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 lr_server_data *fsd = class_server_data(exp->exp_obd);
126 struct lsd_client_data *lcd;
129 int err, log_pri = D_RPCTRACE;
131 /* Propagate error code. */
135 if (!exp->exp_obd->obd_replayable || oti == NULL)
138 cfs_mutex_down(&fed->fed_lcd_lock);
140 /* if the export has already been disconnected, we have no last_rcvd slot,
141 * update server data with latest transno then */
143 cfs_mutex_up(&fed->fed_lcd_lock);
144 CWARN("commit transaction for disconnected client %s: rc %d\n",
145 exp->exp_client_uuid.uuid, rc);
146 err = filter_update_server_data(exp->exp_obd);
150 /* we don't allocate new transnos for replayed requests */
151 cfs_spin_lock(&filter->fo_translock);
152 if (oti->oti_transno == 0) {
153 last_rcvd = le64_to_cpu(fsd->lsd_last_transno) + 1;
154 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156 last_rcvd = oti->oti_transno;
157 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
158 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160 oti->oti_transno = last_rcvd;
162 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166 cfs_spin_unlock(&filter->fo_translock);
169 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
171 off = fed->fed_lr_off;
173 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174 fed->fed_lr_idx, fed->fed_lr_off);
177 class_export_cb_get(exp); /* released when the cb is called */
179 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
185 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
186 lcd, sizeof(*lcd), &off,
187 force_sync | exp->exp_need_sync);
189 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
197 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
199 cfs_mutex_up(&fed->fed_lcd_lock);
203 void f_dput(struct dentry *dentry)
205 /* Can't go inside filter_ddelete because it can block */
206 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208 LASSERT(atomic_read(&dentry->d_count) > 0);
213 static void init_brw_stats(struct brw_stats *brw_stats)
216 for (i = 0; i < BRW_LAST; i++)
217 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221 struct lprocfs_stats **stats)
225 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226 LPROC_FILTER_LAST - 1;
227 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
231 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241 plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243 struct obd_export *exp,
249 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250 /* Self-export gets no proc entry */
253 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
255 /* Mask error for already created
263 struct nid_stat *tmp = exp->exp_nid_stats;
264 LASSERT(tmp != NULL);
266 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267 if (tmp->nid_brw_stats == NULL)
268 GOTO(clean, rc = -ENOMEM);
270 init_brw_stats(tmp->nid_brw_stats);
271 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272 0644, &filter_per_nid_stats_fops,
275 CWARN("Error adding the brw_stats file\n");
277 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
285 /* Always add in ldlm_stats */
286 tmp->nid_ldlm_stats =
287 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
288 LPROCFS_STATS_FLAG_NOPERCPU);
289 if (tmp->nid_ldlm_stats == NULL)
290 GOTO(clean, rc = -ENOMEM);
292 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
293 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
294 tmp->nid_ldlm_stats);
301 lprocfs_exp_cleanup(exp);
305 /* Add client data to the FILTER. We use a bitmap to locate a free space
306 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
307 * Otherwise, we have just read the data from the last_rcvd file and
308 * we know its offset. */
309 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
312 struct filter_obd *filter = &obd->u.filter;
313 struct filter_export_data *fed = &exp->exp_filter_data;
314 struct lr_server_data *fsd = class_server_data(obd);
315 unsigned long *bitmap = filter->fo_last_rcvd_slots;
316 int new_client = (cl_idx == -1);
320 LASSERT(bitmap != NULL);
321 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
324 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
327 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
328 * there's no need for extra complication here
331 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
333 if (cl_idx >= LR_MAX_CLIENTS) {
334 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
338 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
339 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
344 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
345 CERROR("FILTER client %d: bit already set in bitmap!\n",
351 fed->fed_lr_idx = cl_idx;
352 fed->fed_lr_off = le32_to_cpu(fsd->lsd_client_start) +
353 cl_idx * le16_to_cpu(fsd->lsd_client_size);
354 cfs_init_mutex(&fed->fed_lcd_lock);
355 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
357 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
358 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
361 struct lvfs_run_ctxt saved;
362 loff_t off = fed->fed_lr_off;
366 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
367 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
369 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
372 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
373 /* Transaction needed to fix bug 1403 */
374 handle = fsfilt_start(obd,
375 filter->fo_rcvd_filp->f_dentry->d_inode,
376 FSFILT_OP_SETATTR, NULL);
377 if (IS_ERR(handle)) {
378 rc = PTR_ERR(handle);
379 CERROR("unable to start transaction: rc %d\n", rc);
381 fed->fed_lcd->lcd_last_epoch = fsd->lsd_start_epoch;
382 exp->exp_last_request_time = cfs_time_current_sec();
383 rc = fsfilt_add_journal_cb(obd, 0, handle,
384 target_client_add_cb,
385 class_export_cb_get(exp));
387 cfs_spin_lock(&exp->exp_lock);
388 exp->exp_need_sync = 1;
389 cfs_spin_unlock(&exp->exp_lock);
391 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
393 sizeof(*fed->fed_lcd),
394 &off, rc /* sync if no cb */);
396 filter->fo_rcvd_filp->f_dentry->d_inode,
399 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
402 CERROR("error writing %s client idx %u: rc %d\n",
403 LAST_RCVD, fed->fed_lr_idx, rc);
410 static int filter_client_del(struct obd_export *exp)
412 struct filter_export_data *fed = &exp->exp_filter_data;
413 struct filter_obd *filter = &exp->exp_obd->u.filter;
414 struct obd_device *obd = exp->exp_obd;
415 struct lvfs_run_ctxt saved;
416 struct lsd_client_data *lcd = fed->fed_lcd;
424 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
425 if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
428 LASSERT(filter->fo_last_rcvd_slots != NULL);
430 off = fed->fed_lr_off;
432 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
433 fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
435 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
436 * we leak a client slot that will be cleaned on the next recovery. */
438 CERROR("%s: client idx %d has med_off %lld\n",
439 obd->obd_name, fed->fed_lr_idx, off);
440 GOTO(free, rc = -EINVAL);
443 /* Clear the bit _after_ zeroing out the client so we don't
444 race with filter_client_add and zero out new clients.*/
445 if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
446 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
451 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
452 /* Make sure the server's last_transno is up to date.
453 * This should be done before zeroing client slot so last_transno will
454 * be in server data or in client data in case of failure */
455 filter_update_server_data(obd);
457 cfs_mutex_down(&fed->fed_lcd_lock);
458 memset(fed->fed_lcd->lcd_uuid, 0, sizeof fed->fed_lcd->lcd_uuid);
459 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
461 sizeof(*fed->fed_lcd), &off, 0);
462 cfs_mutex_up(&fed->fed_lcd_lock);
463 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
465 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
466 "zero out client %s at idx %u/%llu in %s, rc %d\n",
467 lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
474 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
475 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
476 struct filter_mod_data *fmd)
478 LASSERT_SPIN_LOCKED(&fed->fed_lock);
479 if (--fmd->fmd_refcount == 0) {
480 /* XXX when we have persistent reservations and the handle
481 * is stored herein we need to drop it here. */
482 fed->fed_mod_count--;
483 cfs_list_del(&fmd->fmd_list);
484 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
488 /* drop fmd reference, free it if last ref */
489 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
491 struct filter_export_data *fed;
496 fed = &exp->exp_filter_data;
497 cfs_spin_lock(&fed->fed_lock);
498 filter_fmd_put_nolock(fed, fmd); /* caller reference */
499 cfs_spin_unlock(&fed->fed_lock);
502 /* expire entries from the end of the list if there are too many
503 * or they are too old */
504 static void filter_fmd_expire_nolock(struct filter_obd *filter,
505 struct filter_export_data *fed,
506 struct filter_mod_data *keep)
508 struct filter_mod_data *fmd, *tmp;
510 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
514 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
515 fed->fed_mod_count < filter->fo_fmd_max_num)
518 cfs_list_del_init(&fmd->fmd_list);
519 filter_fmd_put_nolock(fed, fmd); /* list reference */
523 void filter_fmd_expire(struct obd_export *exp)
525 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
526 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
527 &exp->exp_filter_data, NULL);
528 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
531 /* find specified objid, group in export fmd list.
532 * caller must hold fed_lock and take fmd reference itself */
533 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
534 struct filter_export_data *fed,
535 obd_id objid, obd_gr group)
537 struct filter_mod_data *found = NULL, *fmd;
539 LASSERT_SPIN_LOCKED(&fed->fed_lock);
541 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
542 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
544 cfs_list_del(&fmd->fmd_list);
545 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
546 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
551 filter_fmd_expire_nolock(filter, fed, found);
556 /* Find fmd based on objid and group, or return NULL if not found. */
557 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
558 obd_id objid, obd_gr group)
560 struct filter_mod_data *fmd;
562 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
563 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
564 &exp->exp_filter_data, objid, group);
566 fmd->fmd_refcount++; /* caller reference */
567 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
572 /* Find fmd based on objid and group, or create a new one if none is found.
573 * It is possible for this function to return NULL under memory pressure,
574 * or if objid = 0 is passed (which will only cause old entries to expire).
575 * Currently this is not fatal because any fmd state is transient and
576 * may also be freed when it gets sufficiently old. */
577 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
578 obd_id objid, obd_gr group)
580 struct filter_export_data *fed = &exp->exp_filter_data;
581 struct filter_mod_data *found = NULL, *fmd_new = NULL;
583 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
585 cfs_spin_lock(&fed->fed_lock);
586 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
589 cfs_list_add_tail(&fmd_new->fmd_list,
591 fmd_new->fmd_id = objid;
592 fmd_new->fmd_gr = group;
593 fmd_new->fmd_refcount++; /* list reference */
595 fed->fed_mod_count++;
597 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
601 found->fmd_refcount++; /* caller reference */
602 found->fmd_expire = jiffies +
603 exp->exp_obd->u.filter.fo_fmd_max_age;
606 cfs_spin_unlock(&fed->fed_lock);
612 /* drop fmd list reference so it will disappear when last reference is put.
613 * This isn't so critical because it would in fact only affect the one client
614 * that is doing the unlink and at worst we have an stale entry referencing
615 * an object that should never be used again. */
616 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
618 struct filter_mod_data *found = NULL;
620 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
621 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
623 cfs_list_del_init(&found->fmd_list);
624 filter_fmd_put_nolock(&exp->exp_filter_data, found);
626 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
629 #define filter_fmd_drop(exp, objid, group)
632 /* remove all entries from fmd list */
633 static void filter_fmd_cleanup(struct obd_export *exp)
635 struct filter_export_data *fed = &exp->exp_filter_data;
636 struct filter_mod_data *fmd = NULL, *tmp;
638 cfs_spin_lock(&fed->fed_lock);
639 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
640 cfs_list_del_init(&fmd->fmd_list);
641 filter_fmd_put_nolock(fed, fmd);
643 cfs_spin_unlock(&fed->fed_lock);
646 static int filter_init_export(struct obd_export *exp)
648 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
649 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
651 cfs_spin_lock(&exp->exp_lock);
652 exp->exp_connecting = 1;
653 cfs_spin_unlock(&exp->exp_lock);
655 return ldlm_init_export(exp);
658 static int filter_free_server_data(struct filter_obd *filter)
660 lut_fini(NULL, filter->fo_obt.obt_lut);
661 OBD_FREE_PTR(filter->fo_obt.obt_lut);
665 /* assumes caller is already in kernel ctxt */
666 int filter_update_server_data(struct obd_device *obd)
668 struct file *filp = obd->u.obt.obt_rcvd_filp;
669 struct lr_server_data *fsd = class_server_data(obd);
674 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
675 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
676 le64_to_cpu(fsd->lsd_last_transno));
677 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
678 le64_to_cpu(fsd->lsd_mount_count));
680 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
682 CERROR("error writing lr_server_data: rc = %d\n", rc);
687 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
690 struct filter_obd *filter = &obd->u.filter;
696 if (filter->fo_last_objid_files[group] == NULL) {
697 CERROR("Object group "LPU64" not fully setup; not updating "
698 "last_objid\n", group);
702 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
703 obd->obd_name, group, filter->fo_last_objids[group]);
705 tmp = cpu_to_le64(filter->fo_last_objids[group]);
706 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
707 &tmp, sizeof(tmp), &off, force_sync);
709 CERROR("error writing group "LPU64" last objid: rc = %d\n",
713 extern int ost_handle(struct ptlrpc_request *req);
714 /* assumes caller has already in kernel ctxt */
715 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
717 struct filter_obd *filter = &obd->u.filter;
718 struct lr_server_data *fsd;
719 struct lsd_client_data *lcd = NULL;
720 struct inode *inode = filp->f_dentry->d_inode;
721 unsigned long last_rcvd_size = i_size_read(inode);
722 struct lu_target *lut;
729 /* ensure padding in the struct is the correct size */
730 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
731 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
732 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
733 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
735 /* allocate and initialize lu_target */
739 rc = lut_init(NULL, lut, obd, NULL);
742 fsd = class_server_data(obd);
743 if (last_rcvd_size == 0) {
744 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
746 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
747 fsd->lsd_last_transno = 0;
748 mount_count = fsd->lsd_mount_count = 0;
749 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
750 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
751 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
752 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
753 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
754 /* OBD_COMPAT_OST is set in filter_connect_internal when the
755 * MDS first connects and assigns the OST index number. */
756 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
759 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
761 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
765 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
766 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
767 "using the wrong disk %s. Were the "
768 "/dev/ assignments rearranged?\n",
769 obd->obd_uuid.uuid, fsd->lsd_uuid);
770 GOTO(err_fsd, rc = -EINVAL);
772 mount_count = le64_to_cpu(fsd->lsd_mount_count);
773 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
775 /* Assume old last_rcvd format unless I_C_LR is set */
776 if (!(fsd->lsd_feature_incompat &
777 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
778 fsd->lsd_last_transno = fsd->lsd_compat14;
780 /* OBD_COMPAT_OST is set in filter_connect_internal when the
781 * MDS first connects and assigns the OST index number. */
782 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
786 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
787 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
788 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
789 ~FILTER_INCOMPAT_SUPP);
790 GOTO(err_fsd, rc = -EINVAL);
792 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
793 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
794 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
795 ~FILTER_ROCOMPAT_SUPP);
796 /* Do something like remount filesystem read-only */
797 GOTO(err_fsd, rc = -EINVAL);
800 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
802 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
803 obd->obd_name, start_epoch);
804 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
805 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
806 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
807 obd->obd_name, mount_count + 1);
808 CDEBUG(D_INODE, "%s: server data size: %u\n",
809 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
810 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
811 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
812 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
813 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
814 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
815 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
816 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
817 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
818 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
819 le16_to_cpu(fsd->lsd_client_size));
821 if (!obd->obd_replayable) {
822 CWARN("%s: recovery support OFF\n", obd->obd_name);
826 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
827 off < last_rcvd_size; cl_idx++) {
829 struct obd_export *exp;
830 struct filter_export_data *fed;
835 GOTO(err_client, rc = -ENOMEM);
838 /* Don't assume off is incremented properly by
839 * fsfilt_read_record(), in case sizeof(*lcd)
840 * isn't the same as fsd->lsd_client_size. */
841 off = le32_to_cpu(fsd->lsd_client_start) +
842 cl_idx * le16_to_cpu(fsd->lsd_client_size);
843 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
845 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
846 LAST_RCVD, cl_idx, off, rc);
847 break; /* read error shouldn't cause startup to fail */
850 if (lcd->lcd_uuid[0] == '\0') {
851 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
856 check_lcd(obd->obd_name, cl_idx, lcd);
858 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
860 /* These exports are cleaned up by filter_disconnect(), so they
861 * need to be set up like real exports as filter_connect() does.
