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 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 struct obd_export *exp = cb_data;
91 obd_transno_commit_cb(obd, transno, exp, error);
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 spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct filter_obd *filter = &exp->exp_obd->u.filter;
124 struct filter_export_data *fed = &exp->exp_filter_data;
125 struct lsd_client_data *lcd = fed->fed_lcd;
128 int err, log_pri = D_RPCTRACE;
130 /* Propagate error code. */
134 if (!exp->exp_obd->obd_replayable || oti == NULL)
137 /* we don't allocate new transnos for replayed requests */
138 spin_lock(&filter->fo_translock);
139 if (oti->oti_transno == 0) {
140 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
141 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
143 last_rcvd = oti->oti_transno;
144 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
145 filter->fo_fsd->lsd_last_transno =
146 cpu_to_le64(last_rcvd);
148 oti->oti_transno = last_rcvd;
149 if (last_rcvd <= le64_to_cpu(lcd->lcd_last_transno)) {
150 spin_unlock(&filter->fo_translock);
153 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
154 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
155 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
156 spin_unlock(&filter->fo_translock);
159 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
161 off = fed->fed_lr_off;
163 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
164 fed->fed_lr_idx, fed->fed_lr_off);
168 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
174 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
175 lcd, sizeof(*lcd), &off,
176 force_sync | exp->exp_need_sync);
178 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
186 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
187 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
192 void f_dput(struct dentry *dentry)
194 /* Can't go inside filter_ddelete because it can block */
195 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
196 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
197 LASSERT(atomic_read(&dentry->d_count) > 0);
202 static void init_brw_stats(struct brw_stats *brw_stats)
205 for (i = 0; i < BRW_LAST; i++)
206 spin_lock_init(&brw_stats->hist[i].oh_lock);
209 static int lprocfs_init_rw_stats(struct obd_device *obd,
210 struct lprocfs_stats **stats)
214 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
215 LPROC_FILTER_LAST - 1;
216 *stats = lprocfs_alloc_stats(num_stats, 0);
220 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
221 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
222 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
223 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
224 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
229 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
230 plus the procfs overhead :( */
231 static int filter_export_stats_init(struct obd_device *obd,
232 struct obd_export *exp,
238 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
239 /* Self-export gets no proc entry */
242 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
244 /* Mask error for already created
252 struct nid_stat *tmp = exp->exp_nid_stats;
253 LASSERT(tmp != NULL);
255 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
256 if (tmp->nid_brw_stats == NULL)
259 init_brw_stats(tmp->nid_brw_stats);
260 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
261 0644, &filter_per_nid_stats_fops,
264 CWARN("Error adding the brw_stats file\n");
266 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
270 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
274 /* Always add in ldlm_stats */
275 tmp->nid_ldlm_stats = lprocfs_alloc_stats(LDLM_LAST_OPC -
277 if (tmp->nid_ldlm_stats == NULL)
280 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
282 tmp->nid_ldlm_stats);
290 /* Add client data to the FILTER. We use a bitmap to locate a free space
291 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
292 * Otherwise, we have just read the data from the last_rcvd file and
293 * we know its offset. */
294 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
297 struct filter_obd *filter = &obd->u.filter;
298 struct filter_export_data *fed = &exp->exp_filter_data;
299 unsigned long *bitmap = filter->fo_last_rcvd_slots;
300 int new_client = (cl_idx == -1);
304 LASSERT(bitmap != NULL);
305 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
308 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
311 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
312 * there's no need for extra complication here
315 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
317 if (cl_idx >= LR_MAX_CLIENTS) {
318 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
322 if (test_and_set_bit(cl_idx, bitmap)) {
323 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
328 if (test_and_set_bit(cl_idx, bitmap)) {
329 CERROR("FILTER client %d: bit already set in bitmap!\n",
335 fed->fed_lr_idx = cl_idx;
336 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
337 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
338 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
340 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
341 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
344 struct lvfs_run_ctxt saved;
345 loff_t off = fed->fed_lr_off;
349 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
350 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
352 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
353 /* Transaction needed to fix bug 1403 */
354 handle = fsfilt_start(obd,
355 filter->fo_rcvd_filp->f_dentry->d_inode,
356 FSFILT_OP_SETATTR, NULL);
357 if (IS_ERR(handle)) {
358 rc = PTR_ERR(handle);
359 CERROR("unable to start transaction: rc %d\n", rc);
361 fed->fed_lcd->lcd_last_epoch =
362 filter->fo_fsd->lsd_start_epoch;
363 exp->exp_last_request_time = cfs_time_current_sec();
364 rc = fsfilt_add_journal_cb(obd, 0, handle,
365 target_client_add_cb, exp);
367 spin_lock(&exp->exp_lock);
368 exp->exp_need_sync = 1;
369 spin_unlock(&exp->exp_lock);
371 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
373 sizeof(*fed->fed_lcd),
374 &off, rc /* sync if no cb */);
376 filter->fo_rcvd_filp->f_dentry->d_inode,
379 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
382 CERROR("error writing %s client idx %u: rc %d\n",
383 LAST_RCVD, fed->fed_lr_idx, rc);
390 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
392 static int filter_client_free(struct obd_export *exp)
394 struct filter_export_data *fed = &exp->exp_filter_data;
395 struct filter_obd *filter = &exp->exp_obd->u.filter;
396 struct obd_device *obd = exp->exp_obd;
397 struct lvfs_run_ctxt saved;
402 if (fed->fed_lcd == NULL)
405 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
406 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
409 LASSERT(filter->fo_last_rcvd_slots != NULL);
411 off = fed->fed_lr_off;
413 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
414 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
416 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
417 * we leak a client slot that will be cleaned on the next recovery. */
419 CERROR("%s: client idx %d has med_off %lld\n",
420 obd->obd_name, fed->fed_lr_idx, off);
421 GOTO(free, rc = -EINVAL);
424 /* Clear the bit _after_ zeroing out the client so we don't
425 race with filter_client_add and zero out new clients.*/
426 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
427 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
432 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
433 /* Don't force sync on disconnect if aborting recovery,
434 * or it does num_clients * num_osts. b=17194 */
435 int need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
436 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
437 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
438 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
439 sizeof(zero_lcd), &off, 0);
441 /* Make sure the server's last_transno is up to date. Do this
442 * after the client is freed so we know all the client's
443 * transactions have been committed. */
445 filter_update_server_data(obd, filter->fo_rcvd_filp,
446 filter->fo_fsd, need_sync);
447 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
449 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
450 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
451 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
452 LAST_RCVD, need_sync ? "" : "a", rc);
455 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
456 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
463 OBD_FREE_PTR(fed->fed_lcd);
469 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
470 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
471 struct filter_mod_data *fmd)
473 LASSERT_SPIN_LOCKED(&fed->fed_lock);
474 if (--fmd->fmd_refcount == 0) {
475 /* XXX when we have persistent reservations and the handle
476 * is stored herein we need to drop it here. */
477 fed->fed_mod_count--;
478 list_del(&fmd->fmd_list);
479 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
483 /* drop fmd reference, free it if last ref */
484 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
486 struct filter_export_data *fed;
491 fed = &exp->exp_filter_data;
492 spin_lock(&fed->fed_lock);
493 filter_fmd_put_nolock(fed, fmd); /* caller reference */
494 spin_unlock(&fed->fed_lock);
497 /* expire entries from the end of the list if there are too many
498 * or they are too old */
499 static void filter_fmd_expire_nolock(struct filter_obd *filter,
500 struct filter_export_data *fed,
501 struct filter_mod_data *keep)
503 struct filter_mod_data *fmd, *tmp;
505 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
509 if (time_before(jiffies, fmd->fmd_expire) &&
510 fed->fed_mod_count < filter->fo_fmd_max_num)
513 list_del_init(&fmd->fmd_list);
514 filter_fmd_put_nolock(fed, fmd); /* list reference */
518 void filter_fmd_expire(struct obd_export *exp)
520 spin_lock(&exp->exp_filter_data.fed_lock);
521 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
522 &exp->exp_filter_data, NULL);
523 spin_unlock(&exp->exp_filter_data.fed_lock);
526 /* find specified objid, group in export fmd list.
527 * caller must hold fed_lock and take fmd reference itself */
528 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
529 struct filter_export_data *fed,
530 obd_id objid, obd_gr group)
532 struct filter_mod_data *found = NULL, *fmd;
534 LASSERT_SPIN_LOCKED(&fed->fed_lock);
536 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
537 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
539 list_del(&fmd->fmd_list);
540 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
541 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
546 filter_fmd_expire_nolock(filter, fed, found);
551 /* Find fmd based on objid and group, or return NULL if not found. */
552 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
553 obd_id objid, obd_gr group)
555 struct filter_mod_data *fmd;
557 spin_lock(&exp->exp_filter_data.fed_lock);
558 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
559 &exp->exp_filter_data, objid, group);
561 fmd->fmd_refcount++; /* caller reference */
562 spin_unlock(&exp->exp_filter_data.fed_lock);
567 /* Find fmd based on objid and group, or create a new one if none is found.
568 * It is possible for this function to return NULL under memory pressure,
569 * or if objid = 0 is passed (which will only cause old entries to expire).
570 * Currently this is not fatal because any fmd state is transient and
571 * may also be freed when it gets sufficiently old. */
572 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
573 obd_id objid, obd_gr group)
575 struct filter_export_data *fed = &exp->exp_filter_data;
576 struct filter_mod_data *found = NULL, *fmd_new = NULL;
578 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
580 spin_lock(&fed->fed_lock);
581 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
584 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
585 fmd_new->fmd_id = objid;
586 fmd_new->fmd_gr = group;
587 fmd_new->fmd_refcount++; /* list reference */
589 fed->fed_mod_count++;
591 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
595 found->fmd_refcount++; /* caller reference */
596 found->fmd_expire = jiffies +
597 exp->exp_obd->u.filter.fo_fmd_max_age;
600 spin_unlock(&fed->fed_lock);
606 /* drop fmd list reference so it will disappear when last reference is put.
607 * This isn't so critical because it would in fact only affect the one client
608 * that is doing the unlink and at worst we have an stale entry referencing
609 * an object that should never be used again. */
610 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
612 struct filter_mod_data *found = NULL;
614 spin_lock(&exp->exp_filter_data.fed_lock);
615 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
617 list_del_init(&found->fmd_list);
618 filter_fmd_put_nolock(&exp->exp_filter_data, found);
620 spin_unlock(&exp->exp_filter_data.fed_lock);
623 #define filter_fmd_drop(exp, objid, group)
626 /* remove all entries from fmd list */
627 static void filter_fmd_cleanup(struct obd_export *exp)
629 struct filter_export_data *fed = &exp->exp_filter_data;
630 struct filter_mod_data *fmd = NULL, *tmp;
632 spin_lock(&fed->fed_lock);
633 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
634 list_del_init(&fmd->fmd_list);
635 filter_fmd_put_nolock(fed, fmd);
637 spin_unlock(&fed->fed_lock);
640 static int filter_init_export(struct obd_export *exp)
642 spin_lock_init(&exp->exp_filter_data.fed_lock);
643 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
645 spin_lock(&exp->exp_lock);
646 exp->exp_connecting = 1;
647 spin_unlock(&exp->exp_lock);
649 return ldlm_init_export(exp);
652 static int filter_free_server_data(struct filter_obd *filter)
654 OBD_FREE_PTR(filter->fo_fsd);
655 filter->fo_fsd = NULL;
656 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
657 filter->fo_last_rcvd_slots = NULL;
661 /* assumes caller is already in kernel ctxt */
662 int filter_update_server_data(struct obd_device *obd, struct file *filp,
663 struct lr_server_data *fsd, int force_sync)
669 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
670 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
671 le64_to_cpu(fsd->lsd_last_transno));
672 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
673 le64_to_cpu(fsd->lsd_mount_count));
675 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
677 CERROR("error writing lr_server_data: rc = %d\n", rc);
682 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
685 struct filter_obd *filter = &obd->u.filter;
691 if (filter->fo_last_objid_files[group] == NULL) {
692 CERROR("Object group "LPU64" not fully setup; not updating "
693 "last_objid\n", group);
697 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
698 obd->obd_name, group, filter->fo_last_objids[group]);
700 tmp = cpu_to_le64(filter->fo_last_objids[group]);
701 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
702 &tmp, sizeof(tmp), &off, force_sync);
704 CERROR("error writing group "LPU64" last objid: rc = %d\n",
708 extern int ost_handle(struct ptlrpc_request *req);
709 /* assumes caller has already in kernel ctxt */
710 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
712 struct filter_obd *filter = &obd->u.filter;
713 struct lr_server_data *fsd;
714 struct lsd_client_data *lcd = NULL;
715 struct inode *inode = filp->f_dentry->d_inode;
716 unsigned long last_rcvd_size = i_size_read(inode);
723 /* ensure padding in the struct is the correct size */
724 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
725 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
726 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
727 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
729 OBD_ALLOC(fsd, sizeof(*fsd));
732 filter->fo_fsd = fsd;
734 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
735 if (filter->fo_last_rcvd_slots == NULL) {
736 OBD_FREE(fsd, sizeof(*fsd));
740 if (last_rcvd_size == 0) {
741 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
743 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
744 fsd->lsd_last_transno = 0;
745 mount_count = fsd->lsd_mount_count = 0;
746 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
747 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
748 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
749 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
750 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
751 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
753 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
755 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
759 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
760 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
761 "using the wrong disk %s. Were the "
762 "/dev/ assignments rearranged?\n",
763 obd->obd_uuid.uuid, fsd->lsd_uuid);
764 GOTO(err_fsd, rc = -EINVAL);
766 mount_count = le64_to_cpu(fsd->lsd_mount_count);
767 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
769 /* Assume old last_rcvd format unless I_C_LR is set */
770 if (!(fsd->lsd_feature_incompat &
771 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
772 fsd->lsd_last_transno = fsd->lsd_compat14;
776 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
777 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
778 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
779 ~FILTER_INCOMPAT_SUPP);
780 GOTO(err_fsd, rc = -EINVAL);
782 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
783 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
784 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
785 ~FILTER_ROCOMPAT_SUPP);
786 /* Do something like remount filesystem read-only */
787 GOTO(err_fsd, rc = -EINVAL);
790 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
792 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
793 obd->obd_name, start_epoch);
794 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
795 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
796 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
797 obd->obd_name, mount_count + 1);
798 CDEBUG(D_INODE, "%s: server data size: %u\n",
799 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
800 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
801 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
802 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
803 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
804 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
805 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
806 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
807 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
808 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
809 le16_to_cpu(fsd->lsd_client_size));
811 if (!obd->obd_replayable) {
812 CWARN("%s: recovery support OFF\n", obd->obd_name);
816 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
817 off < last_rcvd_size; cl_idx++) {
819 struct obd_export *exp;
820 struct filter_export_data *fed;
825 GOTO(err_client, rc = -ENOMEM);
828 /* Don't assume off is incremented properly by
829 * fsfilt_read_record(), in case sizeof(*lcd)
830 * isn't the same as fsd->lsd_client_size. */
831 off = le32_to_cpu(fsd->lsd_client_start) +
832 cl_idx * le16_to_cpu(fsd->lsd_client_size);
833 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
835 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
836 LAST_RCVD, cl_idx, off, rc);
837 break; /* read error shouldn't cause startup to fail */
840 if (lcd->lcd_uuid[0] == '\0') {
841 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
846 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
848 /* These exports are cleaned up by filter_disconnect(), so they
849 * need to be set up like real exports as filter_connect() does.
