separately, we check here
3300 * only one of them to avoid double setting. */
3301 if (inode->i_mode & S_ISUID)
3302 filter_update_fidea(exp, inode, handle, oa);
3304 handle = fsfilt_start(exp->exp_obd, inode,
3305 FSFILT_OP_SETATTR, oti);
3307 GOTO(out_unlock, rc = PTR_ERR(handle));
3309 if (oa->o_valid & OBD_MD_FLFLAGS) {
3310 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3311 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3313 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3315 /* set cancel cookie callback function */
3316 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3317 filter_cancel_cookies_cb,
3321 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3322 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3324 /* The truncate might have used up our transaction credits. Make
3325 * sure we have one left for the last_rcvd update. */
3326 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3328 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3330 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3334 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3336 CERROR("error on commit, err = %d\n", err);
3343 /* For a partial-page truncate flush the page to disk immediately
3344 * to avoid data corruption during direct disk write. b=17397 */
3345 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3346 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3347 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3357 page_cache_release(page);
3359 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3360 UNLOCK_INODE_MUTEX(inode);
3361 if (ia_valid & ATTR_SIZE)
3362 up_write(&inode->i_alloc_sem);
3364 OBD_FREE(fcc, sizeof(*fcc));
3366 /* trigger quota release */
3367 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3368 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3369 int rc2 = lquota_adjust(filter_quota_interface_ref,
3370 exp->exp_obd, cur_ids,
3371 orig_ids, rc, FSFILT_OP_SETATTR);
3372 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3373 "filter adjust qunit. (rc:%d)\n", rc2);
3378 /* this is called from filter_truncate() until we have filter_punch() */
3379 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3380 struct obd_trans_info *oti)
3382 struct obdo *oa = oinfo->oi_oa;
3383 struct lustre_capa *capa = oinfo_capa(oinfo);
3384 struct ldlm_res_id res_id;
3385 struct filter_mod_data *fmd;
3386 struct lvfs_run_ctxt saved;
3387 struct filter_obd *filter;
3388 struct ldlm_resource *res;
3389 struct dentry *dentry;
3390 __u64 opc = CAPA_OPC_META_WRITE;
3394 if (oa->o_valid & OBD_FL_TRUNC)
3395 opc |= CAPA_OPC_OSS_TRUNC;
3397 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3398 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3402 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3403 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3408 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3409 /* This would be very bad - accidentally truncating a file when
3410 * changing the time or similar - bug 12203. */
3411 if (oa->o_valid & OBD_MD_FLSIZE &&
3412 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3413 static char mdsinum[48];
3415 if (oa->o_valid & OBD_MD_FLFID)
3416 snprintf(mdsinum, sizeof(mdsinum) - 1,
3417 " of inode "LPU64"/%u", oa->o_fid,
3422 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3424 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3429 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3431 RETURN(PTR_ERR(dentry));
3433 filter = &exp->exp_obd->u.filter;
3434 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3437 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3438 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3439 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3440 fmd->fmd_mactime_xid = oti->oti_xid;
3441 filter_fmd_put(exp, fmd);
3444 /* setting objects attributes (including owner/group) */
3445 rc = filter_setattr_internal(exp, dentry, oa, oti);
3447 GOTO(out_unlock, rc);
3449 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3450 &res_id, LDLM_EXTENT, 0);
3453 LDLM_RESOURCE_ADDREF(res);
3454 rc = ldlm_res_lvbo_update(res, NULL, 0);
3455 LDLM_RESOURCE_DELREF(res);
3456 ldlm_resource_putref(res);
3459 oa->o_valid = OBD_MD_FLID;
3461 /* Quota release need uid/gid info */
3462 obdo_from_inode(oa, dentry->d_inode,
3463 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3468 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3472 /* XXX identical to osc_unpackmd */
3473 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3474 struct lov_mds_md *lmm, int lmm_bytes)
3480 if (lmm_bytes < sizeof (*lmm)) {