863 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
865 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
866 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
867 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
869 if (PTR_ERR(exp) == -EALREADY) {
870 /* export already exists, zero out this one */
871 CERROR("Zeroing out duplicate export due to "
873 lcd->lcd_uuid[0] = '\0';
875 GOTO(err_client, rc = PTR_ERR(exp));
878 fed = &exp->exp_filter_data;
880 fed->fed_group = 0; /* will be assigned at connect */
881 filter_export_stats_init(obd, exp, NULL);
882 rc = filter_client_add(obd, exp, cl_idx);
883 /* can't fail for existing client */
884 LASSERTF(rc == 0, "rc = %d\n", rc);
886 /* VBR: set export last committed */
887 exp->exp_last_committed = last_rcvd;
888 cfs_spin_lock(&exp->exp_lock);
889 exp->exp_connecting = 0;
890 exp->exp_in_recovery = 0;
891 cfs_spin_unlock(&exp->exp_lock);
892 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
893 obd->obd_max_recoverable_clients++;
894 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
896 class_export_put(exp);
899 /* Need to check last_rcvd even for duplicated exports. */
900 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
903 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
904 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
910 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
912 filter->fo_mount_count = mount_count + 1;
913 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
915 /* save it, so mount count and last_transno is current */
916 rc = filter_update_server_data(obd);
918 GOTO(err_client, rc);
923 class_disconnect_exports(obd);
925 filter_free_server_data(filter);
929 static int filter_cleanup_groups(struct obd_device *obd)
931 struct filter_obd *filter = &obd->u.filter;
933 struct dentry *dentry;
937 if (filter->fo_dentry_O_groups != NULL) {
938 for (i = 0; i < filter->fo_group_count; i++) {
939 dentry = filter->fo_dentry_O_groups[i];
943 OBD_FREE(filter->fo_dentry_O_groups,
944 filter->fo_group_count *
945 sizeof(*filter->fo_dentry_O_groups));
946 filter->fo_dentry_O_groups = NULL;
948 if (filter->fo_last_objid_files != NULL) {
949 for (i = 0; i < filter->fo_group_count; i++) {
950 filp = filter->fo_last_objid_files[i];
954 OBD_FREE(filter->fo_last_objid_files,
955 filter->fo_group_count *
956 sizeof(*filter->fo_last_objid_files));
957 filter->fo_last_objid_files = NULL;
959 if (filter->fo_dentry_O_sub != NULL) {
960 for (i = 0; i < filter->fo_group_count; i++) {
961 for (j = 0; j < filter->fo_subdir_count; j++) {
962 dentry = filter->fo_dentry_O_sub[i].dentry[j];
967 OBD_FREE(filter->fo_dentry_O_sub,
968 filter->fo_group_count *
969 sizeof(*filter->fo_dentry_O_sub));
970 filter->fo_dentry_O_sub = NULL;
972 if (filter->fo_last_objids != NULL) {
973 OBD_FREE(filter->fo_last_objids,
974 filter->fo_group_count *
975 sizeof(*filter->fo_last_objids));
976 filter->fo_last_objids = NULL;
978 if (filter->fo_dentry_O != NULL) {
979 f_dput(filter->fo_dentry_O);
980 filter->fo_dentry_O = NULL;
985 static int filter_update_last_group(struct obd_device *obd, int group)
987 struct filter_obd *filter = &obd->u.filter;
988 struct file *filp = NULL;
989 int last_group = 0, rc;
993 if (group <= filter->fo_committed_group)
996 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1000 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1004 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1006 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1010 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1011 obd->obd_name, last_group, group);
1015 /* must be sync: bXXXX */
1016 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1018 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1022 filter->fo_committed_group = group;
1025 filp_close(filp, 0);
1029 static int filter_read_group_internal(struct obd_device *obd, int group,
1032 struct filter_obd *filter = &obd->u.filter;
1033 __u64 *new_objids = NULL;
1034 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1035 struct dentry **new_groups = NULL;
1036 struct file **new_files = NULL;
1037 struct dentry *dentry;
1039 int old_count = filter->fo_group_count, rc, stage = 0, i;
1043 int len = group + 1;
1045 snprintf(name, 24, "%d", group);
1049 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1051 if (IS_ERR(dentry)) {
1052 CERROR("Cannot lookup expected object group %d: %ld\n",
1053 group, PTR_ERR(dentry));
1054 RETURN(PTR_ERR(dentry));
1057 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1059 if (IS_ERR(dentry)) {
1060 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1062 RETURN(PTR_ERR(dentry));
1067 snprintf(name, 24, "O/%d/LAST_ID", group);
1069 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1071 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1072 GOTO(cleanup, rc = PTR_ERR(filp));
1076 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1078 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1082 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1083 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1084 if (tmp_subdirs == NULL)
1085 GOTO(cleanup, rc = -ENOMEM);
1088 for (i = 0; i < filter->fo_subdir_count; i++) {
1090 snprintf(dir, sizeof(dir), "d%u", i);
1092 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1095 if (IS_ERR(tmp_subdirs->dentry[i])) {
1096 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1097 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1102 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1103 tmp_subdirs->dentry[i]);
1107 /* 'group' is an index; we need an array of length 'group + 1' */
1108 if (group + 1 > old_count) {
1109 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1110 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1111 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1112 OBD_ALLOC(new_files, len * sizeof(*new_files));
1114 if (new_objids == NULL || new_subdirs == NULL ||
1115 new_groups == NULL || new_files == NULL)
1116 GOTO(cleanup, rc = -ENOMEM);
1119 memcpy(new_objids, filter->fo_last_objids,
1120 old_count * sizeof(*new_objids));
1121 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1122 old_count * sizeof(*new_subdirs));
1123 memcpy(new_groups, filter->fo_dentry_O_groups,
1124 old_count * sizeof(*new_groups));
1125 memcpy(new_files, filter->fo_last_objid_files,
1126 old_count * sizeof(*new_files));
1128 OBD_FREE(filter->fo_last_objids,
1129 old_count * sizeof(*new_objids));
1130 OBD_FREE(filter->fo_dentry_O_sub,
1131 old_count * sizeof(*new_subdirs));
1132 OBD_FREE(filter->fo_dentry_O_groups,
1133 old_count * sizeof(*new_groups));
1134 OBD_FREE(filter->fo_last_objid_files,
1135 old_count * sizeof(*new_files));
1137 filter->fo_last_objids = new_objids;
1138 filter->fo_dentry_O_sub = new_subdirs;
1139 filter->fo_dentry_O_groups = new_groups;
1140 filter->fo_last_objid_files = new_files;
1141 filter->fo_group_count = len;
1144 filter->fo_dentry_O_groups[group] = dentry;
1145 filter->fo_last_objid_files[group] = filp;
1146 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1147 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1148 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1151 filter_update_last_group(obd, group);
1153 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1154 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1155 rc = filter_update_last_objid(obd, group, 1);
1159 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1160 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1161 obd->obd_name, group, last_objid);
1166 if (new_objids != NULL)
1167 OBD_FREE(new_objids, len * sizeof(*new_objids));
1168 if (new_subdirs != NULL)
1169 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1170 if (new_groups != NULL)
1171 OBD_FREE(new_groups, len * sizeof(*new_groups));
1172 if (new_files != NULL)
1173 OBD_FREE(new_files, len * sizeof(*new_files));
1175 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1176 for (i = 0; i < filter->fo_subdir_count; i++) {
1177 if (tmp_subdirs->dentry[i] != NULL)
1178 dput(tmp_subdirs->dentry[i]);
1180 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1183 filp_close(filp, 0);
1190 static int filter_read_groups(struct obd_device *obd, int last_group,
1193 struct filter_obd *filter = &obd->u.filter;
1194 int old_count, group, rc = 0;
1196 cfs_down(&filter->fo_init_lock);
1197 old_count = filter->fo_group_count;
1198 for (group = old_count; group <= last_group; group++) {
1199 rc = filter_read_group_internal(obd, group, create);
1203 cfs_up(&filter->fo_init_lock);
1207 /* FIXME: object groups */
1208 static int filter_prep_groups(struct obd_device *obd)
1210 struct filter_obd *filter = &obd->u.filter;
1211 struct dentry *O_dentry;
1213 int last_group, rc = 0, cleanup_phase = 0;
1217 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1219 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1220 if (IS_ERR(O_dentry)) {
1221 rc = PTR_ERR(O_dentry);
1222 CERROR("cannot open/create O: rc = %d\n", rc);
1225 filter->fo_dentry_O = O_dentry;
1226 cleanup_phase = 1; /* O_dentry */
1228 /* we have to initialize all groups before first connections from
1229 * clients because they may send create/destroy for any group -bzzz */
1230 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1232 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1233 GOTO(cleanup, rc = PTR_ERR(filp));
1235 cleanup_phase = 2; /* filp */
1237 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1239 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1244 last_group = FILTER_GROUP_MDS0;
1246 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1247 FILTER_GROUP_MDS0, last_group);
1248 filter->fo_committed_group = last_group;
1249 rc = filter_read_groups(obd, last_group, 1);
1253 filp_close(filp, 0);
1257 switch (cleanup_phase) {
1259 filp_close(filp, 0);
1261 filter_cleanup_groups(obd);
1262 f_dput(filter->fo_dentry_O);
1263 filter->fo_dentry_O = NULL;
1271 /* setup the object store with correct subdirectories */
1272 static int filter_prep(struct obd_device *obd)
1274 struct lvfs_run_ctxt saved;
1275 struct filter_obd *filter = &obd->u.filter;
1277 struct inode *inode;
1281 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1282 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1283 if (!file || IS_ERR(file)) {
1285 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1289 filter->fo_rcvd_filp = file;
1290 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1291 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1292 file->f_dentry->d_inode->i_mode);
1293 GOTO(err_filp, rc = -ENOENT);
1296 inode = file->f_dentry->d_parent->d_inode;
1297 /* We use i_op->unlink directly in filter_vfs_unlink() */
1298 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1299 CERROR("%s: filesystem does not support create/unlink ops\n",
1301 GOTO(err_filp, rc = -EOPNOTSUPP);
1304 rc = filter_init_server_data(obd, file);
1306 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1309 LASSERT(obd->u.obt.obt_lut);
1310 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1312 /* open and create health check io file*/
1313 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1316 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1318 GOTO(err_server_data, rc);
1320 filter->fo_obt.obt_health_check_filp = file;
1321 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1322 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1323 file->f_dentry->d_inode->i_mode);
1324 GOTO(err_health_check, rc = -ENOENT);
1326 rc = lvfs_check_io_health(obd, file);
1328 GOTO(err_health_check, rc);
1330 rc = filter_prep_groups(obd);
1332 GOTO(err_health_check, rc);
1334 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1339 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1340 CERROR("can't close %s after error\n", HEALTH_CHECK);
1341 filter->fo_obt.obt_health_check_filp = NULL;
1343 target_recovery_fini(obd);
1344 filter_free_server_data(filter);
1346 if (filp_close(filter->fo_rcvd_filp, 0))
1347 CERROR("can't close %s after error\n", LAST_RCVD);
1348 filter->fo_rcvd_filp = NULL;
1352 /* cleanup the filter: write last used object id to status file */
1353 static void filter_post(struct obd_device *obd)
1355 struct lvfs_run_ctxt saved;
1356 struct filter_obd *filter = &obd->u.filter;
1359 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1360 * best to start a transaction with h_sync, because we removed this
1363 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1364 rc = filter_update_server_data(obd);
1366 CERROR("error writing server data: rc = %d\n", rc);
1368 for (i = 0; i < filter->fo_group_count; i++) {
1369 rc = filter_update_last_objid(obd, i,
1370 (i == filter->fo_group_count - 1));
1372 CERROR("error writing group %d lastobjid: rc = %d\n",
1376 rc = filp_close(filter->fo_rcvd_filp, 0);
1377 filter->fo_rcvd_filp = NULL;
1379 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1381 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1382 filter->fo_obt.obt_health_check_filp = NULL;
1384 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1386 filter_cleanup_groups(obd);
1387 filter_free_server_data(filter);
1388 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1390 filter_free_capa_keys(filter);
1391 cleanup_capa_hash(filter->fo_capa_hash);
1394 static void filter_set_last_id(struct filter_obd *filter,
1395 obd_id id, obd_gr group)
1397 LASSERT(group <= filter->fo_group_count);
1399 cfs_spin_lock(&filter->fo_objidlock);
1400 filter->fo_last_objids[group] = id;
1401 cfs_spin_unlock(&filter->fo_objidlock);
1404 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1407 LASSERT(group <= filter->fo_group_count);
1408 LASSERT(filter->fo_last_objids != NULL);
1410 /* FIXME: object groups */
1411 cfs_spin_lock(&filter->fo_objidlock);
1412 id = filter->fo_last_objids[group];
1413 cfs_spin_unlock(&filter->fo_objidlock);
1417 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1419 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1423 /* We never dget the object parent, so DON'T dput it either */
1424 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1426 struct filter_obd *filter = &obd->u.filter;
1427 struct filter_subdirs *subdirs;
1428 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1430 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1431 return filter->fo_dentry_O_groups[group];
1433 subdirs = &filter->fo_dentry_O_sub[group];
1434 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1437 /* We never dget the object parent, so DON'T dput it either */
1438 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1441 unsigned long now = jiffies;
1442 struct dentry *dparent = filter_parent(obd, group, objid);
1445 if (IS_ERR(dparent))
1447 if (dparent == NULL)
1448 return ERR_PTR(-ENOENT);
1450 rc = filter_lock_dentry(obd, dparent);
1451 fsfilt_check_slow(obd, now, "parent lock");
1452 return rc ? ERR_PTR(rc) : dparent;
1455 /* We never dget the object parent, so DON'T dput it either */
1456 static void filter_parent_unlock(struct dentry *dparent)
1458 UNLOCK_INODE_MUTEX(dparent->d_inode);
1461 /* How to get files, dentries, inodes from object id's.