851 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
853 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
854 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
855 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
857 if (PTR_ERR(exp) == -EALREADY) {
858 /* export already exists, zero out this one */
859 CERROR("Zeroing out duplicate export due to "
861 lcd->lcd_uuid[0] = '\0';
863 GOTO(err_client, rc = PTR_ERR(exp));
866 fed = &exp->exp_filter_data;
868 fed->fed_group = 0; /* will be assigned at connect */
869 filter_export_stats_init(obd, exp, NULL);
870 rc = filter_client_add(obd, exp, cl_idx);
871 /* can't fail for existing client */
872 LASSERTF(rc == 0, "rc = %d\n", rc);
874 /* VBR: set export last committed */
875 exp->exp_last_committed = last_rcvd;
876 spin_lock(&exp->exp_lock);
877 exp->exp_connecting = 0;
878 exp->exp_in_recovery = 0;
879 spin_unlock(&exp->exp_lock);
880 spin_lock_bh(&obd->obd_processing_task_lock);
881 obd->obd_max_recoverable_clients++;
882 spin_unlock_bh(&obd->obd_processing_task_lock);
884 class_export_put(exp);
887 /* Need to check last_rcvd even for duplicated exports. */
888 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
891 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
892 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
898 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
900 target_recovery_init(&filter->fo_lut, ost_handle);
903 filter->fo_mount_count = mount_count + 1;
904 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
906 /* save it, so mount count and last_transno is current */
907 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
909 GOTO(err_client, rc);
914 target_recovery_fini(obd);
916 filter_free_server_data(filter);
920 static int filter_cleanup_groups(struct obd_device *obd)
922 struct filter_obd *filter = &obd->u.filter;
924 struct dentry *dentry;
928 if (filter->fo_dentry_O_groups != NULL) {
929 for (i = 0; i < filter->fo_group_count; i++) {
930 dentry = filter->fo_dentry_O_groups[i];
934 OBD_FREE(filter->fo_dentry_O_groups,
935 filter->fo_group_count *
936 sizeof(*filter->fo_dentry_O_groups));
937 filter->fo_dentry_O_groups = NULL;
939 if (filter->fo_last_objid_files != NULL) {
940 for (i = 0; i < filter->fo_group_count; i++) {
941 filp = filter->fo_last_objid_files[i];
945 OBD_FREE(filter->fo_last_objid_files,
946 filter->fo_group_count *
947 sizeof(*filter->fo_last_objid_files));
948 filter->fo_last_objid_files = NULL;
950 if (filter->fo_dentry_O_sub != NULL) {
951 for (i = 0; i < filter->fo_group_count; i++) {
952 for (j = 0; j < filter->fo_subdir_count; j++) {
953 dentry = filter->fo_dentry_O_sub[i].dentry[j];
958 OBD_FREE(filter->fo_dentry_O_sub,
959 filter->fo_group_count *
960 sizeof(*filter->fo_dentry_O_sub));
961 filter->fo_dentry_O_sub = NULL;
963 if (filter->fo_last_objids != NULL) {
964 OBD_FREE(filter->fo_last_objids,
965 filter->fo_group_count *
966 sizeof(*filter->fo_last_objids));
967 filter->fo_last_objids = NULL;
969 if (filter->fo_dentry_O != NULL) {
970 f_dput(filter->fo_dentry_O);
971 filter->fo_dentry_O = NULL;
976 static int filter_update_last_group(struct obd_device *obd, int group)
978 struct filter_obd *filter = &obd->u.filter;
979 struct file *filp = NULL;
980 int last_group = 0, rc;
984 if (group <= filter->fo_committed_group)
987 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
991 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
995 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
997 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1000 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
1001 "off = %llu and last_group = %d\n", off, last_group);
1003 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1004 obd->obd_name, last_group, group);
1008 /* must be sync: bXXXX */
1009 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1011 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1015 filter->fo_committed_group = group;
1018 filp_close(filp, 0);
1022 static int filter_read_group_internal(struct obd_device *obd, int group,
1025 struct filter_obd *filter = &obd->u.filter;
1026 __u64 *new_objids = NULL;
1027 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1028 struct dentry **new_groups = NULL;
1029 struct file **new_files = NULL;
1030 struct dentry *dentry;
1032 int old_count = filter->fo_group_count, rc, stage = 0, i;
1036 int len = group + 1;
1038 snprintf(name, 24, "%d", group);
1042 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1044 if (IS_ERR(dentry)) {
1045 CERROR("Cannot lookup expected object group %d: %ld\n",
1046 group, PTR_ERR(dentry));
1047 RETURN(PTR_ERR(dentry));
1050 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1052 if (IS_ERR(dentry)) {
1053 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1055 RETURN(PTR_ERR(dentry));
1060 snprintf(name, 24, "O/%d/LAST_ID", group);
1062 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1064 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1065 GOTO(cleanup, rc = PTR_ERR(filp));
1069 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1071 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1075 if (filter->fo_subdir_count) {
1076 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1077 if (tmp_subdirs == NULL)
1078 GOTO(cleanup, rc = -ENOMEM);
1081 for (i = 0; i < filter->fo_subdir_count; i++) {
1083 snprintf(dir, sizeof(dir), "d%u", i);
1085 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1088 if (IS_ERR(tmp_subdirs->dentry[i])) {
1089 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1090 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1095 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1096 tmp_subdirs->dentry[i]);
1100 /* 'group' is an index; we need an array of length 'group + 1' */
1101 if (group + 1 > old_count) {
1102 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1103 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1104 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1105 OBD_ALLOC(new_files, len * sizeof(*new_files));
1107 if (new_objids == NULL || new_subdirs == NULL ||
1108 new_groups == NULL || new_files == NULL)
1109 GOTO(cleanup, rc = -ENOMEM);
1112 memcpy(new_objids, filter->fo_last_objids,
1113 old_count * sizeof(*new_objids));
1114 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1115 old_count * sizeof(*new_subdirs));
1116 memcpy(new_groups, filter->fo_dentry_O_groups,
1117 old_count * sizeof(*new_groups));
1118 memcpy(new_files, filter->fo_last_objid_files,
1119 old_count * sizeof(*new_files));
1121 OBD_FREE(filter->fo_last_objids,
1122 old_count * sizeof(*new_objids));
1123 OBD_FREE(filter->fo_dentry_O_sub,
1124 old_count * sizeof(*new_subdirs));
1125 OBD_FREE(filter->fo_dentry_O_groups,
1126 old_count * sizeof(*new_groups));
1127 OBD_FREE(filter->fo_last_objid_files,
1128 old_count * sizeof(*new_files));
1130 filter->fo_last_objids = new_objids;
1131 filter->fo_dentry_O_sub = new_subdirs;
1132 filter->fo_dentry_O_groups = new_groups;
1133 filter->fo_last_objid_files = new_files;
1134 filter->fo_group_count = len;
1137 filter->fo_dentry_O_groups[group] = dentry;
1138 filter->fo_last_objid_files[group] = filp;
1139 if (filter->fo_subdir_count) {
1140 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1141 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1144 filter_update_last_group(obd, group);
1146 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1147 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1148 rc = filter_update_last_objid(obd, group, 1);
1152 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1153 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1154 obd->obd_name, group, last_objid);
1159 if (new_objids != NULL)
1160 OBD_FREE(new_objids, len * sizeof(*new_objids));
1161 if (new_subdirs != NULL)
1162 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1163 if (new_groups != NULL)
1164 OBD_FREE(new_groups, len * sizeof(*new_groups));
1165 if (new_files != NULL)
1166 OBD_FREE(new_files, len * sizeof(*new_files));
1168 if (filter->fo_subdir_count) {
1169 for (i = 0; i < filter->fo_subdir_count; i++) {
1170 if (tmp_subdirs->dentry[i] != NULL)
1171 dput(tmp_subdirs->dentry[i]);
1173 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1176 filp_close(filp, 0);
1183 static int filter_read_groups(struct obd_device *obd, int last_group,
1186 struct filter_obd *filter = &obd->u.filter;
1187 int old_count, group, rc = 0;
1189 down(&filter->fo_init_lock);
1190 old_count = filter->fo_group_count;
1191 for (group = old_count; group <= last_group; group++) {
1193 rc = filter_read_group_internal(obd, group, create);
1197 up(&filter->fo_init_lock);
1201 /* FIXME: object groups */
1202 static int filter_prep_groups(struct obd_device *obd)
1204 struct filter_obd *filter = &obd->u.filter;
1205 struct dentry *dentry, *O_dentry;
1207 int last_group, rc = 0, cleanup_phase = 0;
1211 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1213 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1214 if (IS_ERR(O_dentry)) {
1215 rc = PTR_ERR(O_dentry);
1216 CERROR("cannot open/create O: rc = %d\n", rc);
1219 filter->fo_dentry_O = O_dentry;
1220 cleanup_phase = 1; /* O_dentry */
1222 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1223 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1224 * some time post 1.0 when all old-style OSTs have converted along
1225 * with the init_objid hack. */
1226 dentry = ll_lookup_one_len("R", O_dentry, 1);
1228 GOTO(cleanup, rc = PTR_ERR(dentry));
1229 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1230 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1233 CWARN("converting OST to new object layout\n");
1234 if (IS_ERR(O0_dentry)) {
1235 rc = PTR_ERR(O0_dentry);
1236 CERROR("error looking up O/0: rc %d\n", rc);
1237 GOTO(cleanup_R, rc);
1240 if (O0_dentry->d_inode) {
1241 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1242 GOTO(cleanup_O0, rc = -EEXIST);
1245 LOCK_INODE_MUTEX(O_dentry->d_inode);
1246 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1247 O_dentry->d_inode, O0_dentry,
1249 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1252 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1253 GOTO(cleanup_O0, rc);
1255 filter->fo_fsd->lsd_feature_incompat |=
1256 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1257 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1259 GOTO(cleanup_O0, rc);
1271 cleanup_phase = 2; /* groups */
1273 /* we have to initialize all groups before first connections from
1274 * clients because they may send create/destroy for any group -bzzz */
1275 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1277 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1278 GOTO(cleanup, rc = PTR_ERR(filp));
1280 cleanup_phase = 3; /* filp */
1282 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1284 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1288 last_group = FILTER_MIN_GROUPS;
1290 LASSERT_MDS_GROUP(last_group);
1293 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1294 FILTER_MIN_GROUPS, last_group);
1295 filter->fo_committed_group = last_group;
1296 rc = filter_read_groups(obd, last_group, 1);
1300 filp_close(filp, 0);
1304 switch (cleanup_phase) {
1306 filp_close(filp, 0);
1308 filter_cleanup_groups(obd);
1310 f_dput(filter->fo_dentry_O);
1311 filter->fo_dentry_O = NULL;
1319 /* setup the object store with correct subdirectories */
1320 static int filter_prep(struct obd_device *obd)
1322 struct lvfs_run_ctxt saved;
1323 struct filter_obd *filter = &obd->u.filter;
1325 struct inode *inode;
1329 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1330 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1331 if (!file || IS_ERR(file)) {
1333 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1337 filter->fo_rcvd_filp = file;
1338 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1339 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1340 file->f_dentry->d_inode->i_mode);
1341 GOTO(err_filp, rc = -ENOENT);
