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 obd_device_target *obt = &exp->exp_obd->u.obt;
124 struct tg_export_data *ted = &exp->exp_target_data;
125 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126 struct lsd_client_data *lcd;
129 int err, log_pri = D_RPCTRACE;
131 /* Propagate error code. */
135 if (!exp->exp_obd->obd_replayable || oti == NULL)
138 cfs_mutex_down(&ted->ted_lcd_lock);
140 /* if the export has already been disconnected, we have no last_rcvd slot,
141 * update server data with latest transno then */
143 cfs_mutex_up(&ted->ted_lcd_lock);
144 CWARN("commit transaction for disconnected client %s: rc %d\n",
145 exp->exp_client_uuid.uuid, rc);
146 err = filter_update_server_data(exp->exp_obd);
150 /* we don't allocate new transnos for replayed requests */
151 cfs_spin_lock(&obt->obt_lut->lut_translock);
152 if (oti->oti_transno == 0) {
153 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156 last_rcvd = oti->oti_transno;
157 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160 oti->oti_transno = last_rcvd;
162 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166 cfs_spin_unlock(&obt->obt_lut->lut_translock);
169 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
171 off = ted->ted_lr_off;
173 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174 ted->ted_lr_idx, ted->ted_lr_off);
177 class_export_cb_get(exp); /* released when the cb is called */
179 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
185 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186 lcd, sizeof(*lcd), &off,
187 force_sync | exp->exp_need_sync);
189 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
197 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199 cfs_mutex_up(&ted->ted_lcd_lock);
203 void f_dput(struct dentry *dentry)
205 /* Can't go inside filter_ddelete because it can block */
206 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208 LASSERT(atomic_read(&dentry->d_count) > 0);
213 static void init_brw_stats(struct brw_stats *brw_stats)
216 for (i = 0; i < BRW_LAST; i++)
217 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221 struct lprocfs_stats **stats)
225 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226 LPROC_FILTER_LAST - 1;
227 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
231 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241 plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243 struct obd_export *exp,
249 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250 /* Self-export gets no proc entry */
253 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
255 /* Mask error for already created
263 struct nid_stat *tmp = exp->exp_nid_stats;
264 LASSERT(tmp != NULL);
266 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267 if (tmp->nid_brw_stats == NULL)
268 GOTO(clean, rc = -ENOMEM);
270 init_brw_stats(tmp->nid_brw_stats);
271 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272 0644, &filter_per_nid_stats_fops,
275 CWARN("Error adding the brw_stats file\n");
277 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
285 /* 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 obd_device_target *obt = &obd->u.obt;
313 struct tg_export_data *ted = &exp->exp_target_data;
314 struct lr_server_data *lsd = class_server_data(obd);
315 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
316 int new_client = (cl_idx == -1);
320 LASSERT(bitmap != NULL);
321 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
324 if (strcmp(ted->ted_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 ted->ted_lr_idx = cl_idx;
352 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
353 cl_idx * le16_to_cpu(lsd->lsd_client_size);
354 cfs_init_mutex(&ted->ted_lcd_lock);
355 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
357 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
358 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
361 struct lvfs_run_ctxt saved;
362 loff_t off = ted->ted_lr_off;
366 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
367 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_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 obt->obt_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 ted->ted_lcd->lcd_last_epoch = lsd->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, obt->obt_rcvd_filp,
393 sizeof(*ted->ted_lcd),
394 &off, rc /* sync if no cb */);
396 obt->obt_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, ted->ted_lr_idx, rc);
410 static int filter_client_del(struct obd_export *exp)
412 struct tg_export_data *ted = &exp->exp_target_data;
413 struct obd_device_target *obt = &exp->exp_obd->u.obt;
414 struct lvfs_run_ctxt saved;
419 if (ted->ted_lcd == NULL)
422 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
423 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
426 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
428 off = ted->ted_lr_off;
430 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
431 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
433 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
434 * we leak a client slot that will be cleaned on the next recovery. */
436 CERROR("%s: client idx %d has med_off %lld\n",
437 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
438 GOTO(free, rc = -EINVAL);
441 /* Clear the bit _after_ zeroing out the client so we don't
442 race with filter_client_add and zero out new clients.*/
443 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
444 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
449 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
450 /* Make sure the server's last_transno is up to date.
451 * This should be done before zeroing client slot so last_transno will
452 * be in server data or in client data in case of failure */
453 filter_update_server_data(exp->exp_obd);
455 cfs_mutex_down(&ted->ted_lcd_lock);
456 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
457 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
459 sizeof(*ted->ted_lcd), &off, 0);
460 cfs_mutex_up(&ted->ted_lcd_lock);
461 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
463 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
464 "zero out client %s at idx %u/%llu in %s, rc %d\n",
465 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
472 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
473 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
474 struct filter_mod_data *fmd)
476 LASSERT_SPIN_LOCKED(&fed->fed_lock);
477 if (--fmd->fmd_refcount == 0) {
478 /* XXX when we have persistent reservations and the handle
479 * is stored herein we need to drop it here. */
480 fed->fed_mod_count--;
481 cfs_list_del(&fmd->fmd_list);
482 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
486 /* drop fmd reference, free it if last ref */
487 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
489 struct filter_export_data *fed;
494 fed = &exp->exp_filter_data;
495 cfs_spin_lock(&fed->fed_lock);
496 filter_fmd_put_nolock(fed, fmd); /* caller reference */
497 cfs_spin_unlock(&fed->fed_lock);
500 /* expire entries from the end of the list if there are too many
501 * or they are too old */
502 static void filter_fmd_expire_nolock(struct filter_obd *filter,
503 struct filter_export_data *fed,
504 struct filter_mod_data *keep)
506 struct filter_mod_data *fmd, *tmp;
508 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
512 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
513 fed->fed_mod_count < filter->fo_fmd_max_num)
516 cfs_list_del_init(&fmd->fmd_list);
517 filter_fmd_put_nolock(fed, fmd); /* list reference */
521 void filter_fmd_expire(struct obd_export *exp)
523 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
524 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
525 &exp->exp_filter_data, NULL);
526 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
529 /* find specified objid, group in export fmd list.
530 * caller must hold fed_lock and take fmd reference itself */
531 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
532 struct filter_export_data *fed,
533 obd_id objid, obd_gr group)
535 struct filter_mod_data *found = NULL, *fmd;
537 LASSERT_SPIN_LOCKED(&fed->fed_lock);
539 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
540 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
542 cfs_list_del(&fmd->fmd_list);
543 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
544 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
549 filter_fmd_expire_nolock(filter, fed, found);
554 /* Find fmd based on objid and group, or return NULL if not found. */
555 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
556 obd_id objid, obd_gr group)
558 struct filter_mod_data *fmd;
560 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
561 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
562 &exp->exp_filter_data, objid, group);
564 fmd->fmd_refcount++; /* caller reference */
565 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
570 /* Find fmd based on objid and group, or create a new one if none is found.
571 * It is possible for this function to return NULL under memory pressure,
572 * or if objid = 0 is passed (which will only cause old entries to expire).
573 * Currently this is not fatal because any fmd state is transient and
574 * may also be freed when it gets sufficiently old. */
575 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
576 obd_id objid, obd_gr group)
578 struct filter_export_data *fed = &exp->exp_filter_data;
579 struct filter_mod_data *found = NULL, *fmd_new = NULL;
581 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
583 cfs_spin_lock(&fed->fed_lock);
584 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
587 cfs_list_add_tail(&fmd_new->fmd_list,
589 fmd_new->fmd_id = objid;
590 fmd_new->fmd_gr = group;
591 fmd_new->fmd_refcount++; /* list reference */
593 fed->fed_mod_count++;
595 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
599 found->fmd_refcount++; /* caller reference */
600 found->fmd_expire = jiffies +
601 exp->exp_obd->u.filter.fo_fmd_max_age;
604 cfs_spin_unlock(&fed->fed_lock);
610 /* drop fmd list reference so it will disappear when last reference is put.
611 * This isn't so critical because it would in fact only affect the one client
612 * that is doing the unlink and at worst we have an stale entry referencing
613 * an object that should never be used again. */
614 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
616 struct filter_mod_data *found = NULL;
618 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
619 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
621 cfs_list_del_init(&found->fmd_list);
622 filter_fmd_put_nolock(&exp->exp_filter_data, found);
624 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
627 #define filter_fmd_drop(exp, objid, group)
630 /* remove all entries from fmd list */
631 static void filter_fmd_cleanup(struct obd_export *exp)
633 struct filter_export_data *fed = &exp->exp_filter_data;
634 struct filter_mod_data *fmd = NULL, *tmp;
636 cfs_spin_lock(&fed->fed_lock);
637 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
638 cfs_list_del_init(&fmd->fmd_list);
639 filter_fmd_put_nolock(fed, fmd);
641 cfs_spin_unlock(&fed->fed_lock);
644 static int filter_init_export(struct obd_export *exp)
647 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
648 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650 cfs_spin_lock(&exp->exp_lock);
651 exp->exp_connecting = 1;
652 cfs_spin_unlock(&exp->exp_lock);
653 rc = lut_client_alloc(exp);
655 rc = ldlm_init_export(exp);
660 static int filter_free_server_data(struct obd_device_target *obt)
662 lut_fini(NULL, obt->obt_lut);
663 OBD_FREE_PTR(obt->obt_lut);
667 /* assumes caller is already in kernel ctxt */
668 int filter_update_server_data(struct obd_device *obd)
670 struct file *filp = obd->u.obt.obt_rcvd_filp;
671 struct lr_server_data *lsd = class_server_data(obd);
676 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
677 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
678 le64_to_cpu(lsd->lsd_last_transno));
679 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
680 le64_to_cpu(lsd->lsd_mount_count));
682 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
684 CERROR("error writing lr_server_data: rc = %d\n", rc);
689 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
692 struct filter_obd *filter = &obd->u.filter;
698 if (filter->fo_last_objid_files[group] == NULL) {
699 CERROR("Object group "LPU64" not fully setup; not updating "
700 "last_objid\n", group);
704 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
705 obd->obd_name, group, filter->fo_last_objids[group]);
707 tmp = cpu_to_le64(filter->fo_last_objids[group]);
708 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
709 &tmp, sizeof(tmp), &off, force_sync);
711 CERROR("error writing group "LPU64" last objid: rc = %d\n",
715 extern int ost_handle(struct ptlrpc_request *req);
716 /* assumes caller has already in kernel ctxt */
717 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
719 struct filter_obd *filter = &obd->u.filter;
720 struct lr_server_data *lsd;
721 struct lsd_client_data *lcd = NULL;
722 struct inode *inode = filp->f_dentry->d_inode;
723 unsigned long last_rcvd_size = i_size_read(inode);
724 struct lu_target *lut;
731 /* ensure padding in the struct is the correct size */
732 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
733 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
734 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
735 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
737 /* allocate and initialize lu_target */
741 rc = lut_init(NULL, lut, obd, NULL);
744 lsd = class_server_data(obd);
745 if (last_rcvd_size == 0) {
746 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
748 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
749 lsd->lsd_last_transno = 0;
750 mount_count = lsd->lsd_mount_count = 0;
751 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
752 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
753 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
754 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
755 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
756 /* OBD_COMPAT_OST is set in filter_connect_internal when the
757 * MDS first connects and assigns the OST index number. */
758 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
761 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
763 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
767 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
768 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
769 "using the wrong disk %s. Were the "
770 "/dev/ assignments rearranged?\n",
771 obd->obd_uuid.uuid, lsd->lsd_uuid);
772 GOTO(err_lut, rc = -EINVAL);
774 mount_count = le64_to_cpu(lsd->lsd_mount_count);
775 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
777 /* Assume old last_rcvd format unless I_C_LR is set */
778 if (!(lsd->lsd_feature_incompat &
779 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
780 lsd->lsd_last_transno = lsd->lsd_compat14;
782 /* OBD_COMPAT_OST is set in filter_connect_internal when the
783 * MDS first connects and assigns the OST index number. */
784 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
788 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
789 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
790 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
791 ~FILTER_INCOMPAT_SUPP);
792 GOTO(err_lut, rc = -EINVAL);
794 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
795 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
796 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
797 ~FILTER_ROCOMPAT_SUPP);
798 /* Do something like remount filesystem read-only */
799 GOTO(err_lut, rc = -EINVAL);
802 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
804 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
805 obd->obd_name, start_epoch);
806 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
807 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
808 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
809 obd->obd_name, mount_count + 1);
810 CDEBUG(D_INODE, "%s: server data size: %u\n",
811 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
812 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
813 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
814 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
815 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
816 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
817 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
818 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
819 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
820 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
821 le16_to_cpu(lsd->lsd_client_size));
823 if (!obd->obd_replayable) {
824 CWARN("%s: recovery support OFF\n", obd->obd_name);
830 GOTO(err_client, rc = -ENOMEM);
832 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
833 off < last_rcvd_size; cl_idx++) {
835 struct obd_export *exp;
836 struct filter_export_data *fed;
838 /* Don't assume off is incremented properly by
839 * fsfilt_read_record(), in case sizeof(*lcd)
840 * isn't the same as lsd->lsd_client_size. */
841 off = le32_to_cpu(lsd->lsd_client_start) +
842 cl_idx * le16_to_cpu(lsd->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 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
861 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
862 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
864 /* These exports are cleaned up by filter_disconnect(), so they
865 * need to be set up like real exports as filter_connect() does.