1463 * If dir_dentry is passed, the caller has already locked the parent
1464 * appropriately for this operation (normally a write lock). If
1465 * dir_dentry is NULL, we do a read lock while we do the lookup to
1466 * avoid races with create/destroy and such changing the directory
1467 * internal to the filesystem code. */
1468 struct dentry *filter_fid2dentry(struct obd_device *obd,
1469 struct dentry *dir_dentry,
1470 obd_gr group, obd_id id)
1472 struct dentry *dparent = dir_dentry;
1473 struct dentry *dchild;
1478 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1479 obd->u.filter.fo_destroys_in_progress == 0) {
1480 /* don't fail lookups for orphan recovery, it causes
1481 * later LBUGs when objects still exist during precreate */
1482 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1483 RETURN(ERR_PTR(-ENOENT));
1486 CERROR("fatal: invalid object id 0\n");
1487 RETURN(ERR_PTR(-ESTALE));
1490 len = sprintf(name, LPU64, id);
1491 if (dir_dentry == NULL) {
1492 dparent = filter_parent_lock(obd, group, id);
1493 if (IS_ERR(dparent)) {
1494 CERROR("%s: error getting object "LPU64":"LPU64
1495 " parent: rc %ld\n", obd->obd_name,
1496 id, group, PTR_ERR(dparent));
1500 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1501 dparent->d_name.len, dparent->d_name.name, name);
1502 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1503 if (dir_dentry == NULL)
1504 filter_parent_unlock(dparent);
1505 if (IS_ERR(dchild)) {
1506 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1511 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1512 CERROR("%s: got bad object "LPU64" inode %lu\n",
1513 obd->obd_name, id, dchild->d_inode->i_ino);
1515 RETURN(ERR_PTR(-ENOENT));
1518 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1519 name, dchild, atomic_read(&dchild->d_count));
1521 LASSERT(atomic_read(&dchild->d_count) > 0);
1526 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1527 obd_id group, struct lustre_handle *lockh)
1529 int flags = LDLM_AST_DISCARD_DATA, rc;
1530 struct ldlm_res_id res_id;
1531 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1534 osc_build_res_name(objid, group, &res_id);
1535 /* Tell the clients that the object is gone now and that they should
1536 * throw away any cached pages. */
1537 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1538 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1539 ldlm_completion_ast, NULL, NULL, 0, NULL,
1546 static void filter_fini_destroy(struct obd_device *obd,
1547 struct lustre_handle *lockh)
1549 if (lustre_handle_is_used(lockh))
1550 ldlm_lock_decref(lockh, LCK_PW);
1553 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1554 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1555 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1556 * reverse. Caller must take i_sem before starting the transaction and we
1557 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1558 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1559 struct vfsmount *mnt)
1564 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1565 * itself we already hold dir->i_mutex for child create/unlink ops */
1566 LASSERT(dentry->d_inode != NULL);
1567 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1568 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1572 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1573 GOTO(out, rc = -ENOENT);
1575 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1580 GOTO(out, rc = -EPERM);
1582 /* check_sticky() */
1583 if ((dentry->d_inode->i_uid != current->fsuid &&
1584 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1585 IS_IMMUTABLE(dentry->d_inode))
1586 GOTO(out, rc = -EPERM);
1588 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1589 * that was done because of journal_start (which is already done
1590 * here) or some other ordering issue. */
1593 rc = ll_security_inode_unlink(dir, dentry, mnt);
1597 rc = dir->i_op->unlink(dir, dentry);
1599 /* need to drop i_mutex before we lose inode reference */
1600 UNLOCK_INODE_MUTEX(dentry->d_inode);
1607 /* Caller must hold LCK_PW on parent and push us into kernel context.
1608 * Caller must hold child i_mutex, we drop it always.
1609 * Caller is also required to ensure that dchild->d_inode exists. */
1610 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1611 obd_gr group, struct dentry *dparent,
1612 struct dentry *dchild)
1614 struct inode *inode = dchild->d_inode;
1615 struct filter_obd *filter = &obd->u.filter;
1618 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1619 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1620 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1621 (unsigned long)inode->i_nlink,
1622 atomic_read(&inode->i_count));
1625 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1627 CERROR("error unlinking objid %.*s: rc %d\n",
1628 dchild->d_name.len, dchild->d_name.name, rc);
1632 struct filter_intent_args {
1633 struct ldlm_lock **victim;
1638 static enum interval_iter filter_intent_cb(struct interval_node *n,
1641 struct ldlm_interval *node = (struct ldlm_interval *)n;
1642 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1643 __u64 size = arg->size;
1644 struct ldlm_lock **v = arg->victim;
1645 struct ldlm_lock *lck;
1647 /* If the interval is lower than the current file size,
1649 if (interval_high(n) <= size)
1650 return INTERVAL_ITER_STOP;
1652 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1653 /* Don't send glimpse ASTs to liblustre clients.
1654 * They aren't listening for them, and they do
1655 * entirely synchronous I/O anyways. */
1656 if (lck->l_export == NULL ||
1657 lck->l_export->exp_libclient == 1)
1660 if (*arg->liblustre)
1661 *arg->liblustre = 0;
1664 *v = LDLM_LOCK_GET(lck);
1665 } else if ((*v)->l_policy_data.l_extent.start <
1666 lck->l_policy_data.l_extent.start) {
1667 LDLM_LOCK_RELEASE(*v);
1668 *v = LDLM_LOCK_GET(lck);
1671 /* the same policy group - every lock has the
1672 * same extent, so needn't do it any more */
1676 return INTERVAL_ITER_CONT;
1679 static int filter_intent_policy(struct ldlm_namespace *ns,
1680 struct ldlm_lock **lockp, void *req_cookie,
1681 ldlm_mode_t mode, int flags, void *data)
1683 CFS_LIST_HEAD(rpc_list);
1684 struct ptlrpc_request *req = req_cookie;
1685 struct ldlm_lock *lock = *lockp, *l = NULL;
1686 struct ldlm_resource *res = lock->l_resource;
1687 ldlm_processing_policy policy;
1688 struct ost_lvb *res_lvb, *reply_lvb;
1689 struct ldlm_reply *rep;
1691 int idx, rc, tmpflags = 0, only_liblustre = 1;
1692 struct ldlm_interval_tree *tree;
1693 struct filter_intent_args arg;
1694 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1695 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1696 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1699 policy = ldlm_get_processing_policy(res);
1700 LASSERT(policy != NULL);
1701 LASSERT(req != NULL);
1703 rc = lustre_pack_reply(req, 3, repsize, NULL);
1705 RETURN(req->rq_status = rc);
1707 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1708 LASSERT(rep != NULL);
1710 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1711 sizeof(*reply_lvb));
1712 LASSERT(reply_lvb != NULL);
1714 //fixup_handle_for_resent_req(req, lock, &lockh);
1716 /* Call the extent policy function to see if our request can be
1717 * granted, or is blocked.
1718 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1719 * lock, and should not be granted if the lock will be blocked.
1722 LASSERT(ns == res->lr_namespace);
1724 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1725 check_res_locked(res);
1727 /* FIXME: we should change the policy function slightly, to not make
1728 * this list at all, since we just turn around and free it */
1729 while (!cfs_list_empty(&rpc_list)) {
1730 struct ldlm_lock *wlock =
1731 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1733 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1734 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1735 lock->l_flags &= ~LDLM_FL_CP_REQD;
1736 cfs_list_del_init(&wlock->l_cp_ast);
1737 LDLM_LOCK_RELEASE(wlock);
1740 /* The lock met with no resistance; we're finished. */
1741 if (rc == LDLM_ITER_CONTINUE) {
1742 /* do not grant locks to the liblustre clients: they cannot
1743 * handle ASTs robustly. We need to do this while still
1744 * holding ns_lock to avoid the lock remaining on the res_link
1745 * list (and potentially being added to l_pending_list by an
1746 * AST) when we are going to drop this lock ASAP. */
1747 if (lock->l_export->exp_libclient ||
1748 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1749 ldlm_resource_unlink_lock(lock);
1750 err = ELDLM_LOCK_ABORTED;
1752 err = ELDLM_LOCK_REPLACED;
1758 /* Do not grant any lock, but instead send GL callbacks. The extent
1759 * policy nicely created a list of all PW locks for us. We will choose
1760 * the highest of those which are larger than the size in the LVB, if
1761 * any, and perform a glimpse callback. */
1762 res_lvb = res->lr_lvb_data;
1763 LASSERT(res_lvb != NULL);
1764 *reply_lvb = *res_lvb;
1767 * ->ns_lock guarantees that no new locks are granted, and,
1768 * therefore, that res->lr_lvb_data cannot increase beyond the
1769 * end of already granted lock. As a result, it is safe to
1770 * check against "stale" reply_lvb->lvb_size value without
1773 arg.size = reply_lvb->lvb_size;
1775 arg.liblustre = &only_liblustre;
1776 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1777 tree = &res->lr_itree[idx];
1778 if (tree->lit_mode == LCK_PR)
1781 interval_iterate_reverse(tree->lit_root,
1782 filter_intent_cb, &arg);
1786 /* There were no PW locks beyond the size in the LVB; finished. */
1788 if (only_liblustre) {
1789 /* If we discovered a liblustre client with a PW lock,
1790 * however, the LVB may be out of date! The LVB is
1791 * updated only on glimpse (which we don't do for
1792 * liblustre clients) and cancel (which the client
1793 * obviously has not yet done). So if it has written
1794 * data but kept the lock, the LVB is stale and needs
1795 * to be updated from disk.
1797 * Of course, this will all disappear when we switch to
1798 * taking liblustre locks on the OST. */
1799 ldlm_res_lvbo_update(res, NULL, 1);
1801 RETURN(ELDLM_LOCK_ABORTED);
1805 * This check is for lock taken in filter_prepare_destroy() that does
1806 * not have l_glimpse_ast set. So the logic is: if there is a lock
1807 * with no l_glimpse_ast set, this object is being destroyed already.
1809 * Hence, if you are grabbing DLM locks on the server, always set
1810 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1812 if (l->l_glimpse_ast == NULL) {
1813 /* We are racing with unlink(); just return -ENOENT */
1814 rep->lock_policy_res1 = -ENOENT;
1818 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1819 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1820 /* Update the LVB from disk if the AST failed (this is a legal race) */
1822 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1823 * sending ast is not handled. This can result in lost client writes.