3481 CERROR("lov_mds_md too small: %d, need %d\n",
3482 lmm_bytes, (int)sizeof(*lmm));
3485 /* XXX LOV_MAGIC etc check? */
3487 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3488 CERROR("lov_mds_md: zero lmm_object_id\n");
3493 lsm_size = lov_stripe_md_size(1);
3497 if (*lsmp != NULL && lmm == NULL) {
3498 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3499 OBD_FREE(*lsmp, lsm_size);
3504 if (*lsmp == NULL) {
3505 OBD_ALLOC(*lsmp, lsm_size);
3509 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3510 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3511 OBD_FREE(*lsmp, lsm_size);
3514 loi_init((*lsmp)->lsm_oinfo[0]);
3518 /* XXX zero *lsmp? */
3519 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3520 LASSERT((*lsmp)->lsm_object_id);
3523 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3528 /* caller must hold fo_create_locks[oa->o_gr] */
3529 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3530 struct filter_obd *filter)
3532 struct obdo doa = { 0 }; /* XXX obdo on stack */
3539 LASSERT_MDS_GROUP(oa->o_gr);
3540 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3541 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3543 memset(&doa, 0, sizeof(doa));
3545 doa.o_valid |= OBD_MD_FLGROUP;
3546 doa.o_gr = oa->o_gr;
3547 doa.o_mode = S_IFREG;
3549 if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3550 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3551 exp->exp_obd->obd_name, doa.o_gr);
3555 last = filter_last_id(filter, doa.o_gr);
3557 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3559 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3560 exp->exp_obd->obd_name, oa->o_id + 1, last,
3561 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3563 for (id = last; id > oa->o_id; id--) {
3565 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3566 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3567 CEMERG("error destroying precreate objid "LPU64": %d\n",
3570 /* update last_id on disk periodically so that if we restart
3571 * we don't need to re-scan all of the just-deleted objects. */
3572 if ((id & 511) == 0 && !skip_orphan) {
3573 filter_set_last_id(filter, id - 1, doa.o_gr);
3574 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3578 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3579 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3582 filter_set_last_id(filter, id, doa.o_gr);
3583 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3585 /* don't reuse orphan object, return last used objid */
3589 cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3594 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3595 obd_gr group, int *num);
3596 /* returns a negative error or a nonnegative number of files to create */
3597 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3598 obd_gr group, struct obd_trans_info *oti)
3600 struct obd_device *obd = exp->exp_obd;
3601 struct filter_obd *filter = &obd->u.filter;
3605 /* delete orphans request */
3606 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3607 obd_id last = filter_last_id(filter, group);
3609 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3610 CERROR("%s: dropping old orphan cleanup request\n",
3614 /* This causes inflight precreates to abort and drop lock */
3615 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3616 cfs_down(&filter->fo_create_locks[group]);
3617 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3618 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3619 exp->exp_obd->obd_name, group);
3620 cfs_up(&filter->fo_create_locks[group]);
3623 diff = oa->o_id - last;
3624 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3627 if (-diff > OST_MAX_PRECREATE) {
3628 CERROR("%s: ignoring bogus orphan destroy request: "
3629 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3631 /* FIXME: should reset precreate_next_id on MDS */
3632 GOTO(out, rc = -EINVAL);
3635 rc = filter_destroy_precreated(exp, oa, filter);
3637 CERROR("%s: unable to write lastobjid, but "
3638 "orphans were deleted\n", obd->obd_name);
3641 /* XXX: Used by MDS for the first time! */
3642 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3645 cfs_down(&filter->fo_create_locks[group]);
3646 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3647 CERROR("%s: dropping old precreate request\n",
3651 /* only precreate if group == 0 and o_id is specfied */
3652 if (!filter_group_is_mds(group) || oa->o_id == 0)
3655 diff = oa->o_id - filter_last_id(filter, group);
3656 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3657 filter_last_id(filter, group), diff);