867 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
869 if (PTR_ERR(exp) == -EALREADY) {
870 /* export already exists, zero out this one */
871 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
875 GOTO(err_client, rc = PTR_ERR(exp));
878 fed = &exp->exp_filter_data;
879 *fed->fed_ted.ted_lcd = *lcd;
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);
895 class_export_put(exp);
897 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
898 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
902 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
904 lut->lut_mount_count = mount_count + 1;
905 lsd->lsd_mount_count = cpu_to_le64(lut->lut_mount_count);
907 /* save it, so mount count and last_transno is current */
908 rc = filter_update_server_data(obd);
910 GOTO(err_client, rc);
915 class_disconnect_exports(obd);
917 filter_free_server_data(&obd->u.obt);
921 static int filter_cleanup_groups(struct obd_device *obd)
923 struct filter_obd *filter = &obd->u.filter;
925 struct dentry *dentry;
929 if (filter->fo_dentry_O_groups != NULL) {
930 for (i = 0; i < filter->fo_group_count; i++) {
931 dentry = filter->fo_dentry_O_groups[i];
935 OBD_FREE(filter->fo_dentry_O_groups,
936 filter->fo_group_count *
937 sizeof(*filter->fo_dentry_O_groups));
938 filter->fo_dentry_O_groups = NULL;
940 if (filter->fo_last_objid_files != NULL) {
941 for (i = 0; i < filter->fo_group_count; i++) {
942 filp = filter->fo_last_objid_files[i];
946 OBD_FREE(filter->fo_last_objid_files,
947 filter->fo_group_count *
948 sizeof(*filter->fo_last_objid_files));
949 filter->fo_last_objid_files = NULL;
951 if (filter->fo_dentry_O_sub != NULL) {
952 for (i = 0; i < filter->fo_group_count; i++) {
953 for (j = 0; j < filter->fo_subdir_count; j++) {
954 dentry = filter->fo_dentry_O_sub[i].dentry[j];
959 OBD_FREE(filter->fo_dentry_O_sub,
960 filter->fo_group_count *
961 sizeof(*filter->fo_dentry_O_sub));
962 filter->fo_dentry_O_sub = NULL;
964 if (filter->fo_last_objids != NULL) {
965 OBD_FREE(filter->fo_last_objids,
966 filter->fo_group_count *
967 sizeof(*filter->fo_last_objids));
968 filter->fo_last_objids = NULL;
970 if (filter->fo_dentry_O != NULL) {
971 f_dput(filter->fo_dentry_O);
972 filter->fo_dentry_O = NULL;
977 static int filter_update_last_group(struct obd_device *obd, int group)
979 struct filter_obd *filter = &obd->u.filter;
980 struct file *filp = NULL;
981 int last_group = 0, rc;
985 if (group <= filter->fo_committed_group)
988 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
992 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
996 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
998 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1002 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1003 obd->obd_name, last_group, group);
1007 /* must be sync: bXXXX */
1008 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1010 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1014 filter->fo_committed_group = group;
1017 filp_close(filp, 0);
1021 static int filter_read_group_internal(struct obd_device *obd, int group,
1024 struct filter_obd *filter = &obd->u.filter;
1025 __u64 *new_objids = NULL;
1026 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1027 struct dentry **new_groups = NULL;
1028 struct file **new_files = NULL;
1029 struct dentry *dentry;
1031 int old_count = filter->fo_group_count, rc, stage = 0, i;
1035 int len = group + 1;
1037 snprintf(name, 24, "%d", group);
1041 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1043 if (IS_ERR(dentry)) {
1044 CERROR("Cannot lookup expected object group %d: %ld\n",
1045 group, PTR_ERR(dentry));
1046 RETURN(PTR_ERR(dentry));
1049 dentry = simple_mkdir(filter->fo_dentry_O,
1050 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1051 if (IS_ERR(dentry)) {
1052 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1054 RETURN(PTR_ERR(dentry));
1059 snprintf(name, 24, "O/%d/LAST_ID", group);
1061 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1063 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1064 GOTO(cleanup, rc = PTR_ERR(filp));
1068 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1070 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1074 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1075 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1076 if (tmp_subdirs == NULL)
1077 GOTO(cleanup, rc = -ENOMEM);
1080 for (i = 0; i < filter->fo_subdir_count; i++) {
1082 snprintf(dir, sizeof(dir), "d%u", i);
1084 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1085 obd->u.obt.obt_vfsmnt,
1087 if (IS_ERR(tmp_subdirs->dentry[i])) {
1088 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1089 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1094 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1095 tmp_subdirs->dentry[i]);
1099 /* 'group' is an index; we need an array of length 'group + 1' */
1100 if (group + 1 > old_count) {
1101 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1102 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1103 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1104 OBD_ALLOC(new_files, len * sizeof(*new_files));
1106 if (new_objids == NULL || new_subdirs == NULL ||
1107 new_groups == NULL || new_files == NULL)
1108 GOTO(cleanup, rc = -ENOMEM);
1111 memcpy(new_objids, filter->fo_last_objids,
1112 old_count * sizeof(*new_objids));
1113 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1114 old_count * sizeof(*new_subdirs));
1115 memcpy(new_groups, filter->fo_dentry_O_groups,
1116 old_count * sizeof(*new_groups));
1117 memcpy(new_files, filter->fo_last_objid_files,
1118 old_count * sizeof(*new_files));
1120 OBD_FREE(filter->fo_last_objids,
1121 old_count * sizeof(*new_objids));
1122 OBD_FREE(filter->fo_dentry_O_sub,
1123 old_count * sizeof(*new_subdirs));
1124 OBD_FREE(filter->fo_dentry_O_groups,
1125 old_count * sizeof(*new_groups));
1126 OBD_FREE(filter->fo_last_objid_files,
1127 old_count * sizeof(*new_files));
1129 filter->fo_last_objids = new_objids;
1130 filter->fo_dentry_O_sub = new_subdirs;
1131 filter->fo_dentry_O_groups = new_groups;
1132 filter->fo_last_objid_files = new_files;
1133 filter->fo_group_count = len;
1136 filter->fo_dentry_O_groups[group] = dentry;
1137 filter->fo_last_objid_files[group] = filp;
1138 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1139 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1140 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1143 filter_update_last_group(obd, group);
1145 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1146 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1147 rc = filter_update_last_objid(obd, group, 1);
1151 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1152 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1153 obd->obd_name, group, last_objid);
1158 if (new_objids != NULL)
1159 OBD_FREE(new_objids, len * sizeof(*new_objids));
1160 if (new_subdirs != NULL)
1161 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1162 if (new_groups != NULL)
1163 OBD_FREE(new_groups, len * sizeof(*new_groups));
1164 if (new_files != NULL)
1165 OBD_FREE(new_files, len * sizeof(*new_files));
1167 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1168 for (i = 0; i < filter->fo_subdir_count; i++) {
1169 if (tmp_subdirs->dentry[i] != NULL)
1170 dput(tmp_subdirs->dentry[i]);
1172 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1175 filp_close(filp, 0);
1182 static int filter_read_groups(struct obd_device *obd, int last_group,
1185 struct filter_obd *filter = &obd->u.filter;
1186 int old_count, group, rc = 0;
1188 cfs_down(&filter->fo_init_lock);
1189 old_count = filter->fo_group_count;
1190 for (group = old_count; group <= last_group; group++) {
1191 rc = filter_read_group_internal(obd, group, create);
1195 cfs_up(&filter->fo_init_lock);
1199 /* FIXME: object groups */
1200 static int filter_prep_groups(struct obd_device *obd)
1202 struct filter_obd *filter = &obd->u.filter;
1203 struct dentry *O_dentry;
1205 int last_group, rc = 0, cleanup_phase = 0;
1209 O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1211 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1212 if (IS_ERR(O_dentry)) {
1213 rc = PTR_ERR(O_dentry);
1214 CERROR("cannot open/create O: rc = %d\n", rc);
1217 filter->fo_dentry_O = O_dentry;
1218 cleanup_phase = 1; /* O_dentry */
1220 /* we have to initialize all groups before first connections from
1221 * clients because they may send create/destroy for any group -bzzz */
1222 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1224 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1225 GOTO(cleanup, rc = PTR_ERR(filp));
1227 cleanup_phase = 2; /* filp */
1229 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1231 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1236 last_group = FILTER_GROUP_MDS0;
1238 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1239 FILTER_GROUP_MDS0, last_group);
1240 filter->fo_committed_group = last_group;
1241 rc = filter_read_groups(obd, last_group, 1);
1245 filp_close(filp, 0);
1249 switch (cleanup_phase) {
1251 filp_close(filp, 0);
1253 filter_cleanup_groups(obd);
1254 f_dput(filter->fo_dentry_O);
1255 filter->fo_dentry_O = NULL;
1263 /* setup the object store with correct subdirectories */
1264 static int filter_prep(struct obd_device *obd)
1266 struct lvfs_run_ctxt saved;
1267 struct filter_obd *filter = &obd->u.filter;
1269 struct inode *inode;
1273 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1274 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1275 if (!file || IS_ERR(file)) {
1277 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1281 obd->u.obt.obt_rcvd_filp = file;
1282 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1283 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1284 file->f_dentry->d_inode->i_mode);
1285 GOTO(err_filp, rc = -ENOENT);
1288 inode = file->f_dentry->d_parent->d_inode;
1289 /* We use i_op->unlink directly in filter_vfs_unlink() */
1290 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1291 CERROR("%s: filesystem does not support create/unlink ops\n",
1293 GOTO(err_filp, rc = -EOPNOTSUPP);
1296 rc = filter_init_server_data(obd, file);
1298 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1301 LASSERT(obd->u.obt.obt_lut);
1302 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1304 /* open and create health check io file*/
1305 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1308 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1310 GOTO(err_server_data, rc);
1312 filter->fo_obt.obt_health_check_filp = file;
1313 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1314 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1315 file->f_dentry->d_inode->i_mode);
1316 GOTO(err_health_check, rc = -ENOENT);
1318 rc = lvfs_check_io_health(obd, file);
1320 GOTO(err_health_check, rc);
1322 rc = filter_prep_groups(obd);
1324 GOTO(err_health_check, rc);
1326 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1331 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1332 CERROR("can't close %s after error\n", HEALTH_CHECK);
1333 filter->fo_obt.obt_health_check_filp = NULL;
1335 target_recovery_fini(obd);
1336 filter_free_server_data(&obd->u.obt);
1338 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1339 CERROR("can't close %s after error\n", LAST_RCVD);
1340 obd->u.obt.obt_rcvd_filp = NULL;
1344 /* cleanup the filter: write last used object id to status file */
1345 static void filter_post(struct obd_device *obd)
1347 struct lvfs_run_ctxt saved;
1348 struct filter_obd *filter = &obd->u.filter;
1351 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1352 * best to start a transaction with h_sync, because we removed this
1355 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1356 rc = filter_update_server_data(obd);
1358 CERROR("error writing server data: rc = %d\n", rc);
1360 for (i = 0; i < filter->fo_group_count; i++) {
1361 rc = filter_update_last_objid(obd, i,
1362 (i == filter->fo_group_count - 1));
1364 CERROR("error writing group %d lastobjid: rc = %d\n",
1368 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1369 obd->u.obt.obt_rcvd_filp = NULL;
1371 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1373 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1374 filter->fo_obt.obt_health_check_filp = NULL;
1376 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1378 filter_cleanup_groups(obd);
1379 filter_free_server_data(&obd->u.obt);
1380 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1382 filter_free_capa_keys(filter);
1383 cleanup_capa_hash(filter->fo_capa_hash);
1386 static void filter_set_last_id(struct filter_obd *filter,
1387 obd_id id, obd_gr group)
1389 LASSERT(group <= filter->fo_group_count);
1391 cfs_spin_lock(&filter->fo_objidlock);
1392 filter->fo_last_objids[group] = id;
1393 cfs_spin_unlock(&filter->fo_objidlock);
1396 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1399 LASSERT(group <= filter->fo_group_count);
1400 LASSERT(filter->fo_last_objids != NULL);
1402 /* FIXME: object groups */
1403 cfs_spin_lock(&filter->fo_objidlock);
1404 id = filter->fo_last_objids[group];
1405 cfs_spin_unlock(&filter->fo_objidlock);
1409 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1411 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1415 /* We never dget the object parent, so DON'T dput it either */
1416 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1418 struct filter_obd *filter = &obd->u.filter;
1419 struct filter_subdirs *subdirs;
1420 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1422 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1423 return filter->fo_dentry_O_groups[group];
1425 subdirs = &filter->fo_dentry_O_sub[group];
1426 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1429 /* We never dget the object parent, so DON'T dput it either */
1430 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1433 unsigned long now = jiffies;
1434 struct dentry *dparent = filter_parent(obd, group, objid);
1437 if (IS_ERR(dparent))
1439 if (dparent == NULL)
1440 return ERR_PTR(-ENOENT);
1442 rc = filter_lock_dentry(obd, dparent);
1443 fsfilt_check_slow(obd, now, "parent lock");
1444 return rc ? ERR_PTR(rc) : dparent;
1447 /* We never dget the object parent, so DON'T dput it either */
1448 static void filter_parent_unlock(struct dentry *dparent)
1450 UNLOCK_INODE_MUTEX(dparent->d_inode);
1453 /* How to get files, dentries, inodes from object id's.