1344 inode = file->f_dentry->d_parent->d_inode;
1345 /* We use i_op->unlink directly in filter_vfs_unlink() */
1346 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1347 CERROR("%s: filesystem does not support create/unlink ops\n",
1349 GOTO(err_filp, rc = -EOPNOTSUPP);
1352 /** lu_target has very limited use in filter now */
1353 lut_init(NULL, &filter->fo_lut, obd, NULL);
1355 rc = filter_init_server_data(obd, file);
1357 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1360 /* open and create health check io file*/
1361 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1364 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1366 GOTO(err_server_data, rc);
1368 filter->fo_health_check_filp = file;
1369 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1370 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1371 file->f_dentry->d_inode->i_mode);
1372 GOTO(err_health_check, rc = -ENOENT);
1374 rc = lvfs_check_io_health(obd, file);
1376 GOTO(err_health_check, rc);
1378 rc = filter_prep_groups(obd);
1380 GOTO(err_health_check, rc);
1382 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1387 if (filp_close(filter->fo_health_check_filp, 0))
1388 CERROR("can't close %s after error\n", HEALTH_CHECK);
1389 filter->fo_health_check_filp = NULL;
1391 target_recovery_fini(obd);
1392 filter_free_server_data(filter);
1394 if (filp_close(filter->fo_rcvd_filp, 0))
1395 CERROR("can't close %s after error\n", LAST_RCVD);
1396 filter->fo_rcvd_filp = NULL;
1400 /* cleanup the filter: write last used object id to status file */
1401 static void filter_post(struct obd_device *obd)
1403 struct lvfs_run_ctxt saved;
1404 struct filter_obd *filter = &obd->u.filter;
1407 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1408 * best to start a transaction with h_sync, because we removed this
1411 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1412 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1415 CERROR("error writing server data: rc = %d\n", rc);
1417 for (i = 0; i < filter->fo_group_count; i++) {
1418 rc = filter_update_last_objid(obd, i,
1419 (i == filter->fo_group_count - 1));
1421 CERROR("error writing group %d lastobjid: rc = %d\n",
1425 rc = filp_close(filter->fo_rcvd_filp, 0);
1426 filter->fo_rcvd_filp = NULL;
1428 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1430 rc = filp_close(filter->fo_health_check_filp, 0);
1431 filter->fo_health_check_filp = NULL;
1433 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1435 filter_cleanup_groups(obd);
1436 filter_free_server_data(filter);
1437 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1439 filter_free_capa_keys(filter);
1440 cleanup_capa_hash(filter->fo_capa_hash);
1443 static void filter_set_last_id(struct filter_obd *filter,
1444 obd_id id, obd_gr group)
1446 LASSERT(filter->fo_fsd != NULL);
1447 LASSERT(group <= filter->fo_group_count);
1449 spin_lock(&filter->fo_objidlock);
1450 filter->fo_last_objids[group] = id;
1451 spin_unlock(&filter->fo_objidlock);
1454 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1457 LASSERT(filter->fo_fsd != NULL);
1458 LASSERT(group <= filter->fo_group_count);
1460 /* FIXME: object groups */
1461 spin_lock(&filter->fo_objidlock);
1462 id = filter->fo_last_objids[group];
1463 spin_unlock(&filter->fo_objidlock);
1467 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1469 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1473 /* We never dget the object parent, so DON'T dput it either */
1474 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1476 struct filter_obd *filter = &obd->u.filter;
1477 struct filter_subdirs *subdirs;
1478 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1480 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1481 filter->fo_subdir_count == 0)
1482 return filter->fo_dentry_O_groups[group];
1484 subdirs = &filter->fo_dentry_O_sub[group];
1485 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1488 /* We never dget the object parent, so DON'T dput it either */
1489 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1492 unsigned long now = jiffies;
1493 struct dentry *dparent = filter_parent(obd, group, objid);
1496 if (IS_ERR(dparent))
1498 if (dparent == NULL)
1499 return ERR_PTR(-ENOENT);
1501 rc = filter_lock_dentry(obd, dparent);
1502 fsfilt_check_slow(obd, now, "parent lock");
1503 return rc ? ERR_PTR(rc) : dparent;
1506 /* We never dget the object parent, so DON'T dput it either */
1507 static void filter_parent_unlock(struct dentry *dparent)
1509 UNLOCK_INODE_MUTEX(dparent->d_inode);
1512 /* How to get files, dentries, inodes from object id's.
1514 * If dir_dentry is passed, the caller has already locked the parent
1515 * appropriately for this operation (normally a write lock). If
1516 * dir_dentry is NULL, we do a read lock while we do the lookup to
1517 * avoid races with create/destroy and such changing the directory
1518 * internal to the filesystem code. */
1519 struct dentry *filter_fid2dentry(struct obd_device *obd,
1520 struct dentry *dir_dentry,
1521 obd_gr group, obd_id id)
1523 struct dentry *dparent = dir_dentry;
1524 struct dentry *dchild;
1529 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1530 obd->u.filter.fo_destroys_in_progress == 0) {
1531 /* don't fail lookups for orphan recovery, it causes
1532 * later LBUGs when objects still exist during precreate */
1533 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1534 RETURN(ERR_PTR(-ENOENT));
1537 CERROR("fatal: invalid object id 0\n");
1538 RETURN(ERR_PTR(-ESTALE));
1541 len = sprintf(name, LPU64, id);
1542 if (dir_dentry == NULL) {
1543 dparent = filter_parent_lock(obd, group, id);
1544 if (IS_ERR(dparent)) {
1545 CERROR("%s: error getting object "LPU64":"LPU64
1546 " parent: rc %ld\n", obd->obd_name,
1547 id, group, PTR_ERR(dparent));
1551 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1552 dparent->d_name.len, dparent->d_name.name, name);
1553 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1554 if (dir_dentry == NULL)
1555 filter_parent_unlock(dparent);
1556 if (IS_ERR(dchild)) {
1557 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1562 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1563 CERROR("%s: got bad object "LPU64" inode %lu\n",
1564 obd->obd_name, id, dchild->d_inode->i_ino);
1566 RETURN(ERR_PTR(-ENOENT));
1569 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1570 name, dchild, atomic_read(&dchild->d_count));
1572 LASSERT(atomic_read(&dchild->d_count) > 0);
1577 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1578 obd_id group, struct lustre_handle *lockh)
1580 int flags = LDLM_AST_DISCARD_DATA, rc;
1581 struct ldlm_res_id res_id;
1582 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1585 osc_build_res_name(objid, group, &res_id);
1586 /* Tell the clients that the object is gone now and that they should
1587 * throw away any cached pages. */
1588 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1589 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1590 ldlm_completion_ast, NULL, NULL, 0, NULL,
1597 static void filter_fini_destroy(struct obd_device *obd,
1598 struct lustre_handle *lockh)
1601 ldlm_lock_decref(lockh, LCK_PW);
1604 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1605 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1606 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1607 * reverse. Caller must take i_sem before starting the transaction and we
1608 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1609 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1610 struct vfsmount *mnt)
1615 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1616 * itself we already hold dir->i_mutex for child create/unlink ops */
1617 LASSERT(dentry->d_inode != NULL);
1618 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1619 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1623 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1624 GOTO(out, rc = -ENOENT);
1626 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1631 GOTO(out, rc = -EPERM);
1633 /* check_sticky() */
1634 if ((dentry->d_inode->i_uid != current->fsuid &&
1635 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1636 IS_IMMUTABLE(dentry->d_inode))
1637 GOTO(out, rc = -EPERM);
1639 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1640 * that was done because of journal_start (which is already done
1641 * here) or some other ordering issue. */
1644 rc = ll_security_inode_unlink(dir, dentry, mnt);
1648 rc = dir->i_op->unlink(dir, dentry);
1650 /* need to drop i_mutex before we lose inode reference */
1651 UNLOCK_INODE_MUTEX(dentry->d_inode);
1658 /* Caller must hold LCK_PW on parent and push us into kernel context.
1659 * Caller must hold child i_mutex, we drop it always.
1660 * Caller is also required to ensure that dchild->d_inode exists. */
1661 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1662 obd_gr group, struct dentry *dparent,
1663 struct dentry *dchild)
1665 struct inode *inode = dchild->d_inode;
1666 struct filter_obd *filter = &obd->u.filter;
1669 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1670 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1671 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1672 (unsigned long)inode->i_nlink,
1673 atomic_read(&inode->i_count));
1676 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1678 CERROR("error unlinking objid %.*s: rc %d\n",
1679 dchild->d_name.len, dchild->d_name.name, rc);
1683 struct filter_intent_args {
1684 struct ldlm_lock **victim;
1689 static enum interval_iter filter_intent_cb(struct interval_node *n,
1692 struct ldlm_interval *node = (struct ldlm_interval *)n;
1693 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1694 __u64 size = arg->size;
1695 struct ldlm_lock **v = arg->victim;
1696 struct ldlm_lock *lck;
1698 /* If the interval is lower than the current file size,
1700 if (interval_high(n) <= size)
1701 return INTERVAL_ITER_STOP;
1703 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1704 /* Don't send glimpse ASTs to liblustre clients.
1705 * They aren't listening for them, and they do
1706 * entirely synchronous I/O anyways. */
1707 if (lck->l_export == NULL ||
1708 lck->l_export->exp_libclient == 1)
1711 if (*arg->liblustre)
1712 *arg->liblustre = 0;
1715 *v = LDLM_LOCK_GET(lck);
1716 } else if ((*v)->l_policy_data.l_extent.start <
1717 lck->l_policy_data.l_extent.start) {
1718 LDLM_LOCK_RELEASE(*v);
1719 *v = LDLM_LOCK_GET(lck);
1722 /* the same policy group - every lock has the
1723 * same extent, so needn't do it any more */
1727 return INTERVAL_ITER_CONT;
1730 static int filter_intent_policy(struct ldlm_namespace *ns,
1731 struct ldlm_lock **lockp, void *req_cookie,
1732 ldlm_mode_t mode, int flags, void *data)
1734 CFS_LIST_HEAD(rpc_list);
1735 struct ptlrpc_request *req = req_cookie;
1736 struct ldlm_lock *lock = *lockp, *l = NULL;
1737 struct ldlm_resource *res = lock->l_resource;
1738 ldlm_processing_policy policy;
1739 struct ost_lvb *res_lvb, *reply_lvb;
1740 struct ldlm_reply *rep;
1742 int idx, rc, tmpflags = 0, only_liblustre = 1;
1743 struct ldlm_interval_tree *tree;
1744 struct filter_intent_args arg;
1745 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1746 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1747 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1750 policy = ldlm_get_processing_policy(res);
1751 LASSERT(policy != NULL);
1752 LASSERT(req != NULL);
1754 rc = lustre_pack_reply(req, 3, repsize, NULL);
1756 RETURN(req->rq_status = rc);
1758 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1759 LASSERT(rep != NULL);
1761 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1762 sizeof(*reply_lvb));
1763 LASSERT(reply_lvb != NULL);
1765 //fixup_handle_for_resent_req(req, lock, &lockh);
1767 /* Call the extent policy function to see if our request can be
1768 * granted, or is blocked.
1769 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1770 * lock, and should not be granted if the lock will be blocked.