3659 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3660 oa->o_id, filter_last_id(filter, group), diff);
3664 oa->o_id = filter_last_id(&obd->u.filter, group);
3665 rc = filter_precreate(obd, oa, group, &diff);
3666 oa->o_id = filter_last_id(&obd->u.filter, group);
3668 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3671 /* else diff == 0 */
3674 cfs_up(&filter->fo_create_locks[group]);
3678 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3679 __u64 max_age, __u32 flags)
3681 struct filter_obd *filter = &obd->u.filter;
3682 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3686 /* at least try to account for cached pages. its still racey and
3687 * might be under-reporting if clients haven't announced their
3688 * caches with brw recently */
3689 cfs_spin_lock(&obd->obd_osfs_lock);
3690 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3691 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3692 cfs_spin_unlock(&obd->obd_osfs_lock);
3694 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3695 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3696 filter->fo_tot_dirty, filter->fo_tot_granted,
3697 filter->fo_tot_pending,
3698 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3700 filter_grant_sanity_check(obd, __FUNCTION__);
3702 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3703 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3704 osfs->os_bsize - 1) >> blockbits));
3706 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3707 struct lr_server_data *lsd = filter->fo_fsd;
3708 int index = le32_to_cpu(lsd->lsd_ost_index);
3710 if (obd_fail_val == -1 ||
3711 index == obd_fail_val)
3712 osfs->os_bfree = osfs->os_bavail = 2;
3713 else if (obd_fail_loc & OBD_FAIL_ONCE)
3714 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3717 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3718 * stop creating files on MDS if OST is not good shape to create
3722 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3723 osfs->os_state = OS_STATE_READONLY;
3725 if (filter->fo_raid_degraded)
3726 osfs->os_state |= OS_STATE_DEGRADED;
3730 static int filter_use_existing_obj(struct obd_device *obd,
3731 struct dentry *dchild, void **handle,
3734 struct inode *inode = dchild->d_inode;
3738 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3741 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3742 if (IS_ERR(*handle))
3743 return PTR_ERR(*handle);
3745 iattr.ia_valid = ATTR_MODE;
3746 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3747 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3754 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3757 __u64 os_ffree = -1;
3759 cfs_spin_lock(&obd->obd_osfs_lock);
3760 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3762 os_ffree = obd->obd_osfs.os_ffree;
3763 cfs_spin_unlock(&obd->obd_osfs_lock);
3768 /* We rely on the fact that only one thread will be creating files in a given
3769 * group at a time, which is why we don't need an atomic filter_get_new_id.
3770 * Even if we had that atomic function, the following race would exist:
3772 * thread 1: gets id x from filter_next_id
3773 * thread 2: gets id (x + 1) from filter_next_id
3774 * thread 2: creates object (x + 1)
3775 * thread 1: tries to create object x, gets -ENOSPC
3777 * Caller must hold fo_create_locks[group]
3779 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3780 obd_gr group, int *num)
3782 struct dentry *dchild = NULL, *dparent = NULL;
3783 struct filter_obd *filter;
3784 struct obd_statfs *osfs;
3785 int err = 0, rc = 0, recreate_obj = 0, i;
3786 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3789 void *handle = NULL;
3792 filter = &obd->u.filter;
3794 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3796 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3798 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3799 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3802 OBD_ALLOC(osfs, sizeof(*osfs));
3805 rc = filter_statfs(obd, osfs, cfs_time_current_64() - CFS_HZ,
3807 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3808 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3809 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3810 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3814 OBD_FREE(osfs, sizeof(*osfs));
3819 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3820 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3822 for (i = 0; i < *num && err == 0; i++) {
3823 int cleanup_phase = 0;
3825 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3826 CWARN("%s: create aborted by destroy\n",
3835 last_id = filter_last_id(filter, group);
3836 if (next_id > last_id) {