1455 * If dir_dentry is passed, the caller has already locked the parent
1456 * appropriately for this operation (normally a write lock). If
1457 * dir_dentry is NULL, we do a read lock while we do the lookup to
1458 * avoid races with create/destroy and such changing the directory
1459 * internal to the filesystem code. */
1460 struct dentry *filter_fid2dentry(struct obd_device *obd,
1461 struct dentry *dir_dentry,
1462 obd_gr group, obd_id id)
1464 struct dentry *dparent = dir_dentry;
1465 struct dentry *dchild;
1470 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1471 obd->u.filter.fo_destroys_in_progress == 0) {
1472 /* don't fail lookups for orphan recovery, it causes
1473 * later LBUGs when objects still exist during precreate */
1474 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1475 RETURN(ERR_PTR(-ENOENT));
1478 CERROR("fatal: invalid object id 0\n");
1479 RETURN(ERR_PTR(-ESTALE));
1482 len = sprintf(name, LPU64, id);
1483 if (dir_dentry == NULL) {
1484 dparent = filter_parent_lock(obd, group, id);
1485 if (IS_ERR(dparent)) {
1486 CERROR("%s: error getting object "LPU64":"LPU64
1487 " parent: rc %ld\n", obd->obd_name,
1488 id, group, PTR_ERR(dparent));
1492 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1493 dparent->d_name.len, dparent->d_name.name, name);
1494 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1495 if (dir_dentry == NULL)
1496 filter_parent_unlock(dparent);
1497 if (IS_ERR(dchild)) {
1498 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1503 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1504 CERROR("%s: got bad object "LPU64" inode %lu\n",
1505 obd->obd_name, id, dchild->d_inode->i_ino);
1507 RETURN(ERR_PTR(-ENOENT));
1510 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1511 name, dchild, atomic_read(&dchild->d_count));
1513 LASSERT(atomic_read(&dchild->d_count) > 0);
1518 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1519 obd_id group, struct lustre_handle *lockh)
1521 int flags = LDLM_AST_DISCARD_DATA, rc;
1522 struct ldlm_res_id res_id;
1523 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1526 osc_build_res_name(objid, group, &res_id);
1527 /* Tell the clients that the object is gone now and that they should
1528 * throw away any cached pages. */
1529 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1530 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1531 ldlm_completion_ast, NULL, NULL, 0, NULL,
1538 static void filter_fini_destroy(struct obd_device *obd,
1539 struct lustre_handle *lockh)
1541 if (lustre_handle_is_used(lockh))
1542 ldlm_lock_decref(lockh, LCK_PW);
1545 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1546 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1547 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1548 * reverse. Caller must take i_sem before starting the transaction and we
1549 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1550 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1551 struct vfsmount *mnt)
1556 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1557 * itself we already hold dir->i_mutex for child create/unlink ops */
1558 LASSERT(dentry->d_inode != NULL);
1559 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1560 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1564 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1565 GOTO(out, rc = -ENOENT);
1567 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1572 GOTO(out, rc = -EPERM);
1574 /* check_sticky() */
1575 if ((dentry->d_inode->i_uid != current->fsuid &&
1576 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1577 IS_IMMUTABLE(dentry->d_inode))
1578 GOTO(out, rc = -EPERM);
1580 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1581 * that was done because of journal_start (which is already done
1582 * here) or some other ordering issue. */
1585 rc = ll_security_inode_unlink(dir, dentry, mnt);
1589 rc = dir->i_op->unlink(dir, dentry);
1591 /* need to drop i_mutex before we lose inode reference */
1592 UNLOCK_INODE_MUTEX(dentry->d_inode);
1599 /* Caller must hold LCK_PW on parent and push us into kernel context.
1600 * Caller must hold child i_mutex, we drop it always.
1601 * Caller is also required to ensure that dchild->d_inode exists. */
1602 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1603 obd_gr group, struct dentry *dparent,
1604 struct dentry *dchild)
1606 struct inode *inode = dchild->d_inode;
1609 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1610 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1611 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1612 (unsigned long)inode->i_nlink,
1613 atomic_read(&inode->i_count));
1616 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1618 CERROR("error unlinking objid %.*s: rc %d\n",
1619 dchild->d_name.len, dchild->d_name.name, rc);
1623 struct filter_intent_args {
1624 struct ldlm_lock **victim;
1629 static enum interval_iter filter_intent_cb(struct interval_node *n,
1632 struct ldlm_interval *node = (struct ldlm_interval *)n;
1633 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1634 __u64 size = arg->size;
1635 struct ldlm_lock **v = arg->victim;
1636 struct ldlm_lock *lck;
1638 /* If the interval is lower than the current file size,
1640 if (interval_high(n) <= size)
1641 return INTERVAL_ITER_STOP;
1643 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1644 /* Don't send glimpse ASTs to liblustre clients.
1645 * They aren't listening for them, and they do
1646 * entirely synchronous I/O anyways. */
1647 if (lck->l_export == NULL ||
1648 lck->l_export->exp_libclient == 1)
1651 if (*arg->liblustre)
1652 *arg->liblustre = 0;
1655 *v = LDLM_LOCK_GET(lck);
1656 } else if ((*v)->l_policy_data.l_extent.start <
1657 lck->l_policy_data.l_extent.start) {
1658 LDLM_LOCK_RELEASE(*v);
1659 *v = LDLM_LOCK_GET(lck);
1662 /* the same policy group - every lock has the
1663 * same extent, so needn't do it any more */
1667 return INTERVAL_ITER_CONT;
1670 static int filter_intent_policy(struct ldlm_namespace *ns,
1671 struct ldlm_lock **lockp, void *req_cookie,
1672 ldlm_mode_t mode, int flags, void *data)
1674 CFS_LIST_HEAD(rpc_list);
1675 struct ptlrpc_request *req = req_cookie;
1676 struct ldlm_lock *lock = *lockp, *l = NULL;
1677 struct ldlm_resource *res = lock->l_resource;
1678 ldlm_processing_policy policy;
1679 struct ost_lvb *res_lvb, *reply_lvb;
1680 struct ldlm_reply *rep;
1682 int idx, rc, tmpflags = 0, only_liblustre = 1;
1683 struct ldlm_interval_tree *tree;
1684 struct filter_intent_args arg;
1685 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1686 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1687 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1690 policy = ldlm_get_processing_policy(res);
1691 LASSERT(policy != NULL);
1692 LASSERT(req != NULL);
1694 rc = lustre_pack_reply(req, 3, repsize, NULL);
1696 RETURN(req->rq_status = rc);
1698 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1699 LASSERT(rep != NULL);
1701 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1702 sizeof(*reply_lvb));
1703 LASSERT(reply_lvb != NULL);
1705 //fixup_handle_for_resent_req(req, lock, &lockh);
1707 /* Call the extent policy function to see if our request can be
1708 * granted, or is blocked.
1709 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1710 * lock, and should not be granted if the lock will be blocked.
1713 LASSERT(ns == res->lr_namespace);
1715 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1716 check_res_locked(res);
1718 /* FIXME: we should change the policy function slightly, to not make
1719 * this list at all, since we just turn around and free it */
1720 while (!cfs_list_empty(&rpc_list)) {
1721 struct ldlm_lock *wlock =
1722 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1724 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1725 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1726 lock->l_flags &= ~LDLM_FL_CP_REQD;
1727 cfs_list_del_init(&wlock->l_cp_ast);
1728 LDLM_LOCK_RELEASE(wlock);
1731 /* The lock met with no resistance; we're finished. */
1732 if (rc == LDLM_ITER_CONTINUE) {
1733 /* do not grant locks to the liblustre clients: they cannot
1734 * handle ASTs robustly. We need to do this while still
1735 * holding ns_lock to avoid the lock remaining on the res_link
1736 * list (and potentially being added to l_pending_list by an
1737 * AST) when we are going to drop this lock ASAP. */
1738 if (lock->l_export->exp_libclient ||
1739 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1740 ldlm_resource_unlink_lock(lock);
1741 err = ELDLM_LOCK_ABORTED;
1743 err = ELDLM_LOCK_REPLACED;
1749 /* Do not grant any lock, but instead send GL callbacks. The extent
1750 * policy nicely created a list of all PW locks for us. We will choose
1751 * the highest of those which are larger than the size in the LVB, if
1752 * any, and perform a glimpse callback. */
1753 res_lvb = res->lr_lvb_data;
1754 LASSERT(res_lvb != NULL);
1755 *reply_lvb = *res_lvb;
1758 * ->ns_lock guarantees that no new locks are granted, and,
1759 * therefore, that res->lr_lvb_data cannot increase beyond the
1760 * end of already granted lock. As a result, it is safe to
1761 * check against "stale" reply_lvb->lvb_size value without
1764 arg.size = reply_lvb->lvb_size;
1766 arg.liblustre = &only_liblustre;
1767 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1768 tree = &res->lr_itree[idx];
1769 if (tree->lit_mode == LCK_PR)
1772 interval_iterate_reverse(tree->lit_root,
1773 filter_intent_cb, &arg);
1777 /* There were no PW locks beyond the size in the LVB; finished. */
1779 if (only_liblustre) {
1780 /* If we discovered a liblustre client with a PW lock,
1781 * however, the LVB may be out of date! The LVB is
1782 * updated only on glimpse (which we don't do for
1783 * liblustre clients) and cancel (which the client
1784 * obviously has not yet done). So if it has written
1785 * data but kept the lock, the LVB is stale and needs
1786 * to be updated from disk.
1788 * Of course, this will all disappear when we switch to
1789 * taking liblustre locks on the OST. */
1790 ldlm_res_lvbo_update(res, NULL, 1);
1792 RETURN(ELDLM_LOCK_ABORTED);
1796 * This check is for lock taken in filter_prepare_destroy() that does
1797 * not have l_glimpse_ast set. So the logic is: if there is a lock
1798 * with no l_glimpse_ast set, this object is being destroyed already.
1800 * Hence, if you are grabbing DLM locks on the server, always set
1801 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1803 if (l->l_glimpse_ast == NULL) {
1804 /* We are racing with unlink(); just return -ENOENT */
1805 rep->lock_policy_res1 = -ENOENT;
1809 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1810 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1811 /* Update the LVB from disk if the AST failed (this is a legal race) */
1813 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1814 * sending ast is not handled. This can result in lost client writes.
1817 ldlm_res_lvbo_update(res, NULL, 1);
1820 *reply_lvb = *res_lvb;
1824 LDLM_LOCK_RELEASE(l);
1826 RETURN(ELDLM_LOCK_ABORTED);
1830 * per-obd_device iobuf pool.
1832 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1833 * allocations in write-path has to be minimized (see bug 5137).
1835 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1836 * OST threads (see ost_thread_{init,done}()).
1838 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1839 * at the OST layer there are only (potentially) multiple obd_device of type
1840 * unknown at the time of OST thread creation.
1842 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1843 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1844 * it's very own iobuf.
1848 * filter_kiobuf_pool_init()
1850 * filter_kiobuf_pool_done()
1852 * filter_iobuf_get()
1854 * operate on this array. They are "generic" in a sense that they don't depend
1855 * on actual type of iobuf's (the latter depending on Linux kernel version).
1859 * destroy pool created by filter_iobuf_pool_init
1861 static void filter_iobuf_pool_done(struct filter_obd *filter)
1863 struct filter_iobuf **pool;
1868 pool = filter->fo_iobuf_pool;
1870 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1871 if (pool[i] != NULL)
1872 filter_free_iobuf(pool[i]);
1874 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1875 filter->fo_iobuf_pool = NULL;
1880 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1882 struct filter_obd *filter = &obd->u.filter;
1883 struct sptlrpc_rule_set tmp_rset;
1886 sptlrpc_rule_set_init(&tmp_rset);
1887 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1889 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1894 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1896 cfs_write_lock(&filter->fo_sptlrpc_lock);
1897 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1898 filter->fo_sptlrpc_rset = tmp_rset;
1899 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1905 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1907 static int filter_iobuf_pool_init(struct filter_obd *filter)
1914 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1916 if (filter->fo_iobuf_pool == NULL)
1919 filter->fo_iobuf_count = OSS_THREADS_MAX;
1924 /* Return iobuf allocated for @thread_id. We don't know in advance how
1925 * many threads there will be so we allocate a large empty array and only
1926 * fill in those slots that are actually in use.