1773 LASSERT(ns == res->lr_namespace);
1775 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1776 check_res_locked(res);
1778 /* FIXME: we should change the policy function slightly, to not make
1779 * this list at all, since we just turn around and free it */
1780 while (!list_empty(&rpc_list)) {
1781 struct ldlm_lock *wlock =
1782 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1783 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1784 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1785 lock->l_flags &= ~LDLM_FL_CP_REQD;
1786 list_del_init(&wlock->l_cp_ast);
1787 LDLM_LOCK_RELEASE(wlock);
1790 /* The lock met with no resistance; we're finished. */
1791 if (rc == LDLM_ITER_CONTINUE) {
1792 /* do not grant locks to the liblustre clients: they cannot
1793 * handle ASTs robustly. We need to do this while still
1794 * holding ns_lock to avoid the lock remaining on the res_link
1795 * list (and potentially being added to l_pending_list by an
1796 * AST) when we are going to drop this lock ASAP. */
1797 if (lock->l_export->exp_libclient ||
1798 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1799 ldlm_resource_unlink_lock(lock);
1800 err = ELDLM_LOCK_ABORTED;
1802 err = ELDLM_LOCK_REPLACED;
1808 /* Do not grant any lock, but instead send GL callbacks. The extent
1809 * policy nicely created a list of all PW locks for us. We will choose
1810 * the highest of those which are larger than the size in the LVB, if
1811 * any, and perform a glimpse callback. */
1812 res_lvb = res->lr_lvb_data;
1813 LASSERT(res_lvb != NULL);
1814 *reply_lvb = *res_lvb;
1817 * ->ns_lock guarantees that no new locks are granted, and,
1818 * therefore, that res->lr_lvb_data cannot increase beyond the
1819 * end of already granted lock. As a result, it is safe to
1820 * check against "stale" reply_lvb->lvb_size value without
1823 arg.size = reply_lvb->lvb_size;
1825 arg.liblustre = &only_liblustre;
1826 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1827 tree = &res->lr_itree[idx];
1828 if (tree->lit_mode == LCK_PR)
1831 interval_iterate_reverse(tree->lit_root,
1832 filter_intent_cb, &arg);
1836 /* There were no PW locks beyond the size in the LVB; finished. */
1838 if (only_liblustre) {
1839 /* If we discovered a liblustre client with a PW lock,
1840 * however, the LVB may be out of date! The LVB is
1841 * updated only on glimpse (which we don't do for
1842 * liblustre clients) and cancel (which the client
1843 * obviously has not yet done). So if it has written
1844 * data but kept the lock, the LVB is stale and needs
1845 * to be updated from disk.
1847 * Of course, this will all disappear when we switch to
1848 * taking liblustre locks on the OST. */
1849 ldlm_res_lvbo_update(res, NULL, 0, 1);
1851 RETURN(ELDLM_LOCK_ABORTED);
1855 * This check is for lock taken in filter_prepare_destroy() that does
1856 * not have l_glimpse_ast set. So the logic is: if there is a lock
1857 * with no l_glimpse_ast set, this object is being destroyed already.
1859 * Hence, if you are grabbing DLM locks on the server, always set
1860 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1862 if (l->l_glimpse_ast == NULL) {
1863 /* We are racing with unlink(); just return -ENOENT */
1864 rep->lock_policy_res1 = -ENOENT;
1868 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1869 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1870 /* Update the LVB from disk if the AST failed (this is a legal race) */
1872 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1873 * sending ast is not handled. This can result in lost client writes.
1876 ldlm_res_lvbo_update(res, NULL, 0, 1);
1879 *reply_lvb = *res_lvb;
1883 LDLM_LOCK_RELEASE(l);
1885 RETURN(ELDLM_LOCK_ABORTED);
1889 * per-obd_device iobuf pool.
1891 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1892 * allocations in write-path has to be minimized (see bug 5137).
1894 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1895 * OST threads (see ost_thread_{init,done}()).
1897 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1898 * at the OST layer there are only (potentially) multiple obd_device of type
1899 * unknown at the time of OST thread creation.
1901 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1902 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1903 * it's very own iobuf.
1907 * filter_kiobuf_pool_init()
1909 * filter_kiobuf_pool_done()
1911 * filter_iobuf_get()
1913 * operate on this array. They are "generic" in a sense that they don't depend
1914 * on actual type of iobuf's (the latter depending on Linux kernel version).
1918 * destroy pool created by filter_iobuf_pool_init
1920 static void filter_iobuf_pool_done(struct filter_obd *filter)
1922 struct filter_iobuf **pool;
1927 pool = filter->fo_iobuf_pool;
1929 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1930 if (pool[i] != NULL)
1931 filter_free_iobuf(pool[i]);
1933 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1934 filter->fo_iobuf_pool = NULL;
1939 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1941 struct filter_obd *filter = &obd->u.filter;
1942 struct sptlrpc_rule_set tmp_rset;
1945 sptlrpc_rule_set_init(&tmp_rset);
1946 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1948 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1953 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1955 write_lock(&filter->fo_sptlrpc_lock);
1956 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1957 filter->fo_sptlrpc_rset = tmp_rset;
1958 write_unlock(&filter->fo_sptlrpc_lock);
1964 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1966 static int filter_iobuf_pool_init(struct filter_obd *filter)
1973 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1975 if (filter->fo_iobuf_pool == NULL)
1978 filter->fo_iobuf_count = OSS_THREADS_MAX;
1983 /* Return iobuf allocated for @thread_id. We don't know in advance how
1984 * many threads there will be so we allocate a large empty array and only
1985 * fill in those slots that are actually in use.
1986 * If we haven't allocated a pool entry for this thread before, do so now. */
1987 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1989 int thread_id = (oti && oti->oti_thread) ?
1990 oti->oti_thread->t_id : -1;
1991 struct filter_iobuf *pool = NULL;
1992 struct filter_iobuf **pool_place = NULL;
1994 if (thread_id >= 0) {
1995 LASSERT(thread_id < filter->fo_iobuf_count);
1996 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1999 if (unlikely(pool == NULL)) {
2000 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
2001 PTLRPC_MAX_BRW_PAGES);
2002 if (pool_place != NULL)
2009 /* mount the file system (secretly). lustre_cfg parameters are:
2012 * 3 = flags: failover=f, failout=n
2015 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2018 struct filter_obd *filter = &obd->u.filter;
2019 struct vfsmount *mnt;
2020 struct lustre_mount_info *lmi;
2021 struct obd_uuid uuid;
2029 if (lcfg->lcfg_bufcount < 3 ||
2030 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2031 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2034 lmi = server_get_mount(obd->obd_name);
2036 /* We already mounted in lustre_fill_super.
2037 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2038 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2040 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2042 /* old path - used by lctl */
2043 CERROR("Using old MDS mount method\n");
2044 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2045 MS_NOATIME|MS_NODIRATIME,
2046 lustre_cfg_string(lcfg, 1), option);
2049 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2050 lustre_cfg_string(lcfg, 1), rc);
2054 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2056 if (IS_ERR(obd->obd_fsops))
2057 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2059 rc = filter_iobuf_pool_init(filter);
2063 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2064 CERROR("%s: Underlying device is marked as read-only. "
2065 "Setup failed\n", obd->obd_name);
2066 GOTO(err_ops, rc = -EROFS);
2069 /* failover is the default */
2070 obd->obd_replayable = 1;
2072 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2073 str = lustre_cfg_string(lcfg, 3);
2074 if (strchr(str, 'n')) {
2075 CWARN("%s: recovery disabled\n", obd->obd_name);
2076 obd->obd_replayable = 0;
2080 filter->fo_vfsmnt = mnt;
2081 obd->u.obt.obt_sb = mnt->mnt_sb;
2082 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2083 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2085 fsfilt_setup(obd, obd->u.obt.obt_sb);
2087 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2088 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2089 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2090 obd->obd_lvfs_ctxt.fs = get_ds();
2091 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2093 init_mutex(&filter->fo_init_lock);
2094 filter->fo_committed_group = 0;
2096 rc = filter_prep(obd);
2100 filter->fo_destroys_in_progress = 0;
2101 for (i = 0; i < 32; i++)
2102 sema_init(&filter->fo_create_locks[i], 1);
2104 spin_lock_init(&filter->fo_translock);
2105 spin_lock_init(&filter->fo_objidlock);
2106 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2107 sema_init(&filter->fo_alloc_lock, 1);
2108 init_brw_stats(&filter->fo_filter_stats);
2109 filter->fo_read_cache = 1; /* enable read-only cache by default */
2110 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2111 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2112 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2113 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2115 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2116 spin_lock_init(&filter->fo_llog_list_lock);
2118 filter->fo_fl_oss_capa = 1;
2120 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2121 filter->fo_capa_hash = init_capa_hash();
2122 if (filter->fo_capa_hash == NULL)
2123 GOTO(err_post, rc = -ENOMEM);
2125 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2126 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2127 LDLM_NAMESPACE_GREEDY);
2128 if (obd->obd_namespace == NULL)
2129 GOTO(err_post, rc = -ENOMEM);
2130 obd->obd_namespace->ns_lvbp = obd;
2131 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2132 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2134 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2135 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2137 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2139 CERROR("failed to setup llogging subsystems\n");
2143 rwlock_init(&filter->fo_sptlrpc_lock);
2144 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2145 /* do this after llog being initialized */
2146 filter_adapt_sptlrpc_conf(obd, 1);
2148 rc = lquota_setup(filter_quota_interface_ref, obd);
2152 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2153 if (q->max_sectors < q->max_hw_sectors &&
2154 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2155 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2156 "for larger I/O requests: max_sectors = %u "
2157 "could be up to max_hw_sectors=%u\n",
2158 obd->obd_name, mnt->mnt_sb->s_id,
2159 q->max_sectors, q->max_hw_sectors);
2161 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2162 if (uuid_ptr != NULL) {
2163 class_uuid_unparse(uuid_ptr, &uuid);
2169 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2171 if (obd->obd_recovering) {
2172 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2173 "recovery for at least %d:%.02d, or until %d "
2174 "client%s reconnect. During this time new clients"
2175 " will not be allowed to connect. "
2176 "Recovery progress can be monitored by watching "
2177 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2178 obd->obd_name, lustre_cfg_string(lcfg, 1),
2179 label ?: "", label ? "/" : "", str,
2180 obd->obd_recovery_timeout / 60,
2181 obd->obd_recovery_timeout % 60,
2182 obd->obd_max_recoverable_clients,
2183 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2186 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2187 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2188 label ?: "", label ? "/" : "", str,
2189 obd->obd_replayable ? "enabled" : "disabled");
2197 fsfilt_put_ops(obd->obd_fsops);
2198 filter_iobuf_pool_done(filter);
2200 server_put_mount(obd->obd_name, mnt);
2201 obd->u.obt.obt_sb = 0;
2205 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2207 struct lprocfs_static_vars lvars;
2212 CLASSERT(offsetof(struct obd_device, u.obt) ==
2213 offsetof(struct obd_device, u.filter.fo_obt));
2215 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2218 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2219 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2222 addr = (unsigned long)cfs_page_address(page);
2223 clear_page((void *)addr);
2225 /* lprocfs must be setup before the filter so state can be safely added
2226 * to /proc incrementally as the filter is setup */
2227 lprocfs_filter_init_vars(&lvars);
2228 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2229 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2230 /* Init obdfilter private stats here */
2231 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2232 LPROCFS_CNTR_AVGMINMAX,
2233 "read_bytes", "bytes");
2234 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2235 LPROCFS_CNTR_AVGMINMAX,
2236 "write_bytes", "bytes");
2237 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2238 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2239 "get_page", "usec");
2240 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2241 LPROCFS_CNTR_AVGMINMAX,
2242 "get_page_failures", "num");
2243 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2244 LPROCFS_CNTR_AVGMINMAX,
2245 "cache_access", "pages");
2246 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2247 LPROCFS_CNTR_AVGMINMAX,
2248 "cache_hit", "pages");
2249 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2250 LPROCFS_CNTR_AVGMINMAX,
2251 "cache_miss", "pages");
2253 lproc_filter_attach_seqstat(obd);
2254 obd->obd_proc_exports_entry = lprocfs_register("exports",
2255 obd->obd_proc_entry,
2257 if (IS_ERR(obd->obd_proc_exports_entry)) {
2258 rc = PTR_ERR(obd->obd_proc_exports_entry);
2259 CERROR("error %d setting up lprocfs for %s\n",
2261 obd->obd_proc_exports_entry = NULL;
2264 if (obd->obd_proc_exports_entry)
2265 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2266 lprocfs_nid_stats_clear_read,
2267 lprocfs_nid_stats_clear_write, obd, NULL);
2269 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2270 LUSTRE_CFG_BUFLEN(lcfg, 4));
2271 rc = filter_common_setup(obd, lcfg, (void *)addr);
2272 OBD_PAGE_FREE(page);
2275 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2276 lprocfs_free_per_client_stats(obd);
2277 lprocfs_free_obd_stats(obd);
2278 lprocfs_obd_cleanup(obd);
2284 static struct llog_operations filter_mds_ost_repl_logops;
2286 static struct llog_operations filter_size_orig_logops = {
2287 .lop_setup = llog_obd_origin_setup,
2288 .lop_cleanup = llog_obd_origin_cleanup,
2289 .lop_add = llog_obd_origin_add
2292 static int filter_olg_fini(struct obd_llog_group *olg)
2294 struct llog_ctxt *ctxt;
2295 int rc = 0, rc2 = 0;
2298 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2300 rc = llog_cleanup(ctxt);
2302 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2304 rc2 = llog_cleanup(ctxt);
2309 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2311 rc2 = llog_cleanup(ctxt);
2320 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2321 struct obd_device *tgt)
2326 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2327 &filter_mds_ost_repl_logops);
2331 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2332 &filter_size_orig_logops);
2338 filter_olg_fini(olg);
2343 * Init the default olg, which is embeded in the obd_device, for filter.
