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)
646 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
647 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
649 cfs_spin_lock(&exp->exp_lock);
650 exp->exp_connecting = 1;
651 cfs_spin_unlock(&exp->exp_lock);
653 return ldlm_init_export(exp);
656 static int filter_free_server_data(struct obd_device_target *obt)
658 lut_fini(NULL, obt->obt_lut);
659 OBD_FREE_PTR(obt->obt_lut);
663 /* assumes caller is already in kernel ctxt */
664 int filter_update_server_data(struct obd_device *obd)
666 struct file *filp = obd->u.obt.obt_rcvd_filp;
667 struct lr_server_data *lsd = class_server_data(obd);
672 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
673 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
674 le64_to_cpu(lsd->lsd_last_transno));
675 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
676 le64_to_cpu(lsd->lsd_mount_count));
678 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
680 CERROR("error writing lr_server_data: rc = %d\n", rc);
685 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
688 struct filter_obd *filter = &obd->u.filter;
694 if (filter->fo_last_objid_files[group] == NULL) {
695 CERROR("Object group "LPU64" not fully setup; not updating "
696 "last_objid\n", group);
700 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
701 obd->obd_name, group, filter->fo_last_objids[group]);
703 tmp = cpu_to_le64(filter->fo_last_objids[group]);
704 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
705 &tmp, sizeof(tmp), &off, force_sync);
707 CERROR("error writing group "LPU64" last objid: rc = %d\n",
711 extern int ost_handle(struct ptlrpc_request *req);
712 /* assumes caller has already in kernel ctxt */
713 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
715 struct filter_obd *filter = &obd->u.filter;
716 struct lr_server_data *lsd;
717 struct lsd_client_data *lcd = NULL;
718 struct inode *inode = filp->f_dentry->d_inode;
719 unsigned long last_rcvd_size = i_size_read(inode);
720 struct lu_target *lut;
727 /* ensure padding in the struct is the correct size */
728 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
729 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
730 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
731 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
733 /* allocate and initialize lu_target */
737 rc = lut_init(NULL, lut, obd, NULL);
740 lsd = class_server_data(obd);
741 if (last_rcvd_size == 0) {
742 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
744 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
745 lsd->lsd_last_transno = 0;
746 mount_count = lsd->lsd_mount_count = 0;
747 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
748 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
749 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
750 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
751 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
752 /* OBD_COMPAT_OST is set in filter_connect_internal when the
753 * MDS first connects and assigns the OST index number. */
754 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
757 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
759 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
763 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
764 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
765 "using the wrong disk %s. Were the "
766 "/dev/ assignments rearranged?\n",
767 obd->obd_uuid.uuid, lsd->lsd_uuid);
768 GOTO(err_lut, rc = -EINVAL);
770 mount_count = le64_to_cpu(lsd->lsd_mount_count);
771 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
773 /* Assume old last_rcvd format unless I_C_LR is set */
774 if (!(lsd->lsd_feature_incompat &
775 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
776 lsd->lsd_last_transno = lsd->lsd_compat14;
778 /* OBD_COMPAT_OST is set in filter_connect_internal when the
779 * MDS first connects and assigns the OST index number. */
780 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
784 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
785 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
786 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
787 ~FILTER_INCOMPAT_SUPP);
788 GOTO(err_lut, rc = -EINVAL);
790 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
791 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
792 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
793 ~FILTER_ROCOMPAT_SUPP);
794 /* Do something like remount filesystem read-only */
795 GOTO(err_lut, rc = -EINVAL);
798 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
800 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
801 obd->obd_name, start_epoch);
802 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
803 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
804 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
805 obd->obd_name, mount_count + 1);
806 CDEBUG(D_INODE, "%s: server data size: %u\n",
807 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
808 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
809 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
810 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
811 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
812 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
813 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
814 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
815 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
816 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
817 le16_to_cpu(lsd->lsd_client_size));
819 if (!obd->obd_replayable) {
820 CWARN("%s: recovery support OFF\n", obd->obd_name);
824 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
825 off < last_rcvd_size; cl_idx++) {
827 struct obd_export *exp;
828 struct filter_export_data *fed;
833 GOTO(err_client, rc = -ENOMEM);
836 /* Don't assume off is incremented properly by
837 * fsfilt_read_record(), in case sizeof(*lcd)
838 * isn't the same as lsd->lsd_client_size. */
839 off = le32_to_cpu(lsd->lsd_client_start) +
840 cl_idx * le16_to_cpu(lsd->lsd_client_size);
841 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
843 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
844 LAST_RCVD, cl_idx, off, rc);
845 break; /* read error shouldn't cause startup to fail */
848 if (lcd->lcd_uuid[0] == '\0') {
849 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
854 check_lcd(obd->obd_name, cl_idx, lcd);
856 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
858 /* These exports are cleaned up by filter_disconnect(), so they
859 * need to be set up like real exports as filter_connect() does.
861 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
863 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
864 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
865 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
867 if (PTR_ERR(exp) == -EALREADY) {
868 /* export already exists, zero out this one */
869 CERROR("Zeroing out duplicate export due to "
871 lcd->lcd_uuid[0] = '\0';
873 GOTO(err_client, rc = PTR_ERR(exp));
876 fed = &exp->exp_filter_data;
877 fed->fed_ted.ted_lcd = lcd;
878 fed->fed_group = 0; /* will be assigned at connect */
879 filter_export_stats_init(obd, exp, NULL);
880 rc = filter_client_add(obd, exp, cl_idx);
881 /* can't fail for existing client */
882 LASSERTF(rc == 0, "rc = %d\n", rc);
884 /* VBR: set export last committed */
885 exp->exp_last_committed = last_rcvd;
886 cfs_spin_lock(&exp->exp_lock);
887 exp->exp_connecting = 0;
888 exp->exp_in_recovery = 0;
889 cfs_spin_unlock(&exp->exp_lock);
890 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
891 obd->obd_max_recoverable_clients++;
892 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
894 class_export_put(exp);
897 /* Need to check last_rcvd even for duplicated exports. */
898 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
901 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
902 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
908 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
910 lut->lut_mount_count = mount_count + 1;
911 lsd->lsd_mount_count = cpu_to_le64(lut->lut_mount_count);
913 /* save it, so mount count and last_transno is current */
914 rc = filter_update_server_data(obd);
916 GOTO(err_client, rc);
921 class_disconnect_exports(obd);
923 filter_free_server_data(&obd->u.obt);
927 static int filter_cleanup_groups(struct obd_device *obd)
929 struct filter_obd *filter = &obd->u.filter;
931 struct dentry *dentry;
935 if (filter->fo_dentry_O_groups != NULL) {
936 for (i = 0; i < filter->fo_group_count; i++) {
937 dentry = filter->fo_dentry_O_groups[i];
941 OBD_FREE(filter->fo_dentry_O_groups,
942 filter->fo_group_count *
943 sizeof(*filter->fo_dentry_O_groups));
944 filter->fo_dentry_O_groups = NULL;
946 if (filter->fo_last_objid_files != NULL) {
947 for (i = 0; i < filter->fo_group_count; i++) {
948 filp = filter->fo_last_objid_files[i];
952 OBD_FREE(filter->fo_last_objid_files,
953 filter->fo_group_count *
954 sizeof(*filter->fo_last_objid_files));
955 filter->fo_last_objid_files = NULL;
957 if (filter->fo_dentry_O_sub != NULL) {
958 for (i = 0; i < filter->fo_group_count; i++) {
959 for (j = 0; j < filter->fo_subdir_count; j++) {
960 dentry = filter->fo_dentry_O_sub[i].dentry[j];
965 OBD_FREE(filter->fo_dentry_O_sub,
966 filter->fo_group_count *
967 sizeof(*filter->fo_dentry_O_sub));
968 filter->fo_dentry_O_sub = NULL;
970 if (filter->fo_last_objids != NULL) {
971 OBD_FREE(filter->fo_last_objids,
972 filter->fo_group_count *
973 sizeof(*filter->fo_last_objids));
974 filter->fo_last_objids = NULL;
976 if (filter->fo_dentry_O != NULL) {
977 f_dput(filter->fo_dentry_O);
978 filter->fo_dentry_O = NULL;
983 static int filter_update_last_group(struct obd_device *obd, int group)
985 struct filter_obd *filter = &obd->u.filter;
986 struct file *filp = NULL;
987 int last_group = 0, rc;
991 if (group <= filter->fo_committed_group)
994 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
998 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1002 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1004 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1008 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1009 obd->obd_name, last_group, group);
1013 /* must be sync: bXXXX */
1014 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1016 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1020 filter->fo_committed_group = group;
1023 filp_close(filp, 0);
1027 static int filter_read_group_internal(struct obd_device *obd, int group,
1030 struct filter_obd *filter = &obd->u.filter;
1031 __u64 *new_objids = NULL;
1032 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1033 struct dentry **new_groups = NULL;
1034 struct file **new_files = NULL;
1035 struct dentry *dentry;
1037 int old_count = filter->fo_group_count, rc, stage = 0, i;
1041 int len = group + 1;
1043 snprintf(name, 24, "%d", group);
1047 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1049 if (IS_ERR(dentry)) {
1050 CERROR("Cannot lookup expected object group %d: %ld\n",
1051 group, PTR_ERR(dentry));
1052 RETURN(PTR_ERR(dentry));
1055 dentry = simple_mkdir(filter->fo_dentry_O,
1056 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1057 if (IS_ERR(dentry)) {
1058 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1060 RETURN(PTR_ERR(dentry));
1065 snprintf(name, 24, "O/%d/LAST_ID", group);
1067 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1069 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1070 GOTO(cleanup, rc = PTR_ERR(filp));
1074 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1076 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1080 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1081 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1082 if (tmp_subdirs == NULL)
1083 GOTO(cleanup, rc = -ENOMEM);
1086 for (i = 0; i < filter->fo_subdir_count; i++) {
1088 snprintf(dir, sizeof(dir), "d%u", i);
1090 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1091 obd->u.obt.obt_vfsmnt,
1093 if (IS_ERR(tmp_subdirs->dentry[i])) {
1094 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1095 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1100 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1101 tmp_subdirs->dentry[i]);
1105 /* 'group' is an index; we need an array of length 'group + 1' */
1106 if (group + 1 > old_count) {
1107 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1108 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1109 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1110 OBD_ALLOC(new_files, len * sizeof(*new_files));
1112 if (new_objids == NULL || new_subdirs == NULL ||
1113 new_groups == NULL || new_files == NULL)
1114 GOTO(cleanup, rc = -ENOMEM);
1117 memcpy(new_objids, filter->fo_last_objids,
1118 old_count * sizeof(*new_objids));
1119 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1120 old_count * sizeof(*new_subdirs));
1121 memcpy(new_groups, filter->fo_dentry_O_groups,
1122 old_count * sizeof(*new_groups));
1123 memcpy(new_files, filter->fo_last_objid_files,
1124 old_count * sizeof(*new_files));
1126 OBD_FREE(filter->fo_last_objids,
1127 old_count * sizeof(*new_objids));
1128 OBD_FREE(filter->fo_dentry_O_sub,
1129 old_count * sizeof(*new_subdirs));
1130 OBD_FREE(filter->fo_dentry_O_groups,
1131 old_count * sizeof(*new_groups));
1132 OBD_FREE(filter->fo_last_objid_files,
1133 old_count * sizeof(*new_files));
1135 filter->fo_last_objids = new_objids;
1136 filter->fo_dentry_O_sub = new_subdirs;
1137 filter->fo_dentry_O_groups = new_groups;
1138 filter->fo_last_objid_files = new_files;
1139 filter->fo_group_count = len;