1927 * If we haven't allocated a pool entry for this thread before, do so now. */
1928 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1930 int thread_id = (oti && oti->oti_thread) ?
1931 oti->oti_thread->t_id : -1;
1932 struct filter_iobuf *pool = NULL;
1933 struct filter_iobuf **pool_place = NULL;
1935 if (thread_id >= 0) {
1936 LASSERT(thread_id < filter->fo_iobuf_count);
1937 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1940 if (unlikely(pool == NULL)) {
1941 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1942 PTLRPC_MAX_BRW_PAGES);
1943 if (pool_place != NULL)
1950 /* mount the file system (secretly). lustre_cfg parameters are:
1953 * 3 = flags: failover=f, failout=n
1956 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1959 struct filter_obd *filter = &obd->u.filter;
1960 struct vfsmount *mnt;
1961 struct lustre_mount_info *lmi;
1962 struct obd_uuid uuid;
1970 if (lcfg->lcfg_bufcount < 3 ||
1971 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1972 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1975 lmi = server_get_mount(obd->obd_name);
1977 /* We already mounted in lustre_fill_super.
1978 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1979 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1981 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1983 /* gets recovery timeouts from mount data */
1984 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1985 obd->obd_recovery_timeout =
1986 lsi->lsi_lmd->lmd_recovery_time_soft;
1987 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1988 obd->obd_recovery_time_hard =
1989 lsi->lsi_lmd->lmd_recovery_time_hard;
1991 /* old path - used by lctl */
1992 CERROR("Using old MDS mount method\n");
1993 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1994 MS_NOATIME|MS_NODIRATIME,
1995 lustre_cfg_string(lcfg, 1), option);
1998 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1999 lustre_cfg_string(lcfg, 1), rc);
2003 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2005 if (IS_ERR(obd->obd_fsops))
2006 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2008 rc = filter_iobuf_pool_init(filter);
2012 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2013 CERROR("%s: Underlying device is marked as read-only. "
2014 "Setup failed\n", obd->obd_name);
2015 GOTO(err_ops, rc = -EROFS);
2018 /* failover is the default */
2019 obd->obd_replayable = 1;
2021 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2022 str = lustre_cfg_string(lcfg, 3);
2023 if (strchr(str, 'n')) {
2024 CWARN("%s: recovery disabled\n", obd->obd_name);
2025 obd->obd_replayable = 0;
2029 obd->u.obt.obt_vfsmnt = mnt;
2030 obd->u.obt.obt_sb = mnt->mnt_sb;
2031 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2032 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2034 fsfilt_setup(obd, obd->u.obt.obt_sb);
2036 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2037 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2038 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2039 obd->obd_lvfs_ctxt.fs = get_ds();
2040 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2042 cfs_init_mutex(&filter->fo_init_lock);
2043 filter->fo_committed_group = 0;
2044 filter->fo_destroys_in_progress = 0;
2045 for (i = 0; i < 32; i++)
2046 cfs_sema_init(&filter->fo_create_locks[i], 1);
2048 cfs_spin_lock_init(&filter->fo_objidlock);
2049 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2050 cfs_sema_init(&filter->fo_alloc_lock, 1);
2051 init_brw_stats(&filter->fo_filter_stats);
2052 filter->fo_read_cache = 1; /* enable read-only cache by default */
2053 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2054 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2055 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2056 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2058 rc = filter_prep(obd);
2062 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2063 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2065 filter->fo_fl_oss_capa = 1;
2067 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2068 filter->fo_capa_hash = init_capa_hash();
2069 if (filter->fo_capa_hash == NULL)
2070 GOTO(err_post, rc = -ENOMEM);
2072 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2073 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2074 LDLM_NAMESPACE_GREEDY);
2075 if (obd->obd_namespace == NULL)
2076 GOTO(err_post, rc = -ENOMEM);
2077 obd->obd_namespace->ns_lvbp = obd;
2078 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2079 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2081 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2082 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2084 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2086 CERROR("failed to setup llogging subsystems\n");
2090 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2091 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2092 /* do this after llog being initialized */
2093 filter_adapt_sptlrpc_conf(obd, 1);
2095 rc = lquota_setup(filter_quota_interface_ref, obd);
2099 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2100 if (q->max_sectors < q->max_hw_sectors &&
2101 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2102 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2103 "for larger I/O requests: max_sectors = %u "
2104 "could be up to max_hw_sectors=%u\n",
2105 obd->obd_name, mnt->mnt_sb->s_id,
2106 q->max_sectors, q->max_hw_sectors);
2108 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2109 if (uuid_ptr != NULL) {
2110 class_uuid_unparse(uuid_ptr, &uuid);
2116 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2117 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2118 obd->obd_name, label ?: str, lmi ? "on " : "",
2119 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2120 obd->obd_replayable ? "enabled" : "disabled");
2122 if (obd->obd_recovering)
2123 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2124 "or until %d client%s reconnect%s\n",
2126 obd->obd_recovery_timeout / 60,
2127 obd->obd_recovery_timeout % 60,
2128 obd->obd_max_recoverable_clients,
2129 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2130 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2138 fsfilt_put_ops(obd->obd_fsops);
2139 filter_iobuf_pool_done(filter);
2141 server_put_mount(obd->obd_name, mnt);
2142 obd->u.obt.obt_sb = 0;
2146 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2148 struct lprocfs_static_vars lvars;
2153 CLASSERT(offsetof(struct obd_device, u.obt) ==
2154 offsetof(struct obd_device, u.filter.fo_obt));
2156 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2159 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2160 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2163 addr = (unsigned long)cfs_page_address(page);
2164 clear_page((void *)addr);
2166 /* lprocfs must be setup before the filter so state can be safely added
2167 * to /proc incrementally as the filter is setup */
2168 lprocfs_filter_init_vars(&lvars);
2169 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2170 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2171 /* Init obdfilter private stats here */
2172 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2173 LPROCFS_CNTR_AVGMINMAX,
2174 "read_bytes", "bytes");
2175 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2176 LPROCFS_CNTR_AVGMINMAX,
2177 "write_bytes", "bytes");
2178 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2179 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2180 "get_page", "usec");
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "get_page_failures", "num");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2185 LPROCFS_CNTR_AVGMINMAX,
2186 "cache_access", "pages");
2187 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2188 LPROCFS_CNTR_AVGMINMAX,
2189 "cache_hit", "pages");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2191 LPROCFS_CNTR_AVGMINMAX,
2192 "cache_miss", "pages");
2194 lproc_filter_attach_seqstat(obd);
2195 obd->obd_proc_exports_entry = lprocfs_register("exports",
2196 obd->obd_proc_entry,
2198 if (IS_ERR(obd->obd_proc_exports_entry)) {
2199 rc = PTR_ERR(obd->obd_proc_exports_entry);
2200 CERROR("error %d setting up lprocfs for %s\n",
2202 obd->obd_proc_exports_entry = NULL;
2205 if (obd->obd_proc_exports_entry)
2206 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2207 lprocfs_nid_stats_clear_read,
2208 lprocfs_nid_stats_clear_write, obd, NULL);
2210 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2211 LUSTRE_CFG_BUFLEN(lcfg, 4));
2212 rc = filter_common_setup(obd, lcfg, (void *)addr);
2213 OBD_PAGE_FREE(page);
2216 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2217 lprocfs_free_per_client_stats(obd);
2218 lprocfs_free_obd_stats(obd);
2219 lprocfs_obd_cleanup(obd);
2225 static struct llog_operations filter_mds_ost_repl_logops;
2227 static struct llog_operations filter_size_orig_logops = {
2228 .lop_setup = llog_obd_origin_setup,
2229 .lop_cleanup = llog_obd_origin_cleanup,
2230 .lop_add = llog_obd_origin_add
2233 static int filter_olg_fini(struct obd_llog_group *olg)
2235 struct llog_ctxt *ctxt;
2236 int rc = 0, rc2 = 0;
2239 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2241 rc = llog_cleanup(ctxt);
2243 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2245 rc2 = llog_cleanup(ctxt);
2250 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2252 rc2 = llog_cleanup(ctxt);
2261 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2262 struct obd_device *tgt)
2267 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2268 &filter_mds_ost_repl_logops);
2272 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2273 &filter_size_orig_logops);
2279 filter_olg_fini(olg);
2284 * Init the default olg, which is embeded in the obd_device, for filter.
2287 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2288 struct obd_device *tgt)
2290 struct filter_obd *filter = &obd->u.filter;
2291 struct llog_ctxt *ctxt;
2295 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2296 if (!filter->fo_lcm)
2299 filter_mds_ost_repl_logops = llog_client_ops;
2300 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2301 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2302 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2304 rc = filter_olg_init(obd, olg, tgt);
2306 GOTO(cleanup_lcm, rc);
2308 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2311 GOTO(cleanup_olg, rc);
2313 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2315 CERROR("Can't get ctxt for %p:%x\n", olg,
2316 LLOG_MDS_OST_REPL_CTXT);
2317 GOTO(cleanup_olg, rc = -ENODEV);
2319 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2320 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2321 llog_ctxt_put(ctxt);
2325 filter_olg_fini(olg);
2327 llog_recov_thread_fini(filter->fo_lcm, 1);
2328 filter->fo_lcm = NULL;
2333 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2334 struct obd_device *tgt, int *index)
2336 struct filter_obd *filter = &obd->u.filter;
2337 struct llog_ctxt *ctxt;
2341 LASSERT(olg != NULL);
2342 if (olg == &obd->obd_olg)
2343 return filter_default_olg_init(obd, olg, tgt);
2345 LASSERT(filter->fo_lcm != NULL);
2346 rc = filter_olg_init(obd, olg, tgt);
2349 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2351 CERROR("Can't get ctxt for %p:%x\n", olg,
2352 LLOG_MDS_OST_REPL_CTXT);
2353 filter_olg_fini(olg);
2356 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2357 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2358 llog_ctxt_put(ctxt);
2362 static int filter_llog_finish(struct obd_device *obd, int count)
2364 struct filter_obd *filter = &obd->u.filter;
2365 struct llog_ctxt *ctxt;
2368 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2371 * Make sure that no cached llcds left in recov_thread.
2372 * We actually do sync in disconnect time, but disconnect
2373 * may not come being marked rq_no_resend = 1.
2375 llog_sync(ctxt, NULL);
2378 * Balance class_import_get() in llog_receptor_accept().
2379 * This is safe to do, as llog is already synchronized
2380 * and its import may go.
2382 cfs_mutex_down(&ctxt->loc_sem);
2383 if (ctxt->loc_imp) {
2384 class_import_put(ctxt->loc_imp);
2385 ctxt->loc_imp = NULL;
2387 cfs_mutex_up(&ctxt->loc_sem);
2388 llog_ctxt_put(ctxt);
2391 if (filter->fo_lcm) {
2392 cfs_mutex_down(&ctxt->loc_sem);
2393 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2394 filter->fo_lcm = NULL;
2395 cfs_mutex_up(&ctxt->loc_sem);
2397 RETURN(filter_olg_fini(&obd->obd_olg));
2401 * Find the group llog according to group index in the llog group list.