3837 CERROR("Error: Trying to recreate obj greater"
3838 "than last id "LPD64" > "LPD64"\n",
3840 GOTO(cleanup, rc = -EINVAL);
3843 next_id = filter_last_id(filter, group) + 1;
3845 dparent = filter_parent_lock(obd, group, next_id);
3846 if (IS_ERR(dparent))
3847 GOTO(cleanup, rc = PTR_ERR(dparent));
3848 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3850 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3852 GOTO(cleanup, rc = PTR_ERR(dchild));
3853 cleanup_phase = 2; /* f_dput(dchild) */
3855 if (dchild->d_inode != NULL) {
3856 /* This would only happen if lastobjid was bad on disk*/
3857 /* Could also happen if recreating missing obj but it
3858 * already exists. */
3860 CERROR("%s: recreating existing object %.*s?\n",
3861 obd->obd_name, dchild->d_name.len,
3862 dchild->d_name.name);
3864 /* Use these existing objects if they are
3866 if (dchild->d_inode->i_size == 0) {
3867 rc = filter_use_existing_obj(obd,dchild,
3868 &handle, &cleanup_phase);
3875 CERROR("%s: Serious error: objid %.*s already "
3876 "exists; is this filesystem corrupt?\n",
3877 obd->obd_name, dchild->d_name.len,
3878 dchild->d_name.name);
3881 GOTO(cleanup, rc = -EEXIST);
3884 handle = fsfilt_start_log(obd, dparent->d_inode,
3885 FSFILT_OP_CREATE, NULL, 1);
3887 GOTO(cleanup, rc = PTR_ERR(handle));
3890 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3891 ",od->o_id="LPU64")\n", obd->obd_name, group,
3894 /* We mark object SUID+SGID to flag it for accepting UID+GID
3895 * from client on first write. Currently the permission bits
3896 * on the OST are never used, so this is OK. */
3897 rc = ll_vfs_create(dparent->d_inode, dchild,
3898 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3900 CERROR("create failed rc = %d\n", rc);
3901 if (rc == -ENOSPC) {
3902 os_ffree = filter_calc_free_inodes(obd);
3904 CERROR("%s: free inode "LPU64"\n",
3905 obd->obd_name, os_ffree);
3910 if (dchild->d_inode)
3911 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3912 dchild->d_inode->i_ino);
3915 if (!recreate_obj) {
3916 filter_set_last_id(filter, next_id, group);
3917 err = filter_update_last_objid(obd, group, 0);
3919 CERROR("unable to write lastobjid "
3920 "but file created\n");
3924 switch(cleanup_phase) {
3926 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3928 CERROR("error on commit, err = %d\n", err);
3935 filter_parent_unlock(dparent);
3942 if (cfs_time_after(jiffies, enough_time)) {
3944 "%s: precreate slow - want %d got %d \n",
3945 obd->obd_name, *num, i);
3952 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3953 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3958 static int filter_create(struct obd_export *exp, struct obdo *oa,
3959 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3961 struct obd_device *obd = exp->exp_obd;
3962 struct filter_export_data *fed;
3963 struct filter_obd *filter;
3964 struct lvfs_run_ctxt saved;
3965 struct lov_stripe_md *lsm = NULL;
3966 int rc = 0, diff, group = oa->o_gr;
3969 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3970 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3972 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3973 CERROR("!!! nid %s sent invalid object group %d\n",
3974 obd_export_nid2str(exp), group);
3978 fed = &exp->exp_filter_data;
3979 filter = &obd->u.filter;
3981 if (fed->fed_group != group) {
3982 CERROR("!!! this export (nid %s) used object group %d "
3983 "earlier; now it's trying to use group %d! This could "
3984 "be a bug in the MDS. Please report to "
3985 "http://bugzilla.lustre.org/\n",
3986 obd_export_nid2str(exp), fed->fed_group, group);
3993 rc = obd_alloc_memmd(exp, &lsm);
4000 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4002 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4003 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4004 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4005 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4006 oa->o_id, filter_last_id(filter,
4011 cfs_down(&filter->fo_create_locks[oa->o_gr]);
4012 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4013 cfs_up(&filter->fo_create_locks[oa->o_gr]);
4016 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4019 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4020 if (rc && ea != NULL && *ea != lsm) {
4021 obd_free_memmd(exp, &lsm);
4022 } else if (rc == 0 && ea != NULL) {
4023 /* XXX LOV STACKING: the lsm that is passed to us from
4024 * LOV does not have valid lsm_oinfo data structs, so
4025 * don't go touching that. This needs to be fixed in a
4027 lsm->lsm_object_id = oa->o_id;