1142 filter->fo_dentry_O_groups[group] = dentry;
1143 filter->fo_last_objid_files[group] = filp;
1144 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1145 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1146 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1149 filter_update_last_group(obd, group);
1151 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1152 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1153 rc = filter_update_last_objid(obd, group, 1);
1157 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1158 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1159 obd->obd_name, group, last_objid);
1164 if (new_objids != NULL)
1165 OBD_FREE(new_objids, len * sizeof(*new_objids));
1166 if (new_subdirs != NULL)
1167 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1168 if (new_groups != NULL)
1169 OBD_FREE(new_groups, len * sizeof(*new_groups));
1170 if (new_files != NULL)
1171 OBD_FREE(new_files, len * sizeof(*new_files));
1173 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1174 for (i = 0; i < filter->fo_subdir_count; i++) {
1175 if (tmp_subdirs->dentry[i] != NULL)
1176 dput(tmp_subdirs->dentry[i]);
1178 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1181 filp_close(filp, 0);
1188 static int filter_read_groups(struct obd_device *obd, int last_group,
1191 struct filter_obd *filter = &obd->u.filter;
1192 int old_count, group, rc = 0;
1194 cfs_down(&filter->fo_init_lock);
1195 old_count = filter->fo_group_count;
1196 for (group = old_count; group <= last_group; group++) {
1197 rc = filter_read_group_internal(obd, group, create);
1201 cfs_up(&filter->fo_init_lock);
1205 /* FIXME: object groups */
1206 static int filter_prep_groups(struct obd_device *obd)
1208 struct filter_obd *filter = &obd->u.filter;
1209 struct dentry *O_dentry;
1211 int last_group, rc = 0, cleanup_phase = 0;
1215 O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1217 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1218 if (IS_ERR(O_dentry)) {
1219 rc = PTR_ERR(O_dentry);
1220 CERROR("cannot open/create O: rc = %d\n", rc);
1223 filter->fo_dentry_O = O_dentry;
1224 cleanup_phase = 1; /* O_dentry */
1226 /* we have to initialize all groups before first connections from
1227 * clients because they may send create/destroy for any group -bzzz */
1228 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1230 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1231 GOTO(cleanup, rc = PTR_ERR(filp));
1233 cleanup_phase = 2; /* filp */
1235 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1237 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1242 last_group = FILTER_GROUP_MDS0;
1244 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1245 FILTER_GROUP_MDS0, last_group);
1246 filter->fo_committed_group = last_group;
1247 rc = filter_read_groups(obd, last_group, 1);
1251 filp_close(filp, 0);
1255 switch (cleanup_phase) {
1257 filp_close(filp, 0);
1259 filter_cleanup_groups(obd);
1260 f_dput(filter->fo_dentry_O);
1261 filter->fo_dentry_O = NULL;
1269 /* setup the object store with correct subdirectories */
1270 static int filter_prep(struct obd_device *obd)
1272 struct lvfs_run_ctxt saved;
1273 struct filter_obd *filter = &obd->u.filter;
1275 struct inode *inode;
1279 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1280 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1281 if (!file || IS_ERR(file)) {
1283 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1287 obd->u.obt.obt_rcvd_filp = file;
1288 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1289 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1290 file->f_dentry->d_inode->i_mode);
1291 GOTO(err_filp, rc = -ENOENT);
1294 inode = file->f_dentry->d_parent->d_inode;
1295 /* We use i_op->unlink directly in filter_vfs_unlink() */
1296 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1297 CERROR("%s: filesystem does not support create/unlink ops\n",
1299 GOTO(err_filp, rc = -EOPNOTSUPP);
1302 rc = filter_init_server_data(obd, file);
1304 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1307 LASSERT(obd->u.obt.obt_lut);
1308 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1310 /* open and create health check io file*/
1311 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1314 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1316 GOTO(err_server_data, rc);
1318 filter->fo_obt.obt_health_check_filp = file;
1319 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1320 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1321 file->f_dentry->d_inode->i_mode);
1322 GOTO(err_health_check, rc = -ENOENT);
1324 rc = lvfs_check_io_health(obd, file);
1326 GOTO(err_health_check, rc);
1328 rc = filter_prep_groups(obd);
1330 GOTO(err_health_check, rc);
1332 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1337 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1338 CERROR("can't close %s after error\n", HEALTH_CHECK);
1339 filter->fo_obt.obt_health_check_filp = NULL;
1341 target_recovery_fini(obd);
1342 filter_free_server_data(&obd->u.obt);
1344 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1345 CERROR("can't close %s after error\n", LAST_RCVD);
1346 obd->u.obt.obt_rcvd_filp = NULL;
1350 /* cleanup the filter: write last used object id to status file */
1351 static void filter_post(struct obd_device *obd)
1353 struct lvfs_run_ctxt saved;
1354 struct filter_obd *filter = &obd->u.filter;
1357 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1358 * best to start a transaction with h_sync, because we removed this
1361 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1362 rc = filter_update_server_data(obd);
1364 CERROR("error writing server data: rc = %d\n", rc);
1366 for (i = 0; i < filter->fo_group_count; i++) {
1367 rc = filter_update_last_objid(obd, i,
1368 (i == filter->fo_group_count - 1));
1370 CERROR("error writing group %d lastobjid: rc = %d\n",
1374 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1375 obd->u.obt.obt_rcvd_filp = NULL;
1377 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1379 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1380 filter->fo_obt.obt_health_check_filp = NULL;
1382 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1384 filter_cleanup_groups(obd);
1385 filter_free_server_data(&obd->u.obt);
1386 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1388 filter_free_capa_keys(filter);
1389 cleanup_capa_hash(filter->fo_capa_hash);
1392 static void filter_set_last_id(struct filter_obd *filter,
1393 obd_id id, obd_gr group)
1395 LASSERT(group <= filter->fo_group_count);
1397 cfs_spin_lock(&filter->fo_objidlock);
1398 filter->fo_last_objids[group] = id;
1399 cfs_spin_unlock(&filter->fo_objidlock);
1402 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1405 LASSERT(group <= filter->fo_group_count);
1406 LASSERT(filter->fo_last_objids != NULL);
1408 /* FIXME: object groups */
1409 cfs_spin_lock(&filter->fo_objidlock);
1410 id = filter->fo_last_objids[group];
1411 cfs_spin_unlock(&filter->fo_objidlock);
1415 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1417 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1421 /* We never dget the object parent, so DON'T dput it either */
1422 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1424 struct filter_obd *filter = &obd->u.filter;
1425 struct filter_subdirs *subdirs;
1426 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1428 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1429 return filter->fo_dentry_O_groups[group];
1431 subdirs = &filter->fo_dentry_O_sub[group];
1432 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1435 /* We never dget the object parent, so DON'T dput it either */
1436 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1439 unsigned long now = jiffies;
1440 struct dentry *dparent = filter_parent(obd, group, objid);
1443 if (IS_ERR(dparent))
1445 if (dparent == NULL)
1446 return ERR_PTR(-ENOENT);
1448 rc = filter_lock_dentry(obd, dparent);
1449 fsfilt_check_slow(obd, now, "parent lock");
1450 return rc ? ERR_PTR(rc) : dparent;
1453 /* We never dget the object parent, so DON'T dput it either */
1454 static void filter_parent_unlock(struct dentry *dparent)
1456 UNLOCK_INODE_MUTEX(dparent->d_inode);
1459 /* How to get files, dentries, inodes from object id's.
1461 * If dir_dentry is passed, the caller has already locked the parent
1462 * appropriately for this operation (normally a write lock). If
1463 * dir_dentry is NULL, we do a read lock while we do the lookup to
1464 * avoid races with create/destroy and such changing the directory
1465 * internal to the filesystem code. */
1466 struct dentry *filter_fid2dentry(struct obd_device *obd,
1467 struct dentry *dir_dentry,
1468 obd_gr group, obd_id id)
1470 struct dentry *dparent = dir_dentry;
1471 struct dentry *dchild;
1476 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1477 obd->u.filter.fo_destroys_in_progress == 0) {
1478 /* don't fail lookups for orphan recovery, it causes
1479 * later LBUGs when objects still exist during precreate */
1480 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1481 RETURN(ERR_PTR(-ENOENT));
1484 CERROR("fatal: invalid object id 0\n");
1485 RETURN(ERR_PTR(-ESTALE));
1488 len = sprintf(name, LPU64, id);
1489 if (dir_dentry == NULL) {
1490 dparent = filter_parent_lock(obd, group, id);
1491 if (IS_ERR(dparent)) {
1492 CERROR("%s: error getting object "LPU64":"LPU64
1493 " parent: rc %ld\n", obd->obd_name,
1494 id, group, PTR_ERR(dparent));
1498 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1499 dparent->d_name.len, dparent->d_name.name, name);
1500 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1501 if (dir_dentry == NULL)
1502 filter_parent_unlock(dparent);
1503 if (IS_ERR(dchild)) {
1504 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1509 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1510 CERROR("%s: got bad object "LPU64" inode %lu\n",
1511 obd->obd_name, id, dchild->d_inode->i_ino);
1513 RETURN(ERR_PTR(-ENOENT));
1516 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1517 name, dchild, atomic_read(&dchild->d_count));
1519 LASSERT(atomic_read(&dchild->d_count) > 0);
1524 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1525 obd_id group, struct lustre_handle *lockh)
1527 int flags = LDLM_AST_DISCARD_DATA, rc;
1528 struct ldlm_res_id res_id;
1529 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1532 osc_build_res_name(objid, group, &res_id);
1533 /* Tell the clients that the object is gone now and that they should
1534 * throw away any cached pages. */
1535 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1536 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1537 ldlm_completion_ast, NULL, NULL, 0, NULL,
1544 static void filter_fini_destroy(struct obd_device *obd,
1545 struct lustre_handle *lockh)
1547 if (lustre_handle_is_used(lockh))
1548 ldlm_lock_decref(lockh, LCK_PW);
1551 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1552 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1553 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1554 * reverse. Caller must take i_sem before starting the transaction and we
1555 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1556 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1557 struct vfsmount *mnt)
1562 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1563 * itself we already hold dir->i_mutex for child create/unlink ops */
1564 LASSERT(dentry->d_inode != NULL);
1565 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1566 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1570 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1571 GOTO(out, rc = -ENOENT);
1573 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1578 GOTO(out, rc = -EPERM);
1580 /* check_sticky() */
1581 if ((dentry->d_inode->i_uid != current->fsuid &&
1582 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1583 IS_IMMUTABLE(dentry->d_inode))
1584 GOTO(out, rc = -EPERM);
1586 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1587 * that was done because of journal_start (which is already done
1588 * here) or some other ordering issue. */
1591 rc = ll_security_inode_unlink(dir, dentry, mnt);
1595 rc = dir->i_op->unlink(dir, dentry);
1597 /* need to drop i_mutex before we lose inode reference */
1598 UNLOCK_INODE_MUTEX(dentry->d_inode);
1605 /* Caller must hold LCK_PW on parent and push us into kernel context.
1606 * Caller must hold child i_mutex, we drop it always.
1607 * Caller is also required to ensure that dchild->d_inode exists. */
1608 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1609 obd_gr group, struct dentry *dparent,
1610 struct dentry *dchild)
1612 struct inode *inode = dchild->d_inode;
1615 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1616 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1617 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1618 (unsigned long)inode->i_nlink,
1619 atomic_read(&inode->i_count));
1622 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1624 CERROR("error unlinking objid %.*s: rc %d\n",
1625 dchild->d_name.len, dchild->d_name.name, rc);
1629 struct filter_intent_args {
1630 struct ldlm_lock **victim;
1635 static enum interval_iter filter_intent_cb(struct interval_node *n,
1638 struct ldlm_interval *node = (struct ldlm_interval *)n;
1639 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1640 __u64 size = arg->size;
1641 struct ldlm_lock **v = arg->victim;
1642 struct ldlm_lock *lck;
1644 /* If the interval is lower than the current file size,
1646 if (interval_high(n) <= size)
1647 return INTERVAL_ITER_STOP;
1649 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1650 /* Don't send glimpse ASTs to liblustre clients.