1826 ldlm_res_lvbo_update(res, NULL, 1);
1829 *reply_lvb = *res_lvb;
1833 LDLM_LOCK_RELEASE(l);
1835 RETURN(ELDLM_LOCK_ABORTED);
1839 * per-obd_device iobuf pool.
1841 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1842 * allocations in write-path has to be minimized (see bug 5137).
1844 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1845 * OST threads (see ost_thread_{init,done}()).
1847 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1848 * at the OST layer there are only (potentially) multiple obd_device of type
1849 * unknown at the time of OST thread creation.
1851 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1852 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1853 * it's very own iobuf.
1857 * filter_kiobuf_pool_init()
1859 * filter_kiobuf_pool_done()
1861 * filter_iobuf_get()
1863 * operate on this array. They are "generic" in a sense that they don't depend
1864 * on actual type of iobuf's (the latter depending on Linux kernel version).
1868 * destroy pool created by filter_iobuf_pool_init
1870 static void filter_iobuf_pool_done(struct filter_obd *filter)
1872 struct filter_iobuf **pool;
1877 pool = filter->fo_iobuf_pool;
1879 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1880 if (pool[i] != NULL)
1881 filter_free_iobuf(pool[i]);
1883 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1884 filter->fo_iobuf_pool = NULL;
1889 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1891 struct filter_obd *filter = &obd->u.filter;
1892 struct sptlrpc_rule_set tmp_rset;
1895 sptlrpc_rule_set_init(&tmp_rset);
1896 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1898 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1903 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1905 cfs_write_lock(&filter->fo_sptlrpc_lock);
1906 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1907 filter->fo_sptlrpc_rset = tmp_rset;
1908 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1914 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1916 static int filter_iobuf_pool_init(struct filter_obd *filter)
1923 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1925 if (filter->fo_iobuf_pool == NULL)
1928 filter->fo_iobuf_count = OSS_THREADS_MAX;
1933 /* Return iobuf allocated for @thread_id. We don't know in advance how
1934 * many threads there will be so we allocate a large empty array and only
1935 * fill in those slots that are actually in use.
1936 * If we haven't allocated a pool entry for this thread before, do so now. */
1937 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1939 int thread_id = (oti && oti->oti_thread) ?
1940 oti->oti_thread->t_id : -1;
1941 struct filter_iobuf *pool = NULL;
1942 struct filter_iobuf **pool_place = NULL;
1944 if (thread_id >= 0) {
1945 LASSERT(thread_id < filter->fo_iobuf_count);
1946 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1949 if (unlikely(pool == NULL)) {
1950 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1951 PTLRPC_MAX_BRW_PAGES);
1952 if (pool_place != NULL)
1959 /* mount the file system (secretly). lustre_cfg parameters are:
1962 * 3 = flags: failover=f, failout=n
1965 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1968 struct filter_obd *filter = &obd->u.filter;
1969 struct vfsmount *mnt;
1970 struct lustre_mount_info *lmi;
1971 struct obd_uuid uuid;
1979 if (lcfg->lcfg_bufcount < 3 ||
1980 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1981 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1984 lmi = server_get_mount(obd->obd_name);
1986 /* We already mounted in lustre_fill_super.
1987 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1988 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1990 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1992 /* gets recovery timeouts from mount data */
1993 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1994 obd->obd_recovery_timeout =
1995 lsi->lsi_lmd->lmd_recovery_time_soft;
1996 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1997 obd->obd_recovery_time_hard =
1998 lsi->lsi_lmd->lmd_recovery_time_hard;
2000 /* old path - used by lctl */
2001 CERROR("Using old MDS mount method\n");
2002 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2003 MS_NOATIME|MS_NODIRATIME,
2004 lustre_cfg_string(lcfg, 1), option);
2007 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2008 lustre_cfg_string(lcfg, 1), rc);
2012 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2014 if (IS_ERR(obd->obd_fsops))
2015 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2017 rc = filter_iobuf_pool_init(filter);
2021 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2022 CERROR("%s: Underlying device is marked as read-only. "
2023 "Setup failed\n", obd->obd_name);
2024 GOTO(err_ops, rc = -EROFS);
2027 /* failover is the default */
2028 obd->obd_replayable = 1;
2030 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2031 str = lustre_cfg_string(lcfg, 3);
2032 if (strchr(str, 'n')) {
2033 CWARN("%s: recovery disabled\n", obd->obd_name);
2034 obd->obd_replayable = 0;
2038 filter->fo_vfsmnt = mnt;
2039 obd->u.obt.obt_sb = mnt->mnt_sb;
2040 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2041 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2043 fsfilt_setup(obd, obd->u.obt.obt_sb);
2045 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2046 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2047 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2048 obd->obd_lvfs_ctxt.fs = get_ds();
2049 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2051 cfs_init_mutex(&filter->fo_init_lock);
2052 filter->fo_committed_group = 0;
2053 filter->fo_destroys_in_progress = 0;
2054 for (i = 0; i < 32; i++)
2055 cfs_sema_init(&filter->fo_create_locks[i], 1);
2057 cfs_spin_lock_init(&filter->fo_objidlock);
2058 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2059 cfs_sema_init(&filter->fo_alloc_lock, 1);
2060 init_brw_stats(&filter->fo_filter_stats);
2061 filter->fo_read_cache = 1; /* enable read-only cache by default */
2062 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2063 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2064 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2065 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2067 rc = filter_prep(obd);
2071 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2072 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2074 filter->fo_fl_oss_capa = 1;
2076 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2077 filter->fo_capa_hash = init_capa_hash();
2078 if (filter->fo_capa_hash == NULL)
2079 GOTO(err_post, rc = -ENOMEM);
2081 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2082 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2083 LDLM_NAMESPACE_GREEDY);
2084 if (obd->obd_namespace == NULL)
2085 GOTO(err_post, rc = -ENOMEM);
2086 obd->obd_namespace->ns_lvbp = obd;
2087 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2088 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2090 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2091 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2093 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2095 CERROR("failed to setup llogging subsystems\n");
2099 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2100 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2101 /* do this after llog being initialized */
2102 filter_adapt_sptlrpc_conf(obd, 1);
2104 rc = lquota_setup(filter_quota_interface_ref, obd);
2108 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2109 if (q->max_sectors < q->max_hw_sectors &&
2110 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2111 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2112 "for larger I/O requests: max_sectors = %u "
2113 "could be up to max_hw_sectors=%u\n",
2114 obd->obd_name, mnt->mnt_sb->s_id,
2115 q->max_sectors, q->max_hw_sectors);
2117 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2118 if (uuid_ptr != NULL) {
2119 class_uuid_unparse(uuid_ptr, &uuid);
2125 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2126 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2127 obd->obd_name, label ?: str, lmi ? "on " : "",
2128 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2129 obd->obd_replayable ? "enabled" : "disabled");
2131 if (obd->obd_recovering)
2132 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2133 "or until %d client%s reconnect%s\n",
2135 obd->obd_recovery_timeout / 60,
2136 obd->obd_recovery_timeout % 60,
2137 obd->obd_max_recoverable_clients,
2138 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2139 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2147 fsfilt_put_ops(obd->obd_fsops);
2148 filter_iobuf_pool_done(filter);
2150 server_put_mount(obd->obd_name, mnt);
2151 obd->u.obt.obt_sb = 0;
2155 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2157 struct lprocfs_static_vars lvars;
2162 CLASSERT(offsetof(struct obd_device, u.obt) ==
2163 offsetof(struct obd_device, u.filter.fo_obt));
2165 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2168 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2169 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2172 addr = (unsigned long)cfs_page_address(page);
2173 clear_page((void *)addr);
2175 /* lprocfs must be setup before the filter so state can be safely added
2176 * to /proc incrementally as the filter is setup */
2177 lprocfs_filter_init_vars(&lvars);
2178 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2179 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2180 /* Init obdfilter private stats here */
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "read_bytes", "bytes");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2185 LPROCFS_CNTR_AVGMINMAX,
2186 "write_bytes", "bytes");
2187 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2188 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2189 "get_page", "usec");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2191 LPROCFS_CNTR_AVGMINMAX,
2192 "get_page_failures", "num");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2194 LPROCFS_CNTR_AVGMINMAX,
2195 "cache_access", "pages");
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2197 LPROCFS_CNTR_AVGMINMAX,
2198 "cache_hit", "pages");
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2200 LPROCFS_CNTR_AVGMINMAX,
2201 "cache_miss", "pages");
2203 lproc_filter_attach_seqstat(obd);
2204 obd->obd_proc_exports_entry = lprocfs_register("exports",
2205 obd->obd_proc_entry,
2207 if (IS_ERR(obd->obd_proc_exports_entry)) {
2208 rc = PTR_ERR(obd->obd_proc_exports_entry);
2209 CERROR("error %d setting up lprocfs for %s\n",
2211 obd->obd_proc_exports_entry = NULL;
2214 if (obd->obd_proc_exports_entry)
2215 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2216 lprocfs_nid_stats_clear_read,
2217 lprocfs_nid_stats_clear_write, obd, NULL);
2219 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2220 LUSTRE_CFG_BUFLEN(lcfg, 4));
2221 rc = filter_common_setup(obd, lcfg, (void *)addr);
2222 OBD_PAGE_FREE(page);
2225 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2226 lprocfs_free_per_client_stats(obd);
2227 lprocfs_free_obd_stats(obd);
2228 lprocfs_obd_cleanup(obd);
2234 static struct llog_operations filter_mds_ost_repl_logops;
2236 static struct llog_operations filter_size_orig_logops = {
2237 .lop_setup = llog_obd_origin_setup,
2238 .lop_cleanup = llog_obd_origin_cleanup,
2239 .lop_add = llog_obd_origin_add
2242 static int filter_olg_fini(struct obd_llog_group *olg)
2244 struct llog_ctxt *ctxt;
2245 int rc = 0, rc2 = 0;
2248 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2250 rc = llog_cleanup(ctxt);
2252 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2254 rc2 = llog_cleanup(ctxt);
2259 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2261 rc2 = llog_cleanup(ctxt);
2270 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2271 struct obd_device *tgt)
2276 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2277 &filter_mds_ost_repl_logops);
2281 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2282 &filter_size_orig_logops);
2288 filter_olg_fini(olg);
2293 * Init the default olg, which is embeded in the obd_device, for filter.
2296 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2297 struct obd_device *tgt)
2299 struct filter_obd *filter = &obd->u.filter;
2300 struct llog_ctxt *ctxt;
2304 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2305 if (!filter->fo_lcm)
2308 filter_mds_ost_repl_logops = llog_client_ops;
2309 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2310 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2311 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2313 rc = filter_olg_init(obd, olg, tgt);
2315 GOTO(cleanup_lcm, rc);
2317 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2320 GOTO(cleanup_olg, rc);
2322 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2324 CERROR("Can't get ctxt for %p:%x\n", olg,
2325 LLOG_MDS_OST_REPL_CTXT);
2326 GOTO(cleanup_olg, rc = -ENODEV);
2328 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2329 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2330 llog_ctxt_put(ctxt);
2334 filter_olg_fini(olg);
2336 llog_recov_thread_fini(filter->fo_lcm, 1);
2337 filter->fo_lcm = NULL;
2342 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2343 struct obd_device *tgt, int *index)
2345 struct filter_obd *filter = &obd->u.filter;
2346 struct llog_ctxt *ctxt;
2350 LASSERT(olg != NULL);
2351 if (olg == &obd->obd_olg)
2352 return filter_default_olg_init(obd, olg, tgt);
2354 LASSERT(filter->fo_lcm != NULL);
2355 rc = filter_olg_init(obd, olg, tgt);
2358 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2360 CERROR("Can't get ctxt for %p:%x\n", olg,
2361 LLOG_MDS_OST_REPL_CTXT);
2362 filter_olg_fini(olg);
2365 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2366 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2367 llog_ctxt_put(ctxt);
2371 static int filter_llog_finish(struct obd_device *obd, int count)
2373 struct filter_obd *filter = &obd->u.filter;
2374 struct llog_ctxt *ctxt;
2377 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2380 * Make sure that no cached llcds left in recov_thread.
2381 * We actually do sync in disconnect time, but disconnect
2382 * may not come being marked rq_no_resend = 1.
2384 llog_sync(ctxt, NULL);
2387 * Balance class_import_get() in llog_receptor_accept().
2388 * This is safe to do, as llog is already synchronized
2389 * and its import may go.
2391 cfs_mutex_down(&ctxt->loc_sem);
2392 if (ctxt->loc_imp) {
2393 class_import_put(ctxt->loc_imp);
2394 ctxt->loc_imp = NULL;
2396 cfs_mutex_up(&ctxt->loc_sem);
2397 llog_ctxt_put(ctxt);
2400 if (filter->fo_lcm) {
2401 cfs_mutex_down(&ctxt->loc_sem);
2402 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2403 filter->fo_lcm = NULL;
2404 cfs_mutex_up(&ctxt->loc_sem);
2406 RETURN(filter_olg_fini(&obd->obd_olg));
2410 * Find the group llog according to group index in the llog group list.