4034 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4035 struct lov_stripe_md *md, struct obd_trans_info *oti,
4036 struct obd_export *md_exp, void *capa)
4038 unsigned int qcids[MAXQUOTAS] = {0, 0};
4039 struct obd_device *obd;
4040 struct filter_obd *filter;
4041 struct dentry *dchild = NULL, *dparent = NULL;
4042 struct lustre_handle lockh = { 0 };
4043 struct lvfs_run_ctxt saved;
4044 void *handle = NULL;
4045 struct llog_cookie *fcc = NULL;
4046 int rc, rc2, cleanup_phase = 0, sync = 0;
4050 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4051 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4052 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4057 filter = &obd->u.filter;
4059 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4062 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4063 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4065 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4067 GOTO(cleanup, rc = PTR_ERR(dchild));
4070 if (dchild->d_inode == NULL) {
4071 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4073 /* If object already gone, cancel cookie right now */
4074 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4075 struct llog_ctxt *ctxt;
4076 struct obd_llog_group *olg;
4078 olg = filter_find_olg(obd, oa->o_gr);
4080 CERROR(" %s: can not find olg of group %d\n",
4081 obd->obd_name, (int)oa->o_gr);
4082 GOTO(cleanup, rc = PTR_ERR(olg));
4084 fcc = &oa->o_lcookie;
4085 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4086 llog_cancel(ctxt, NULL, 1, fcc, 0);
4087 llog_ctxt_put(ctxt);
4088 fcc = NULL; /* we didn't allocate fcc, don't free it */
4090 GOTO(cleanup, rc = -ENOENT);
4093 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4097 /* Our MDC connection is established by the MDS to us */
4098 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4099 OBD_ALLOC(fcc, sizeof(*fcc));
4101 *fcc = oa->o_lcookie;
4103 DQUOT_INIT(dchild->d_inode);
4105 /* we're gonna truncate it first in order to avoid possible deadlock:
4107 * open trasaction open transaction
4108 * down(i_zombie) down(i_zombie)
4109 * restart transaction
4110 * (see BUG 4180) -bzzz
4112 * take i_alloc_sem too to prevent other threads from writing to the
4113 * file while we are truncating it. This can cause lock ordering issue
4114 * between page lock, i_mutex & starting new journal handle.
4115 * (see bug 20321) -johann
4117 down_write(&dchild->d_inode->i_alloc_sem);
4118 LOCK_INODE_MUTEX(dchild->d_inode);
4120 /* VBR: version recovery check */
4121 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4123 UNLOCK_INODE_MUTEX(dchild->d_inode);
4124 up_write(&dchild->d_inode->i_alloc_sem);
4128 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4130 if (IS_ERR(handle)) {
4131 UNLOCK_INODE_MUTEX(dchild->d_inode);
4132 up_write(&dchild->d_inode->i_alloc_sem);
4133 GOTO(cleanup, rc = PTR_ERR(handle));
4136 iattr.ia_valid = ATTR_SIZE;
4138 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4139 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4140 UNLOCK_INODE_MUTEX(dchild->d_inode);
4141 up_write(&dchild->d_inode->i_alloc_sem);
4145 GOTO(cleanup, rc = rc2);
4147 /* We don't actually need to lock the parent until we are unlinking
4148 * here, and not while truncating above. That avoids holding the
4149 * parent lock for a long time during truncate, which can block other
4150 * threads from doing anything to objects in that directory. bug 7171 */
4151 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4152 if (IS_ERR(dparent))
4153 GOTO(cleanup, rc = PTR_ERR(dparent));
4154 cleanup_phase = 3; /* filter_parent_unlock */
4156 LOCK_INODE_MUTEX(dchild->d_inode);
4157 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4158 if (IS_ERR(handle)) {
4159 UNLOCK_INODE_MUTEX(dchild->d_inode);
4160 GOTO(cleanup, rc = PTR_ERR(handle));
4162 cleanup_phase = 4; /* fsfilt_commit */
4164 /* Quota release need uid/gid of inode */
4165 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4167 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4169 /* this drops dchild->d_inode->i_mutex unconditionally */
4170 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4174 switch(cleanup_phase) {
4177 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4178 oti->oti_handle : handle,
4179 filter_cancel_cookies_cb,
4181 /* If add_journal_cb failed, then filter_finish_transno
4182 * will commit the handle and we will do a sync
4183 * on commit. then we call callback directly to free
4186 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4188 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4191 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4193 CERROR("error on commit, err = %d\n", rc2);