1651 * They aren't listening for them, and they do
1652 * entirely synchronous I/O anyways. */
1653 if (lck->l_export == NULL ||
1654 lck->l_export->exp_libclient == 1)
1657 if (*arg->liblustre)
1658 *arg->liblustre = 0;
1661 *v = LDLM_LOCK_GET(lck);
1662 } else if ((*v)->l_policy_data.l_extent.start <
1663 lck->l_policy_data.l_extent.start) {
1664 LDLM_LOCK_RELEASE(*v);
1665 *v = LDLM_LOCK_GET(lck);
1668 /* the same policy group - every lock has the
1669 * same extent, so needn't do it any more */
1673 return INTERVAL_ITER_CONT;
1676 static int filter_intent_policy(struct ldlm_namespace *ns,
1677 struct ldlm_lock **lockp, void *req_cookie,
1678 ldlm_mode_t mode, int flags, void *data)
1680 CFS_LIST_HEAD(rpc_list);
1681 struct ptlrpc_request *req = req_cookie;
1682 struct ldlm_lock *lock = *lockp, *l = NULL;
1683 struct ldlm_resource *res = lock->l_resource;
1684 ldlm_processing_policy policy;
1685 struct ost_lvb *res_lvb, *reply_lvb;
1686 struct ldlm_reply *rep;
1688 int idx, rc, tmpflags = 0, only_liblustre = 1;
1689 struct ldlm_interval_tree *tree;
1690 struct filter_intent_args arg;
1691 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1692 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1693 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1696 policy = ldlm_get_processing_policy(res);
1697 LASSERT(policy != NULL);
1698 LASSERT(req != NULL);
1700 rc = lustre_pack_reply(req, 3, repsize, NULL);
1702 RETURN(req->rq_status = rc);
1704 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1705 LASSERT(rep != NULL);
1707 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1708 sizeof(*reply_lvb));
1709 LASSERT(reply_lvb != NULL);
1711 //fixup_handle_for_resent_req(req, lock, &lockh);
1713 /* Call the extent policy function to see if our request can be
1714 * granted, or is blocked.
1715 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1716 * lock, and should not be granted if the lock will be blocked.
1719 LASSERT(ns == res->lr_namespace);
1721 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1722 check_res_locked(res);
1724 /* FIXME: we should change the policy function slightly, to not make
1725 * this list at all, since we just turn around and free it */
1726 while (!cfs_list_empty(&rpc_list)) {
1727 struct ldlm_lock *wlock =
1728 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1730 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1731 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1732 lock->l_flags &= ~LDLM_FL_CP_REQD;
1733 cfs_list_del_init(&wlock->l_cp_ast);
1734 LDLM_LOCK_RELEASE(wlock);
1737 /* The lock met with no resistance; we're finished. */
1738 if (rc == LDLM_ITER_CONTINUE) {
1739 /* do not grant locks to the liblustre clients: they cannot
1740 * handle ASTs robustly. We need to do this while still
1741 * holding ns_lock to avoid the lock remaining on the res_link
1742 * list (and potentially being added to l_pending_list by an
1743 * AST) when we are going to drop this lock ASAP. */
1744 if (lock->l_export->exp_libclient ||
1745 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1746 ldlm_resource_unlink_lock(lock);
1747 err = ELDLM_LOCK_ABORTED;
1749 err = ELDLM_LOCK_REPLACED;
1755 /* Do not grant any lock, but instead send GL callbacks. The extent
1756 * policy nicely created a list of all PW locks for us. We will choose
1757 * the highest of those which are larger than the size in the LVB, if
1758 * any, and perform a glimpse callback. */
1759 res_lvb = res->lr_lvb_data;
1760 LASSERT(res_lvb != NULL);
1761 *reply_lvb = *res_lvb;
1764 * ->ns_lock guarantees that no new locks are granted, and,
1765 * therefore, that res->lr_lvb_data cannot increase beyond the
1766 * end of already granted lock. As a result, it is safe to
1767 * check against "stale" reply_lvb->lvb_size value without
1770 arg.size = reply_lvb->lvb_size;
1772 arg.liblustre = &only_liblustre;
1773 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1774 tree = &res->lr_itree[idx];
1775 if (tree->lit_mode == LCK_PR)
1778 interval_iterate_reverse(tree->lit_root,
1779 filter_intent_cb, &arg);
1783 /* There were no PW locks beyond the size in the LVB; finished. */
1785 if (only_liblustre) {
1786 /* If we discovered a liblustre client with a PW lock,
1787 * however, the LVB may be out of date! The LVB is
1788 * updated only on glimpse (which we don't do for
1789 * liblustre clients) and cancel (which the client
1790 * obviously has not yet done). So if it has written
1791 * data but kept the lock, the LVB is stale and needs
1792 * to be updated from disk.
1794 * Of course, this will all disappear when we switch to
1795 * taking liblustre locks on the OST. */
1796 ldlm_res_lvbo_update(res, NULL, 1);
1798 RETURN(ELDLM_LOCK_ABORTED);
1802 * This check is for lock taken in filter_prepare_destroy() that does
1803 * not have l_glimpse_ast set. So the logic is: if there is a lock
1804 * with no l_glimpse_ast set, this object is being destroyed already.
1806 * Hence, if you are grabbing DLM locks on the server, always set
1807 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1809 if (l->l_glimpse_ast == NULL) {
1810 /* We are racing with unlink(); just return -ENOENT */
1811 rep->lock_policy_res1 = -ENOENT;
1815 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1816 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1817 /* Update the LVB from disk if the AST failed (this is a legal race) */
1819 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1820 * sending ast is not handled. This can result in lost client writes.
1823 ldlm_res_lvbo_update(res, NULL, 1);
1826 *reply_lvb = *res_lvb;
1830 LDLM_LOCK_RELEASE(l);
1832 RETURN(ELDLM_LOCK_ABORTED);
1836 * per-obd_device iobuf pool.
1838 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1839 * allocations in write-path has to be minimized (see bug 5137).
1841 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1842 * OST threads (see ost_thread_{init,done}()).
1844 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1845 * at the OST layer there are only (potentially) multiple obd_device of type
1846 * unknown at the time of OST thread creation.
1848 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1849 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1850 * it's very own iobuf.
1854 * filter_kiobuf_pool_init()
1856 * filter_kiobuf_pool_done()
1858 * filter_iobuf_get()
1860 * operate on this array. They are "generic" in a sense that they don't depend
1861 * on actual type of iobuf's (the latter depending on Linux kernel version).
1865 * destroy pool created by filter_iobuf_pool_init
1867 static void filter_iobuf_pool_done(struct filter_obd *filter)
1869 struct filter_iobuf **pool;
1874 pool = filter->fo_iobuf_pool;
1876 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1877 if (pool[i] != NULL)
1878 filter_free_iobuf(pool[i]);
1880 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1881 filter->fo_iobuf_pool = NULL;
1886 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1888 struct filter_obd *filter = &obd->u.filter;
1889 struct sptlrpc_rule_set tmp_rset;
1892 sptlrpc_rule_set_init(&tmp_rset);
1893 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1895 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1900 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1902 cfs_write_lock(&filter->fo_sptlrpc_lock);
1903 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1904 filter->fo_sptlrpc_rset = tmp_rset;
1905 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1911 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1913 static int filter_iobuf_pool_init(struct filter_obd *filter)
1920 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1922 if (filter->fo_iobuf_pool == NULL)
1925 filter->fo_iobuf_count = OSS_THREADS_MAX;
1930 /* Return iobuf allocated for @thread_id. We don't know in advance how
1931 * many threads there will be so we allocate a large empty array and only
1932 * fill in those slots that are actually in use.
1933 * If we haven't allocated a pool entry for this thread before, do so now. */
1934 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1936 int thread_id = (oti && oti->oti_thread) ?
1937 oti->oti_thread->t_id : -1;
1938 struct filter_iobuf *pool = NULL;
1939 struct filter_iobuf **pool_place = NULL;
1941 if (thread_id >= 0) {
1942 LASSERT(thread_id < filter->fo_iobuf_count);
1943 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1946 if (unlikely(pool == NULL)) {
1947 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1948 PTLRPC_MAX_BRW_PAGES);
1949 if (pool_place != NULL)
1956 /* mount the file system (secretly). lustre_cfg parameters are:
1959 * 3 = flags: failover=f, failout=n
1962 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1965 struct filter_obd *filter = &obd->u.filter;
1966 struct vfsmount *mnt;
1967 struct lustre_mount_info *lmi;
1968 struct obd_uuid uuid;
1976 if (lcfg->lcfg_bufcount < 3 ||
1977 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1978 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1981 lmi = server_get_mount(obd->obd_name);
1983 /* We already mounted in lustre_fill_super.