2412 static struct obd_llog_group *
2413 filter_find_olg_internal(struct filter_obd *filter, int group)
2415 struct obd_llog_group *olg;
2417 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2418 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2419 if (olg->olg_group == group)
2426 * Find the group llog according to group index on the filter
2428 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2430 struct obd_llog_group *olg = NULL;
2431 struct filter_obd *filter;
2433 filter = &obd->u.filter;
2435 if (group == FILTER_GROUP_LLOG)
2436 RETURN(&obd->obd_olg);
2438 cfs_spin_lock(&filter->fo_llog_list_lock);
2439 olg = filter_find_olg_internal(filter, group);
2440 cfs_spin_unlock(&filter->fo_llog_list_lock);
2445 * Find the llog_group of the filter according to the group. If it can not
2446 * find, create the llog_group, which only happens when mds is being synced
2449 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2451 struct obd_llog_group *olg = NULL;
2452 struct filter_obd *filter;
2455 filter = &obd->u.filter;
2457 if (group == FILTER_GROUP_LLOG)
2458 RETURN(&obd->obd_olg);
2460 cfs_spin_lock(&filter->fo_llog_list_lock);
2461 olg = filter_find_olg_internal(filter, group);
2463 if (olg->olg_initializing) {
2464 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2466 GOTO(out_unlock, olg);
2471 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2473 llog_group_init(olg, group);
2474 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2475 olg->olg_initializing = 1;
2476 cfs_spin_unlock(&filter->fo_llog_list_lock);
2478 rc = obd_llog_init(obd, olg, obd, NULL);
2480 cfs_spin_lock(&filter->fo_llog_list_lock);
2481 cfs_list_del(&olg->olg_list);
2482 cfs_spin_unlock(&filter->fo_llog_list_lock);
2484 GOTO(out, olg = ERR_PTR(-ENOMEM));
2486 cfs_spin_lock(&filter->fo_llog_list_lock);
2487 olg->olg_initializing = 0;
2488 cfs_spin_unlock(&filter->fo_llog_list_lock);
2489 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2490 obd->obd_name, group, olg);
2495 cfs_spin_unlock(&filter->fo_llog_list_lock);
2499 static int filter_llog_connect(struct obd_export *exp,
2500 struct llogd_conn_body *body)
2502 struct obd_device *obd = exp->exp_obd;
2503 struct llog_ctxt *ctxt;
2504 struct obd_llog_group *olg;
2508 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2509 obd->obd_name, body->lgdc_logid.lgl_oid,
2510 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2512 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2514 CERROR(" %s: can not find olg of group %d\n",
2515 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2518 llog_group_set_export(olg, exp);
2520 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2521 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2522 body->lgdc_ctxt_idx);
2524 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2525 obd->obd_name, body->lgdc_logid.lgl_oid,
2526 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2528 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2529 obd->u.filter.fo_mds_ost_sync = 1;
2530 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2531 rc = llog_connect(ctxt, &body->lgdc_logid,
2532 &body->lgdc_gen, NULL);
2533 llog_ctxt_put(ctxt);
2535 CERROR("failed to connect rc %d idx %d\n", rc,
2536 body->lgdc_ctxt_idx);
2541 static int filter_llog_preclean(struct obd_device *obd)
2543 struct obd_llog_group *olg, *tmp;
2544 struct filter_obd *filter;
2545 cfs_list_t remove_list;
2549 rc = obd_llog_finish(obd, 0);
2551 CERROR("failed to cleanup llogging subsystem\n");
2553 filter = &obd->u.filter;
2554 CFS_INIT_LIST_HEAD(&remove_list);
2556 cfs_spin_lock(&filter->fo_llog_list_lock);
2557 while (!cfs_list_empty(&filter->fo_llog_list)) {
2558 olg = cfs_list_entry(filter->fo_llog_list.next,
2559 struct obd_llog_group, olg_list);
2560 cfs_list_del(&olg->olg_list);
2561 cfs_list_add(&olg->olg_list, &remove_list);
2563 cfs_spin_unlock(&filter->fo_llog_list_lock);
2565 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2566 cfs_list_del_init(&olg->olg_list);
2567 rc = filter_olg_fini(olg);
2569 CERROR("failed to cleanup llogging subsystem for %u\n",
2577 static int filter_precleanup(struct obd_device *obd,
2578 enum obd_cleanup_stage stage)
2584 case OBD_CLEANUP_EARLY:
2586 case OBD_CLEANUP_EXPORTS:
2587 /* Stop recovery before namespace cleanup. */
2588 target_recovery_fini(obd);
2589 rc = filter_llog_preclean(obd);
2595 static int filter_cleanup(struct obd_device *obd)
2597 struct filter_obd *filter = &obd->u.filter;
2601 LCONSOLE_WARN("%s: shutting down for failover; client state "
2602 "will be preserved.\n", obd->obd_name);
2604 obd_exports_barrier(obd);
2605 obd_zombie_barrier();
2607 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2608 lprocfs_free_per_client_stats(obd);
2609 lprocfs_free_obd_stats(obd);
2610 lprocfs_obd_cleanup(obd);
2611 lquota_cleanup(filter_quota_interface_ref, obd);
2613 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2614 obd->obd_namespace = NULL;
2616 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2618 if (obd->u.obt.obt_sb == NULL)
2623 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2624 shrink_dcache_sb(obd->u.obt.obt_sb);
2626 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2627 obd->u.obt.obt_sb = NULL;
2629 fsfilt_put_ops(obd->obd_fsops);
2631 filter_iobuf_pool_done(filter);
2633 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2638 static int filter_connect_internal(struct obd_export *exp,
2639 struct obd_connect_data *data,
2642 struct filter_export_data *fed = &exp->exp_filter_data;
2647 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2648 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2649 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2650 data->ocd_connect_flags, data->ocd_version,
2651 data->ocd_grant, data->ocd_index);
2653 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2654 CWARN("!!! This export (nid %s) used object group %d "
2655 "earlier; now it's trying to use group %d! This could "
2656 "be a bug in the MDS. Please report to "
2657 "http://bugzilla.lustre.org/\n",
2658 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2661 fed->fed_group = data->ocd_group;
2663 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2664 exp->exp_connect_flags = data->ocd_connect_flags;
2665 data->ocd_version = LUSTRE_VERSION_CODE;
2667 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2668 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2669 data->ocd_connect_flags & OBD_CONNECT_MDS))
2672 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2673 struct filter_obd *filter = &exp->exp_obd->u.filter;
2674 obd_size left, want;
2676 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2677 left = filter_grant_space_left(exp);
2678 want = data->ocd_grant;
2679 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2680 data->ocd_grant = fed->fed_grant;
2681 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2683 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2684 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2685 exp->exp_client_uuid.uuid, exp,
2686 data->ocd_grant, want, left);
2688 filter->fo_tot_granted_clients ++;
2691 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2692 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2693 int index = le32_to_cpu(lsd->lsd_ost_index);
2695 if (!(lsd->lsd_feature_compat &
2696 cpu_to_le32(OBD_COMPAT_OST))) {
2697 /* this will only happen on the first connect */
2698 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2699 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2700 /* sync is not needed here as filter_client_add will
2701 * set exp_need_sync flag */
2702 filter_update_server_data(exp->exp_obd);
2703 } else if (index != data->ocd_index) {
2704 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2705 " %u doesn't match actual OST index"
2706 " %u in last_rcvd file, bad "
2708 obd_export_nid2str(exp), index,
2712 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2713 * FIXME: We don't strictly need the COMPAT flag for that,
2714 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2715 * FIXME: We needed it for the index, as index 0 is valid. */
2718 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2719 data->ocd_brw_size = 65536;
2720 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2721 data->ocd_brw_size = min(data->ocd_brw_size,
2722 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2723 LASSERT(data->ocd_brw_size);
2726 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2727 __u32 cksum_types = data->ocd_cksum_types;
2729 /* The client set in ocd_cksum_types the checksum types it
2730 * supports. We have to mask off the algorithms that we don't
2732 if (cksum_types & OBD_CKSUM_ALL)
2733 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2735 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2737 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2738 "%x\n", exp->exp_obd->obd_name,
2739 obd_export_nid2str(exp), cksum_types,
2740 data->ocd_cksum_types);
2742 /* This client does not support OBD_CONNECT_CKSUM
2743 * fall back to CRC32 */
2744 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2745 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2746 exp->exp_obd->obd_name,
2747 obd_export_nid2str(exp));
2753 static int filter_reconnect(const struct lu_env *env,
2754 struct obd_export *exp, struct obd_device *obd,
2755 struct obd_uuid *cluuid,
2756 struct obd_connect_data *data,
2762 if (exp == NULL || obd == NULL || cluuid == NULL)
2765 rc = filter_connect_internal(exp, data, 1);
2767 filter_export_stats_init(obd, exp, localdata);
2772 /* nearly identical to mds_connect */
2773 static int filter_connect(const struct lu_env *env,
2774 struct obd_export **exp, struct obd_device *obd,
2775 struct obd_uuid *cluuid,
2776 struct obd_connect_data *data, void *localdata)
2778 struct lvfs_run_ctxt saved;
2779 struct lustre_handle conn = { 0 };
2780 struct obd_export *lexp;
2781 struct filter_export_data *fed;
2782 struct lsd_client_data *lcd = NULL;
2787 if (exp == NULL || obd == NULL || cluuid == NULL)
2790 rc = class_connect(&conn, obd, cluuid);
2793 lexp = class_conn2export(&conn);
2794 LASSERT(lexp != NULL);
2796 fed = &lexp->exp_filter_data;
2798 rc = filter_connect_internal(lexp, data, 0);
2802 filter_export_stats_init(obd, lexp, localdata);
2803 if (obd->obd_replayable) {
2804 OBD_ALLOC(lcd, sizeof(*lcd));
2806 CERROR("filter: out of memory for client data\n");
2807 GOTO(cleanup, rc = -ENOMEM);
2810 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2812 rc = filter_client_add(obd, lexp, -1);
2817 group = data->ocd_group;
2819 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2820 obd->obd_name, lexp->exp_handle.h_cookie, group);
2822 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2823 rc = filter_read_groups(obd, group, 1);
2824 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2826 CERROR("can't read group %u\n", group);
2834 class_disconnect(lexp);
2835 lprocfs_exp_cleanup(lexp);
2844 /* Do extra sanity checks for grant accounting. We do this at connect,
2845 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2846 * always get rid of it or turn it off when we know accounting is good. */
2847 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2849 struct filter_export_data *fed;
2850 struct obd_export *exp;
2851 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2852 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2853 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2855 if (cfs_list_empty(&obd->obd_exports))
2858 /* We don't want to do this for large machines that do lots of
2859 mounts or unmounts. It burns... */
2860 if (obd->obd_num_exports > 100)
2863 cfs_spin_lock(&obd->obd_osfs_lock);
2864 cfs_spin_lock(&obd->obd_dev_lock);
2865 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2867 fed = &exp->exp_filter_data;
2868 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2871 if (maxsize > 0) { /* we may not have done a statfs yet */
2872 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2873 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2874 exp->exp_client_uuid.uuid, exp,
2875 fed->fed_grant, fed->fed_pending, maxsize);
2876 LASSERTF(fed->fed_dirty <= maxsize,
2877 "%s: cli %s/%p %ld > "LPU64"\n", func,
2878 exp->exp_client_uuid.uuid, exp,
2879 fed->fed_dirty, maxsize);
2882 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2883 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2884 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2886 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2887 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2888 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2889 tot_granted += fed->fed_grant + fed->fed_pending;
2890 tot_pending += fed->fed_pending;
2891 tot_dirty += fed->fed_dirty;
2893 fo_tot_granted = obd->u.filter.fo_tot_granted;
2894 fo_tot_pending = obd->u.filter.fo_tot_pending;
2895 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2896 cfs_spin_unlock(&obd->obd_dev_lock);
2897 cfs_spin_unlock(&obd->obd_osfs_lock);
2899 /* Do these assertions outside the spinlocks so we don't kill system */
2900 if (tot_granted != fo_tot_granted)
2901 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2902 func, tot_granted, fo_tot_granted);
2903 if (tot_pending != fo_tot_pending)
2904 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2905 func, tot_pending, fo_tot_pending);
2906 if (tot_dirty != fo_tot_dirty)
2907 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2908 func, tot_dirty, fo_tot_dirty);
2909 if (tot_pending > tot_granted)
2910 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2911 func, tot_pending, tot_granted);
2912 if (tot_granted > maxsize)
2913 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2914 func, tot_granted, maxsize);
2915 if (tot_dirty > maxsize)
2916 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2917 func, tot_dirty, maxsize);
2920 /* Remove this client from the grant accounting totals. We also remove
2921 * the export from the obd device under the osfs and dev locks to ensure
2922 * that the filter_grant_sanity_check() calculations are always valid.