2403 static struct obd_llog_group *
2404 filter_find_olg_internal(struct filter_obd *filter, int group)
2406 struct obd_llog_group *olg;
2408 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2409 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2410 if (olg->olg_group == group)
2417 * Find the group llog according to group index on the filter
2419 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2421 struct obd_llog_group *olg = NULL;
2422 struct filter_obd *filter;
2424 filter = &obd->u.filter;
2426 if (group == FILTER_GROUP_LLOG)
2427 RETURN(&obd->obd_olg);
2429 cfs_spin_lock(&filter->fo_llog_list_lock);
2430 olg = filter_find_olg_internal(filter, group);
2431 cfs_spin_unlock(&filter->fo_llog_list_lock);
2436 * Find the llog_group of the filter according to the group. If it can not
2437 * find, create the llog_group, which only happens when mds is being synced
2440 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2442 struct obd_llog_group *olg = NULL;
2443 struct filter_obd *filter;
2446 filter = &obd->u.filter;
2448 if (group == FILTER_GROUP_LLOG)
2449 RETURN(&obd->obd_olg);
2451 cfs_spin_lock(&filter->fo_llog_list_lock);
2452 olg = filter_find_olg_internal(filter, group);
2454 if (olg->olg_initializing) {
2455 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2457 GOTO(out_unlock, olg);
2462 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2464 llog_group_init(olg, group);
2465 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2466 olg->olg_initializing = 1;
2467 cfs_spin_unlock(&filter->fo_llog_list_lock);
2469 rc = obd_llog_init(obd, olg, obd, NULL);
2471 cfs_spin_lock(&filter->fo_llog_list_lock);
2472 cfs_list_del(&olg->olg_list);
2473 cfs_spin_unlock(&filter->fo_llog_list_lock);
2475 GOTO(out, olg = ERR_PTR(-ENOMEM));
2477 cfs_spin_lock(&filter->fo_llog_list_lock);
2478 olg->olg_initializing = 0;
2479 cfs_spin_unlock(&filter->fo_llog_list_lock);
2480 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2481 obd->obd_name, group, olg);
2486 cfs_spin_unlock(&filter->fo_llog_list_lock);
2490 static int filter_llog_connect(struct obd_export *exp,
2491 struct llogd_conn_body *body)
2493 struct obd_device *obd = exp->exp_obd;
2494 struct llog_ctxt *ctxt;
2495 struct obd_llog_group *olg;
2499 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2500 obd->obd_name, body->lgdc_logid.lgl_oid,
2501 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2503 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2505 CERROR(" %s: can not find olg of group %d\n",
2506 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2509 llog_group_set_export(olg, exp);
2511 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2512 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2513 body->lgdc_ctxt_idx);
2515 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2516 obd->obd_name, body->lgdc_logid.lgl_oid,
2517 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2519 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2520 obd->u.filter.fo_mds_ost_sync = 1;
2521 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2522 rc = llog_connect(ctxt, &body->lgdc_logid,
2523 &body->lgdc_gen, NULL);
2524 llog_ctxt_put(ctxt);
2526 CERROR("failed to connect rc %d idx %d\n", rc,
2527 body->lgdc_ctxt_idx);
2532 static int filter_llog_preclean(struct obd_device *obd)
2534 struct obd_llog_group *olg, *tmp;
2535 struct filter_obd *filter;
2536 cfs_list_t remove_list;
2540 rc = obd_llog_finish(obd, 0);
2542 CERROR("failed to cleanup llogging subsystem\n");
2544 filter = &obd->u.filter;
2545 CFS_INIT_LIST_HEAD(&remove_list);
2547 cfs_spin_lock(&filter->fo_llog_list_lock);
2548 while (!cfs_list_empty(&filter->fo_llog_list)) {
2549 olg = cfs_list_entry(filter->fo_llog_list.next,
2550 struct obd_llog_group, olg_list);
2551 cfs_list_del(&olg->olg_list);
2552 cfs_list_add(&olg->olg_list, &remove_list);
2554 cfs_spin_unlock(&filter->fo_llog_list_lock);
2556 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2557 cfs_list_del_init(&olg->olg_list);
2558 rc = filter_olg_fini(olg);
2560 CERROR("failed to cleanup llogging subsystem for %u\n",
2568 static int filter_precleanup(struct obd_device *obd,
2569 enum obd_cleanup_stage stage)
2575 case OBD_CLEANUP_EARLY:
2577 case OBD_CLEANUP_EXPORTS:
2578 /* Stop recovery before namespace cleanup. */
2579 target_recovery_fini(obd);
2580 rc = filter_llog_preclean(obd);
2586 static int filter_cleanup(struct obd_device *obd)
2588 struct filter_obd *filter = &obd->u.filter;
2592 LCONSOLE_WARN("%s: shutting down for failover; client state "
2593 "will be preserved.\n", obd->obd_name);
2595 obd_exports_barrier(obd);
2596 obd_zombie_barrier();
2598 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2599 lprocfs_free_per_client_stats(obd);
2600 lprocfs_free_obd_stats(obd);
2601 lprocfs_obd_cleanup(obd);
2602 lquota_cleanup(filter_quota_interface_ref, obd);
2604 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2605 obd->obd_namespace = NULL;
2607 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2609 if (obd->u.obt.obt_sb == NULL)
2614 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2615 shrink_dcache_sb(obd->u.obt.obt_sb);
2617 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2618 obd->u.obt.obt_sb = NULL;
2620 fsfilt_put_ops(obd->obd_fsops);
2622 filter_iobuf_pool_done(filter);
2624 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2629 static int filter_connect_internal(struct obd_export *exp,
2630 struct obd_connect_data *data,
2633 struct filter_export_data *fed = &exp->exp_filter_data;
2638 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2639 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2640 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2641 data->ocd_connect_flags, data->ocd_version,
2642 data->ocd_grant, data->ocd_index);
2644 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2645 CWARN("!!! This export (nid %s) used object group %d "
2646 "earlier; now it's trying to use group %d! This could "
2647 "be a bug in the MDS. Please report to "
2648 "http://bugzilla.lustre.org/\n",
2649 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2652 fed->fed_group = data->ocd_group;
2654 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2655 exp->exp_connect_flags = data->ocd_connect_flags;
2656 data->ocd_version = LUSTRE_VERSION_CODE;
2658 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2659 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2660 data->ocd_connect_flags & OBD_CONNECT_MDS))
2663 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2664 struct filter_obd *filter = &exp->exp_obd->u.filter;
2665 obd_size left, want;
2667 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2668 left = filter_grant_space_left(exp);
2669 want = data->ocd_grant;
2670 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2671 data->ocd_grant = fed->fed_grant;
2672 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2674 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2675 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2676 exp->exp_client_uuid.uuid, exp,
2677 data->ocd_grant, want, left);
2679 filter->fo_tot_granted_clients ++;
2682 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2683 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2684 int index = le32_to_cpu(lsd->lsd_ost_index);
2686 if (!(lsd->lsd_feature_compat &
2687 cpu_to_le32(OBD_COMPAT_OST))) {
2688 /* this will only happen on the first connect */
2689 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2690 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2691 /* sync is not needed here as filter_client_add will
2692 * set exp_need_sync flag */
2693 filter_update_server_data(exp->exp_obd);
2694 } else if (index != data->ocd_index) {
2695 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2696 " %u doesn't match actual OST index"
2697 " %u in last_rcvd file, bad "
2699 obd_export_nid2str(exp), index,
2703 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2704 * FIXME: We don't strictly need the COMPAT flag for that,
2705 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2706 * FIXME: We needed it for the index, as index 0 is valid. */
2709 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2710 data->ocd_brw_size = 65536;
2711 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2712 data->ocd_brw_size = min(data->ocd_brw_size,
2713 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2714 LASSERT(data->ocd_brw_size);
2717 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2718 __u32 cksum_types = data->ocd_cksum_types;
2720 /* The client set in ocd_cksum_types the checksum types it
2721 * supports. We have to mask off the algorithms that we don't
2723 if (cksum_types & OBD_CKSUM_ALL)
2724 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2726 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2728 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2729 "%x\n", exp->exp_obd->obd_name,
2730 obd_export_nid2str(exp), cksum_types,
2731 data->ocd_cksum_types);
2733 /* This client does not support OBD_CONNECT_CKSUM
2734 * fall back to CRC32 */
2735 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2736 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2737 exp->exp_obd->obd_name,
2738 obd_export_nid2str(exp));
2744 static int filter_reconnect(const struct lu_env *env,
2745 struct obd_export *exp, struct obd_device *obd,
2746 struct obd_uuid *cluuid,
2747 struct obd_connect_data *data,
2753 if (exp == NULL || obd == NULL || cluuid == NULL)
2756 rc = filter_connect_internal(exp, data, 1);
2758 filter_export_stats_init(obd, exp, localdata);
2763 /* nearly identical to mds_connect */
2764 static int filter_connect(const struct lu_env *env,
2765 struct obd_export **exp, struct obd_device *obd,
2766 struct obd_uuid *cluuid,
2767 struct obd_connect_data *data, void *localdata)
2769 struct lvfs_run_ctxt saved;
2770 struct lustre_handle conn = { 0 };
2771 struct obd_export *lexp;
2776 if (exp == NULL || obd == NULL || cluuid == NULL)
2779 rc = class_connect(&conn, obd, cluuid);
2782 lexp = class_conn2export(&conn);
2783 LASSERT(lexp != NULL);
2785 rc = filter_connect_internal(lexp, data, 0);
2789 filter_export_stats_init(obd, lexp, localdata);
2790 if (obd->obd_replayable) {
2791 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2793 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2794 rc = filter_client_add(obd, lexp, -1);
2799 group = data->ocd_group;
2801 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2802 obd->obd_name, lexp->exp_handle.h_cookie, group);
2804 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2805 rc = filter_read_groups(obd, group, 1);
2806 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2808 CERROR("can't read group %u\n", group);
2816 class_disconnect(lexp);
2817 lprocfs_exp_cleanup(lexp);
2826 /* Do extra sanity checks for grant accounting. We do this at connect,
2827 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2828 * always get rid of it or turn it off when we know accounting is good. */
2829 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2831 struct filter_export_data *fed;
2832 struct obd_export *exp;
2833 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2834 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2835 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2837 if (cfs_list_empty(&obd->obd_exports))
2840 /* We don't want to do this for large machines that do lots of
2841 mounts or unmounts. It burns... */
2842 if (obd->obd_num_exports > 100)
2845 cfs_spin_lock(&obd->obd_osfs_lock);
2846 cfs_spin_lock(&obd->obd_dev_lock);
2847 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2849 fed = &exp->exp_filter_data;
2850 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2853 if (maxsize > 0) { /* we may not have done a statfs yet */
2854 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2855 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2856 exp->exp_client_uuid.uuid, exp,
2857 fed->fed_grant, fed->fed_pending, maxsize);
2858 LASSERTF(fed->fed_dirty <= maxsize,
2859 "%s: cli %s/%p %ld > "LPU64"\n", func,
2860 exp->exp_client_uuid.uuid, exp,
2861 fed->fed_dirty, maxsize);
2864 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2865 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2866 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2868 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2869 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2870 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2871 tot_granted += fed->fed_grant + fed->fed_pending;
2872 tot_pending += fed->fed_pending;
2873 tot_dirty += fed->fed_dirty;
2875 fo_tot_granted = obd->u.filter.fo_tot_granted;
2876 fo_tot_pending = obd->u.filter.fo_tot_pending;
2877 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2878 cfs_spin_unlock(&obd->obd_dev_lock);
2879 cfs_spin_unlock(&obd->obd_osfs_lock);
2881 /* Do these assertions outside the spinlocks so we don't kill system */
2882 if (tot_granted != fo_tot_granted)
2883 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2884 func, tot_granted, fo_tot_granted);
2885 if (tot_pending != fo_tot_pending)
2886 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2887 func, tot_pending, fo_tot_pending);
2888 if (tot_dirty != fo_tot_dirty)
2889 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2890 func, tot_dirty, fo_tot_dirty);
2891 if (tot_pending > tot_granted)
2892 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2893 func, tot_pending, tot_granted);
2894 if (tot_granted > maxsize)
2895 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2896 func, tot_granted, maxsize);
2897 if (tot_dirty > maxsize)
2898 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2899 func, tot_dirty, maxsize);
2902 /* Remove this client from the grant accounting totals. We also remove
2903 * the export from the obd device under the osfs and dev locks to ensure
2904 * that the filter_grant_sanity_check() calculations are always valid.