1984 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1985 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1987 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1989 /* gets recovery timeouts from mount data */
1990 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1991 obd->obd_recovery_timeout =
1992 lsi->lsi_lmd->lmd_recovery_time_soft;
1993 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1994 obd->obd_recovery_time_hard =
1995 lsi->lsi_lmd->lmd_recovery_time_hard;
1997 /* old path - used by lctl */
1998 CERROR("Using old MDS mount method\n");
1999 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2000 MS_NOATIME|MS_NODIRATIME,
2001 lustre_cfg_string(lcfg, 1), option);
2004 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2005 lustre_cfg_string(lcfg, 1), rc);
2009 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2011 if (IS_ERR(obd->obd_fsops))
2012 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2014 rc = filter_iobuf_pool_init(filter);
2018 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2019 CERROR("%s: Underlying device is marked as read-only. "
2020 "Setup failed\n", obd->obd_name);
2021 GOTO(err_ops, rc = -EROFS);
2024 /* failover is the default */
2025 obd->obd_replayable = 1;
2027 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2028 str = lustre_cfg_string(lcfg, 3);
2029 if (strchr(str, 'n')) {
2030 CWARN("%s: recovery disabled\n", obd->obd_name);
2031 obd->obd_replayable = 0;
2035 obd->u.obt.obt_vfsmnt = mnt;
2036 obd->u.obt.obt_sb = mnt->mnt_sb;
2037 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2038 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2040 fsfilt_setup(obd, obd->u.obt.obt_sb);
2042 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2043 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2044 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2045 obd->obd_lvfs_ctxt.fs = get_ds();
2046 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2048 cfs_init_mutex(&filter->fo_init_lock);
2049 filter->fo_committed_group = 0;
2050 filter->fo_destroys_in_progress = 0;
2051 for (i = 0; i < 32; i++)
2052 cfs_sema_init(&filter->fo_create_locks[i], 1);
2054 cfs_spin_lock_init(&filter->fo_objidlock);
2055 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2056 cfs_sema_init(&filter->fo_alloc_lock, 1);
2057 init_brw_stats(&filter->fo_filter_stats);
2058 filter->fo_read_cache = 1; /* enable read-only cache by default */
2059 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2060 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2061 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2062 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2064 rc = filter_prep(obd);
2068 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2069 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2071 filter->fo_fl_oss_capa = 1;
2073 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2074 filter->fo_capa_hash = init_capa_hash();
2075 if (filter->fo_capa_hash == NULL)
2076 GOTO(err_post, rc = -ENOMEM);
2078 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2079 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2080 LDLM_NAMESPACE_GREEDY);
2081 if (obd->obd_namespace == NULL)
2082 GOTO(err_post, rc = -ENOMEM);
2083 obd->obd_namespace->ns_lvbp = obd;
2084 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2085 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2087 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2088 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2090 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2092 CERROR("failed to setup llogging subsystems\n");
2096 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2097 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2098 /* do this after llog being initialized */
2099 filter_adapt_sptlrpc_conf(obd, 1);
2101 rc = lquota_setup(filter_quota_interface_ref, obd);
2105 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2106 if (q->max_sectors < q->max_hw_sectors &&
2107 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2108 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2109 "for larger I/O requests: max_sectors = %u "
2110 "could be up to max_hw_sectors=%u\n",
2111 obd->obd_name, mnt->mnt_sb->s_id,
2112 q->max_sectors, q->max_hw_sectors);
2114 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2115 if (uuid_ptr != NULL) {
2116 class_uuid_unparse(uuid_ptr, &uuid);
2122 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2123 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2124 obd->obd_name, label ?: str, lmi ? "on " : "",
2125 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2126 obd->obd_replayable ? "enabled" : "disabled");
2128 if (obd->obd_recovering)
2129 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2130 "or until %d client%s reconnect%s\n",
2132 obd->obd_recovery_timeout / 60,
2133 obd->obd_recovery_timeout % 60,
2134 obd->obd_max_recoverable_clients,
2135 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2136 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2144 fsfilt_put_ops(obd->obd_fsops);
2145 filter_iobuf_pool_done(filter);
2147 server_put_mount(obd->obd_name, mnt);
2148 obd->u.obt.obt_sb = 0;
2152 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2154 struct lprocfs_static_vars lvars;
2159 CLASSERT(offsetof(struct obd_device, u.obt) ==
2160 offsetof(struct obd_device, u.filter.fo_obt));
2162 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2165 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2166 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2169 addr = (unsigned long)cfs_page_address(page);
2170 clear_page((void *)addr);
2172 /* lprocfs must be setup before the filter so state can be safely added
2173 * to /proc incrementally as the filter is setup */
2174 lprocfs_filter_init_vars(&lvars);
2175 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2176 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2177 /* Init obdfilter private stats here */
2178 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2179 LPROCFS_CNTR_AVGMINMAX,
2180 "read_bytes", "bytes");
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "write_bytes", "bytes");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2185 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2186 "get_page", "usec");
2187 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2188 LPROCFS_CNTR_AVGMINMAX,
2189 "get_page_failures", "num");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2191 LPROCFS_CNTR_AVGMINMAX,
2192 "cache_access", "pages");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2194 LPROCFS_CNTR_AVGMINMAX,
2195 "cache_hit", "pages");
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2197 LPROCFS_CNTR_AVGMINMAX,
2198 "cache_miss", "pages");
2200 lproc_filter_attach_seqstat(obd);
2201 obd->obd_proc_exports_entry = lprocfs_register("exports",
2202 obd->obd_proc_entry,
2204 if (IS_ERR(obd->obd_proc_exports_entry)) {
2205 rc = PTR_ERR(obd->obd_proc_exports_entry);
2206 CERROR("error %d setting up lprocfs for %s\n",
2208 obd->obd_proc_exports_entry = NULL;
2211 if (obd->obd_proc_exports_entry)
2212 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2213 lprocfs_nid_stats_clear_read,
2214 lprocfs_nid_stats_clear_write, obd, NULL);
2216 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2217 LUSTRE_CFG_BUFLEN(lcfg, 4));
2218 rc = filter_common_setup(obd, lcfg, (void *)addr);
2219 OBD_PAGE_FREE(page);
2222 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2223 lprocfs_free_per_client_stats(obd);
2224 lprocfs_free_obd_stats(obd);
2225 lprocfs_obd_cleanup(obd);
2231 static struct llog_operations filter_mds_ost_repl_logops;
2233 static struct llog_operations filter_size_orig_logops = {
2234 .lop_setup = llog_obd_origin_setup,
2235 .lop_cleanup = llog_obd_origin_cleanup,
2236 .lop_add = llog_obd_origin_add
2239 static int filter_olg_fini(struct obd_llog_group *olg)
2241 struct llog_ctxt *ctxt;
2242 int rc = 0, rc2 = 0;
2245 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2247 rc = llog_cleanup(ctxt);
2249 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2251 rc2 = llog_cleanup(ctxt);
2256 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2258 rc2 = llog_cleanup(ctxt);
2267 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2268 struct obd_device *tgt)
2273 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2274 &filter_mds_ost_repl_logops);
2278 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2279 &filter_size_orig_logops);
2285 filter_olg_fini(olg);
2290 * Init the default olg, which is embeded in the obd_device, for filter.
2293 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2294 struct obd_device *tgt)
2296 struct filter_obd *filter = &obd->u.filter;
2297 struct llog_ctxt *ctxt;
2301 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2302 if (!filter->fo_lcm)
2305 filter_mds_ost_repl_logops = llog_client_ops;
2306 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2307 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2308 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2310 rc = filter_olg_init(obd, olg, tgt);
2312 GOTO(cleanup_lcm, rc);
2314 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2317 GOTO(cleanup_olg, rc);
2319 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2321 CERROR("Can't get ctxt for %p:%x\n", olg,
2322 LLOG_MDS_OST_REPL_CTXT);
2323 GOTO(cleanup_olg, rc = -ENODEV);
2325 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2326 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2327 llog_ctxt_put(ctxt);
2331 filter_olg_fini(olg);
2333 llog_recov_thread_fini(filter->fo_lcm, 1);
2334 filter->fo_lcm = NULL;
2339 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2340 struct obd_device *tgt, int *index)
2342 struct filter_obd *filter = &obd->u.filter;
2343 struct llog_ctxt *ctxt;
2347 LASSERT(olg != NULL);
2348 if (olg == &obd->obd_olg)
2349 return filter_default_olg_init(obd, olg, tgt);
2351 LASSERT(filter->fo_lcm != NULL);
2352 rc = filter_olg_init(obd, olg, tgt);
2355 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2357 CERROR("Can't get ctxt for %p:%x\n", olg,
2358 LLOG_MDS_OST_REPL_CTXT);
2359 filter_olg_fini(olg);
2362 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2363 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2364 llog_ctxt_put(ctxt);
2368 static int filter_llog_finish(struct obd_device *obd, int count)
2370 struct filter_obd *filter = &obd->u.filter;
2371 struct llog_ctxt *ctxt;
2374 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2377 * Make sure that no cached llcds left in recov_thread.
2378 * We actually do sync in disconnect time, but disconnect
2379 * may not come being marked rq_no_resend = 1.
2381 llog_sync(ctxt, NULL);
2384 * Balance class_import_get() in llog_receptor_accept().
2385 * This is safe to do, as llog is already synchronized
2386 * and its import may go.
2388 cfs_mutex_down(&ctxt->loc_sem);
2389 if (ctxt->loc_imp) {
2390 class_import_put(ctxt->loc_imp);
2391 ctxt->loc_imp = NULL;
2393 cfs_mutex_up(&ctxt->loc_sem);
2394 llog_ctxt_put(ctxt);
2397 if (filter->fo_lcm) {
2398 cfs_mutex_down(&ctxt->loc_sem);
2399 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2400 filter->fo_lcm = NULL;
2401 cfs_mutex_up(&ctxt->loc_sem);
2403 RETURN(filter_olg_fini(&obd->obd_olg));
2407 * Find the group llog according to group index in the llog group list.
2409 static struct obd_llog_group *
2410 filter_find_olg_internal(struct filter_obd *filter, int group)
2412 struct obd_llog_group *olg;
2414 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2415 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2416 if (olg->olg_group == group)
2423 * Find the group llog according to group index on the filter
2425 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2427 struct obd_llog_group *olg = NULL;
2428 struct filter_obd *filter;
2430 filter = &obd->u.filter;
2432 if (group == FILTER_GROUP_LLOG)
2433 RETURN(&obd->obd_olg);
2435 cfs_spin_lock(&filter->fo_llog_list_lock);
2436 olg = filter_find_olg_internal(filter, group);
2437 cfs_spin_unlock(&filter->fo_llog_list_lock);
2442 * Find the llog_group of the filter according to the group. If it can not
2443 * find, create the llog_group, which only happens when mds is being synced
2446 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2448 struct obd_llog_group *olg = NULL;
2449 struct filter_obd *filter;
2452 filter = &obd->u.filter;
2454 if (group == FILTER_GROUP_LLOG)
2455 RETURN(&obd->obd_olg);
2457 cfs_spin_lock(&filter->fo_llog_list_lock);
2458 olg = filter_find_olg_internal(filter, group);
2460 if (olg->olg_initializing) {
2461 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2463 GOTO(out_unlock, olg);
2468 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2470 llog_group_init(olg, group);
2471 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2472 olg->olg_initializing = 1;
2473 cfs_spin_unlock(&filter->fo_llog_list_lock);
2475 rc = obd_llog_init(obd, olg, obd, NULL);
2477 cfs_spin_lock(&filter->fo_llog_list_lock);
2478 cfs_list_del(&olg->olg_list);
2479 cfs_spin_unlock(&filter->fo_llog_list_lock);
2481 GOTO(out, olg = ERR_PTR(-ENOMEM));
2483 cfs_spin_lock(&filter->fo_llog_list_lock);
2484 olg->olg_initializing = 0;
2485 cfs_spin_unlock(&filter->fo_llog_list_lock);
2486 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2487 obd->obd_name, group, olg);
2492 cfs_spin_unlock(&filter->fo_llog_list_lock);
2496 static int filter_llog_connect(struct obd_export *exp,
2497 struct llogd_conn_body *body)
2499 struct obd_device *obd = exp->exp_obd;
2500 struct llog_ctxt *ctxt;
2501 struct obd_llog_group *olg;
2505 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2506 obd->obd_name, body->lgdc_logid.lgl_oid,
2507 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2509 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2511 CERROR(" %s: can not find olg of group %d\n",
2512 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2515 llog_group_set_export(olg, exp);
2517 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2518 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2519 body->lgdc_ctxt_idx);
2521 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2522 obd->obd_name, body->lgdc_logid.lgl_oid,
2523 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2525 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2526 obd->u.filter.fo_mds_ost_sync = 1;
2527 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2528 rc = llog_connect(ctxt, &body->lgdc_logid,
2529 &body->lgdc_gen, NULL);
2530 llog_ctxt_put(ctxt);
2532 CERROR("failed to connect rc %d idx %d\n", rc,
2533 body->lgdc_ctxt_idx);
2538 static int filter_llog_preclean(struct obd_device *obd)
2540 struct obd_llog_group *olg, *tmp;
2541 struct filter_obd *filter;
2542 cfs_list_t remove_list;
2546 rc = obd_llog_finish(obd, 0);
2548 CERROR("failed to cleanup llogging subsystem\n");
2550 filter = &obd->u.filter;
2551 CFS_INIT_LIST_HEAD(&remove_list);
2553 cfs_spin_lock(&filter->fo_llog_list_lock);
2554 while (!cfs_list_empty(&filter->fo_llog_list)) {
2555 olg = cfs_list_entry(filter->fo_llog_list.next,
2556 struct obd_llog_group, olg_list);
2557 cfs_list_del(&olg->olg_list);
2558 cfs_list_add(&olg->olg_list, &remove_list);
2560 cfs_spin_unlock(&filter->fo_llog_list_lock);
2562 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2563 cfs_list_del_init(&olg->olg_list);
2564 rc = filter_olg_fini(olg);
2566 CERROR("failed to cleanup llogging subsystem for %u\n",
2574 static int filter_precleanup(struct obd_device *obd,
2575 enum obd_cleanup_stage stage)
2581 case OBD_CLEANUP_EARLY:
2583 case OBD_CLEANUP_EXPORTS:
2584 /* Stop recovery before namespace cleanup. */
2585 target_recovery_fini(obd);
2586 rc = filter_llog_preclean(obd);
2592 static int filter_cleanup(struct obd_device *obd)