2923 * The client should do something similar when it invalidates its import. */
2924 static void filter_grant_discard(struct obd_export *exp)
2926 struct obd_device *obd = exp->exp_obd;
2927 struct filter_obd *filter = &obd->u.filter;
2928 struct filter_export_data *fed = &exp->exp_filter_data;
2930 cfs_spin_lock(&obd->obd_osfs_lock);
2931 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2932 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2933 obd->obd_name, filter->fo_tot_granted,
2934 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2935 filter->fo_tot_granted -= fed->fed_grant;
2936 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2937 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2938 obd->obd_name, filter->fo_tot_pending,
2939 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2940 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2941 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2942 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2943 obd->obd_name, filter->fo_tot_dirty,
2944 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2945 filter->fo_tot_dirty -= fed->fed_dirty;
2949 cfs_spin_unlock(&obd->obd_osfs_lock);
2952 static int filter_destroy_export(struct obd_export *exp)
2954 struct filter_export_data *fed = &exp->exp_filter_data;
2957 if (fed->fed_pending)
2958 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2959 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2960 exp, fed->fed_pending);
2962 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2964 target_destroy_export(exp);
2965 ldlm_destroy_export(exp);
2967 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2970 lut_client_free(exp);
2972 if (!exp->exp_obd->obd_replayable)
2973 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2975 filter_grant_discard(exp);
2976 filter_fmd_cleanup(exp);
2978 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2979 struct filter_obd *filter = &exp->exp_obd->u.filter;
2980 if (filter->fo_tot_granted_clients > 0)
2981 filter->fo_tot_granted_clients --;
2984 if (!(exp->exp_flags & OBD_OPT_FORCE))
2985 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2990 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2992 struct obd_llog_group *olg_min, *olg;
2993 struct filter_obd *filter;
2994 int worked = -1, group;
2995 struct llog_ctxt *ctxt;
2998 filter = &obd->u.filter;
3000 /* we can't sync log holding spinlock. also, we do not want to get
3001 * into livelock. so we do following: loop over MDS's exports in
3002 * group order and skip already synced llogs -bzzz */
3004 /* look for group with min. number, but > worked */
3007 cfs_spin_lock(&filter->fo_llog_list_lock);
3008 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3009 if (olg->olg_group <= worked) {
3010 /* this group is already synced */
3013 if (group < olg->olg_group) {
3014 /* we have group with smaller number to sync */
3017 /* store current minimal group */
3019 group = olg->olg_group;
3021 cfs_spin_unlock(&filter->fo_llog_list_lock);
3023 if (olg_min == NULL)
3026 worked = olg_min->olg_group;
3027 if (olg_min->olg_exp &&
3028 (dexp == olg_min->olg_exp || dexp == NULL)) {
3030 ctxt = llog_group_get_ctxt(olg_min,
3031 LLOG_MDS_OST_REPL_CTXT);
3033 err = llog_sync(ctxt, olg_min->olg_exp);
3034 llog_ctxt_put(ctxt);
3036 CERROR("error flushing logs to MDS: "
3041 } while (olg_min != NULL);
3044 /* Also incredibly similar to mds_disconnect */
3045 static int filter_disconnect(struct obd_export *exp)
3047 struct obd_device *obd = exp->exp_obd;
3052 class_export_get(exp);
3054 if (!(exp->exp_flags & OBD_OPT_FORCE))
3055 filter_grant_sanity_check(obd, __FUNCTION__);
3056 filter_grant_discard(exp);
3058 /* Flush any remaining cancel messages out to the target */
3059 filter_sync_llogs(obd, exp);
3061 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3063 rc = server_disconnect_export(exp);
3065 /* Do not erase record for recoverable client. */
3066 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3067 filter_client_del(exp);
3069 fsfilt_sync(obd, obd->u.obt.obt_sb);
3071 class_export_put(exp);
3075 /* reverse import is changed, sync all cancels */
3076 static void filter_revimp_update(struct obd_export *exp)
3081 class_export_get(exp);
3083 /* flush any remaining cancel messages out to the target */
3084 filter_sync_llogs(exp->exp_obd, exp);
3085 class_export_put(exp);
3089 static int filter_ping(struct obd_export *exp)
3091 filter_fmd_expire(exp);
3095 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3096 const char *what, int quiet)
3098 struct dentry *dchild = NULL;
3101 if (oa->o_valid & OBD_MD_FLGROUP)
3104 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3106 if (IS_ERR(dchild)) {
3107 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3108 what, group, oa->o_id);
3112 if (dchild->d_inode == NULL) {
3114 CERROR("%s: %s on non-existent object: "LPU64"\n",
3115 obd->obd_name, what, oa->o_id);
3117 RETURN(ERR_PTR(-ENOENT));
3123 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3125 struct dentry *dentry = NULL;
3126 struct obd_device *obd;
3130 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3131 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3132 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3136 obd = class_exp2obd(exp);
3138 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3142 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3144 RETURN(PTR_ERR(dentry));
3146 /* Limit the valid bits in the return data to what we actually use */
3147 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3148 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3154 /* this should be enabled/disabled in condition to enabled/disabled large
3155 * inodes (fast EAs) in backing store FS. */
3156 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3157 void *handle, struct obdo *oa)
3159 struct obd_device *obd = exp->exp_obd;
3163 if (oa->o_valid & OBD_MD_FLFID) {
3164 struct filter_fid ff;
3166 if (!(oa->o_valid & OBD_MD_FLGROUP))
3168 /* packing fid and converting it to LE for storing into EA.
3169 * Here ->o_stripe_idx should be filled by LOV and rest of
3170 * fields - by client. */
3171 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3172 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3173 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3174 ff.ff_objid = cpu_to_le64(oa->o_id);
3175 ff.ff_group = cpu_to_le64(oa->o_gr);
3177 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3178 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3179 oa->o_generation, oa->o_id, oa->o_gr);
3181 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3183 CERROR("store fid in object failed! rc: %d\n", rc);
3185 CDEBUG(D_HA, "OSS object without fid info!\n");
3191 /* this is called from filter_truncate() until we have filter_punch() */
3192 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3193 struct obdo *oa, struct obd_trans_info *oti)
3195 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3196 struct llog_cookie *fcc = NULL;
3197 struct filter_obd *filter;
3198 int rc, err, sync = 0;
3199 loff_t old_size = 0;
3200 unsigned int ia_valid;
3201 struct inode *inode;
3202 struct page *page = NULL;
3207 LASSERT(dentry != NULL);
3208 LASSERT(!IS_ERR(dentry));
3210 inode = dentry->d_inode;
3211 LASSERT(inode != NULL);
3213 filter = &exp->exp_obd->u.filter;
3214 iattr_from_obdo(&iattr, oa, oa->o_valid);
3215 ia_valid = iattr.ia_valid;
3217 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3218 OBD_ALLOC(fcc, sizeof(*fcc));
3220 *fcc = oa->o_lcookie;
3222 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3224 /* Filter truncates and writes are serialized by
3225 * i_alloc_sem, see the comment in
3226 * filter_preprw_write.*/
3227 if (ia_valid & ATTR_SIZE)
3228 down_write(&inode->i_alloc_sem);
3229 LOCK_INODE_MUTEX(inode);
3230 old_size = i_size_read(inode);
3233 /* VBR: version recovery check */
3234 rc = filter_version_get_check(exp, oti, inode);
3236 GOTO(out_unlock, rc);
3238 /* Let's pin the last page so that ldiskfs_truncate
3239 * should not start GFP_FS allocation. */
3240 if (ia_valid & ATTR_SIZE) {
3241 page = grab_cache_page(inode->i_mapping,
3242 iattr.ia_size >> PAGE_CACHE_SHIFT);
3244 GOTO(out_unlock, rc = -ENOMEM);
3249 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3250 * then we will accept the UID+GID sent by the client during write for
3251 * initializing the ownership of this inode. We only allow this to
3252 * happen once so clear these bits in setattr. In 2.6 kernels it is
3253 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3254 * the flags that are actually being set. */
3255 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3256 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3257 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3259 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3260 if (!(ia_valid & ATTR_MODE)) {
3261 iattr.ia_mode = inode->i_mode;
3262 iattr.ia_valid |= ATTR_MODE;
3264 iattr.ia_mode &= ~S_ISUID;
3266 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3267 if (!(iattr.ia_valid & ATTR_MODE)) {
3268 iattr.ia_mode = inode->i_mode;
3269 iattr.ia_valid |= ATTR_MODE;
3271 iattr.ia_mode &= ~S_ISGID;
3274 orig_ids[USRQUOTA] = inode->i_uid;
3275 orig_ids[GRPQUOTA] = inode->i_gid;
3276 handle = fsfilt_start_log(exp->exp_obd, inode,
3277 FSFILT_OP_SETATTR, oti, 1);
3279 GOTO(out_unlock, rc = PTR_ERR(handle));
3281 /* update inode EA only once when inode is suid bit marked. As
3282 * on 2.6.x UID and GID may be set separately, we check here
3283 * only one of them to avoid double setting. */
3284 if (inode->i_mode & S_ISUID)
3285 filter_update_fidea(exp, inode, handle, oa);
3287 handle = fsfilt_start(exp->exp_obd, inode,
3288 FSFILT_OP_SETATTR, oti);
3290 GOTO(out_unlock, rc = PTR_ERR(handle));
3292 if (oa->o_valid & OBD_MD_FLFLAGS) {
3293 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3294 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3296 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3298 /* set cancel cookie callback function */
3299 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3300 filter_cancel_cookies_cb,
3304 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3305 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3307 /* The truncate might have used up our transaction credits. Make
3308 * sure we have one left for the last_rcvd update. */
3309 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3311 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3313 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3317 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3319 CERROR("error on commit, err = %d\n", err);
3326 /* For a partial-page truncate flush the page to disk immediately
3327 * to avoid data corruption during direct disk write. b=17397 */
3328 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3329 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3330 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3340 page_cache_release(page);
3342 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3343 UNLOCK_INODE_MUTEX(inode);
3344 if (ia_valid & ATTR_SIZE)
3345 up_write(&inode->i_alloc_sem);
3347 OBD_FREE(fcc, sizeof(*fcc));
3349 /* trigger quota release */
3350 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3351 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3352 int rc2 = lquota_adjust(filter_quota_interface_ref,
3353 exp->exp_obd, cur_ids,
3354 orig_ids, rc, FSFILT_OP_SETATTR);
3355 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3356 "filter adjust qunit. (rc:%d)\n", rc2);
3361 /* this is called from filter_truncate() until we have filter_punch() */
3362 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3363 struct obd_trans_info *oti)
3365 struct obdo *oa = oinfo->oi_oa;
3366 struct lustre_capa *capa = oinfo_capa(oinfo);
3367 struct ldlm_res_id res_id;
3368 struct filter_mod_data *fmd;
3369 struct lvfs_run_ctxt saved;
3370 struct filter_obd *filter;
3371 struct ldlm_resource *res;
3372 struct dentry *dentry;
3373 __u64 opc = CAPA_OPC_META_WRITE;
3377 if (oa->o_valid & OBD_FL_TRUNC)
3378 opc |= CAPA_OPC_OSS_TRUNC;
3380 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3381 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3385 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3386 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3391 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3392 /* This would be very bad - accidentally truncating a file when
3393 * changing the time or similar - bug 12203. */
3394 if (oa->o_valid & OBD_MD_FLSIZE &&
3395 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3396 static char mdsinum[48];
3398 if (oa->o_valid & OBD_MD_FLFID)
3399 snprintf(mdsinum, sizeof(mdsinum) - 1,
3400 " of inode "LPU64"/%u", oa->o_fid,
3405 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3407 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3412 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3414 RETURN(PTR_ERR(dentry));
3416 filter = &exp->exp_obd->u.filter;
3417 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3420 * We need to be atomic against a concurrent write
3421 * (which takes the semaphore for reading). fmd_mactime_xid
3422 * checks will have no effect if a write request with lower
3423 * xid starts just before a setattr and finishes later than
3424 * the setattr (see bug 21489, comment 27).