2905 * The client should do something similar when it invalidates its import. */
2906 static void filter_grant_discard(struct obd_export *exp)
2908 struct obd_device *obd = exp->exp_obd;
2909 struct filter_obd *filter = &obd->u.filter;
2910 struct filter_export_data *fed = &exp->exp_filter_data;
2912 cfs_spin_lock(&obd->obd_osfs_lock);
2913 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2914 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2915 obd->obd_name, filter->fo_tot_granted,
2916 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2917 filter->fo_tot_granted -= fed->fed_grant;
2918 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2919 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2920 obd->obd_name, filter->fo_tot_pending,
2921 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2922 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2923 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2924 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2925 obd->obd_name, filter->fo_tot_dirty,
2926 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2927 filter->fo_tot_dirty -= fed->fed_dirty;
2931 cfs_spin_unlock(&obd->obd_osfs_lock);
2934 static int filter_destroy_export(struct obd_export *exp)
2936 struct filter_export_data *fed = &exp->exp_filter_data;
2939 if (fed->fed_pending)
2940 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2941 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2942 exp, fed->fed_pending);
2944 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2946 target_destroy_export(exp);
2947 ldlm_destroy_export(exp);
2948 lut_client_free(exp);
2950 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2953 if (!exp->exp_obd->obd_replayable)
2954 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2956 filter_grant_discard(exp);
2957 filter_fmd_cleanup(exp);
2959 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2960 struct filter_obd *filter = &exp->exp_obd->u.filter;
2961 if (filter->fo_tot_granted_clients > 0)
2962 filter->fo_tot_granted_clients --;
2965 if (!(exp->exp_flags & OBD_OPT_FORCE))
2966 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2971 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2973 struct obd_llog_group *olg_min, *olg;
2974 struct filter_obd *filter;
2975 int worked = -1, group;
2976 struct llog_ctxt *ctxt;
2979 filter = &obd->u.filter;
2981 /* we can't sync log holding spinlock. also, we do not want to get
2982 * into livelock. so we do following: loop over MDS's exports in
2983 * group order and skip already synced llogs -bzzz */
2985 /* look for group with min. number, but > worked */
2988 cfs_spin_lock(&filter->fo_llog_list_lock);
2989 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2990 if (olg->olg_group <= worked) {
2991 /* this group is already synced */
2994 if (group < olg->olg_group) {
2995 /* we have group with smaller number to sync */
2998 /* store current minimal group */
3000 group = olg->olg_group;
3002 cfs_spin_unlock(&filter->fo_llog_list_lock);
3004 if (olg_min == NULL)
3007 worked = olg_min->olg_group;
3008 if (olg_min->olg_exp &&
3009 (dexp == olg_min->olg_exp || dexp == NULL)) {
3011 ctxt = llog_group_get_ctxt(olg_min,
3012 LLOG_MDS_OST_REPL_CTXT);
3014 err = llog_sync(ctxt, olg_min->olg_exp);
3015 llog_ctxt_put(ctxt);
3017 CERROR("error flushing logs to MDS: "
3022 } while (olg_min != NULL);
3025 /* Also incredibly similar to mds_disconnect */
3026 static int filter_disconnect(struct obd_export *exp)
3028 struct obd_device *obd = exp->exp_obd;
3033 class_export_get(exp);
3035 if (!(exp->exp_flags & OBD_OPT_FORCE))
3036 filter_grant_sanity_check(obd, __FUNCTION__);
3037 filter_grant_discard(exp);
3039 /* Flush any remaining cancel messages out to the target */
3040 filter_sync_llogs(obd, exp);
3042 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3044 rc = server_disconnect_export(exp);
3046 /* Do not erase record for recoverable client. */
3047 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3048 filter_client_del(exp);
3050 fsfilt_sync(obd, obd->u.obt.obt_sb);
3052 class_export_put(exp);
3056 /* reverse import is changed, sync all cancels */
3057 static void filter_revimp_update(struct obd_export *exp)
3062 class_export_get(exp);
3064 /* flush any remaining cancel messages out to the target */
3065 filter_sync_llogs(exp->exp_obd, exp);
3066 class_export_put(exp);
3070 static int filter_ping(struct obd_export *exp)
3072 filter_fmd_expire(exp);
3076 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3077 const char *what, int quiet)
3079 struct dentry *dchild = NULL;
3082 if (oa->o_valid & OBD_MD_FLGROUP)
3085 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3087 if (IS_ERR(dchild)) {
3088 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3089 what, group, oa->o_id);
3093 if (dchild->d_inode == NULL) {
3095 CERROR("%s: %s on non-existent object: "LPU64"\n",
3096 obd->obd_name, what, oa->o_id);
3098 RETURN(ERR_PTR(-ENOENT));
3104 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3106 struct dentry *dentry = NULL;
3107 struct obd_device *obd;
3111 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3112 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3113 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3117 obd = class_exp2obd(exp);
3119 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3123 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3125 RETURN(PTR_ERR(dentry));
3127 /* Limit the valid bits in the return data to what we actually use */
3128 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3129 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3135 /* this should be enabled/disabled in condition to enabled/disabled large
3136 * inodes (fast EAs) in backing store FS. */
3137 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3138 void *handle, struct obdo *oa)
3140 struct obd_device *obd = exp->exp_obd;
3144 if (oa->o_valid & OBD_MD_FLFID) {
3145 struct filter_fid ff;
3147 if (!(oa->o_valid & OBD_MD_FLGROUP))
3149 /* packing fid and converting it to LE for storing into EA.
3150 * Here ->o_stripe_idx should be filled by LOV and rest of
3151 * fields - by client. */
3152 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3153 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3154 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3155 ff.ff_objid = cpu_to_le64(oa->o_id);
3156 ff.ff_group = cpu_to_le64(oa->o_gr);
3158 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3159 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3160 oa->o_generation, oa->o_id, oa->o_gr);
3162 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3164 CERROR("store fid in object failed! rc: %d\n", rc);
3166 CDEBUG(D_HA, "OSS object without fid info!\n");
3172 /* this is called from filter_truncate() until we have filter_punch() */
3173 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3174 struct obdo *oa, struct obd_trans_info *oti)
3176 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3177 struct llog_cookie *fcc = NULL;
3178 struct filter_obd *filter;
3179 int rc, err, sync = 0;
3180 loff_t old_size = 0;
3181 unsigned int ia_valid;
3182 struct inode *inode;
3183 struct page *page = NULL;
3188 LASSERT(dentry != NULL);
3189 LASSERT(!IS_ERR(dentry));
3191 inode = dentry->d_inode;
3192 LASSERT(inode != NULL);
3194 filter = &exp->exp_obd->u.filter;
3195 iattr_from_obdo(&iattr, oa, oa->o_valid);
3196 ia_valid = iattr.ia_valid;
3198 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3199 OBD_ALLOC(fcc, sizeof(*fcc));
3201 *fcc = oa->o_lcookie;
3203 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3205 /* Filter truncates and writes are serialized by
3206 * i_alloc_sem, see the comment in
3207 * filter_preprw_write.*/
3208 if (ia_valid & ATTR_SIZE)
3209 down_write(&inode->i_alloc_sem);
3210 LOCK_INODE_MUTEX(inode);
3211 old_size = i_size_read(inode);
3214 /* VBR: version recovery check */
3215 rc = filter_version_get_check(exp, oti, inode);
3217 GOTO(out_unlock, rc);
3219 /* Let's pin the last page so that ldiskfs_truncate
3220 * should not start GFP_FS allocation. */
3221 if (ia_valid & ATTR_SIZE) {
3222 page = grab_cache_page(inode->i_mapping,
3223 iattr.ia_size >> PAGE_CACHE_SHIFT);
3225 GOTO(out_unlock, rc = -ENOMEM);
3230 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3231 * then we will accept the UID+GID sent by the client during write for
3232 * initializing the ownership of this inode. We only allow this to
3233 * happen once so clear these bits in setattr. In 2.6 kernels it is
3234 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3235 * the flags that are actually being set. */
3236 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3237 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3238 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3240 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3241 if (!(ia_valid & ATTR_MODE)) {
3242 iattr.ia_mode = inode->i_mode;
3243 iattr.ia_valid |= ATTR_MODE;
3245 iattr.ia_mode &= ~S_ISUID;
3247 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3248 if (!(iattr.ia_valid & ATTR_MODE)) {
3249 iattr.ia_mode = inode->i_mode;
3250 iattr.ia_valid |= ATTR_MODE;
3252 iattr.ia_mode &= ~S_ISGID;
3255 orig_ids[USRQUOTA] = inode->i_uid;
3256 orig_ids[GRPQUOTA] = inode->i_gid;
3257 handle = fsfilt_start_log(exp->exp_obd, inode,
3258 FSFILT_OP_SETATTR, oti, 1);
3260 GOTO(out_unlock, rc = PTR_ERR(handle));
3262 /* update inode EA only once when inode is suid bit marked. As
3263 * on 2.6.x UID and GID may be set separately, we check here
3264 * only one of them to avoid double setting. */
3265 if (inode->i_mode & S_ISUID)
3266 filter_update_fidea(exp, inode, handle, oa);
3268 handle = fsfilt_start(exp->exp_obd, inode,
3269 FSFILT_OP_SETATTR, oti);
3271 GOTO(out_unlock, rc = PTR_ERR(handle));
3273 if (oa->o_valid & OBD_MD_FLFLAGS) {
3274 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3275 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3277 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3279 /* set cancel cookie callback function */
3280 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3281 filter_cancel_cookies_cb,
3285 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3286 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3288 /* The truncate might have used up our transaction credits. Make
3289 * sure we have one left for the last_rcvd update. */
3290 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3292 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3294 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3298 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3300 CERROR("error on commit, err = %d\n", err);
3307 /* For a partial-page truncate flush the page to disk immediately
3308 * to avoid data corruption during direct disk write. b=17397 */
3309 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3310 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3311 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3321 page_cache_release(page);
3323 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3324 UNLOCK_INODE_MUTEX(inode);
3325 if (ia_valid & ATTR_SIZE)
3326 up_write(&inode->i_alloc_sem);
3328 OBD_FREE(fcc, sizeof(*fcc));
3330 /* trigger quota release */
3331 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3332 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3333 int rc2 = lquota_adjust(filter_quota_interface_ref,
3334 exp->exp_obd, cur_ids,
3335 orig_ids, rc, FSFILT_OP_SETATTR);
3336 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3337 "filter adjust qunit. (rc:%d)\n", rc2);
3342 /* this is called from filter_truncate() until we have filter_punch() */
3343 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3344 struct obd_trans_info *oti)
3346 struct obdo *oa = oinfo->oi_oa;
3347 struct lustre_capa *capa = oinfo_capa(oinfo);
3348 struct ldlm_res_id res_id;
3349 struct filter_mod_data *fmd;
3350 struct lvfs_run_ctxt saved;
3351 struct filter_obd *filter;
3352 struct ldlm_resource *res;
3353 struct dentry *dentry;
3354 __u64 opc = CAPA_OPC_META_WRITE;
3358 if (oa->o_valid & OBD_FL_TRUNC)
3359 opc |= CAPA_OPC_OSS_TRUNC;
3361 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3362 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3366 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3367 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3372 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3373 /* This would be very bad - accidentally truncating a file when
3374 * changing the time or similar - bug 12203. */
3375 if (oa->o_valid & OBD_MD_FLSIZE &&
3376 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3377 static char mdsinum[48];
3379 if (oa->o_valid & OBD_MD_FLFID)
3380 snprintf(mdsinum, sizeof(mdsinum) - 1,
3381 " of inode "LPU64"/%u", oa->o_fid,
3386 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3388 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3393 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3395 RETURN(PTR_ERR(dentry));
3397 filter = &exp->exp_obd->u.filter;
3398 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3401 * We need to be atomic against a concurrent write
3402 * (which takes the semaphore for reading). fmd_mactime_xid
3403 * checks will have no effect if a write request with lower
3404 * xid starts just before a setattr and finishes later than
3405 * the setattr (see bug 21489, comment 27).