2594 struct filter_obd *filter = &obd->u.filter;
2598 LCONSOLE_WARN("%s: shutting down for failover; client state "
2599 "will be preserved.\n", obd->obd_name);
2601 obd_exports_barrier(obd);
2602 obd_zombie_barrier();
2604 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2605 lprocfs_free_per_client_stats(obd);
2606 lprocfs_free_obd_stats(obd);
2607 lprocfs_obd_cleanup(obd);
2608 lquota_cleanup(filter_quota_interface_ref, obd);
2610 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2611 obd->obd_namespace = NULL;
2613 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2615 if (obd->u.obt.obt_sb == NULL)
2620 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2621 shrink_dcache_sb(obd->u.obt.obt_sb);
2623 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2624 obd->u.obt.obt_sb = NULL;
2626 fsfilt_put_ops(obd->obd_fsops);
2628 filter_iobuf_pool_done(filter);
2630 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2635 static int filter_connect_internal(struct obd_export *exp,
2636 struct obd_connect_data *data,
2639 struct filter_export_data *fed = &exp->exp_filter_data;
2644 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2645 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2646 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2647 data->ocd_connect_flags, data->ocd_version,
2648 data->ocd_grant, data->ocd_index);
2650 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2651 CWARN("!!! This export (nid %s) used object group %d "
2652 "earlier; now it's trying to use group %d! This could "
2653 "be a bug in the MDS. Please report to "
2654 "http://bugzilla.lustre.org/\n",
2655 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2658 fed->fed_group = data->ocd_group;
2660 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2661 exp->exp_connect_flags = data->ocd_connect_flags;
2662 data->ocd_version = LUSTRE_VERSION_CODE;
2664 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2665 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2666 data->ocd_connect_flags & OBD_CONNECT_MDS))
2669 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2670 struct filter_obd *filter = &exp->exp_obd->u.filter;
2671 obd_size left, want;
2673 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2674 left = filter_grant_space_left(exp);
2675 want = data->ocd_grant;
2676 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2677 data->ocd_grant = fed->fed_grant;
2678 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2680 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2681 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2682 exp->exp_client_uuid.uuid, exp,
2683 data->ocd_grant, want, left);
2685 filter->fo_tot_granted_clients ++;
2688 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2689 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2690 int index = le32_to_cpu(lsd->lsd_ost_index);
2692 if (!(lsd->lsd_feature_compat &
2693 cpu_to_le32(OBD_COMPAT_OST))) {
2694 /* this will only happen on the first connect */
2695 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2696 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2697 /* sync is not needed here as filter_client_add will
2698 * set exp_need_sync flag */
2699 filter_update_server_data(exp->exp_obd);
2700 } else if (index != data->ocd_index) {
2701 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2702 " %u doesn't match actual OST index"
2703 " %u in last_rcvd file, bad "
2705 obd_export_nid2str(exp), index,
2709 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2710 * FIXME: We don't strictly need the COMPAT flag for that,
2711 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2712 * FIXME: We needed it for the index, as index 0 is valid. */
2715 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2716 data->ocd_brw_size = 65536;
2717 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2718 data->ocd_brw_size = min(data->ocd_brw_size,
2719 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2720 LASSERT(data->ocd_brw_size);
2723 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2724 __u32 cksum_types = data->ocd_cksum_types;
2726 /* The client set in ocd_cksum_types the checksum types it
2727 * supports. We have to mask off the algorithms that we don't
2729 if (cksum_types & OBD_CKSUM_ALL)
2730 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2732 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2734 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2735 "%x\n", exp->exp_obd->obd_name,
2736 obd_export_nid2str(exp), cksum_types,
2737 data->ocd_cksum_types);
2739 /* This client does not support OBD_CONNECT_CKSUM
2740 * fall back to CRC32 */
2741 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2742 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2743 exp->exp_obd->obd_name,
2744 obd_export_nid2str(exp));
2750 static int filter_reconnect(const struct lu_env *env,
2751 struct obd_export *exp, struct obd_device *obd,
2752 struct obd_uuid *cluuid,
2753 struct obd_connect_data *data,
2759 if (exp == NULL || obd == NULL || cluuid == NULL)
2762 rc = filter_connect_internal(exp, data, 1);
2764 filter_export_stats_init(obd, exp, localdata);
2769 /* nearly identical to mds_connect */
2770 static int filter_connect(const struct lu_env *env,
2771 struct obd_export **exp, struct obd_device *obd,
2772 struct obd_uuid *cluuid,
2773 struct obd_connect_data *data, void *localdata)
2775 struct lvfs_run_ctxt saved;
2776 struct lustre_handle conn = { 0 };
2777 struct obd_export *lexp;
2778 struct tg_export_data *ted;
2779 struct lsd_client_data *lcd = NULL;
2784 if (exp == NULL || obd == NULL || cluuid == NULL)
2787 rc = class_connect(&conn, obd, cluuid);
2790 lexp = class_conn2export(&conn);
2791 LASSERT(lexp != NULL);
2793 ted = &lexp->exp_target_data;
2795 rc = filter_connect_internal(lexp, data, 0);
2799 filter_export_stats_init(obd, lexp, localdata);
2800 if (obd->obd_replayable) {
2801 OBD_ALLOC(lcd, sizeof(*lcd));
2803 CERROR("filter: out of memory for client data\n");
2804 GOTO(cleanup, rc = -ENOMEM);
2807 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2809 rc = filter_client_add(obd, lexp, -1);
2814 group = data->ocd_group;
2816 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2817 obd->obd_name, lexp->exp_handle.h_cookie, group);
2819 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2820 rc = filter_read_groups(obd, group, 1);
2821 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2823 CERROR("can't read group %u\n", group);
2831 class_disconnect(lexp);
2832 lprocfs_exp_cleanup(lexp);
2841 /* Do extra sanity checks for grant accounting. We do this at connect,
2842 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2843 * always get rid of it or turn it off when we know accounting is good. */
2844 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2846 struct filter_export_data *fed;
2847 struct obd_export *exp;
2848 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2849 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2850 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2852 if (cfs_list_empty(&obd->obd_exports))
2855 /* We don't want to do this for large machines that do lots of
2856 mounts or unmounts. It burns... */
2857 if (obd->obd_num_exports > 100)
2860 cfs_spin_lock(&obd->obd_osfs_lock);
2861 cfs_spin_lock(&obd->obd_dev_lock);
2862 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2864 fed = &exp->exp_filter_data;
2865 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2868 if (maxsize > 0) { /* we may not have done a statfs yet */
2869 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2870 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2871 exp->exp_client_uuid.uuid, exp,
2872 fed->fed_grant, fed->fed_pending, maxsize);
2873 LASSERTF(fed->fed_dirty <= maxsize,
2874 "%s: cli %s/%p %ld > "LPU64"\n", func,
2875 exp->exp_client_uuid.uuid, exp,
2876 fed->fed_dirty, maxsize);
2879 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2880 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2881 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2883 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2884 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2885 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2886 tot_granted += fed->fed_grant + fed->fed_pending;
2887 tot_pending += fed->fed_pending;
2888 tot_dirty += fed->fed_dirty;
2890 fo_tot_granted = obd->u.filter.fo_tot_granted;
2891 fo_tot_pending = obd->u.filter.fo_tot_pending;
2892 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2893 cfs_spin_unlock(&obd->obd_dev_lock);
2894 cfs_spin_unlock(&obd->obd_osfs_lock);
2896 /* Do these assertions outside the spinlocks so we don't kill system */
2897 if (tot_granted != fo_tot_granted)
2898 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2899 func, tot_granted, fo_tot_granted);
2900 if (tot_pending != fo_tot_pending)
2901 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2902 func, tot_pending, fo_tot_pending);
2903 if (tot_dirty != fo_tot_dirty)
2904 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2905 func, tot_dirty, fo_tot_dirty);
2906 if (tot_pending > tot_granted)
2907 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2908 func, tot_pending, tot_granted);
2909 if (tot_granted > maxsize)
2910 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2911 func, tot_granted, maxsize);
2912 if (tot_dirty > maxsize)
2913 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2914 func, tot_dirty, maxsize);
2917 /* Remove this client from the grant accounting totals. We also remove
2918 * the export from the obd device under the osfs and dev locks to ensure
2919 * that the filter_grant_sanity_check() calculations are always valid.
2920 * The client should do something similar when it invalidates its import. */
2921 static void filter_grant_discard(struct obd_export *exp)
2923 struct obd_device *obd = exp->exp_obd;
2924 struct filter_obd *filter = &obd->u.filter;
2925 struct filter_export_data *fed = &exp->exp_filter_data;
2927 cfs_spin_lock(&obd->obd_osfs_lock);
2928 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2929 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2930 obd->obd_name, filter->fo_tot_granted,
2931 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2932 filter->fo_tot_granted -= fed->fed_grant;
2933 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2934 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2935 obd->obd_name, filter->fo_tot_pending,
2936 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2937 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2938 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2939 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2940 obd->obd_name, filter->fo_tot_dirty,
2941 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2942 filter->fo_tot_dirty -= fed->fed_dirty;
2946 cfs_spin_unlock(&obd->obd_osfs_lock);
2949 static int filter_destroy_export(struct obd_export *exp)
2951 struct filter_export_data *fed = &exp->exp_filter_data;
2954 if (fed->fed_pending)
2955 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2956 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2957 exp, fed->fed_pending);
2959 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2961 target_destroy_export(exp);
2962 ldlm_destroy_export(exp);
2964 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2967 lut_client_free(exp);
2969 if (!exp->exp_obd->obd_replayable)
2970 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2972 filter_grant_discard(exp);
2973 filter_fmd_cleanup(exp);
2975 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2976 struct filter_obd *filter = &exp->exp_obd->u.filter;
2977 if (filter->fo_tot_granted_clients > 0)
2978 filter->fo_tot_granted_clients --;
2981 if (!(exp->exp_flags & OBD_OPT_FORCE))
2982 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2987 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2989 struct obd_llog_group *olg_min, *olg;
2990 struct filter_obd *filter;
2991 int worked = -1, group;
2992 struct llog_ctxt *ctxt;
2995 filter = &obd->u.filter;
2997 /* we can't sync log holding spinlock. also, we do not want to get
2998 * into livelock. so we do following: loop over MDS's exports in
2999 * group order and skip already synced llogs -bzzz */
3001 /* look for group with min. number, but > worked */
3004 cfs_spin_lock(&filter->fo_llog_list_lock);
3005 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3006 if (olg->olg_group <= worked) {
3007 /* this group is already synced */
3010 if (group < olg->olg_group) {
3011 /* we have group with smaller number to sync */
3014 /* store current minimal group */
3016 group = olg->olg_group;
3018 cfs_spin_unlock(&filter->fo_llog_list_lock);
3020 if (olg_min == NULL)
3023 worked = olg_min->olg_group;
3024 if (olg_min->olg_exp &&
3025 (dexp == olg_min->olg_exp || dexp == NULL)) {
3027 ctxt = llog_group_get_ctxt(olg_min,
3028 LLOG_MDS_OST_REPL_CTXT);
3030 err = llog_sync(ctxt, olg_min->olg_exp);
3031 llog_ctxt_put(ctxt);
3033 CERROR("error flushing logs to MDS: "
3038 } while (olg_min != NULL);
3041 /* Also incredibly similar to mds_disconnect */
3042 static int filter_disconnect(struct obd_export *exp)
3044 struct obd_device *obd = exp->exp_obd;
3049 class_export_get(exp);
3051 if (!(exp->exp_flags & OBD_OPT_FORCE))
3052 filter_grant_sanity_check(obd, __FUNCTION__);
3053 filter_grant_discard(exp);
3055 /* Flush any remaining cancel messages out to the target */
3056 filter_sync_llogs(obd, exp);
3058 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3060 rc = server_disconnect_export(exp);
3062 /* Do not erase record for recoverable client. */
3063 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3064 filter_client_del(exp);
3066 fsfilt_sync(obd, obd->u.obt.obt_sb);
3068 class_export_put(exp);
3072 /* reverse import is changed, sync all cancels */
3073 static void filter_revimp_update(struct obd_export *exp)
3078 class_export_get(exp);
3080 /* flush any remaining cancel messages out to the target */
3081 filter_sync_llogs(exp->exp_obd, exp);
3082 class_export_put(exp);
3086 static int filter_ping(struct obd_export *exp)
3088 filter_fmd_expire(exp);
3092 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3093 const char *what, int quiet)
3095 struct dentry *dchild = NULL;
3098 if (oa->o_valid & OBD_MD_FLGROUP)
3101 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3103 if (IS_ERR(dchild)) {
3104 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3105 what, group, oa->o_id);
3109 if (dchild->d_inode == NULL) {
3111 CERROR("%s: %s on non-existent object: "LPU64"\n",
3112 obd->obd_name, what, oa->o_id);
3114 RETURN(ERR_PTR(-ENOENT));
3120 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3122 struct dentry *dentry = NULL;
3123 struct obd_device *obd;
3127 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3128 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3129 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3133 obd = class_exp2obd(exp);
3135 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3139 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3141 RETURN(PTR_ERR(dentry));
3143 /* Limit the valid bits in the return data to what we actually use */
3144 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3145 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3151 /* this should be enabled/disabled in condition to enabled/disabled large
3152 * inodes (fast EAs) in backing store FS. */
3153 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3154 void *handle, struct obdo *oa)
3156 struct obd_device *obd = exp->exp_obd;
3160 if (oa->o_valid & OBD_MD_FLFID) {
3161 struct filter_fid ff;
3163 if (!(oa->o_valid & OBD_MD_FLGROUP))
3165 /* packing fid and converting it to LE for storing into EA.