2346 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2347 struct obd_device *tgt)
2349 struct filter_obd *filter = &obd->u.filter;
2350 struct llog_ctxt *ctxt;
2354 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2355 if (!filter->fo_lcm)
2358 filter_mds_ost_repl_logops = llog_client_ops;
2359 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2360 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2361 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2363 rc = filter_olg_init(obd, olg, tgt);
2365 GOTO(cleanup_lcm, rc);
2367 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2370 GOTO(cleanup_olg, rc);
2372 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2374 CERROR("Can't get ctxt for %p:%x\n", olg,
2375 LLOG_MDS_OST_REPL_CTXT);
2376 GOTO(cleanup_olg, rc = -ENODEV);
2378 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2379 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2380 llog_ctxt_put(ctxt);
2384 filter_olg_fini(olg);
2386 llog_recov_thread_fini(filter->fo_lcm, 1);
2387 filter->fo_lcm = NULL;
2392 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2393 struct obd_device *tgt, int count, struct llog_catid *catid,
2394 struct obd_uuid *uuid)
2396 struct filter_obd *filter = &obd->u.filter;
2397 struct llog_ctxt *ctxt;
2401 LASSERT(olg != NULL);
2402 if (olg == &obd->obd_olg)
2403 return filter_default_olg_init(obd, olg, tgt);
2405 LASSERT(filter->fo_lcm != NULL);
2406 rc = filter_olg_init(obd, olg, tgt);
2409 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2411 CERROR("Can't get ctxt for %p:%x\n", olg,
2412 LLOG_MDS_OST_REPL_CTXT);
2413 filter_olg_fini(olg);
2416 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2417 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2418 llog_ctxt_put(ctxt);
2422 static int filter_llog_finish(struct obd_device *obd, int count)
2424 struct filter_obd *filter = &obd->u.filter;
2425 struct llog_ctxt *ctxt;
2428 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2431 * Make sure that no cached llcds left in recov_thread.
2432 * We actually do sync in disconnect time, but disconnect
2433 * may not come being marked rq_no_resend = 1.
2435 llog_sync(ctxt, NULL);
2438 * Balance class_import_get() in llog_receptor_accept().
2439 * This is safe to do, as llog is already synchronized
2440 * and its import may go.
2442 mutex_down(&ctxt->loc_sem);
2443 if (ctxt->loc_imp) {
2444 class_import_put(ctxt->loc_imp);
2445 ctxt->loc_imp = NULL;
2447 mutex_up(&ctxt->loc_sem);
2448 llog_ctxt_put(ctxt);
2451 if (filter->fo_lcm) {
2452 mutex_down(&ctxt->loc_sem);
2453 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2454 filter->fo_lcm = NULL;
2455 mutex_up(&ctxt->loc_sem);
2457 RETURN(filter_olg_fini(&obd->obd_olg));
2461 * Find the group llog according to group index in the llog group list.
2463 static struct obd_llog_group *
2464 filter_find_olg_internal(struct filter_obd *filter, int group)
2466 struct obd_llog_group *olg;
2468 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2469 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2470 if (olg->olg_group == group)
2477 * Find the group llog according to group index on the filter
2479 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2481 struct obd_llog_group *olg = NULL;
2482 struct filter_obd *filter;
2484 filter = &obd->u.filter;
2486 if (group == FILTER_GROUP_LLOG)
2487 RETURN(&obd->obd_olg);
2489 spin_lock(&filter->fo_llog_list_lock);
2490 olg = filter_find_olg_internal(filter, group);
2491 spin_unlock(&filter->fo_llog_list_lock);
2496 * Find the llog_group of the filter according to the group. If it can not
2497 * find, create the llog_group, which only happens when mds is being synced
2500 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2502 struct obd_llog_group *olg = NULL;
2503 struct filter_obd *filter;
2506 filter = &obd->u.filter;
2508 if (group == FILTER_GROUP_LLOG)
2509 RETURN(&obd->obd_olg);
2511 spin_lock(&filter->fo_llog_list_lock);
2512 olg = filter_find_olg_internal(filter, group);
2514 if (olg->olg_initializing) {
2515 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2517 GOTO(out_unlock, olg);
2522 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2524 llog_group_init(olg, group);
2525 list_add(&olg->olg_list, &filter->fo_llog_list);
2526 olg->olg_initializing = 1;
2527 spin_unlock(&filter->fo_llog_list_lock);
2529 rc = llog_cat_initialize(obd, olg, 1, NULL);
2531 spin_lock(&filter->fo_llog_list_lock);
2532 list_del(&olg->olg_list);
2533 spin_unlock(&filter->fo_llog_list_lock);
2535 GOTO(out, olg = ERR_PTR(-ENOMEM));
2537 spin_lock(&filter->fo_llog_list_lock);
2538 olg->olg_initializing = 0;
2539 spin_unlock(&filter->fo_llog_list_lock);
2540 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2541 obd->obd_name, group, olg);
2546 spin_unlock(&filter->fo_llog_list_lock);
2550 static int filter_llog_connect(struct obd_export *exp,
2551 struct llogd_conn_body *body)
2553 struct obd_device *obd = exp->exp_obd;
2554 struct llog_ctxt *ctxt;
2555 struct obd_llog_group *olg;
2559 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2560 obd->obd_name, body->lgdc_logid.lgl_oid,
2561 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2563 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2565 CERROR(" %s: can not find olg of group %d\n",
2566 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2569 llog_group_set_export(olg, exp);
2571 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2572 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2573 body->lgdc_ctxt_idx);
2575 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2576 obd->obd_name, body->lgdc_logid.lgl_oid,
2577 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2579 rc = llog_connect(ctxt, &body->lgdc_logid,
2580 &body->lgdc_gen, NULL);
2581 llog_ctxt_put(ctxt);
2583 CERROR("failed to connect rc %d idx %d\n", rc,
2584 body->lgdc_ctxt_idx);
2589 static int filter_llog_preclean(struct obd_device *obd)
2591 struct obd_llog_group *olg, *tmp;
2592 struct filter_obd *filter;
2593 struct list_head remove_list;
2597 rc = obd_llog_finish(obd, 0);
2599 CERROR("failed to cleanup llogging subsystem\n");
2601 filter = &obd->u.filter;
2602 CFS_INIT_LIST_HEAD(&remove_list);
2604 spin_lock(&filter->fo_llog_list_lock);
2605 while (!list_empty(&filter->fo_llog_list)) {
2606 olg = list_entry(filter->fo_llog_list.next,
2607 struct obd_llog_group, olg_list);
2608 list_del(&olg->olg_list);
2609 list_add(&olg->olg_list, &remove_list);
2611 spin_unlock(&filter->fo_llog_list_lock);
2613 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2614 list_del_init(&olg->olg_list);
2615 rc = filter_olg_fini(olg);
2617 CERROR("failed to cleanup llogging subsystem for %u\n",
2625 static int filter_precleanup(struct obd_device *obd,
2626 enum obd_cleanup_stage stage)
2632 case OBD_CLEANUP_EARLY:
2634 case OBD_CLEANUP_EXPORTS:
2635 /* Stop recovery before namespace cleanup. */
2636 target_stop_recovery_thread(obd);
2637 target_cleanup_recovery(obd);
2638 rc = filter_llog_preclean(obd);
2644 static int filter_cleanup(struct obd_device *obd)
2646 struct filter_obd *filter = &obd->u.filter;
2650 LCONSOLE_WARN("%s: shutting down for failover; client state "
2651 "will be preserved.\n", obd->obd_name);
2653 if (!list_empty(&obd->obd_exports)) {
2654 CERROR("%s: still has clients!\n", obd->obd_name);
2655 class_disconnect_exports(obd);
2656 if (!list_empty(&obd->obd_exports)) {
2657 CERROR("still has exports after forced cleanup?\n");
2662 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2663 lprocfs_free_per_client_stats(obd);
2664 lprocfs_free_obd_stats(obd);
2665 lprocfs_obd_cleanup(obd);
2666 lquota_cleanup(filter_quota_interface_ref, obd);
2668 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2669 obd->obd_namespace = NULL;
2671 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2673 if (obd->u.obt.obt_sb == NULL)
2678 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2679 shrink_dcache_sb(obd->u.obt.obt_sb);
2681 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2682 obd->u.obt.obt_sb = NULL;
2684 fsfilt_put_ops(obd->obd_fsops);
2686 filter_iobuf_pool_done(filter);
2688 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2693 static int filter_connect_internal(struct obd_export *exp,
2694 struct obd_connect_data *data)
2696 struct filter_export_data *fed = &exp->exp_filter_data;
2701 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2702 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2703 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2704 data->ocd_connect_flags, data->ocd_version,
2705 data->ocd_grant, data->ocd_index);
2707 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2708 CWARN("!!! This export (nid %s) used object group %d "
2709 "earlier; now it's trying to use group %d! This could "
2710 "be a bug in the MDS. Please report to "
2711 "http://bugzilla.lustre.org/\n",
2712 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2715 fed->fed_group = data->ocd_group;
2717 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2718 exp->exp_connect_flags = data->ocd_connect_flags;
2719 data->ocd_version = LUSTRE_VERSION_CODE;
2721 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2722 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2723 data->ocd_connect_flags & OBD_CONNECT_MDS))
2726 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2727 struct filter_obd *filter = &exp->exp_obd->u.filter;
2728 obd_size left, want;
2730 spin_lock(&exp->exp_obd->obd_osfs_lock);
2731 left = filter_grant_space_left(exp);
2732 want = data->ocd_grant;
2733 filter_grant(exp, fed->fed_grant, want, left);
2734 data->ocd_grant = fed->fed_grant;
2735 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2737 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2738 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2739 exp->exp_client_uuid.uuid, exp,
2740 data->ocd_grant, want, left);
2742 filter->fo_tot_granted_clients ++;
2745 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2746 struct filter_obd *filter = &exp->exp_obd->u.filter;
2747 struct lr_server_data *lsd = filter->fo_fsd;
2748 int index = le32_to_cpu(lsd->lsd_ost_index);
2750 if (!(lsd->lsd_feature_compat &
2751 cpu_to_le32(OBD_COMPAT_OST))) {
2752 /* this will only happen on the first connect */
2753 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2754 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2755 filter_update_server_data(exp->exp_obd,
2756 filter->fo_rcvd_filp, lsd, 1);
2757 } else if (index != data->ocd_index) {
2758 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2759 " %u doesn't match actual OST index"
2760 " %u in last_rcvd file, bad "
2762 obd_export_nid2str(exp), index,
2766 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2767 * FIXME: We don't strictly need the COMPAT flag for that,
2768 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2769 * FIXME: We needed it for the index, as index 0 is valid. */
2772 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2773 data->ocd_brw_size = 65536;
2774 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2775 data->ocd_brw_size = min(data->ocd_brw_size,
2776 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2777 LASSERT(data->ocd_brw_size);
2780 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2781 __u32 cksum_types = data->ocd_cksum_types;
2783 /* The client set in ocd_cksum_types the checksum types it
2784 * supports. We have to mask off the algorithms that we don't
2786 if (cksum_types & OBD_CKSUM_ALL)
2787 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2789 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2791 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2792 "%x\n", exp->exp_obd->obd_name,
2793 obd_export_nid2str(exp), cksum_types,
2794 data->ocd_cksum_types);
2796 /* This client does not support OBD_CONNECT_CKSUM
2797 * fall back to CRC32 */
2798 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2799 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2800 exp->exp_obd->obd_name,
2801 obd_export_nid2str(exp));
2807 static int filter_reconnect(const struct lu_env *env,
2808 struct obd_export *exp, struct obd_device *obd,
2809 struct obd_uuid *cluuid,
2810 struct obd_connect_data *data,
2816 if (exp == NULL || obd == NULL || cluuid == NULL)
2819 rc = filter_connect_internal(exp, data);
2821 filter_export_stats_init(obd, exp, localdata);
2826 /* nearly identical to mds_connect */
2827 static int filter_connect(const struct lu_env *env,
2828 struct obd_export **exp, struct obd_device *obd,
2829 struct obd_uuid *cluuid,
2830 struct obd_connect_data *data, void *localdata)
2832 struct lvfs_run_ctxt saved;
2833 struct lustre_handle conn = { 0 };
2834 struct obd_export *lexp;
2835 struct filter_export_data *fed;
2836 struct lsd_client_data *lcd = NULL;
2841 if (exp == NULL || obd == NULL || cluuid == NULL)
2844 rc = class_connect(&conn, obd, cluuid);
2847 lexp = class_conn2export(&conn);
2848 LASSERT(lexp != NULL);
2850 fed = &lexp->exp_filter_data;
2852 rc = filter_connect_internal(lexp, data);
2856 filter_export_stats_init(obd, lexp, localdata);
2857 if (obd->obd_replayable) {
2858 OBD_ALLOC(lcd, sizeof(*lcd));
2860 CERROR("filter: out of memory for client data\n");
2861 GOTO(cleanup, rc = -ENOMEM);
2864 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2866 rc = filter_client_add(obd, lexp, -1);
2871 group = data->ocd_group;
2873 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2874 obd->obd_name, lexp->exp_handle.h_cookie, group);
2876 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2877 rc = filter_read_groups(obd, group, 1);
2878 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2880 CERROR("can't read group %u\n", group);
2890 fed->fed_lcd = NULL;
2892 class_disconnect(lexp);
2901 /* Do extra sanity checks for grant accounting. We do this at connect,
2902 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2903 * always get rid of it or turn it off when we know accounting is good. */
2904 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2906 struct filter_export_data *fed;
2907 struct obd_export *exp;
2908 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2909 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2910 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2912 if (list_empty(&obd->obd_exports))
2915 /* We don't want to do this for large machines that do lots of
2916 mounts or unmounts. It burns... */
2917 if (obd->obd_num_exports > 100)
2920 spin_lock(&obd->obd_osfs_lock);
2921 spin_lock(&obd->obd_dev_lock);
2922 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2924 fed = &exp->exp_filter_data;
2925 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2928 if (maxsize > 0) { /* we may not have done a statfs yet */
2929 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2930 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2931 exp->exp_client_uuid.uuid, exp,
2932 fed->fed_grant, fed->fed_pending, maxsize);
2933 LASSERTF(fed->fed_dirty <= maxsize,
2934 "%s: cli %s/%p %ld > "LPU64"\n", func,
2935 exp->exp_client_uuid.uuid, exp,
2936 fed->fed_dirty, maxsize);
2939 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2940 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2941 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2943 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2944 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2945 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2946 tot_granted += fed->fed_grant + fed->fed_pending;
2947 tot_pending += fed->fed_pending;
2948 tot_dirty += fed->fed_dirty;
2950 fo_tot_granted = obd->u.filter.fo_tot_granted;
2951 fo_tot_pending = obd->u.filter.fo_tot_pending;
2952 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2953 spin_unlock(&obd->obd_dev_lock);
2954 spin_unlock(&obd->obd_osfs_lock);
2956 /* Do these assertions outside the spinlocks so we don't kill system */
2957 if (tot_granted != fo_tot_granted)
2958 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2959 func, tot_granted, fo_tot_granted);
2960 if (tot_pending != fo_tot_pending)
2961 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2962 func, tot_pending, fo_tot_pending);
2963 if (tot_dirty != fo_tot_dirty)
2964 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2965 func, tot_dirty, fo_tot_dirty);
2966 if (tot_pending > tot_granted)
2967 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2968 func, tot_pending, tot_granted);
2969 if (tot_granted > maxsize)
2970 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2971 func, tot_granted, maxsize);
2972 if (tot_dirty > maxsize)
2973 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2974 func, tot_dirty, maxsize);
2977 /* Remove this client from the grant accounting totals. We also remove
2978 * the export from the obd device under the osfs and dev locks to ensure
2979 * that the filter_grant_sanity_check() calculations are always valid.