3408 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3409 down_write(&dentry->d_inode->i_alloc_sem);
3410 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3411 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3412 fmd->fmd_mactime_xid = oti->oti_xid;
3413 filter_fmd_put(exp, fmd);
3414 up_write(&dentry->d_inode->i_alloc_sem);
3417 /* setting objects attributes (including owner/group) */
3418 rc = filter_setattr_internal(exp, dentry, oa, oti);
3420 GOTO(out_unlock, rc);
3422 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3423 &res_id, LDLM_EXTENT, 0);
3426 LDLM_RESOURCE_ADDREF(res);
3427 rc = ldlm_res_lvbo_update(res, NULL, 0);
3428 LDLM_RESOURCE_DELREF(res);
3429 ldlm_resource_putref(res);
3432 oa->o_valid = OBD_MD_FLID;
3434 /* Quota release need uid/gid info */
3435 obdo_from_inode(oa, dentry->d_inode,
3436 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3441 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3445 /* XXX identical to osc_unpackmd */
3446 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3447 struct lov_mds_md *lmm, int lmm_bytes)
3453 if (lmm_bytes < sizeof (*lmm)) {
3454 CERROR("lov_mds_md too small: %d, need %d\n",
3455 lmm_bytes, (int)sizeof(*lmm));
3458 /* XXX LOV_MAGIC etc check? */
3460 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3461 CERROR("lov_mds_md: zero lmm_object_id\n");
3466 lsm_size = lov_stripe_md_size(1);
3470 if (*lsmp != NULL && lmm == NULL) {
3471 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3472 OBD_FREE(*lsmp, lsm_size);
3477 if (*lsmp == NULL) {
3478 OBD_ALLOC(*lsmp, lsm_size);
3482 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3483 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3484 OBD_FREE(*lsmp, lsm_size);
3487 loi_init((*lsmp)->lsm_oinfo[0]);
3491 /* XXX zero *lsmp? */
3492 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3493 LASSERT((*lsmp)->lsm_object_id);
3496 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3501 /* caller must hold fo_create_locks[oa->o_gr] */
3502 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3503 struct filter_obd *filter)
3505 struct obdo doa = { 0 }; /* XXX obdo on stack */
3512 LASSERT_MDS_GROUP(oa->o_gr);
3513 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3514 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3516 memset(&doa, 0, sizeof(doa));
3518 doa.o_valid |= OBD_MD_FLGROUP;
3519 doa.o_gr = oa->o_gr;
3520 doa.o_mode = S_IFREG;
3522 if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3523 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3524 exp->exp_obd->obd_name, doa.o_gr);
3528 last = filter_last_id(filter, doa.o_gr);
3530 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3532 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3533 exp->exp_obd->obd_name, oa->o_id + 1, last,
3534 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3536 for (id = last; id > oa->o_id; id--) {
3538 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3539 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3540 CEMERG("error destroying precreate objid "LPU64": %d\n",
3543 /* update last_id on disk periodically so that if we restart
3544 * we don't need to re-scan all of the just-deleted objects. */
3545 if ((id & 511) == 0 && !skip_orphan) {
3546 filter_set_last_id(filter, id - 1, doa.o_gr);
3547 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3551 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3552 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3555 filter_set_last_id(filter, id, doa.o_gr);
3556 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3558 /* don't reuse orphan object, return last used objid */
3562 cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3567 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3568 obd_gr group, int *num);
3569 /* returns a negative error or a nonnegative number of files to create */
3570 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3571 obd_gr group, struct obd_trans_info *oti)
3573 struct obd_device *obd = exp->exp_obd;
3574 struct filter_obd *filter = &obd->u.filter;
3578 /* delete orphans request */
3579 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3580 obd_id last = filter_last_id(filter, group);
3582 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3583 CERROR("%s: dropping old orphan cleanup request\n",
3587 /* This causes inflight precreates to abort and drop lock */
3588 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3589 cfs_down(&filter->fo_create_locks[group]);
3590 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3591 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3592 exp->exp_obd->obd_name, group);
3593 cfs_up(&filter->fo_create_locks[group]);
3596 diff = oa->o_id - last;
3597 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3600 if (-diff > OST_MAX_PRECREATE) {
3601 CERROR("%s: ignoring bogus orphan destroy request: "
3602 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3604 /* FIXME: should reset precreate_next_id on MDS */
3605 GOTO(out, rc = -EINVAL);
3608 rc = filter_destroy_precreated(exp, oa, filter);
3610 CERROR("%s: unable to write lastobjid, but "
3611 "orphans were deleted\n", obd->obd_name);
3614 /* XXX: Used by MDS for the first time! */
3615 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3618 cfs_down(&filter->fo_create_locks[group]);
3619 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3620 CERROR("%s: dropping old precreate request\n",
3624 /* only precreate if group == 0 and o_id is specfied */
3625 if (!filter_group_is_mds(group) || oa->o_id == 0)
3628 diff = oa->o_id - filter_last_id(filter, group);
3629 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3630 filter_last_id(filter, group), diff);
3632 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3633 oa->o_id, filter_last_id(filter, group), diff);
3637 oa->o_id = filter_last_id(&obd->u.filter, group);
3638 rc = filter_precreate(obd, oa, group, &diff);
3639 oa->o_id = filter_last_id(&obd->u.filter, group);
3641 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3644 /* else diff == 0 */
3647 cfs_up(&filter->fo_create_locks[group]);
3651 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3652 __u64 max_age, __u32 flags)
3654 struct filter_obd *filter = &obd->u.filter;
3655 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3659 /* at least try to account for cached pages. its still racey and
3660 * might be under-reporting if clients haven't announced their
3661 * caches with brw recently */
3662 cfs_spin_lock(&obd->obd_osfs_lock);
3663 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3664 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3665 cfs_spin_unlock(&obd->obd_osfs_lock);
3667 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3668 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3669 filter->fo_tot_dirty, filter->fo_tot_granted,
3670 filter->fo_tot_pending,
3671 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3673 filter_grant_sanity_check(obd, __FUNCTION__);
3675 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3676 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3677 osfs->os_bsize - 1) >> blockbits));
3679 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3680 struct lr_server_data *lsd = class_server_data(obd);
3681 int index = le32_to_cpu(lsd->lsd_ost_index);
3683 if (obd_fail_val == -1 ||
3684 index == obd_fail_val)
3685 osfs->os_bfree = osfs->os_bavail = 2;
3686 else if (obd_fail_loc & OBD_FAIL_ONCE)
3687 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3690 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3691 * stop creating files on MDS if OST is not good shape to create
3695 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3696 osfs->os_state = OS_STATE_READONLY;
3698 if (filter->fo_raid_degraded)
3699 osfs->os_state |= OS_STATE_DEGRADED;
3703 static int filter_use_existing_obj(struct obd_device *obd,
3704 struct dentry *dchild, void **handle,
3707 struct inode *inode = dchild->d_inode;
3711 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3714 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3715 if (IS_ERR(*handle))
3716 return PTR_ERR(*handle);
3718 iattr.ia_valid = ATTR_MODE;
3719 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3720 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3727 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3730 __u64 os_ffree = -1;
3732 cfs_spin_lock(&obd->obd_osfs_lock);
3733 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3735 os_ffree = obd->obd_osfs.os_ffree;
3736 cfs_spin_unlock(&obd->obd_osfs_lock);
3741 /* We rely on the fact that only one thread will be creating files in a given
3742 * group at a time, which is why we don't need an atomic filter_get_new_id.
3743 * Even if we had that atomic function, the following race would exist:
3745 * thread 1: gets id x from filter_next_id
3746 * thread 2: gets id (x + 1) from filter_next_id
3747 * thread 2: creates object (x + 1)
3748 * thread 1: tries to create object x, gets -ENOSPC
3750 * Caller must hold fo_create_locks[group]
3752 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3753 obd_gr group, int *num)
3755 struct dentry *dchild = NULL, *dparent = NULL;
3756 struct filter_obd *filter;
3757 struct obd_statfs *osfs;
3758 int err = 0, rc = 0, recreate_obj = 0, i;
3759 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3762 void *handle = NULL;
3765 filter = &obd->u.filter;
3767 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3769 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3771 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3772 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3775 OBD_ALLOC(osfs, sizeof(*osfs));
3778 rc = filter_statfs(obd, osfs, cfs_time_current_64() - CFS_HZ,
3780 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3781 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3782 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3783 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3787 OBD_FREE(osfs, sizeof(*osfs));
3792 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3793 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3795 for (i = 0; i < *num && err == 0; i++) {
3796 int cleanup_phase = 0;
3798 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3799 CWARN("%s: create aborted by destroy\n",
3808 last_id = filter_last_id(filter, group);
3809 if (next_id > last_id) {
3810 CERROR("Error: Trying to recreate obj greater"
3811 "than last id "LPD64" > "LPD64"\n",
3813 GOTO(cleanup, rc = -EINVAL);
3816 next_id = filter_last_id(filter, group) + 1;
3818 dparent = filter_parent_lock(obd, group, next_id);
3819 if (IS_ERR(dparent))
3820 GOTO(cleanup, rc = PTR_ERR(dparent));
3821 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3823 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3825 GOTO(cleanup, rc = PTR_ERR(dchild));
3826 cleanup_phase = 2; /* f_dput(dchild) */
3828 if (dchild->d_inode != NULL) {
3829 /* This would only happen if lastobjid was bad on disk*/
3830 /* Could also happen if recreating missing obj but it
3831 * already exists. */
3833 CERROR("%s: recreating existing object %.*s?\n",
3834 obd->obd_name, dchild->d_name.len,
3835 dchild->d_name.name);
3837 /* Use these existing objects if they are
3839 if (dchild->d_inode->i_size == 0) {
3840 rc = filter_use_existing_obj(obd,dchild,
3841 &handle, &cleanup_phase);
3848 CERROR("%s: Serious error: objid %.*s already "
3849 "exists; is this filesystem corrupt?\n",
3850 obd->obd_name, dchild->d_name.len,
3851 dchild->d_name.name);
3854 GOTO(cleanup, rc = -EEXIST);
3857 handle = fsfilt_start_log(obd, dparent->d_inode,
3858 FSFILT_OP_CREATE, NULL, 1);
3860 GOTO(cleanup, rc = PTR_ERR(handle));
3863 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3864 ",od->o_id="LPU64")\n", obd->obd_name, group,
3867 /* We mark object SUID+SGID to flag it for accepting UID+GID
3868 * from client on first write. Currently the permission bits
3869 * on the OST are never used, so this is OK. */
3870 rc = ll_vfs_create(dparent->d_inode, dchild,
3871 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3873 CERROR("create failed rc = %d\n", rc);
3874 if (rc == -ENOSPC) {
3875 os_ffree = filter_calc_free_inodes(obd);
3877 CERROR("%s: free inode "LPU64"\n",
3878 obd->obd_name, os_ffree);
3883 if (dchild->d_inode)
3884 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3885 dchild->d_inode->i_ino);
3888 if (!recreate_obj) {
3889 filter_set_last_id(filter, next_id, group);
3890 err = filter_update_last_objid(obd, group, 0);
3892 CERROR("unable to write lastobjid "
3893 "but file created\n");
3897 switch(cleanup_phase) {
3899 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3901 CERROR("error on commit, err = %d\n", err);
3908 filter_parent_unlock(dparent);
3915 if (cfs_time_after(jiffies, enough_time)) {
3917 "%s: precreate slow - want %d got %d \n",
3918 obd->obd_name, *num, i);
3925 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3926 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3931 static int filter_create(struct obd_export *exp, struct obdo *oa,
3932 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3934 struct obd_device *obd = exp->exp_obd;
3935 struct filter_export_data *fed;
3936 struct filter_obd *filter;
3937 struct lvfs_run_ctxt saved;
3938 struct lov_stripe_md *lsm = NULL;
3939 int rc = 0, diff, group = oa->o_gr;
3942 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3943 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3945 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3946 CERROR("!!! nid %s sent invalid object group %d\n",
3947 obd_export_nid2str(exp), group);
3951 fed = &exp->exp_filter_data;
3952 filter = &obd->u.filter;
3954 if (fed->fed_group != group) {
3955 CERROR("!!! this export (nid %s) used object group %d "
3956 "earlier; now it's trying to use group %d! This could "
3957 "be a bug in the MDS. Please report to "
3958 "http://bugzilla.lustre.org/\n",
3959 obd_export_nid2str(exp), fed->fed_group, group);
3966 rc = obd_alloc_memmd(exp, &lsm);
3973 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3975 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3976 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3977 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3978 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3979 oa->o_id, filter_last_id(filter,
3984 cfs_down(&filter->fo_create_locks[oa->o_gr]);
3985 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3986 cfs_up(&filter->fo_create_locks[oa->o_gr]);
3989 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3992 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3993 if (rc && ea != NULL && *ea != lsm) {
3994 obd_free_memmd(exp, &lsm);
3995 } else if (rc == 0 && ea != NULL) {
3996 /* XXX LOV STACKING: the lsm that is passed to us from
3997 * LOV does not have valid lsm_oinfo data structs, so
3998 * don't go touching that. This needs to be fixed in a
4000 lsm->lsm_object_id = oa->o_id;
4007 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4008 struct lov_stripe_md *md, struct obd_trans_info *oti,
4009 struct obd_export *md_exp, void *capa)
4011 unsigned int qcids[MAXQUOTAS] = {0, 0};
4012 struct obd_device *obd;
4013 struct filter_obd *filter;
4014 struct dentry *dchild = NULL, *dparent = NULL;
4015 struct lustre_handle lockh = { 0 };
4016 struct lvfs_run_ctxt saved;
4017 void *handle = NULL;
4018 struct llog_cookie *fcc = NULL;
4019 int rc, rc2, cleanup_phase = 0, sync = 0;
4023 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4024 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4025 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4030 filter = &obd->u.filter;
4032 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4035 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4036 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4038 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4040 GOTO(cleanup, rc = PTR_ERR(dchild));
4043 if (dchild->d_inode == NULL) {
4044 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4046 /* If object already gone, cancel cookie right now */
4047 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4048 struct llog_ctxt *ctxt;
4049 struct obd_llog_group *olg;
4051 olg = filter_find_olg(obd, oa->o_gr);
4053 CERROR(" %s: can not find olg of group %d\n",
4054 obd->obd_name, (int)oa->o_gr);
4055 GOTO(cleanup, rc = PTR_ERR(olg));
4057 fcc = &oa->o_lcookie;
4058 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4059 llog_cancel(ctxt, NULL, 1, fcc, 0);
4060 llog_ctxt_put(ctxt);
4061 fcc = NULL; /* we didn't allocate fcc, don't free it */
4063 GOTO(cleanup, rc = -ENOENT);
4066 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4070 /* Our MDC connection is established by the MDS to us */
4071 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4072 OBD_ALLOC(fcc, sizeof(*fcc));
4074 *fcc = oa->o_lcookie;
4076 DQUOT_INIT(dchild->d_inode);
4078 /* we're gonna truncate it first in order to avoid possible deadlock:
4080 * open trasaction open transaction
4081 * down(i_zombie) down(i_zombie)
4082 * restart transaction
4083 * (see BUG 4180) -bzzz
4085 * take i_alloc_sem too to prevent other threads from writing to the
4086 * file while we are truncating it. This can cause lock ordering issue
4087 * between page lock, i_mutex & starting new journal handle.