3166 * Here ->o_stripe_idx should be filled by LOV and rest of
3167 * fields - by client. */
3168 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3169 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3170 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3171 ff.ff_objid = cpu_to_le64(oa->o_id);
3172 ff.ff_group = cpu_to_le64(oa->o_gr);
3174 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3175 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3176 oa->o_generation, oa->o_id, oa->o_gr);
3178 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3180 CERROR("store fid in object failed! rc: %d\n", rc);
3182 CDEBUG(D_HA, "OSS object without fid info!\n");
3188 /* this is called from filter_truncate() until we have filter_punch() */
3189 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3190 struct obdo *oa, struct obd_trans_info *oti)
3192 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3193 struct llog_cookie *fcc = NULL;
3194 struct filter_obd *filter;
3195 int rc, err, sync = 0;
3196 loff_t old_size = 0;
3197 unsigned int ia_valid;
3198 struct inode *inode;
3199 struct page *page = NULL;
3204 LASSERT(dentry != NULL);
3205 LASSERT(!IS_ERR(dentry));
3207 inode = dentry->d_inode;
3208 LASSERT(inode != NULL);
3210 filter = &exp->exp_obd->u.filter;
3211 iattr_from_obdo(&iattr, oa, oa->o_valid);
3212 ia_valid = iattr.ia_valid;
3214 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3215 OBD_ALLOC(fcc, sizeof(*fcc));
3217 *fcc = oa->o_lcookie;
3219 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3221 /* Filter truncates and writes are serialized by
3222 * i_alloc_sem, see the comment in
3223 * filter_preprw_write.*/
3224 if (ia_valid & ATTR_SIZE)
3225 down_write(&inode->i_alloc_sem);
3226 LOCK_INODE_MUTEX(inode);
3227 old_size = i_size_read(inode);
3230 /* VBR: version recovery check */
3231 rc = filter_version_get_check(exp, oti, inode);
3233 GOTO(out_unlock, rc);
3235 /* Let's pin the last page so that ldiskfs_truncate
3236 * should not start GFP_FS allocation. */
3237 if (ia_valid & ATTR_SIZE) {
3238 page = grab_cache_page(inode->i_mapping,
3239 iattr.ia_size >> PAGE_CACHE_SHIFT);
3241 GOTO(out_unlock, rc = -ENOMEM);
3246 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3247 * then we will accept the UID+GID sent by the client during write for
3248 * initializing the ownership of this inode. We only allow this to
3249 * happen once so clear these bits in setattr. In 2.6 kernels it is
3250 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3251 * the flags that are actually being set. */
3252 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3253 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3254 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3256 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3257 if (!(ia_valid & ATTR_MODE)) {
3258 iattr.ia_mode = inode->i_mode;
3259 iattr.ia_valid |= ATTR_MODE;
3261 iattr.ia_mode &= ~S_ISUID;
3263 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3264 if (!(iattr.ia_valid & ATTR_MODE)) {
3265 iattr.ia_mode = inode->i_mode;
3266 iattr.ia_valid |= ATTR_MODE;
3268 iattr.ia_mode &= ~S_ISGID;
3271 orig_ids[USRQUOTA] = inode->i_uid;
3272 orig_ids[GRPQUOTA] = inode->i_gid;
3273 handle = fsfilt_start_log(exp->exp_obd, inode,
3274 FSFILT_OP_SETATTR, oti, 1);
3276 GOTO(out_unlock, rc = PTR_ERR(handle));
3278 /* update inode EA only once when inode is suid bit marked. As
3279 * on 2.6.x UID and GID may be set separately, we check here
3280 * only one of them to avoid double setting. */
3281 if (inode->i_mode & S_ISUID)
3282 filter_update_fidea(exp, inode, handle, oa);
3284 handle = fsfilt_start(exp->exp_obd, inode,
3285 FSFILT_OP_SETATTR, oti);
3287 GOTO(out_unlock, rc = PTR_ERR(handle));
3289 if (oa->o_valid & OBD_MD_FLFLAGS) {
3290 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3291 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3293 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3295 /* set cancel cookie callback function */
3296 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3297 filter_cancel_cookies_cb,
3301 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3302 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3304 /* The truncate might have used up our transaction credits. Make
3305 * sure we have one left for the last_rcvd update. */
3306 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3308 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3310 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3314 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3316 CERROR("error on commit, err = %d\n", err);
3323 /* For a partial-page truncate flush the page to disk immediately
3324 * to avoid data corruption during direct disk write. b=17397 */
3325 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3326 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3327 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3337 page_cache_release(page);
3339 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3340 UNLOCK_INODE_MUTEX(inode);
3341 if (ia_valid & ATTR_SIZE)
3342 up_write(&inode->i_alloc_sem);
3344 OBD_FREE(fcc, sizeof(*fcc));
3346 /* trigger quota release */
3347 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3348 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3349 int rc2 = lquota_adjust(filter_quota_interface_ref,
3350 exp->exp_obd, cur_ids,
3351 orig_ids, rc, FSFILT_OP_SETATTR);
3352 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3353 "filter adjust qunit. (rc:%d)\n", rc2);
3358 /* this is called from filter_truncate() until we have filter_punch() */
3359 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3360 struct obd_trans_info *oti)
3362 struct obdo *oa = oinfo->oi_oa;
3363 struct lustre_capa *capa = oinfo_capa(oinfo);
3364 struct ldlm_res_id res_id;
3365 struct filter_mod_data *fmd;
3366 struct lvfs_run_ctxt saved;
3367 struct filter_obd *filter;
3368 struct ldlm_resource *res;
3369 struct dentry *dentry;
3370 __u64 opc = CAPA_OPC_META_WRITE;
3374 if (oa->o_valid & OBD_FL_TRUNC)
3375 opc |= CAPA_OPC_OSS_TRUNC;
3377 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3378 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3382 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3383 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3388 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3389 /* This would be very bad - accidentally truncating a file when
3390 * changing the time or similar - bug 12203. */
3391 if (oa->o_valid & OBD_MD_FLSIZE &&
3392 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3393 static char mdsinum[48];
3395 if (oa->o_valid & OBD_MD_FLFID)
3396 snprintf(mdsinum, sizeof(mdsinum) - 1,
3397 " of inode "LPU64"/%u", oa->o_fid,
3402 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3404 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3409 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3411 RETURN(PTR_ERR(dentry));
3413 filter = &exp->exp_obd->u.filter;
3414 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3417 * We need to be atomic against a concurrent write
3418 * (which takes the semaphore for reading). fmd_mactime_xid
3419 * checks will have no effect if a write request with lower
3420 * xid starts just before a setattr and finishes later than
3421 * the setattr (see bug 21489, comment 27).
3424 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3425 down_write(&dentry->d_inode->i_alloc_sem);
3426 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3427 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3428 fmd->fmd_mactime_xid = oti->oti_xid;
3429 filter_fmd_put(exp, fmd);
3430 up_write(&dentry->d_inode->i_alloc_sem);
3433 /* setting objects attributes (including owner/group) */
3434 rc = filter_setattr_internal(exp, dentry, oa, oti);
3436 GOTO(out_unlock, rc);
3438 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3439 &res_id, LDLM_EXTENT, 0);
3442 LDLM_RESOURCE_ADDREF(res);
3443 rc = ldlm_res_lvbo_update(res, NULL, 0);
3444 LDLM_RESOURCE_DELREF(res);
3445 ldlm_resource_putref(res);
3448 oa->o_valid = OBD_MD_FLID;
3450 /* Quota release need uid/gid info */
3451 obdo_from_inode(oa, dentry->d_inode,
3452 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3457 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3461 /* XXX identical to osc_unpackmd */
3462 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3463 struct lov_mds_md *lmm, int lmm_bytes)
3469 if (lmm_bytes < sizeof (*lmm)) {
3470 CERROR("lov_mds_md too small: %d, need %d\n",
3471 lmm_bytes, (int)sizeof(*lmm));
3474 /* XXX LOV_MAGIC etc check? */
3476 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3477 CERROR("lov_mds_md: zero lmm_object_id\n");
3482 lsm_size = lov_stripe_md_size(1);
3486 if (*lsmp != NULL && lmm == NULL) {
3487 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3488 OBD_FREE(*lsmp, lsm_size);
3493 if (*lsmp == NULL) {
3494 OBD_ALLOC(*lsmp, lsm_size);
3498 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3499 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3500 OBD_FREE(*lsmp, lsm_size);
3503 loi_init((*lsmp)->lsm_oinfo[0]);
3507 /* XXX zero *lsmp? */
3508 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3509 LASSERT((*lsmp)->lsm_object_id);
3512 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3517 /* caller must hold fo_create_locks[oa->o_gr] */
3518 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3519 struct filter_obd *filter)
3521 struct obdo doa = { 0 }; /* XXX obdo on stack */
3528 LASSERT_MDS_GROUP(oa->o_gr);
3529 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3530 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3532 memset(&doa, 0, sizeof(doa));
3534 doa.o_valid |= OBD_MD_FLGROUP;
3535 doa.o_gr = oa->o_gr;
3536 doa.o_mode = S_IFREG;
3538 if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3539 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3540 exp->exp_obd->obd_name, doa.o_gr);
3544 last = filter_last_id(filter, doa.o_gr);
3546 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3548 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3549 exp->exp_obd->obd_name, oa->o_id + 1, last,
3550 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3552 for (id = last; id > oa->o_id; id--) {
3554 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3555 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3556 CEMERG("error destroying precreate objid "LPU64": %d\n",
3559 /* update last_id on disk periodically so that if we restart
3560 * we don't need to re-scan all of the just-deleted objects. */
3561 if ((id & 511) == 0 && !skip_orphan) {
3562 filter_set_last_id(filter, id - 1, doa.o_gr);
3563 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3567 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3568 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3571 filter_set_last_id(filter, id, doa.o_gr);
3572 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3574 /* don't reuse orphan object, return last used objid */
3578 cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3583 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3584 obd_gr group, int *num);
3585 /* returns a negative error or a nonnegative number of files to create */
3586 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3587 obd_gr group, struct obd_trans_info *oti)
3589 struct obd_device *obd = exp->exp_obd;
3590 struct filter_obd *filter = &obd->u.filter;
3594 /* delete orphans request */
3595 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3596 obd_id last = filter_last_id(filter, group);
3598 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3599 CERROR("%s: dropping old orphan cleanup request\n",
3603 /* This causes inflight precreates to abort and drop lock */
3604 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3605 cfs_down(&filter->fo_create_locks[group]);
3606 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3607 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3608 exp->exp_obd->obd_name, group);
3609 cfs_up(&filter->fo_create_locks[group]);
3612 diff = oa->o_id - last;
3613 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3616 if (-diff > OST_MAX_PRECREATE) {
3617 CERROR("%s: ignoring bogus orphan destroy request: "
3618 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3620 /* FIXME: should reset precreate_next_id on MDS */
3621 GOTO(out, rc = -EINVAL);
3624 rc = filter_destroy_precreated(exp, oa, filter);
3626 CERROR("%s: unable to write lastobjid, but "
3627 "orphans were deleted\n", obd->obd_name);
3630 /* XXX: Used by MDS for the first time! */
3631 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3634 cfs_down(&filter->fo_create_locks[group]);
3635 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3636 CERROR("%s: dropping old precreate request\n",
3640 /* only precreate if group == 0 and o_id is specfied */
3641 if (!filter_group_is_mds(group) || oa->o_id == 0)
3644 diff = oa->o_id - filter_last_id(filter, group);
3645 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3646 filter_last_id(filter, group), diff);
3648 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3649 oa->o_id, filter_last_id(filter, group), diff);
3653 oa->o_id = filter_last_id(&obd->u.filter, group);
3654 rc = filter_precreate(obd, oa, group, &diff);
3655 oa->o_id = filter_last_id(&obd->u.filter, group);
3657 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3660 /* else diff == 0 */
3663 cfs_up(&filter->fo_create_locks[group]);
3667 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3668 __u64 max_age, __u32 flags)
3670 struct filter_obd *filter = &obd->u.filter;
3671 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3675 /* at least try to account for cached pages. its still racey and
3676 * might be under-reporting if clients haven't announced their
3677 * caches with brw recently */
3678 cfs_spin_lock(&obd->obd_osfs_lock);
3679 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3680 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3681 cfs_spin_unlock(&obd->obd_osfs_lock);
3683 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3684 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3685 filter->fo_tot_dirty, filter->fo_tot_granted,
3686 filter->fo_tot_pending,
3687 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3689 filter_grant_sanity_check(obd, __FUNCTION__);
3691 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3692 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3693 osfs->os_bsize - 1) >> blockbits));
3695 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3696 struct lr_server_data *lsd = class_server_data(obd);
3697 int index = le32_to_cpu(lsd->lsd_ost_index);
3699 if (obd_fail_val == -1 ||
3700 index == obd_fail_val)
3701 osfs->os_bfree = osfs->os_bavail = 2;
3702 else if (obd_fail_loc & OBD_FAIL_ONCE)
3703 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3706 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3707 * stop creating files on MDS if OST is not good shape to create
3711 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3712 osfs->os_state = OS_STATE_READONLY;
3714 if (filter->fo_raid_degraded)
3715 osfs->os_state |= OS_STATE_DEGRADED;
3719 static int filter_use_existing_obj(struct obd_device *obd,
3720 struct dentry *dchild, void **handle,
3723 struct inode *inode = dchild->d_inode;
3727 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3730 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3731 if (IS_ERR(*handle))
3732 return PTR_ERR(*handle);
3734 iattr.ia_valid = ATTR_MODE;
3735 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3736 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3743 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3746 __u64 os_ffree = -1;
3748 cfs_spin_lock(&obd->obd_osfs_lock);
3749 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3751 os_ffree = obd->obd_osfs.os_ffree;
3752 cfs_spin_unlock(&obd->obd_osfs_lock);
3757 /* We rely on the fact that only one thread will be creating files in a given
3758 * group at a time, which is why we don't need an atomic filter_get_new_id.