2980 * The client should do something similar when it invalidates its import. */
2981 static void filter_grant_discard(struct obd_export *exp)
2983 struct obd_device *obd = exp->exp_obd;
2984 struct filter_obd *filter = &obd->u.filter;
2985 struct filter_export_data *fed = &exp->exp_filter_data;
2987 spin_lock(&obd->obd_osfs_lock);
2988 spin_lock(&obd->obd_dev_lock);
2989 list_del_init(&exp->exp_obd_chain);
2990 spin_unlock(&obd->obd_dev_lock);
2992 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2993 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2994 obd->obd_name, filter->fo_tot_granted,
2995 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2996 filter->fo_tot_granted -= fed->fed_grant;
2997 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2998 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2999 obd->obd_name, filter->fo_tot_pending,
3000 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3001 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3002 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3003 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3004 obd->obd_name, filter->fo_tot_dirty,
3005 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3006 filter->fo_tot_dirty -= fed->fed_dirty;
3010 spin_unlock(&obd->obd_osfs_lock);
3013 static int filter_destroy_export(struct obd_export *exp)
3017 if (exp->exp_filter_data.fed_pending)
3018 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3019 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3020 exp, exp->exp_filter_data.fed_pending);
3022 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3024 target_destroy_export(exp);
3025 ldlm_destroy_export(exp);
3027 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3031 if (exp->exp_obd->obd_replayable)
3032 filter_client_free(exp);
3034 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3036 filter_grant_discard(exp);
3037 filter_fmd_cleanup(exp);
3039 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3040 struct filter_obd *filter = &exp->exp_obd->u.filter;
3041 if (filter->fo_tot_granted_clients > 0)
3042 filter->fo_tot_granted_clients --;
3045 if (!(exp->exp_flags & OBD_OPT_FORCE))
3046 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3051 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3053 struct obd_llog_group *olg_min, *olg;
3054 struct filter_obd *filter;
3055 int worked = -1, group;
3056 struct llog_ctxt *ctxt;
3059 filter = &obd->u.filter;
3061 /* we can't sync log holding spinlock. also, we do not want to get
3062 * into livelock. so we do following: loop over MDS's exports in
3063 * group order and skip already synced llogs -bzzz */
3065 /* look for group with min. number, but > worked */
3068 spin_lock(&filter->fo_llog_list_lock);
3069 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3070 if (olg->olg_group <= worked) {
3071 /* this group is already synced */
3074 if (group < olg->olg_group) {
3075 /* we have group with smaller number to sync */
3078 /* store current minimal group */
3080 group = olg->olg_group;
3082 spin_unlock(&filter->fo_llog_list_lock);
3084 if (olg_min == NULL)
3087 worked = olg_min->olg_group;
3088 if (olg_min->olg_exp &&
3089 (dexp == olg_min->olg_exp || dexp == NULL)) {
3091 ctxt = llog_group_get_ctxt(olg_min,
3092 LLOG_MDS_OST_REPL_CTXT);
3094 err = llog_sync(ctxt, olg_min->olg_exp);
3095 llog_ctxt_put(ctxt);
3097 CERROR("error flushing logs to MDS: "
3102 } while (olg_min != NULL);
3105 /* Also incredibly similar to mds_disconnect */
3106 static int filter_disconnect(struct obd_export *exp)
3108 struct obd_device *obd = exp->exp_obd;
3113 class_export_get(exp);
3115 if (!(exp->exp_flags & OBD_OPT_FORCE))
3116 filter_grant_sanity_check(obd, __FUNCTION__);
3117 filter_grant_discard(exp);
3119 /* Flush any remaining cancel messages out to the target */
3120 filter_sync_llogs(obd, exp);
3122 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3124 /* Disconnect early so that clients can't keep using export */
3125 rc = class_disconnect(exp);
3126 if (exp->exp_obd->obd_namespace != NULL)
3127 ldlm_cancel_locks_for_export(exp);
3129 fsfilt_sync(obd, obd->u.obt.obt_sb);
3131 lprocfs_exp_cleanup(exp);
3133 class_export_put(exp);
3137 /* reverse import is changed, sync all cancels */
3138 static void filter_revimp_update(struct obd_export *exp)
3143 class_export_get(exp);
3145 /* flush any remaining cancel messages out to the target */
3146 filter_sync_llogs(exp->exp_obd, exp);
3147 class_export_put(exp);
3151 static int filter_ping(struct obd_export *exp)
3153 filter_fmd_expire(exp);
3157 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3158 const char *what, int quiet)
3160 struct dentry *dchild = NULL;
3163 if (oa->o_valid & OBD_MD_FLGROUP)
3166 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3168 if (IS_ERR(dchild)) {
3169 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3170 what, group, oa->o_id);
3174 if (dchild->d_inode == NULL) {
3176 CERROR("%s: %s on non-existent object: "LPU64"\n",
3177 obd->obd_name, what, oa->o_id);
3179 RETURN(ERR_PTR(-ENOENT));
3185 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3187 struct dentry *dentry = NULL;
3188 struct obd_device *obd;
3192 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3193 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3197 obd = class_exp2obd(exp);
3199 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3203 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3205 RETURN(PTR_ERR(dentry));
3207 /* Limit the valid bits in the return data to what we actually use */
3208 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3209 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3215 /* this should be enabled/disabled in condition to enabled/disabled large
3216 * inodes (fast EAs) in backing store FS. */
3217 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3218 void *handle, struct obdo *oa)
3220 struct obd_device *obd = exp->exp_obd;
3224 if (oa->o_valid & OBD_MD_FLFID) {
3225 struct filter_fid ff;
3227 if (!(oa->o_valid & OBD_MD_FLGROUP))
3229 /* packing fid and converting it to LE for storing into EA.
3230 * Here ->o_stripe_idx should be filled by LOV and rest of
3231 * fields - by client. */
3232 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3233 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3234 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3235 ff.ff_objid = cpu_to_le64(oa->o_id);
3236 ff.ff_group = cpu_to_le64(oa->o_gr);
3238 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3239 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3240 oa->o_generation, oa->o_id, oa->o_gr);
3242 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3244 CERROR("store fid in object failed! rc: %d\n", rc);
3246 CDEBUG(D_HA, "OSS object without fid info!\n");
3252 /* this is called from filter_truncate() until we have filter_punch() */
3253 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3254 struct obdo *oa, struct obd_trans_info *oti)
3256 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3257 struct llog_cookie *fcc = NULL;
3258 struct filter_obd *filter;
3259 int rc, err, sync = 0;
3260 loff_t old_size = 0;
3261 unsigned int ia_valid;
3262 struct inode *inode;
3267 LASSERT(dentry != NULL);
3268 LASSERT(!IS_ERR(dentry));
3270 inode = dentry->d_inode;
3271 LASSERT(inode != NULL);
3273 filter = &exp->exp_obd->u.filter;
3274 iattr_from_obdo(&iattr, oa, oa->o_valid);
3275 ia_valid = iattr.ia_valid;
3277 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3278 OBD_ALLOC(fcc, sizeof(*fcc));
3280 *fcc = oa->o_lcookie;
3282 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3284 /* Filter truncates and writes are serialized by
3285 * i_alloc_sem, see the comment in
3286 * filter_preprw_write.*/
3287 if (ia_valid & ATTR_SIZE)
3288 down_write(&inode->i_alloc_sem);
3289 LOCK_INODE_MUTEX(inode);
3290 old_size = i_size_read(inode);
3293 /* VBR: version recovery check */
3294 rc = filter_version_get_check(exp, oti, inode);
3296 GOTO(out_unlock, rc);
3298 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3299 * then we will accept the UID+GID sent by the client during write for
3300 * initializing the ownership of this inode. We only allow this to
3301 * happen once so clear these bits in setattr. In 2.6 kernels it is
3302 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3303 * the flags that are actually being set. */
3304 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3305 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3306 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3308 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3309 if (!(ia_valid & ATTR_MODE)) {
3310 iattr.ia_mode = inode->i_mode;
3311 iattr.ia_valid |= ATTR_MODE;
3313 iattr.ia_mode &= ~S_ISUID;
3315 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3316 if (!(iattr.ia_valid & ATTR_MODE)) {
3317 iattr.ia_mode = inode->i_mode;
3318 iattr.ia_valid |= ATTR_MODE;
3320 iattr.ia_mode &= ~S_ISGID;
3323 orig_ids[USRQUOTA] = inode->i_uid;
3324 orig_ids[GRPQUOTA] = inode->i_gid;
3325 handle = fsfilt_start_log(exp->exp_obd, inode,
3326 FSFILT_OP_SETATTR, oti, 1);
3328 GOTO(out_unlock, rc = PTR_ERR(handle));
3330 /* update inode EA only once when inode is suid bit marked. As
3331 * on 2.6.x UID and GID may be set separately, we check here
3332 * only one of them to avoid double setting. */
3333 if (inode->i_mode & S_ISUID)
3334 filter_update_fidea(exp, inode, handle, oa);
3336 handle = fsfilt_start(exp->exp_obd, inode,
3337 FSFILT_OP_SETATTR, oti);
3339 GOTO(out_unlock, rc = PTR_ERR(handle));
3341 if (oa->o_valid & OBD_MD_FLFLAGS) {
3342 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3343 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3345 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3347 /* set cancel cookie callback function */
3348 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3349 filter_cancel_cookies_cb,
3353 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3354 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3356 /* The truncate might have used up our transaction credits. Make
3357 * sure we have one left for the last_rcvd update. */
3358 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3360 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3362 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3366 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3368 CERROR("error on commit, err = %d\n", err);
3375 /* For a partial-page truncate flush the page to disk immediately
3376 * to avoid data corruption during direct disk write. b=17397 */
3377 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3378 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3379 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3387 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3388 UNLOCK_INODE_MUTEX(inode);
3389 if (ia_valid & ATTR_SIZE)
3390 up_write(&inode->i_alloc_sem);
3392 OBD_FREE(fcc, sizeof(*fcc));
3394 /* trigger quota release */
3395 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3396 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3397 int rc2 = lquota_adjust(filter_quota_interface_ref,
3398 exp->exp_obd, cur_ids,
3399 orig_ids, rc, FSFILT_OP_SETATTR);
3400 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3401 "filter adjust qunit. (rc:%d)\n", rc2);
3406 /* this is called from filter_truncate() until we have filter_punch() */
3407 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3408 struct obd_trans_info *oti)
3410 struct obdo *oa = oinfo->oi_oa;
3411 struct lustre_capa *capa = oinfo_capa(oinfo);
3412 struct ldlm_res_id res_id;
3413 struct filter_mod_data *fmd;
3414 struct lvfs_run_ctxt saved;
3415 struct filter_obd *filter;
3416 struct ldlm_resource *res;
3417 struct dentry *dentry;
3418 __u64 opc = CAPA_OPC_META_WRITE;
3422 if (oa->o_valid & OBD_FL_TRUNC)
3423 opc |= CAPA_OPC_OSS_TRUNC;
3424 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3428 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3429 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3434 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3435 /* This would be very bad - accidentally truncating a file when
3436 * changing the time or similar - bug 12203. */
3437 if (oa->o_valid & OBD_MD_FLSIZE &&
3438 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3439 static char mdsinum[48];
3441 if (oa->o_valid & OBD_MD_FLFID)
3442 snprintf(mdsinum, sizeof(mdsinum) - 1,
3443 " of inode "LPU64"/%u", oa->o_fid,
3448 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3450 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3455 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3457 RETURN(PTR_ERR(dentry));
3459 filter = &exp->exp_obd->u.filter;
3460 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3464 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3465 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3466 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3467 fmd->fmd_mactime_xid = oti->oti_xid;
3468 filter_fmd_put(exp, fmd);
3471 /* setting objects attributes (including owner/group) */
3472 rc = filter_setattr_internal(exp, dentry, oa, oti);
3474 GOTO(out_unlock, rc);
3476 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3477 &res_id, LDLM_EXTENT, 0);
3480 LDLM_RESOURCE_ADDREF(res);
3481 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3482 LDLM_RESOURCE_DELREF(res);
3483 ldlm_resource_putref(res);
3486 oa->o_valid = OBD_MD_FLID;
3488 /* Quota release need uid/gid info */
3489 obdo_from_inode(oa, dentry->d_inode,