4088 * (see bug 20321) -johann
4090 down_write(&dchild->d_inode->i_alloc_sem);
4091 LOCK_INODE_MUTEX(dchild->d_inode);
4093 /* VBR: version recovery check */
4094 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4096 UNLOCK_INODE_MUTEX(dchild->d_inode);
4097 up_write(&dchild->d_inode->i_alloc_sem);
4101 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4103 if (IS_ERR(handle)) {
4104 UNLOCK_INODE_MUTEX(dchild->d_inode);
4105 up_write(&dchild->d_inode->i_alloc_sem);
4106 GOTO(cleanup, rc = PTR_ERR(handle));
4109 iattr.ia_valid = ATTR_SIZE;
4111 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4112 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4113 UNLOCK_INODE_MUTEX(dchild->d_inode);
4114 up_write(&dchild->d_inode->i_alloc_sem);
4118 GOTO(cleanup, rc = rc2);
4120 /* We don't actually need to lock the parent until we are unlinking
4121 * here, and not while truncating above. That avoids holding the
4122 * parent lock for a long time during truncate, which can block other
4123 * threads from doing anything to objects in that directory. bug 7171 */
4124 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4125 if (IS_ERR(dparent))
4126 GOTO(cleanup, rc = PTR_ERR(dparent));
4127 cleanup_phase = 3; /* filter_parent_unlock */
4129 LOCK_INODE_MUTEX(dchild->d_inode);
4130 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4131 if (IS_ERR(handle)) {
4132 UNLOCK_INODE_MUTEX(dchild->d_inode);
4133 GOTO(cleanup, rc = PTR_ERR(handle));
4135 cleanup_phase = 4; /* fsfilt_commit */
4137 /* Quota release need uid/gid of inode */
4138 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4140 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4142 /* this drops dchild->d_inode->i_mutex unconditionally */
4143 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4147 switch(cleanup_phase) {
4150 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4151 oti->oti_handle : handle,
4152 filter_cancel_cookies_cb,
4154 /* If add_journal_cb failed, then filter_finish_transno
4155 * will commit the handle and we will do a sync
4156 * on commit. then we call callback directly to free
4159 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4161 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4164 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4166 CERROR("error on commit, err = %d\n", rc2);
4173 filter_parent_unlock(dparent);
4175 filter_fini_destroy(obd, &lockh);
4179 OBD_FREE(fcc, sizeof(*fcc));
4181 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4184 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4188 /* trigger quota release */
4189 qcids[USRQUOTA] = oa->o_uid;
4190 qcids[GRPQUOTA] = oa->o_gid;
4191 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4194 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4198 /* NB start and end are used for punch, but not truncate */
4199 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4200 struct obd_trans_info *oti,
4201 struct ptlrpc_request_set *rqset)
4206 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4207 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4208 oinfo->oi_policy.l_extent.end);
4212 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4213 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4214 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4216 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4217 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4218 rc = filter_setattr(exp, oinfo, oti);
4219 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4223 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4224 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4227 struct lvfs_run_ctxt saved;
4228 struct obd_device_target *obt;
4229 struct dentry *dentry;
4233 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4234 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4235 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4239 obt = &exp->exp_obd->u.obt;
4241 /* An objid of zero is taken to mean "sync whole filesystem" */
4242 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4243 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4244 /* Flush any remaining cancel messages out to the target */
4245 filter_sync_llogs(exp->exp_obd, exp);
4249 dentry = filter_oa2dentry(exp->exp_obd, oa);
4251 RETURN(PTR_ERR(dentry));
4253 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4255 LOCK_INODE_MUTEX(dentry->d_inode);
4257 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4259 /* just any file to grab fsync method - "file" arg unused */
4260 struct file *file = obt->obt_rcvd_filp;
4262 if (file->f_op && file->f_op->fsync)
4263 rc = file->f_op->fsync(NULL, dentry, 1);
4265 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4269 UNLOCK_INODE_MUTEX(dentry->d_inode);
4271 oa->o_valid = OBD_MD_FLID;
4272 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4274 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4280 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4281 void *key, __u32 *vallen, void *val,
4282 struct lov_stripe_md *lsm)
4284 struct obd_device *obd;
4287 obd = class_exp2obd(exp);
4289 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4293 if (KEY_IS(KEY_BLOCKSIZE)) {
4294 __u32 *blocksize = val;
4296 if (*vallen < sizeof(*blocksize))
4298 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4300 *vallen = sizeof(*blocksize);
4304 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4305 __u32 *blocksize_bits = val;
4306 if (blocksize_bits) {
4307 if (*vallen < sizeof(*blocksize_bits))
4309 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4311 *vallen = sizeof(*blocksize_bits);
4315 if (KEY_IS(KEY_LAST_ID)) {
4316 obd_id *last_id = val;
4317 /* FIXME: object groups */
4319 if (*vallen < sizeof(*last_id))
4321 *last_id = filter_last_id(&obd->u.filter,
4322 exp->exp_filter_data.fed_group);
4324 *vallen = sizeof(*last_id);
4328 if (KEY_IS(KEY_FIEMAP)) {
4329 struct ll_fiemap_info_key *fm_key = key;
4330 struct dentry *dentry;
4331 struct ll_user_fiemap *fiemap = val;
4332 struct lvfs_run_ctxt saved;
4335 if (fiemap == NULL) {
4336 *vallen = fiemap_count_to_size(
4337 fm_key->fiemap.fm_extent_count);
4341 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4344 RETURN(PTR_ERR(dentry));
4346 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4347 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4348 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4350 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4356 CDEBUG(D_IOCTL, "invalid key\n");
4360 static inline int filter_setup_llog_group(struct obd_export *exp,
4361 struct obd_device *obd,
4364 struct obd_llog_group *olg;
4365 struct llog_ctxt *ctxt;
4368 olg = filter_find_create_olg(obd, group);
4370 RETURN(PTR_ERR(olg));
4372 llog_group_set_export(olg, exp);
4374 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4375 LASSERTF(ctxt != NULL, "ctxt is null\n");
4377 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4378 llog_ctxt_put(ctxt);
4382 static int filter_set_grant_shrink(struct obd_export *exp,
4383 struct ost_body *body)
4385 /* handle shrink grant */
4386 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4387 filter_grant_incoming(exp, &body->oa);
4388 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4394 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4396 struct obd_device *obd;
4402 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4406 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4407 obd_export_nid2str(exp));
4408 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4410 /* setup llog imports */
4412 group = (int)(*(__u32 *)val);
4414 group = 0; /* default value */
4416 LASSERT_MDS_GROUP(group);
4417 rc = filter_setup_llog_group(exp, obd, group);
4421 if (group == FILTER_GROUP_MDS0) {
4422 /* setup llog group 1 for interop */
4423 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4426 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4431 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4432 void *key, __u32 vallen, void *val,
4433 struct ptlrpc_request_set *set)
4435 struct obd_device *obd;
4440 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4444 if (KEY_IS(KEY_CAPA_KEY)) {
4446 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4448 CERROR("filter update capability key failed: %d\n", rc);
4452 if (KEY_IS(KEY_REVIMP_UPD)) {
4453 filter_revimp_update(exp);
4454 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4458 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4459 filter_adapt_sptlrpc_conf(obd, 0);
4463 if (KEY_IS(KEY_MDS_CONN))
4464 RETURN(filter_set_mds_conn(exp, val));
4466 if (KEY_IS(KEY_GRANT_SHRINK))
4467 RETURN(filter_set_grant_shrink(exp, val));
4472 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4473 int len, void *karg, void *uarg)
4475 struct obd_device *obd = exp->exp_obd;
4476 struct obd_ioctl_data *data = karg;
4480 case OBD_IOC_ABORT_RECOVERY: {
4481 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4482 target_stop_recovery_thread(obd);
4486 case OBD_IOC_SYNC: {
4487 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4488 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4492 case OBD_IOC_SET_READONLY: {
4494 struct super_block *sb = obd->u.obt.obt_sb;
4495 struct inode *inode = sb->s_root->d_inode;
4496 BDEVNAME_DECLARE_STORAGE(tmp);
4497 CERROR("*** setting device %s read-only ***\n",
4498 ll_bdevname(sb, tmp));
4500 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4501 if (!IS_ERR(handle))
4502 rc = fsfilt_commit(obd, inode, handle, 1);
4504 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4505 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4507 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4511 case OBD_IOC_CATLOGLIST: {
4512 rc = llog_catalog_list(obd, 1, data);
4516 case OBD_IOC_LLOG_CANCEL:
4517 case OBD_IOC_LLOG_REMOVE:
4518 case OBD_IOC_LLOG_INFO:
4519 case OBD_IOC_LLOG_PRINT: {
4520 /* FIXME to be finished */
4521 RETURN(-EOPNOTSUPP);
4523 struct llog_ctxt *ctxt = NULL;
4525 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4526 rc = llog_ioctl(ctxt, cmd, data);
4527 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4540 static int filter_health_check(struct obd_device *obd)
4542 #ifdef USE_HEALTH_CHECK_WRITE
4543 struct filter_obd *filter = &obd->u.filter;
4548 * health_check to return 0 on healthy
4549 * and 1 on unhealthy.
4551 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4554 #ifdef USE_HEALTH_CHECK_WRITE
4555 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4556 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4561 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4564 return filter_fid2dentry(data, NULL, gr, id);
4567 static int filter_process_config(struct obd_device *obd, obd_count len,
4570 struct lustre_cfg *lcfg = buf;
4571 struct lprocfs_static_vars lvars;
4574 switch (lcfg->lcfg_command) {
4576 lprocfs_filter_init_vars(&lvars);
4578 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4588 static struct lvfs_callback_ops filter_lvfs_ops = {
4589 l_fid2dentry: filter_lvfs_fid2dentry,
4592 static struct obd_ops filter_obd_ops = {
4593 .o_owner = THIS_MODULE,
4594 .o_get_info = filter_get_info,
4595 .o_set_info_async = filter_set_info_async,
4596 .o_setup = filter_setup,
4597 .o_precleanup = filter_precleanup,
4598 .o_cleanup = filter_cleanup,
4599 .o_connect = filter_connect,
4600 .o_reconnect = filter_reconnect,
4601 .o_disconnect = filter_disconnect,
4602 .o_ping = filter_ping,
4603 .o_init_export = filter_init_export,
4604 .o_destroy_export = filter_destroy_export,
4605 .o_statfs = filter_statfs,
4606 .o_getattr = filter_getattr,
4607 .o_unpackmd = filter_unpackmd,
4608 .o_create = filter_create,
4609 .o_setattr = filter_setattr,
4610 .o_destroy = filter_destroy,
4611 .o_brw = filter_brw,
4612 .o_punch = filter_truncate,
4613 .o_sync = filter_sync,
4614 .o_preprw = filter_preprw,
4615 .o_commitrw = filter_commitrw,
4616 .o_llog_init = filter_llog_init,
4617 .o_llog_connect = filter_llog_connect,
4618 .o_llog_finish = filter_llog_finish,
4619 .o_iocontrol = filter_iocontrol,
4620 .o_health_check = filter_health_check,
4621 .o_process_config = filter_process_config,
4624 quota_interface_t *filter_quota_interface_ref;
4625 extern quota_interface_t filter_quota_interface;
4627 static int __init obdfilter_init(void)
4629 struct lprocfs_static_vars lvars;
4632 /** sanity check for group<->mdsno conversion */
4633 for (i = 0; i < 32; i++)
4634 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4636 lprocfs_filter_init_vars(&lvars);
4638 cfs_request_module("%s", "lquota");
4639 OBD_ALLOC(obdfilter_created_scratchpad,
4640 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4641 sizeof(*obdfilter_created_scratchpad));
4642 if (obdfilter_created_scratchpad == NULL)
4645 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4646 sizeof(struct filter_mod_data),
4649 GOTO(out, rc = -ENOMEM);
4651 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4652 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4654 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4655 LUSTRE_OST_NAME, NULL);
4659 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4660 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4661 ll_fmd_cachep = NULL;
4663 if (filter_quota_interface_ref)
4664 PORTAL_SYMBOL_PUT(filter_quota_interface);
4666 OBD_FREE(obdfilter_created_scratchpad,
4667 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4668 sizeof(*obdfilter_created_scratchpad));
4674 static void __exit obdfilter_exit(void)
4676 if (filter_quota_interface_ref)
4677 PORTAL_SYMBOL_PUT(filter_quota_interface);
4679 if (ll_fmd_cachep) {
4680 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4681 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4682 ll_fmd_cachep = NULL;
4685 class_unregister_type(LUSTRE_OST_NAME);
4686 OBD_FREE(obdfilter_created_scratchpad,
4687 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4688 sizeof(*obdfilter_created_scratchpad));
4691 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4692 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4693 MODULE_LICENSE("GPL");
4695 module_init(obdfilter_init);
4696 module_exit(obdfilter_exit);