3759 * Even if we had that atomic function, the following race would exist:
3761 * thread 1: gets id x from filter_next_id
3762 * thread 2: gets id (x + 1) from filter_next_id
3763 * thread 2: creates object (x + 1)
3764 * thread 1: tries to create object x, gets -ENOSPC
3766 * Caller must hold fo_create_locks[group]
3768 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3769 obd_gr group, int *num)
3771 struct dentry *dchild = NULL, *dparent = NULL;
3772 struct filter_obd *filter;
3773 struct obd_statfs *osfs;
3774 int err = 0, rc = 0, recreate_obj = 0, i;
3775 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3778 void *handle = NULL;
3781 filter = &obd->u.filter;
3783 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3785 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3787 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3788 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3791 OBD_ALLOC(osfs, sizeof(*osfs));
3794 rc = filter_statfs(obd, osfs, cfs_time_current_64() - CFS_HZ,
3796 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3797 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3798 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3799 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3803 OBD_FREE(osfs, sizeof(*osfs));
3808 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3809 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3811 for (i = 0; i < *num && err == 0; i++) {
3812 int cleanup_phase = 0;
3814 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3815 CWARN("%s: create aborted by destroy\n",
3824 last_id = filter_last_id(filter, group);
3825 if (next_id > last_id) {
3826 CERROR("Error: Trying to recreate obj greater"
3827 "than last id "LPD64" > "LPD64"\n",
3829 GOTO(cleanup, rc = -EINVAL);
3832 next_id = filter_last_id(filter, group) + 1;
3834 dparent = filter_parent_lock(obd, group, next_id);
3835 if (IS_ERR(dparent))
3836 GOTO(cleanup, rc = PTR_ERR(dparent));
3837 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3839 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3841 GOTO(cleanup, rc = PTR_ERR(dchild));
3842 cleanup_phase = 2; /* f_dput(dchild) */
3844 if (dchild->d_inode != NULL) {
3845 /* This would only happen if lastobjid was bad on disk*/
3846 /* Could also happen if recreating missing obj but it
3847 * already exists. */
3849 CERROR("%s: recreating existing object %.*s?\n",
3850 obd->obd_name, dchild->d_name.len,
3851 dchild->d_name.name);
3853 /* Use these existing objects if they are
3855 if (dchild->d_inode->i_size == 0) {
3856 rc = filter_use_existing_obj(obd,dchild,
3857 &handle, &cleanup_phase);
3864 CERROR("%s: Serious error: objid %.*s already "
3865 "exists; is this filesystem corrupt?\n",
3866 obd->obd_name, dchild->d_name.len,
3867 dchild->d_name.name);
3870 GOTO(cleanup, rc = -EEXIST);
3873 handle = fsfilt_start_log(obd, dparent->d_inode,
3874 FSFILT_OP_CREATE, NULL, 1);
3876 GOTO(cleanup, rc = PTR_ERR(handle));
3879 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3880 ",od->o_id="LPU64")\n", obd->obd_name, group,
3883 /* We mark object SUID+SGID to flag it for accepting UID+GID
3884 * from client on first write. Currently the permission bits
3885 * on the OST are never used, so this is OK. */
3886 rc = ll_vfs_create(dparent->d_inode, dchild,
3887 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3889 CERROR("create failed rc = %d\n", rc);
3890 if (rc == -ENOSPC) {
3891 os_ffree = filter_calc_free_inodes(obd);
3893 CERROR("%s: free inode "LPU64"\n",
3894 obd->obd_name, os_ffree);
3899 if (dchild->d_inode)
3900 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3901 dchild->d_inode->i_ino);
3904 if (!recreate_obj) {
3905 filter_set_last_id(filter, next_id, group);
3906 err = filter_update_last_objid(obd, group, 0);
3908 CERROR("unable to write lastobjid "
3909 "but file created\n");
3913 switch(cleanup_phase) {
3915 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3917 CERROR("error on commit, err = %d\n", err);
3924 filter_parent_unlock(dparent);
3931 if (cfs_time_after(jiffies, enough_time)) {
3933 "%s: precreate slow - want %d got %d \n",
3934 obd->obd_name, *num, i);
3941 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3942 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3947 static int filter_create(struct obd_export *exp, struct obdo *oa,
3948 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3950 struct obd_device *obd = exp->exp_obd;
3951 struct filter_export_data *fed;
3952 struct filter_obd *filter;
3953 struct lvfs_run_ctxt saved;
3954 struct lov_stripe_md *lsm = NULL;
3955 int rc = 0, diff, group = oa->o_gr;
3958 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3959 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3961 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3962 CERROR("!!! nid %s sent invalid object group %d\n",
3963 obd_export_nid2str(exp), group);
3967 fed = &exp->exp_filter_data;
3968 filter = &obd->u.filter;
3970 if (fed->fed_group != group) {
3971 CERROR("!!! this export (nid %s) used object group %d "
3972 "earlier; now it's trying to use group %d! This could "
3973 "be a bug in the MDS. Please report to "
3974 "http://bugzilla.lustre.org/\n",
3975 obd_export_nid2str(exp), fed->fed_group, group);
3982 rc = obd_alloc_memmd(exp, &lsm);
3989 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3991 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3992 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3993 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3994 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3995 oa->o_id, filter_last_id(filter,
4000 cfs_down(&filter->fo_create_locks[oa->o_gr]);
4001 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4002 cfs_up(&filter->fo_create_locks[oa->o_gr]);
4005 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4008 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4009 if (rc && ea != NULL && *ea != lsm) {
4010 obd_free_memmd(exp, &lsm);
4011 } else if (rc == 0 && ea != NULL) {
4012 /* XXX LOV STACKING: the lsm that is passed to us from
4013 * LOV does not have valid lsm_oinfo data structs, so
4014 * don't go touching that. This needs to be fixed in a
4016 lsm->lsm_object_id = oa->o_id;
4023 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4024 struct lov_stripe_md *md, struct obd_trans_info *oti,
4025 struct obd_export *md_exp, void *capa)
4027 unsigned int qcids[MAXQUOTAS] = {0, 0};
4028 struct obd_device *obd;
4029 struct filter_obd *filter;
4030 struct dentry *dchild = NULL, *dparent = NULL;
4031 struct lustre_handle lockh = { 0 };
4032 struct lvfs_run_ctxt saved;
4033 void *handle = NULL;
4034 struct llog_cookie *fcc = NULL;
4035 int rc, rc2, cleanup_phase = 0, sync = 0;
4039 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4040 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4041 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4046 filter = &obd->u.filter;
4048 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4051 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4052 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4054 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4056 GOTO(cleanup, rc = PTR_ERR(dchild));
4059 if (dchild->d_inode == NULL) {
4060 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4062 /* If object already gone, cancel cookie right now */
4063 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4064 struct llog_ctxt *ctxt;
4065 struct obd_llog_group *olg;
4067 olg = filter_find_olg(obd, oa->o_gr);
4069 CERROR(" %s: can not find olg of group %d\n",
4070 obd->obd_name, (int)oa->o_gr);
4071 GOTO(cleanup, rc = PTR_ERR(olg));
4073 fcc = &oa->o_lcookie;
4074 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4075 llog_cancel(ctxt, NULL, 1, fcc, 0);
4076 llog_ctxt_put(ctxt);
4077 fcc = NULL; /* we didn't allocate fcc, don't free it */
4079 GOTO(cleanup, rc = -ENOENT);
4082 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4086 /* Our MDC connection is established by the MDS to us */
4087 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4088 OBD_ALLOC(fcc, sizeof(*fcc));
4090 *fcc = oa->o_lcookie;
4092 DQUOT_INIT(dchild->d_inode);
4094 /* we're gonna truncate it first in order to avoid possible deadlock:
4096 * open trasaction open transaction
4097 * down(i_zombie) down(i_zombie)
4098 * restart transaction
4099 * (see BUG 4180) -bzzz
4101 * take i_alloc_sem too to prevent other threads from writing to the
4102 * file while we are truncating it. This can cause lock ordering issue
4103 * between page lock, i_mutex & starting new journal handle.