3490 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3496 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3500 /* XXX identical to osc_unpackmd */
3501 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3502 struct lov_mds_md *lmm, int lmm_bytes)
3508 if (lmm_bytes < sizeof (*lmm)) {
3509 CERROR("lov_mds_md too small: %d, need %d\n",
3510 lmm_bytes, (int)sizeof(*lmm));
3513 /* XXX LOV_MAGIC etc check? */
3515 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3516 CERROR("lov_mds_md: zero lmm_object_id\n");
3521 lsm_size = lov_stripe_md_size(1);
3525 if (*lsmp != NULL && lmm == NULL) {
3526 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3527 OBD_FREE(*lsmp, lsm_size);
3532 if (*lsmp == NULL) {
3533 OBD_ALLOC(*lsmp, lsm_size);
3537 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3538 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3539 OBD_FREE(*lsmp, lsm_size);
3542 loi_init((*lsmp)->lsm_oinfo[0]);
3546 /* XXX zero *lsmp? */
3547 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3548 LASSERT((*lsmp)->lsm_object_id);
3551 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3556 /* caller must hold fo_create_locks[oa->o_gr] */
3557 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3558 struct filter_obd *filter)
3560 struct obdo doa; /* XXX obdo on stack */
3567 LASSERT_MDS_GROUP(oa->o_gr);
3568 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3569 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3571 memset(&doa, 0, sizeof(doa));
3573 doa.o_valid |= OBD_MD_FLGROUP;
3574 doa.o_gr = oa->o_gr;
3575 doa.o_mode = S_IFREG;
3577 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3578 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3579 exp->exp_obd->obd_name, doa.o_gr);
3583 last = filter_last_id(filter, doa.o_gr);
3585 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3587 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3588 exp->exp_obd->obd_name, oa->o_id + 1, last,
3589 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3591 for (id = last; id > oa->o_id; id--) {
3593 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3594 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3595 CEMERG("error destroying precreate objid "LPU64": %d\n",
3598 /* update last_id on disk periodically so that if we restart
3599 * we don't need to re-scan all of the just-deleted objects. */
3600 if ((id & 511) == 0 && !skip_orphan) {
3601 filter_set_last_id(filter, id - 1, doa.o_gr);
3602 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3606 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3607 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3610 filter_set_last_id(filter, id, doa.o_gr);
3611 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3613 /* don't reuse orphan object, return last used objid */
3617 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3622 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3623 obd_gr group, int *num);
3624 /* returns a negative error or a nonnegative number of files to create */
3625 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3626 obd_gr group, struct obd_trans_info *oti)
3628 struct obd_device *obd = exp->exp_obd;
3629 struct filter_obd *filter = &obd->u.filter;
3633 /* delete orphans request */
3634 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3635 obd_id last = filter_last_id(filter, group);
3637 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3638 CERROR("%s: dropping old orphan cleanup request\n",
3642 /* This causes inflight precreates to abort and drop lock */
3643 set_bit(group, &filter->fo_destroys_in_progress);
3644 down(&filter->fo_create_locks[group]);
3645 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3646 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3647 exp->exp_obd->obd_name, group);
3648 up(&filter->fo_create_locks[group]);
3651 diff = oa->o_id - last;
3652 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3655 if (-diff > OST_MAX_PRECREATE) {
3656 CERROR("%s: ignoring bogus orphan destroy request: "
3657 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3659 /* FIXME: should reset precreate_next_id on MDS */
3660 GOTO(out, rc = -EINVAL);
3663 rc = filter_destroy_precreated(exp, oa, filter);
3665 CERROR("%s: unable to write lastobjid, but "
3666 "orphans were deleted\n", obd->obd_name);
3669 /* XXX: Used by MDS for the first time! */
3670 clear_bit(group, &filter->fo_destroys_in_progress);
3673 down(&filter->fo_create_locks[group]);
3674 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3675 CERROR("%s: dropping old precreate request\n",
3679 /* only precreate if group == 0 and o_id is specfied */
3680 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3683 diff = oa->o_id - filter_last_id(filter, group);
3684 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3685 filter_last_id(filter, group), diff);
3687 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3688 oa->o_id, filter_last_id(filter, group), diff);
3692 oa->o_id = filter_last_id(&obd->u.filter, group);
3693 rc = filter_precreate(obd, oa, group, &diff);
3694 oa->o_id = filter_last_id(&obd->u.filter, group);
3696 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3699 /* else diff == 0 */
3702 up(&filter->fo_create_locks[group]);
3706 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3707 __u64 max_age, __u32 flags)
3709 struct filter_obd *filter = &obd->u.filter;
3710 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3714 /* at least try to account for cached pages. its still racey and
3715 * might be under-reporting if clients haven't announced their
3716 * caches with brw recently */
3717 spin_lock(&obd->obd_osfs_lock);
3718 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3719 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3720 spin_unlock(&obd->obd_osfs_lock);
3722 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3723 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3724 filter->fo_tot_dirty, filter->fo_tot_granted,
3725 filter->fo_tot_pending,
3726 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3728 filter_grant_sanity_check(obd, __FUNCTION__);
3730 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3731 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3732 osfs->os_bsize - 1) >> blockbits));
3734 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3735 * stop creating files on MDS if OST is not good shape to create
3737 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3738 OS_STATE_READONLY : 0;
3742 static int filter_use_existing_obj(struct obd_device *obd,
3743 struct dentry *dchild, void **handle,
3746 struct inode *inode = dchild->d_inode;
3750 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3753 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3754 if (IS_ERR(*handle))
3755 return PTR_ERR(*handle);
3757 iattr.ia_valid = ATTR_MODE;
3758 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3759 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3766 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3769 __u64 os_ffree = -1;
3771 spin_lock(&obd->obd_osfs_lock);
3772 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3774 os_ffree = obd->obd_osfs.os_ffree;
3775 spin_unlock(&obd->obd_osfs_lock);
3780 /* We rely on the fact that only one thread will be creating files in a given
3781 * group at a time, which is why we don't need an atomic filter_get_new_id.
3782 * Even if we had that atomic function, the following race would exist:
3784 * thread 1: gets id x from filter_next_id
3785 * thread 2: gets id (x + 1) from filter_next_id
3786 * thread 2: creates object (x + 1)
3787 * thread 1: tries to create object x, gets -ENOSPC
3789 * Caller must hold fo_create_locks[group]
3791 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3792 obd_gr group, int *num)
3794 struct dentry *dchild = NULL, *dparent = NULL;
3795 struct filter_obd *filter;
3796 struct obd_statfs *osfs;
3797 int err = 0, rc = 0, recreate_obj = 0, i;
3798 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3801 void *handle = NULL;
3804 filter = &obd->u.filter;
3806 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3808 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3810 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3811 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3814 OBD_ALLOC(osfs, sizeof(*osfs));
3817 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3818 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3819 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3820 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3821 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3825 OBD_FREE(osfs, sizeof(*osfs));
3830 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3831 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3833 for (i = 0; i < *num && err == 0; i++) {
3834 int cleanup_phase = 0;
3836 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3837 CWARN("%s: create aborted by destroy\n",
3846 last_id = filter_last_id(filter, group);
3847 if (next_id > last_id) {
3848 CERROR("Error: Trying to recreate obj greater"
3849 "than last id "LPD64" > "LPD64"\n",
3851 GOTO(cleanup, rc = -EINVAL);
3854 next_id = filter_last_id(filter, group) + 1;
3856 dparent = filter_parent_lock(obd, group, next_id);
3857 if (IS_ERR(dparent))
3858 GOTO(cleanup, rc = PTR_ERR(dparent));
3859 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3861 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3863 GOTO(cleanup, rc = PTR_ERR(dchild));
3864 cleanup_phase = 2; /* f_dput(dchild) */
3866 if (dchild->d_inode != NULL) {
3867 /* This would only happen if lastobjid was bad on disk*/
3868 /* Could also happen if recreating missing obj but it
3869 * already exists. */
3871 CERROR("%s: recreating existing object %.*s?\n",
3872 obd->obd_name, dchild->d_name.len,
3873 dchild->d_name.name);
3875 /* Use these existing objects if they are
3877 if (dchild->d_inode->i_size == 0) {
3878 rc = filter_use_existing_obj(obd,dchild,
3879 &handle, &cleanup_phase);
3886 CERROR("%s: Serious error: objid %.*s already "
3887 "exists; is this filesystem corrupt?\n",
3888 obd->obd_name, dchild->d_name.len,
3889 dchild->d_name.name);
3892 GOTO(cleanup, rc = -EEXIST);
3895 handle = fsfilt_start_log(obd, dparent->d_inode,
3896 FSFILT_OP_CREATE, NULL, 1);
3898 GOTO(cleanup, rc = PTR_ERR(handle));
3901 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3902 ",od->o_id="LPU64")\n", obd->obd_name, group,
3905 /* We mark object SUID+SGID to flag it for accepting UID+GID
3906 * from client on first write. Currently the permission bits
3907 * on the OST are never used, so this is OK. */
3908 rc = ll_vfs_create(dparent->d_inode, dchild,
3909 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3911 CERROR("create failed rc = %d\n", rc);
3912 if (rc == -ENOSPC) {
3913 os_ffree = filter_calc_free_inodes(obd);
3915 CERROR("%s: free inode "LPU64"\n",
3916 obd->obd_name, os_ffree);
3922 if (!recreate_obj) {
3923 filter_set_last_id(filter, next_id, group);
3924 err = filter_update_last_objid(obd, group, 0);
3926 CERROR("unable to write lastobjid "
3927 "but file created\n");
3931 switch(cleanup_phase) {
3933 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3935 CERROR("error on commit, err = %d\n", err);
3942 filter_parent_unlock(dparent);
3949 if (time_after(jiffies, enough_time)) {
3951 "%s: precreate slow - want %d got %d \n",
3952 obd->obd_name, *num, i);
3959 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3960 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3965 static int filter_create(struct obd_export *exp, struct obdo *oa,
3966 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3968 struct obd_device *obd = exp->exp_obd;
3969 struct filter_export_data *fed;
3970 struct filter_obd *filter;
3971 struct lvfs_run_ctxt saved;
3972 struct lov_stripe_md *lsm = NULL;
3973 int rc = 0, diff, group = oa->o_gr;
3976 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3977 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3979 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3980 CERROR("!!! nid %s sent invalid object group %d\n",
3981 obd_export_nid2str(exp), group);
3985 fed = &exp->exp_filter_data;
3986 filter = &obd->u.filter;
3988 if (fed->fed_group != group) {
3989 CERROR("!!! this export (nid %s) used object group %d "
3990 "earlier; now it's trying to use group %d! This could "
3991 "be a bug in the MDS. Please report to "
3992 "http://bugzilla.lustre.org/\n",
3993 obd_export_nid2str(exp), fed->fed_group, group);
4000 rc = obd_alloc_memmd(exp, &lsm);
4007 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4009 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4010 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4011 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4012 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4013 oa->o_id, filter_last_id(filter,
4018 down(&filter->fo_create_locks[oa->o_gr]);
4019 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4020 up(&filter->fo_create_locks[oa->o_gr]);
4023 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4026 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4027 if (rc && ea != NULL && *ea != lsm) {
4028 obd_free_memmd(exp, &lsm);
4029 } else if (rc == 0 && ea != NULL) {
4030 /* XXX LOV STACKING: the lsm that is passed to us from
4031 * LOV does not have valid lsm_oinfo data structs, so
4032 * don't go touching that. This needs to be fixed in a
4034 lsm->lsm_object_id = oa->o_id;
4041 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4042 struct lov_stripe_md *md, struct obd_trans_info *oti,