3427 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3428 down_write(&dentry->d_inode->i_alloc_sem);
3429 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3430 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3431 fmd->fmd_mactime_xid = oti->oti_xid;
3432 filter_fmd_put(exp, fmd);
3433 up_write(&dentry->d_inode->i_alloc_sem);
3436 /* setting objects attributes (including owner/group) */
3437 rc = filter_setattr_internal(exp, dentry, oa, oti);
3439 GOTO(out_unlock, rc);
3441 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3442 &res_id, LDLM_EXTENT, 0);
3445 LDLM_RESOURCE_ADDREF(res);
3446 rc = ldlm_res_lvbo_update(res, NULL, 0);
3447 LDLM_RESOURCE_DELREF(res);
3448 ldlm_resource_putref(res);
3451 oa->o_valid = OBD_MD_FLID;
3453 /* Quota release need uid/gid info */
3454 obdo_from_inode(oa, dentry->d_inode,
3455 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3460 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3464 /* XXX identical to osc_unpackmd */
3465 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3466 struct lov_mds_md *lmm, int lmm_bytes)
3472 if (lmm_bytes < sizeof (*lmm)) {
3473 CERROR("lov_mds_md too small: %d, need %d\n",
3474 lmm_bytes, (int)sizeof(*lmm));
3477 /* XXX LOV_MAGIC etc check? */
3479 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3480 CERROR("lov_mds_md: zero lmm_object_id\n");
3485 lsm_size = lov_stripe_md_size(1);
3489 if (*lsmp != NULL && lmm == NULL) {
3490 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3491 OBD_FREE(*lsmp, lsm_size);
3496 if (*lsmp == NULL) {
3497 OBD_ALLOC(*lsmp, lsm_size);
3501 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3502 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3503 OBD_FREE(*lsmp, lsm_size);
3506 loi_init((*lsmp)->lsm_oinfo[0]);
3510 /* XXX zero *lsmp? */
3511 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3512 LASSERT((*lsmp)->lsm_object_id);
3515 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3520 /* caller must hold fo_create_locks[oa->o_gr] */
3521 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3522 struct filter_obd *filter)
3524 struct obdo doa = { 0 }; /* XXX obdo on stack */
3531 LASSERT_MDS_GROUP(oa->o_gr);
3532 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3533 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3535 memset(&doa, 0, sizeof(doa));
3537 doa.o_valid |= OBD_MD_FLGROUP;
3538 doa.o_gr = oa->o_gr;
3539 doa.o_mode = S_IFREG;
3541 if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3542 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3543 exp->exp_obd->obd_name, doa.o_gr);
3547 last = filter_last_id(filter, doa.o_gr);
3549 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3551 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3552 exp->exp_obd->obd_name, oa->o_id + 1, last,
3553 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3555 for (id = last; id > oa->o_id; id--) {
3557 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3558 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3559 CEMERG("error destroying precreate objid "LPU64": %d\n",
3562 /* update last_id on disk periodically so that if we restart
3563 * we don't need to re-scan all of the just-deleted objects. */
3564 if ((id & 511) == 0 && !skip_orphan) {
3565 filter_set_last_id(filter, id - 1, doa.o_gr);
3566 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3570 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3571 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3574 filter_set_last_id(filter, id, doa.o_gr);
3575 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3577 /* don't reuse orphan object, return last used objid */
3581 cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3586 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3587 obd_gr group, int *num);
3588 /* returns a negative error or a nonnegative number of files to create */
3589 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3590 obd_gr group, struct obd_trans_info *oti)
3592 struct obd_device *obd = exp->exp_obd;
3593 struct filter_obd *filter = &obd->u.filter;
3597 /* delete orphans request */
3598 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3599 obd_id last = filter_last_id(filter, group);
3601 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3602 CERROR("%s: dropping old orphan cleanup request\n",
3606 /* This causes inflight precreates to abort and drop lock */
3607 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3608 cfs_down(&filter->fo_create_locks[group]);
3609 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3610 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3611 exp->exp_obd->obd_name, group);
3612 cfs_up(&filter->fo_create_locks[group]);
3615 diff = oa->o_id - last;
3616 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3619 if (-diff > OST_MAX_PRECREATE) {
3620 CERROR("%s: ignoring bogus orphan destroy request: "
3621 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3623 /* FIXME: should reset precreate_next_id on MDS */
3624 GOTO(out, rc = -EINVAL);
3627 rc = filter_destroy_precreated(exp, oa, filter);
3629 CERROR("%s: unable to write lastobjid, but "
3630 "orphans were deleted\n", obd->obd_name);
3633 /* XXX: Used by MDS for the first time! */
3634 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3637 cfs_down(&filter->fo_create_locks[group]);
3638 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3639 CERROR("%s: dropping old precreate request\n",
3643 /* only precreate if group == 0 and o_id is specfied */
3644 if (!filter_group_is_mds(group) || oa->o_id == 0)
3647 diff = oa->o_id - filter_last_id(filter, group);
3648 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3649 filter_last_id(filter, group), diff);
3651 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3652 oa->o_id, filter_last_id(filter, group), diff);
3656 oa->o_id = filter_last_id(&obd->u.filter, group);
3657 rc = filter_precreate(obd, oa, group, &diff);
3658 oa->o_id = filter_last_id(&obd->u.filter, group);
3660 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3663 /* else diff == 0 */
3666 cfs_up(&filter->fo_create_locks[group]);
3670 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3671 __u64 max_age, __u32 flags)
3673 struct filter_obd *filter = &obd->u.filter;
3674 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3678 /* at least try to account for cached pages. its still racey and
3679 * might be under-reporting if clients haven't announced their
3680 * caches with brw recently */
3681 cfs_spin_lock(&obd->obd_osfs_lock);
3682 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3683 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3684 cfs_spin_unlock(&obd->obd_osfs_lock);
3686 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3687 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3688 filter->fo_tot_dirty, filter->fo_tot_granted,
3689 filter->fo_tot_pending,
3690 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3692 filter_grant_sanity_check(obd, __FUNCTION__);
3694 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3695 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3696 osfs->os_bsize - 1) >> blockbits));
3698 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3699 struct lr_server_data *lsd = class_server_data(obd);
3700 int index = le32_to_cpu(lsd->lsd_ost_index);
3702 if (obd_fail_val == -1 ||
3703 index == obd_fail_val)
3704 osfs->os_bfree = osfs->os_bavail = 2;
3705 else if (obd_fail_loc & OBD_FAIL_ONCE)
3706 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3709 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3710 * stop creating files on MDS if OST is not good shape to create
3714 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3715 osfs->os_state = OS_STATE_READONLY;
3717 if (filter->fo_raid_degraded)
3718 osfs->os_state |= OS_STATE_DEGRADED;
3722 static int filter_use_existing_obj(struct obd_device *obd,
3723 struct dentry *dchild, void **handle,
3726 struct inode *inode = dchild->d_inode;
3730 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3733 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3734 if (IS_ERR(*handle))
3735 return PTR_ERR(*handle);
3737 iattr.ia_valid = ATTR_MODE;
3738 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3739 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3746 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3749 __u64 os_ffree = -1;
3751 cfs_spin_lock(&obd->obd_osfs_lock);
3752 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3754 os_ffree = obd->obd_osfs.os_ffree;
3755 cfs_spin_unlock(&obd->obd_osfs_lock);
3760 /* We rely on the fact that only one thread will be creating files in a given
3761 * group at a time, which is why we don't need an atomic filter_get_new_id.
3762 * Even if we had that atomic function, the following race would exist:
3764 * thread 1: gets id x from filter_next_id
3765 * thread 2: gets id (x + 1) from filter_next_id
3766 * thread 2: creates object (x + 1)
3767 * thread 1: tries to create object x, gets -ENOSPC
3769 * Caller must hold fo_create_locks[group]
3771 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3772 obd_gr group, int *num)
3774 struct dentry *dchild = NULL, *dparent = NULL;
3775 struct filter_obd *filter;
3776 struct obd_statfs *osfs;
3777 int err = 0, rc = 0, recreate_obj = 0, i;
3778 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3781 void *handle = NULL;
3784 filter = &obd->u.filter;
3786 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3788 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3790 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3791 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3794 OBD_ALLOC(osfs, sizeof(*osfs));
3797 rc = filter_statfs(obd, osfs, cfs_time_current_64() - CFS_HZ,
3799 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3800 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3801 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3802 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3806 OBD_FREE(osfs, sizeof(*osfs));
3811 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3812 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3814 for (i = 0; i < *num && err == 0; i++) {
3815 int cleanup_phase = 0;
3817 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3818 CWARN("%s: create aborted by destroy\n",
3827 last_id = filter_last_id(filter, group);
3828 if (next_id > last_id) {
3829 CERROR("Error: Trying to recreate obj greater"
3830 "than last id "LPD64" > "LPD64"\n",
3832 GOTO(cleanup, rc = -EINVAL);
3835 next_id = filter_last_id(filter, group) + 1;
3837 dparent = filter_parent_lock(obd, group, next_id);
3838 if (IS_ERR(dparent))
3839 GOTO(cleanup, rc = PTR_ERR(dparent));
3840 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3842 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3844 GOTO(cleanup, rc = PTR_ERR(dchild));
3845 cleanup_phase = 2; /* f_dput(dchild) */
3847 if (dchild->d_inode != NULL) {
3848 /* This would only happen if lastobjid was bad on disk*/
3849 /* Could also happen if recreating missing obj but it
3850 * already exists. */
3852 CERROR("%s: recreating existing object %.*s?\n",
3853 obd->obd_name, dchild->d_name.len,
3854 dchild->d_name.name);
3856 /* Use these existing objects if they are
3858 if (dchild->d_inode->i_size == 0) {
3859 rc = filter_use_existing_obj(obd,dchild,
3860 &handle, &cleanup_phase);
3867 CERROR("%s: Serious error: objid %.*s already "
3868 "exists; is this filesystem corrupt?\n",
3869 obd->obd_name, dchild->d_name.len,
3870 dchild->d_name.name);
3873 GOTO(cleanup, rc = -EEXIST);
3876 handle = fsfilt_start_log(obd, dparent->d_inode,
3877 FSFILT_OP_CREATE, NULL, 1);
3879 GOTO(cleanup, rc = PTR_ERR(handle));
3882 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3883 ",od->o_id="LPU64")\n", obd->obd_name, group,
3886 /* We mark object SUID+SGID to flag it for accepting UID+GID
3887 * from client on first write. Currently the permission bits
3888 * on the OST are never used, so this is OK. */
3889 rc = ll_vfs_create(dparent->d_inode, dchild,
3890 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3892 CERROR("create failed rc = %d\n", rc);
3893 if (rc == -ENOSPC) {
3894 os_ffree = filter_calc_free_inodes(obd);
3896 CERROR("%s: free inode "LPU64"\n",
3897 obd->obd_name, os_ffree);
3902 if (dchild->d_inode)
3903 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3904 dchild->d_inode->i_ino);
3907 if (!recreate_obj) {
3908 filter_set_last_id(filter, next_id, group);
3909 err = filter_update_last_objid(obd, group, 0);
3911 CERROR("unable to write lastobjid "
3912 "but file created\n");
3916 switch(cleanup_phase) {
3918 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3920 CERROR("error on commit, err = %d\n", err);
3927 filter_parent_unlock(dparent);
3934 if (cfs_time_after(jiffies, enough_time)) {
3936 "%s: precreate slow - want %d got %d \n",
3937 obd->obd_name, *num, i);
3944 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3945 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3950 static int filter_create(struct obd_export *exp, struct obdo *oa,
3951 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3953 struct obd_device *obd = exp->exp_obd;
3954 struct filter_export_data *fed;
3955 struct filter_obd *filter;
3956 struct lvfs_run_ctxt saved;
3957 struct lov_stripe_md *lsm = NULL;
3958 int rc = 0, diff, group = oa->o_gr;
3961 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3962 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3964 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3965 CERROR("!!! nid %s sent invalid object group %d\n",
3966 obd_export_nid2str(exp), group);
3970 fed = &exp->exp_filter_data;
3971 filter = &obd->u.filter;
3973 if (fed->fed_group != group) {
3974 CERROR("!!! this export (nid %s) used object group %d "
3975 "earlier; now it's trying to use group %d! This could "
3976 "be a bug in the MDS. Please report to "
3977 "http://bugzilla.lustre.org/\n",
3978 obd_export_nid2str(exp), fed->fed_group, group);
3985 rc = obd_alloc_memmd(exp, &lsm);
3992 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3994 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3995 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3996 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3997 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3998 oa->o_id, filter_last_id(filter,
4003 cfs_down(&filter->fo_create_locks[oa->o_gr]);
4004 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4005 cfs_up(&filter->fo_create_locks[oa->o_gr]);
4008 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4011 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4012 if (rc && ea != NULL && *ea != lsm) {
4013 obd_free_memmd(exp, &lsm);
4014 } else if (rc == 0 && ea != NULL) {
4015 /* XXX LOV STACKING: the lsm that is passed to us from
4016 * LOV does not have valid lsm_oinfo data structs, so
4017 * don't go touching that. This needs to be fixed in a
4019 lsm->lsm_object_id = oa->o_id;
4026 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4027 struct lov_stripe_md *md, struct obd_trans_info *oti,
4028 struct obd_export *md_exp, void *capa)
4030 unsigned int qcids[MAXQUOTAS] = {0, 0};
4031 struct obd_device *obd;
4032 struct filter_obd *filter;
4033 struct dentry *dchild = NULL, *dparent = NULL;
4034 struct lustre_handle lockh = { 0 };
4035 struct lvfs_run_ctxt saved;
4036 void *handle = NULL;
4037 struct llog_cookie *fcc = NULL;
4038 int rc, rc2, cleanup_phase = 0, sync = 0;
4042 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4043 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4044 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4049 filter = &obd->u.filter;
4051 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4054 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4055 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4057 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4059 GOTO(cleanup, rc = PTR_ERR(dchild));
4062 if (dchild->d_inode == NULL) {
4063 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4065 /* If object already gone, cancel cookie right now */
4066 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4067 struct llog_ctxt *ctxt;
4068 struct obd_llog_group *olg;
4070 olg = filter_find_olg(obd, oa->o_gr);
4072 CERROR(" %s: can not find olg of group %d\n",
4073 obd->obd_name, (int)oa->o_gr);
4074 GOTO(cleanup, rc = PTR_ERR(olg));
4076 fcc = &oa->o_lcookie;
4077 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4078 llog_cancel(ctxt, NULL, 1, fcc, 0);
4079 llog_ctxt_put(ctxt);
4080 fcc = NULL; /* we didn't allocate fcc, don't free it */
4082 GOTO(cleanup, rc = -ENOENT);
4085 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4089 /* Our MDC connection is established by the MDS to us */
4090 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4091 OBD_ALLOC(fcc, sizeof(*fcc));
4093 *fcc = oa->o_lcookie;
4095 DQUOT_INIT(dchild->d_inode);
4097 /* we're gonna truncate it first in order to avoid possible deadlock:
4099 * open trasaction open transaction
4100 * down(i_zombie) down(i_zombie)
4101 * restart transaction
4102 * (see BUG 4180) -bzzz
4104 * take i_alloc_sem too to prevent other threads from writing to the
4105 * file while we are truncating it. This can cause lock ordering issue
4106 * between page lock, i_mutex & starting new journal handle.