4104 * (see bug 20321) -johann
4106 down_write(&dchild->d_inode->i_alloc_sem);
4107 LOCK_INODE_MUTEX(dchild->d_inode);
4109 /* VBR: version recovery check */
4110 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4112 UNLOCK_INODE_MUTEX(dchild->d_inode);
4113 up_write(&dchild->d_inode->i_alloc_sem);
4117 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4119 if (IS_ERR(handle)) {
4120 UNLOCK_INODE_MUTEX(dchild->d_inode);
4121 up_write(&dchild->d_inode->i_alloc_sem);
4122 GOTO(cleanup, rc = PTR_ERR(handle));
4125 iattr.ia_valid = ATTR_SIZE;
4127 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4128 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4129 UNLOCK_INODE_MUTEX(dchild->d_inode);
4130 up_write(&dchild->d_inode->i_alloc_sem);
4134 GOTO(cleanup, rc = rc2);
4136 /* We don't actually need to lock the parent until we are unlinking
4137 * here, and not while truncating above. That avoids holding the
4138 * parent lock for a long time during truncate, which can block other
4139 * threads from doing anything to objects in that directory. bug 7171 */
4140 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4141 if (IS_ERR(dparent))
4142 GOTO(cleanup, rc = PTR_ERR(dparent));
4143 cleanup_phase = 3; /* filter_parent_unlock */
4145 LOCK_INODE_MUTEX(dchild->d_inode);
4146 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4147 if (IS_ERR(handle)) {
4148 UNLOCK_INODE_MUTEX(dchild->d_inode);
4149 GOTO(cleanup, rc = PTR_ERR(handle));
4151 cleanup_phase = 4; /* fsfilt_commit */
4153 /* Quota release need uid/gid of inode */
4154 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4156 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4158 /* this drops dchild->d_inode->i_mutex unconditionally */
4159 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4163 switch(cleanup_phase) {
4166 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4167 oti->oti_handle : handle,
4168 filter_cancel_cookies_cb,
4170 /* If add_journal_cb failed, then filter_finish_transno
4171 * will commit the handle and we will do a sync
4172 * on commit. then we call callback directly to free
4175 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4177 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4180 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4182 CERROR("error on commit, err = %d\n", rc2);
4189 filter_parent_unlock(dparent);
4191 filter_fini_destroy(obd, &lockh);
4195 OBD_FREE(fcc, sizeof(*fcc));
4197 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4200 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4204 /* trigger quota release */
4205 qcids[USRQUOTA] = oa->o_uid;
4206 qcids[GRPQUOTA] = oa->o_gid;
4207 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4210 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4214 /* NB start and end are used for punch, but not truncate */
4215 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4216 struct obd_trans_info *oti,
4217 struct ptlrpc_request_set *rqset)
4222 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4223 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4224 oinfo->oi_policy.l_extent.end);
4228 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4229 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4230 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4232 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4233 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4234 rc = filter_setattr(exp, oinfo, oti);
4235 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4239 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4240 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4243 struct lvfs_run_ctxt saved;
4244 struct obd_device_target *obt;
4245 struct dentry *dentry;
4249 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4250 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4251 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4255 obt = &exp->exp_obd->u.obt;
4257 /* An objid of zero is taken to mean "sync whole filesystem" */
4258 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4259 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4260 /* Flush any remaining cancel messages out to the target */
4261 filter_sync_llogs(exp->exp_obd, exp);
4265 dentry = filter_oa2dentry(exp->exp_obd, oa);
4267 RETURN(PTR_ERR(dentry));
4269 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4271 LOCK_INODE_MUTEX(dentry->d_inode);
4273 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4275 /* just any file to grab fsync method - "file" arg unused */
4276 struct file *file = obt->obt_rcvd_filp;
4278 if (file->f_op && file->f_op->fsync)
4279 rc = file->f_op->fsync(NULL, dentry, 1);
4281 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4285 UNLOCK_INODE_MUTEX(dentry->d_inode);
4287 oa->o_valid = OBD_MD_FLID;
4288 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4290 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4296 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4297 void *key, __u32 *vallen, void *val,
4298 struct lov_stripe_md *lsm)
4300 struct obd_device *obd;
4303 obd = class_exp2obd(exp);
4305 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4309 if (KEY_IS(KEY_BLOCKSIZE)) {
4310 __u32 *blocksize = val;
4312 if (*vallen < sizeof(*blocksize))
4314 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4316 *vallen = sizeof(*blocksize);
4320 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4321 __u32 *blocksize_bits = val;
4322 if (blocksize_bits) {
4323 if (*vallen < sizeof(*blocksize_bits))
4325 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4327 *vallen = sizeof(*blocksize_bits);
4331 if (KEY_IS(KEY_LAST_ID)) {
4332 obd_id *last_id = val;
4333 /* FIXME: object groups */
4335 if (*vallen < sizeof(*last_id))
4337 *last_id = filter_last_id(&obd->u.filter,
4338 exp->exp_filter_data.fed_group);
4340 *vallen = sizeof(*last_id);
4344 if (KEY_IS(KEY_FIEMAP)) {
4345 struct ll_fiemap_info_key *fm_key = key;
4346 struct dentry *dentry;
4347 struct ll_user_fiemap *fiemap = val;
4348 struct lvfs_run_ctxt saved;
4351 if (fiemap == NULL) {
4352 *vallen = fiemap_count_to_size(
4353 fm_key->fiemap.fm_extent_count);
4357 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4360 RETURN(PTR_ERR(dentry));
4362 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4363 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4364 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4366 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4372 CDEBUG(D_IOCTL, "invalid key\n");
4376 static inline int filter_setup_llog_group(struct obd_export *exp,
4377 struct obd_device *obd,
4380 struct obd_llog_group *olg;
4381 struct llog_ctxt *ctxt;
4384 olg = filter_find_create_olg(obd, group);
4386 RETURN(PTR_ERR(olg));
4388 llog_group_set_export(olg, exp);
4390 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4391 LASSERTF(ctxt != NULL, "ctxt is null\n");
4393 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4394 llog_ctxt_put(ctxt);
4398 static int filter_set_grant_shrink(struct obd_export *exp,
4399 struct ost_body *body)
4401 /* handle shrink grant */
4402 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4403 filter_grant_incoming(exp, &body->oa);
4404 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4410 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4412 struct obd_device *obd;
4418 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4422 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4423 obd_export_nid2str(exp));
4424 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4426 /* setup llog imports */
4428 group = (int)(*(__u32 *)val);
4430 group = 0; /* default value */
4432 LASSERT_MDS_GROUP(group);
4433 rc = filter_setup_llog_group(exp, obd, group);
4437 if (group == FILTER_GROUP_MDS0) {
4438 /* setup llog group 1 for interop */
4439 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4442 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4447 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4448 void *key, __u32 vallen, void *val,
4449 struct ptlrpc_request_set *set)
4451 struct obd_device *obd;
4456 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4460 if (KEY_IS(KEY_CAPA_KEY)) {
4462 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4464 CERROR("filter update capability key failed: %d\n", rc);
4468 if (KEY_IS(KEY_REVIMP_UPD)) {
4469 filter_revimp_update(exp);
4470 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4474 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4475 filter_adapt_sptlrpc_conf(obd, 0);
4479 if (KEY_IS(KEY_MDS_CONN))
4480 RETURN(filter_set_mds_conn(exp, val));
4482 if (KEY_IS(KEY_GRANT_SHRINK))
4483 RETURN(filter_set_grant_shrink(exp, val));
4488 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4489 int len, void *karg, void *uarg)
4491 struct obd_device *obd = exp->exp_obd;
4492 struct obd_ioctl_data *data = karg;
4496 case OBD_IOC_ABORT_RECOVERY: {
4497 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4498 target_stop_recovery_thread(obd);
4502 case OBD_IOC_SYNC: {
4503 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4504 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4508 case OBD_IOC_SET_READONLY: {
4510 struct super_block *sb = obd->u.obt.obt_sb;
4511 struct inode *inode = sb->s_root->d_inode;
4512 BDEVNAME_DECLARE_STORAGE(tmp);
4513 CERROR("*** setting device %s read-only ***\n",
4514 ll_bdevname(sb, tmp));
4516 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4517 if (!IS_ERR(handle))
4518 rc = fsfilt_commit(obd, inode, handle, 1);
4520 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4521 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4523 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4527 case OBD_IOC_CATLOGLIST: {
4528 rc = llog_catalog_list(obd, 1, data);
4532 case OBD_IOC_LLOG_CANCEL:
4533 case OBD_IOC_LLOG_REMOVE:
4534 case OBD_IOC_LLOG_INFO:
4535 case OBD_IOC_LLOG_PRINT: {
4536 /* FIXME to be finished */
4537 RETURN(-EOPNOTSUPP);
4539 struct llog_ctxt *ctxt = NULL;
4541 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4542 rc = llog_ioctl(ctxt, cmd, data);
4543 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4556 static int filter_health_check(struct obd_device *obd)
4558 #ifdef USE_HEALTH_CHECK_WRITE
4559 struct filter_obd *filter = &obd->u.filter;
4564 * health_check to return 0 on healthy
4565 * and 1 on unhealthy.
4567 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4570 #ifdef USE_HEALTH_CHECK_WRITE
4571 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4572 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4577 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4580 return filter_fid2dentry(data, NULL, gr, id);
4583 static int filter_process_config(struct obd_device *obd, obd_count len,
4586 struct lustre_cfg *lcfg = buf;
4587 struct lprocfs_static_vars lvars;
4590 switch (lcfg->lcfg_command) {
4592 lprocfs_filter_init_vars(&lvars);
4594 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4604 static struct lvfs_callback_ops filter_lvfs_ops = {
4605 l_fid2dentry: filter_lvfs_fid2dentry,
4608 static struct obd_ops filter_obd_ops = {
4609 .o_owner = THIS_MODULE,
4610 .o_get_info = filter_get_info,
4611 .o_set_info_async = filter_set_info_async,
4612 .o_setup = filter_setup,
4613 .o_precleanup = filter_precleanup,
4614 .o_cleanup = filter_cleanup,
4615 .o_connect = filter_connect,
4616 .o_reconnect = filter_reconnect,
4617 .o_disconnect = filter_disconnect,
4618 .o_ping = filter_ping,
4619 .o_init_export = filter_init_export,
4620 .o_destroy_export = filter_destroy_export,
4621 .o_statfs = filter_statfs,
4622 .o_getattr = filter_getattr,
4623 .o_unpackmd = filter_unpackmd,
4624 .o_create = filter_create,
4625 .o_setattr = filter_setattr,
4626 .o_destroy = filter_destroy,
4627 .o_brw = filter_brw,
4628 .o_punch = filter_truncate,
4629 .o_sync = filter_sync,
4630 .o_preprw = filter_preprw,
4631 .o_commitrw = filter_commitrw,
4632 .o_llog_init = filter_llog_init,
4633 .o_llog_connect = filter_llog_connect,
4634 .o_llog_finish = filter_llog_finish,
4635 .o_iocontrol = filter_iocontrol,
4636 .o_health_check = filter_health_check,
4637 .o_process_config = filter_process_config,
4640 quota_interface_t *filter_quota_interface_ref;
4641 extern quota_interface_t filter_quota_interface;
4643 static int __init obdfilter_init(void)
4645 struct lprocfs_static_vars lvars;
4648 /** sanity check for group<->mdsno conversion */
4649 for (i = 0; i < 32; i++)
4650 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4652 lprocfs_filter_init_vars(&lvars);
4654 cfs_request_module("%s", "lquota");
4655 OBD_ALLOC(obdfilter_created_scratchpad,
4656 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4657 sizeof(*obdfilter_created_scratchpad));
4658 if (obdfilter_created_scratchpad == NULL)
4661 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4662 sizeof(struct filter_mod_data),
4665 GOTO(out, rc = -ENOMEM);
4667 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4668 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4670 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4671 LUSTRE_OST_NAME, NULL);
4675 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4676 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4677 ll_fmd_cachep = NULL;
4679 if (filter_quota_interface_ref)
4680 PORTAL_SYMBOL_PUT(filter_quota_interface);
4682 OBD_FREE(obdfilter_created_scratchpad,
4683 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4684 sizeof(*obdfilter_created_scratchpad));
4690 static void __exit obdfilter_exit(void)
4692 if (filter_quota_interface_ref)
4693 PORTAL_SYMBOL_PUT(filter_quota_interface);
4695 if (ll_fmd_cachep) {
4696 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4697 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4698 ll_fmd_cachep = NULL;
4701 class_unregister_type(LUSTRE_OST_NAME);
4702 OBD_FREE(obdfilter_created_scratchpad,
4703 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4704 sizeof(*obdfilter_created_scratchpad));
4707 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4708 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4709 MODULE_LICENSE("GPL");
4711 module_init(obdfilter_init);
4712 module_exit(obdfilter_exit);