4200 filter_parent_unlock(dparent);
4202 filter_fini_destroy(obd, &lockh);
4206 OBD_FREE(fcc, sizeof(*fcc));
4208 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4211 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4215 /* trigger quota release */
4216 qcids[USRQUOTA] = oa->o_uid;
4217 qcids[GRPQUOTA] = oa->o_gid;
4218 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4221 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4225 /* NB start and end are used for punch, but not truncate */
4226 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4227 struct obd_trans_info *oti,
4228 struct ptlrpc_request_set *rqset)
4233 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4234 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4235 oinfo->oi_policy.l_extent.end);
4239 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4240 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4241 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4243 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4244 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4245 rc = filter_setattr(exp, oinfo, oti);
4246 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4250 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4251 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4254 struct lvfs_run_ctxt saved;
4255 struct filter_obd *filter;
4256 struct dentry *dentry;
4260 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4261 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4262 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4266 filter = &exp->exp_obd->u.filter;
4268 /* An objid of zero is taken to mean "sync whole filesystem" */
4269 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4270 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4271 /* Flush any remaining cancel messages out to the target */
4272 filter_sync_llogs(exp->exp_obd, exp);
4276 dentry = filter_oa2dentry(exp->exp_obd, oa);
4278 RETURN(PTR_ERR(dentry));
4280 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4282 LOCK_INODE_MUTEX(dentry->d_inode);
4284 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4286 /* just any file to grab fsync method - "file" arg unused */
4287 struct file *file = filter->fo_rcvd_filp;
4289 if (file->f_op && file->f_op->fsync)
4290 rc = file->f_op->fsync(NULL, dentry, 1);
4292 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4296 UNLOCK_INODE_MUTEX(dentry->d_inode);
4298 oa->o_valid = OBD_MD_FLID;
4299 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4301 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4307 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4308 void *key, __u32 *vallen, void *val,
4309 struct lov_stripe_md *lsm)
4311 struct obd_device *obd;
4314 obd = class_exp2obd(exp);
4316 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4320 if (KEY_IS(KEY_BLOCKSIZE)) {
4321 __u32 *blocksize = val;
4323 if (*vallen < sizeof(*blocksize))
4325 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4327 *vallen = sizeof(*blocksize);
4331 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4332 __u32 *blocksize_bits = val;
4333 if (blocksize_bits) {
4334 if (*vallen < sizeof(*blocksize_bits))
4336 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4338 *vallen = sizeof(*blocksize_bits);
4342 if (KEY_IS(KEY_LAST_ID)) {
4343 obd_id *last_id = val;
4344 /* FIXME: object groups */
4346 if (*vallen < sizeof(*last_id))
4348 *last_id = filter_last_id(&obd->u.filter,
4349 exp->exp_filter_data.fed_group);
4351 *vallen = sizeof(*last_id);
4355 if (KEY_IS(KEY_FIEMAP)) {
4356 struct ll_fiemap_info_key *fm_key = key;
4357 struct dentry *dentry;
4358 struct ll_user_fiemap *fiemap = val;
4359 struct lvfs_run_ctxt saved;
4362 if (fiemap == NULL) {
4363 *vallen = fiemap_count_to_size(
4364 fm_key->fiemap.fm_extent_count);
4368 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4371 RETURN(PTR_ERR(dentry));
4373 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4374 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4375 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4377 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
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_MDS_GROUP(group);
4444 rc = filter_setup_llog_group(exp, obd, group);
4448 if (group == FILTER_GROUP_MDS0) {
4449 /* setup llog group 1 for interop */
4450 filter_setup_llog_group(exp, obd, FILTER_GROUP_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 < 32; i++)
4661 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(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);