4107 * (see bug 20321) -johann
4109 down_write(&dchild->d_inode->i_alloc_sem);
4110 LOCK_INODE_MUTEX(dchild->d_inode);
4112 /* VBR: version recovery check */
4113 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4115 UNLOCK_INODE_MUTEX(dchild->d_inode);
4116 up_write(&dchild->d_inode->i_alloc_sem);
4120 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4122 if (IS_ERR(handle)) {
4123 UNLOCK_INODE_MUTEX(dchild->d_inode);
4124 up_write(&dchild->d_inode->i_alloc_sem);
4125 GOTO(cleanup, rc = PTR_ERR(handle));
4128 iattr.ia_valid = ATTR_SIZE;
4130 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4131 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4132 UNLOCK_INODE_MUTEX(dchild->d_inode);
4133 up_write(&dchild->d_inode->i_alloc_sem);
4137 GOTO(cleanup, rc = rc2);
4139 /* We don't actually need to lock the parent until we are unlinking
4140 * here, and not while truncating above. That avoids holding the
4141 * parent lock for a long time during truncate, which can block other
4142 * threads from doing anything to objects in that directory. bug 7171 */
4143 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4144 if (IS_ERR(dparent))
4145 GOTO(cleanup, rc = PTR_ERR(dparent));
4146 cleanup_phase = 3; /* filter_parent_unlock */
4148 LOCK_INODE_MUTEX(dchild->d_inode);
4149 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4150 if (IS_ERR(handle)) {
4151 UNLOCK_INODE_MUTEX(dchild->d_inode);
4152 GOTO(cleanup, rc = PTR_ERR(handle));
4154 cleanup_phase = 4; /* fsfilt_commit */
4156 /* Quota release need uid/gid of inode */
4157 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4159 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4161 /* this drops dchild->d_inode->i_mutex unconditionally */
4162 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4166 switch(cleanup_phase) {
4169 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4170 oti->oti_handle : handle,
4171 filter_cancel_cookies_cb,
4173 /* If add_journal_cb failed, then filter_finish_transno
4174 * will commit the handle and we will do a sync
4175 * on commit. then we call callback directly to free
4178 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4180 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4183 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4185 CERROR("error on commit, err = %d\n", rc2);
4192 filter_parent_unlock(dparent);
4194 filter_fini_destroy(obd, &lockh);
4198 OBD_FREE(fcc, sizeof(*fcc));
4200 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4203 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4207 /* trigger quota release */
4208 qcids[USRQUOTA] = oa->o_uid;
4209 qcids[GRPQUOTA] = oa->o_gid;
4210 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4213 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4217 /* NB start and end are used for punch, but not truncate */
4218 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4219 struct obd_trans_info *oti,
4220 struct ptlrpc_request_set *rqset)
4225 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4226 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4227 oinfo->oi_policy.l_extent.end);
4231 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4232 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4233 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4235 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4236 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4237 rc = filter_setattr(exp, oinfo, oti);
4238 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4242 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4243 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4246 struct lvfs_run_ctxt saved;
4247 struct filter_obd *filter;
4248 struct dentry *dentry;
4252 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4253 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4254 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4258 filter = &exp->exp_obd->u.filter;
4260 /* An objid of zero is taken to mean "sync whole filesystem" */
4261 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4262 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4263 /* Flush any remaining cancel messages out to the target */
4264 filter_sync_llogs(exp->exp_obd, exp);
4268 dentry = filter_oa2dentry(exp->exp_obd, oa);
4270 RETURN(PTR_ERR(dentry));
4272 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4274 LOCK_INODE_MUTEX(dentry->d_inode);
4276 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4278 /* just any file to grab fsync method - "file" arg unused */
4279 struct file *file = filter->fo_rcvd_filp;
4281 if (file->f_op && file->f_op->fsync)
4282 rc = file->f_op->fsync(NULL, dentry, 1);
4284 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4288 UNLOCK_INODE_MUTEX(dentry->d_inode);
4290 oa->o_valid = OBD_MD_FLID;
4291 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4293 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4299 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4300 void *key, __u32 *vallen, void *val,
4301 struct lov_stripe_md *lsm)
4303 struct obd_device *obd;
4306 obd = class_exp2obd(exp);
4308 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4312 if (KEY_IS(KEY_BLOCKSIZE)) {
4313 __u32 *blocksize = val;
4315 if (*vallen < sizeof(*blocksize))
4317 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4319 *vallen = sizeof(*blocksize);
4323 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4324 __u32 *blocksize_bits = val;
4325 if (blocksize_bits) {
4326 if (*vallen < sizeof(*blocksize_bits))
4328 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4330 *vallen = sizeof(*blocksize_bits);
4334 if (KEY_IS(KEY_LAST_ID)) {
4335 obd_id *last_id = val;
4336 /* FIXME: object groups */
4338 if (*vallen < sizeof(*last_id))
4340 *last_id = filter_last_id(&obd->u.filter,
4341 exp->exp_filter_data.fed_group);
4343 *vallen = sizeof(*last_id);
4347 if (KEY_IS(KEY_FIEMAP)) {
4348 struct ll_fiemap_info_key *fm_key = key;
4349 struct dentry *dentry;
4350 struct ll_user_fiemap *fiemap = val;
4351 struct lvfs_run_ctxt saved;
4354 if (fiemap == NULL) {
4355 *vallen = fiemap_count_to_size(
4356 fm_key->fiemap.fm_extent_count);
4360 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4363 RETURN(PTR_ERR(dentry));
4365 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4366 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4367 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4369 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4375 CDEBUG(D_IOCTL, "invalid key\n");
4379 static inline int filter_setup_llog_group(struct obd_export *exp,
4380 struct obd_device *obd,
4383 struct obd_llog_group *olg;
4384 struct llog_ctxt *ctxt;
4387 olg = filter_find_create_olg(obd, group);
4389 RETURN(PTR_ERR(olg));
4391 llog_group_set_export(olg, exp);
4393 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4394 LASSERTF(ctxt != NULL, "ctxt is null\n");
4396 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4397 llog_ctxt_put(ctxt);
4401 static int filter_set_grant_shrink(struct obd_export *exp,
4402 struct ost_body *body)
4404 /* handle shrink grant */
4405 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4406 filter_grant_incoming(exp, &body->oa);
4407 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4413 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4415 struct obd_device *obd;
4421 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4425 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4426 obd_export_nid2str(exp));
4427 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4429 /* setup llog imports */
4431 group = (int)(*(__u32 *)val);
4433 group = 0; /* default value */
4435 LASSERT_MDS_GROUP(group);
4436 rc = filter_setup_llog_group(exp, obd, group);
4440 if (group == FILTER_GROUP_MDS0) {
4441 /* setup llog group 1 for interop */
4442 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4445 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4450 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4451 void *key, __u32 vallen, void *val,
4452 struct ptlrpc_request_set *set)
4454 struct obd_device *obd;
4459 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4463 if (KEY_IS(KEY_CAPA_KEY)) {
4465 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4467 CERROR("filter update capability key failed: %d\n", rc);
4471 if (KEY_IS(KEY_REVIMP_UPD)) {
4472 filter_revimp_update(exp);
4473 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4477 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4478 filter_adapt_sptlrpc_conf(obd, 0);
4482 if (KEY_IS(KEY_MDS_CONN))
4483 RETURN(filter_set_mds_conn(exp, val));
4485 if (KEY_IS(KEY_GRANT_SHRINK))
4486 RETURN(filter_set_grant_shrink(exp, val));
4491 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4492 int len, void *karg, void *uarg)
4494 struct obd_device *obd = exp->exp_obd;
4495 struct obd_ioctl_data *data = karg;
4499 case OBD_IOC_ABORT_RECOVERY: {
4500 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4501 target_stop_recovery_thread(obd);
4505 case OBD_IOC_SYNC: {
4506 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4507 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4511 case OBD_IOC_SET_READONLY: {
4513 struct super_block *sb = obd->u.obt.obt_sb;
4514 struct inode *inode = sb->s_root->d_inode;
4515 BDEVNAME_DECLARE_STORAGE(tmp);
4516 CERROR("*** setting device %s read-only ***\n",
4517 ll_bdevname(sb, tmp));
4519 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4520 if (!IS_ERR(handle))
4521 rc = fsfilt_commit(obd, inode, handle, 1);
4523 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4524 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4526 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4530 case OBD_IOC_CATLOGLIST: {
4531 rc = llog_catalog_list(obd, 1, data);
4535 case OBD_IOC_LLOG_CANCEL:
4536 case OBD_IOC_LLOG_REMOVE:
4537 case OBD_IOC_LLOG_INFO:
4538 case OBD_IOC_LLOG_PRINT: {
4539 /* FIXME to be finished */
4540 RETURN(-EOPNOTSUPP);
4542 struct llog_ctxt *ctxt = NULL;
4544 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4545 rc = llog_ioctl(ctxt, cmd, data);
4546 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4559 static int filter_health_check(struct obd_device *obd)
4561 #ifdef USE_HEALTH_CHECK_WRITE
4562 struct filter_obd *filter = &obd->u.filter;
4567 * health_check to return 0 on healthy
4568 * and 1 on unhealthy.
4570 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4573 #ifdef USE_HEALTH_CHECK_WRITE
4574 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4575 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4580 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4583 return filter_fid2dentry(data, NULL, gr, id);
4586 static int filter_process_config(struct obd_device *obd, obd_count len,
4589 struct lustre_cfg *lcfg = buf;
4590 struct lprocfs_static_vars lvars;
4593 switch (lcfg->lcfg_command) {
4595 lprocfs_filter_init_vars(&lvars);
4597 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4607 static struct lvfs_callback_ops filter_lvfs_ops = {
4608 l_fid2dentry: filter_lvfs_fid2dentry,
4611 static struct obd_ops filter_obd_ops = {
4612 .o_owner = THIS_MODULE,
4613 .o_get_info = filter_get_info,
4614 .o_set_info_async = filter_set_info_async,
4615 .o_setup = filter_setup,
4616 .o_precleanup = filter_precleanup,
4617 .o_cleanup = filter_cleanup,
4618 .o_connect = filter_connect,
4619 .o_reconnect = filter_reconnect,
4620 .o_disconnect = filter_disconnect,
4621 .o_ping = filter_ping,
4622 .o_init_export = filter_init_export,
4623 .o_destroy_export = filter_destroy_export,
4624 .o_statfs = filter_statfs,
4625 .o_getattr = filter_getattr,
4626 .o_unpackmd = filter_unpackmd,
4627 .o_create = filter_create,
4628 .o_setattr = filter_setattr,
4629 .o_destroy = filter_destroy,
4630 .o_brw = filter_brw,
4631 .o_punch = filter_truncate,
4632 .o_sync = filter_sync,
4633 .o_preprw = filter_preprw,
4634 .o_commitrw = filter_commitrw,
4635 .o_llog_init = filter_llog_init,
4636 .o_llog_connect = filter_llog_connect,
4637 .o_llog_finish = filter_llog_finish,
4638 .o_iocontrol = filter_iocontrol,
4639 .o_health_check = filter_health_check,
4640 .o_process_config = filter_process_config,
4643 quota_interface_t *filter_quota_interface_ref;
4644 extern quota_interface_t filter_quota_interface;
4646 static int __init obdfilter_init(void)
4648 struct lprocfs_static_vars lvars;
4651 /** sanity check for group<->mdsno conversion */
4652 for (i = 0; i < 32; i++)
4653 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4655 lprocfs_filter_init_vars(&lvars);
4657 cfs_request_module("%s", "lquota");
4658 OBD_ALLOC(obdfilter_created_scratchpad,
4659 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4660 sizeof(*obdfilter_created_scratchpad));
4661 if (obdfilter_created_scratchpad == NULL)
4664 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4665 sizeof(struct filter_mod_data),
4668 GOTO(out, rc = -ENOMEM);
4670 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4671 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4673 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4674 LUSTRE_OST_NAME, NULL);
4678 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4679 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4680 ll_fmd_cachep = NULL;
4682 if (filter_quota_interface_ref)
4683 PORTAL_SYMBOL_PUT(filter_quota_interface);
4685 OBD_FREE(obdfilter_created_scratchpad,
4686 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4687 sizeof(*obdfilter_created_scratchpad));
4693 static void __exit obdfilter_exit(void)
4695 if (filter_quota_interface_ref)
4696 PORTAL_SYMBOL_PUT(filter_quota_interface);
4698 if (ll_fmd_cachep) {
4699 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4700 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4701 ll_fmd_cachep = NULL;
4704 class_unregister_type(LUSTRE_OST_NAME);
4705 OBD_FREE(obdfilter_created_scratchpad,
4706 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4707 sizeof(*obdfilter_created_scratchpad));
4710 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4711 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4712 MODULE_LICENSE("GPL");
4714 module_init(obdfilter_init);
4715 module_exit(obdfilter_exit);