4043 struct obd_export *md_exp, void *capa)
4045 unsigned int qcids[MAXQUOTAS] = {0, 0};
4046 struct obd_device *obd;
4047 struct filter_obd *filter;
4048 struct dentry *dchild = NULL, *dparent = NULL;
4049 struct lustre_handle lockh = { 0 };
4050 struct lvfs_run_ctxt saved;
4051 void *handle = NULL;
4052 struct llog_cookie *fcc = NULL;
4053 int rc, rc2, cleanup_phase = 0, sync = 0;
4057 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4059 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4060 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4065 filter = &obd->u.filter;
4067 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4070 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4071 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4073 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4075 GOTO(cleanup, rc = PTR_ERR(dchild));
4078 if (dchild->d_inode == NULL) {
4079 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4081 /* If object already gone, cancel cookie right now */
4082 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4083 struct llog_ctxt *ctxt;
4084 struct obd_llog_group *olg;
4085 fcc = &oa->o_lcookie;
4086 olg = filter_find_olg(obd, oa->o_gr);
4088 CERROR(" %s: can not find olg of group %d\n",
4089 obd->obd_name, (int)oa->o_gr);
4090 GOTO(cleanup, rc = PTR_ERR(olg));
4092 llog_group_set_export(olg, exp);
4094 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4095 llog_cancel(ctxt, NULL, 1, fcc, 0);
4096 llog_ctxt_put(ctxt);
4097 fcc = NULL; /* we didn't allocate fcc, don't free it */
4099 GOTO(cleanup, rc = -ENOENT);
4102 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4104 /* Our MDC connection is established by the MDS to us */
4105 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4106 OBD_ALLOC(fcc, sizeof(*fcc));
4108 *fcc = oa->o_lcookie;
4110 DQUOT_INIT(dchild->d_inode);
4112 /* we're gonna truncate it first in order to avoid possible deadlock:
4114 * open trasaction open transaction
4115 * down(i_zombie) down(i_zombie)
4116 * restart transaction
4117 * (see BUG 4180) -bzzz
4119 LOCK_INODE_MUTEX(dchild->d_inode);
4121 /* VBR: version recovery check */
4122 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4126 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4128 if (IS_ERR(handle)) {
4129 UNLOCK_INODE_MUTEX(dchild->d_inode);
4130 GOTO(cleanup, rc = PTR_ERR(handle));
4133 iattr.ia_valid = ATTR_SIZE;
4135 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4136 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4137 UNLOCK_INODE_MUTEX(dchild->d_inode);
4141 GOTO(cleanup, rc = rc2);
4143 /* We don't actually need to lock the parent until we are unlinking
4144 * here, and not while truncating above. That avoids holding the
4145 * parent lock for a long time during truncate, which can block other
4146 * threads from doing anything to objects in that directory. bug 7171 */
4147 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4148 if (IS_ERR(dparent))
4149 GOTO(cleanup, rc = PTR_ERR(dparent));
4150 cleanup_phase = 3; /* filter_parent_unlock */
4152 LOCK_INODE_MUTEX(dchild->d_inode);
4153 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4154 if (IS_ERR(handle)) {
4155 UNLOCK_INODE_MUTEX(dchild->d_inode);
4156 GOTO(cleanup, rc = PTR_ERR(handle));
4158 cleanup_phase = 4; /* fsfilt_commit */
4160 /* Quota release need uid/gid of inode */
4161 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4163 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4165 /* this drops dchild->d_inode->i_mutex unconditionally */
4166 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4170 switch(cleanup_phase) {
4173 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4174 oti->oti_handle : handle,
4175 filter_cancel_cookies_cb,
4177 /* If add_journal_cb failed, then filter_finish_transno
4178 * will commit the handle and we will do a sync
4179 * on commit. then we call callback directly to free
4182 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4184 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4187 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4189 CERROR("error on commit, err = %d\n", rc2);
4196 filter_parent_unlock(dparent);
4198 filter_fini_destroy(obd, &lockh);
4202 OBD_FREE(fcc, sizeof(*fcc));
4204 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4207 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4211 /* trigger quota release */
4212 qcids[USRQUOTA] = oa->o_uid;
4213 qcids[GRPQUOTA] = oa->o_gid;
4214 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4217 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4221 /* NB start and end are used for punch, but not truncate */
4222 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4223 struct obd_trans_info *oti,
4224 struct ptlrpc_request_set *rqset)
4229 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4230 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4231 oinfo->oi_policy.l_extent.end);
4235 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4236 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4237 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4239 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4240 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4241 rc = filter_setattr(exp, oinfo, oti);
4242 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4246 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4247 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4250 struct lvfs_run_ctxt saved;
4251 struct filter_obd *filter;
4252 struct dentry *dentry;
4256 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4257 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4261 filter = &exp->exp_obd->u.filter;
4263 /* An objid of zero is taken to mean "sync whole filesystem" */
4264 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4265 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4266 /* Flush any remaining cancel messages out to the target */
4267 filter_sync_llogs(exp->exp_obd, exp);
4271 dentry = filter_oa2dentry(exp->exp_obd, oa);
4273 RETURN(PTR_ERR(dentry));
4275 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4277 LOCK_INODE_MUTEX(dentry->d_inode);
4279 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4281 /* just any file to grab fsync method - "file" arg unused */
4282 struct file *file = filter->fo_rcvd_filp;
4284 if (file->f_op && file->f_op->fsync)
4285 rc = file->f_op->fsync(NULL, dentry, 1);
4287 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4291 UNLOCK_INODE_MUTEX(dentry->d_inode);
4293 oa->o_valid = OBD_MD_FLID;
4294 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4296 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4302 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4303 void *key, __u32 *vallen, void *val,
4304 struct lov_stripe_md *lsm)
4306 struct obd_device *obd;
4309 obd = class_exp2obd(exp);
4311 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4315 if (KEY_IS(KEY_BLOCKSIZE)) {
4316 __u32 *blocksize = val;
4318 if (*vallen < sizeof(*blocksize))
4320 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4322 *vallen = sizeof(*blocksize);
4326 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4327 __u32 *blocksize_bits = val;
4328 if (blocksize_bits) {
4329 if (*vallen < sizeof(*blocksize_bits))
4331 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4333 *vallen = sizeof(*blocksize_bits);
4337 if (KEY_IS(KEY_LAST_ID)) {
4338 obd_id *last_id = val;
4339 /* FIXME: object groups */
4341 if (*vallen < sizeof(*last_id))
4343 *last_id = filter_last_id(&obd->u.filter,
4344 exp->exp_filter_data.fed_group);
4346 *vallen = sizeof(*last_id);
4350 if (KEY_IS(KEY_FIEMAP)) {
4351 struct ll_fiemap_info_key *fm_key = key;
4352 struct dentry *dentry;
4353 struct ll_user_fiemap *fiemap = val;
4354 struct lvfs_run_ctxt saved;
4357 if (fiemap == NULL) {
4358 *vallen = fiemap_count_to_size(
4359 fm_key->fiemap.fm_extent_count);
4363 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4366 RETURN(PTR_ERR(dentry));
4368 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4369 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4370 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4371 EXT3_IOC_FIEMAP, (long)fiemap);
4372 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4378 CDEBUG(D_IOCTL, "invalid key\n");
4382 static inline int filter_setup_llog_group(struct obd_export *exp,
4383 struct obd_device *obd,
4386 struct obd_llog_group *olg;
4387 struct llog_ctxt *ctxt;
4390 olg = filter_find_create_olg(obd, group);
4392 RETURN(PTR_ERR(olg));
4394 llog_group_set_export(olg, exp);
4396 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4397 LASSERTF(ctxt != NULL, "ctxt is null\n");
4399 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4400 llog_ctxt_put(ctxt);
4404 static int filter_set_grant_shrink(struct obd_export *exp,
4405 struct ost_body *body)
4407 /* handle shrink grant */
4408 spin_lock(&exp->exp_obd->obd_osfs_lock);
4409 filter_grant_incoming(exp, &body->oa);
4410 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4416 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4418 struct obd_device *obd;
4424 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4428 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4429 obd_export_nid2str(exp));
4430 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4432 /* setup llog imports */
4434 group = (int)(*(__u32 *)val);
4436 group = 0; /* default value */
4438 LASSERT_MDS_GROUP(group);
4439 rc = filter_setup_llog_group(exp, obd, group);
4443 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4445 if (group == FILTER_GROUP_MDS0) {
4446 /* setup llog group 1 for interop */
4447 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4453 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4454 void *key, __u32 vallen, void *val,
4455 struct ptlrpc_request_set *set)
4457 struct obd_device *obd;
4462 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4466 if (KEY_IS(KEY_CAPA_KEY)) {
4468 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4470 CERROR("filter update capability key failed: %d\n", rc);
4474 if (KEY_IS(KEY_REVIMP_UPD)) {
4475 filter_revimp_update(exp);
4476 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4480 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4481 filter_adapt_sptlrpc_conf(obd, 0);
4485 if (KEY_IS(KEY_MDS_CONN))
4486 RETURN(filter_set_mds_conn(exp, val));
4488 if (KEY_IS(KEY_GRANT_SHRINK))
4489 RETURN(filter_set_grant_shrink(exp, val));
4494 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4495 int len, void *karg, void *uarg)
4497 struct obd_device *obd = exp->exp_obd;
4498 struct obd_ioctl_data *data = karg;
4502 case OBD_IOC_ABORT_RECOVERY: {
4503 CERROR("aborting recovery for device %s\n", obd->obd_name);
4504 target_stop_recovery_thread(obd);
4508 case OBD_IOC_SYNC: {
4509 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4510 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4514 case OBD_IOC_SET_READONLY: {
4516 struct super_block *sb = obd->u.obt.obt_sb;
4517 struct inode *inode = sb->s_root->d_inode;
4518 BDEVNAME_DECLARE_STORAGE(tmp);
4519 CERROR("*** setting device %s read-only ***\n",
4520 ll_bdevname(sb, tmp));
4522 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4523 if (!IS_ERR(handle))
4524 rc = fsfilt_commit(obd, inode, handle, 1);
4526 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4527 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4529 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4533 case OBD_IOC_CATLOGLIST: {
4534 rc = llog_catalog_list(obd, 1, data);
4538 case OBD_IOC_LLOG_CANCEL:
4539 case OBD_IOC_LLOG_REMOVE:
4540 case OBD_IOC_LLOG_INFO:
4541 case OBD_IOC_LLOG_PRINT: {
4542 /* FIXME to be finished */
4543 RETURN(-EOPNOTSUPP);
4545 struct llog_ctxt *ctxt = NULL;
4547 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4548 rc = llog_ioctl(ctxt, cmd, data);
4549 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4562 static int filter_health_check(struct obd_device *obd)
4564 #ifdef USE_HEALTH_CHECK_WRITE
4565 struct filter_obd *filter = &obd->u.filter;
4570 * health_check to return 0 on healthy
4571 * and 1 on unhealthy.
4573 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4576 #ifdef USE_HEALTH_CHECK_WRITE
4577 LASSERT(filter->fo_health_check_filp != NULL);
4578 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4583 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4586 return filter_fid2dentry(data, NULL, gr, id);
4589 static int filter_process_config(struct obd_device *obd, obd_count len,
4592 struct lustre_cfg *lcfg = buf;
4593 struct lprocfs_static_vars lvars;
4596 switch (lcfg->lcfg_command) {
4598 lprocfs_filter_init_vars(&lvars);
4600 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4610 static struct lvfs_callback_ops filter_lvfs_ops = {
4611 l_fid2dentry: filter_lvfs_fid2dentry,
4614 static struct obd_ops filter_obd_ops = {
4615 .o_owner = THIS_MODULE,
4616 .o_get_info = filter_get_info,
4617 .o_set_info_async = filter_set_info_async,
4618 .o_setup = filter_setup,
4619 .o_precleanup = filter_precleanup,
4620 .o_cleanup = filter_cleanup,
4621 .o_connect = filter_connect,
4622 .o_reconnect = filter_reconnect,
4623 .o_disconnect = filter_disconnect,
4624 .o_ping = filter_ping,
4625 .o_init_export = filter_init_export,
4626 .o_destroy_export = filter_destroy_export,
4627 .o_statfs = filter_statfs,
4628 .o_getattr = filter_getattr,
4629 .o_unpackmd = filter_unpackmd,
4630 .o_create = filter_create,
4631 .o_setattr = filter_setattr,
4632 .o_destroy = filter_destroy,
4633 .o_brw = filter_brw,
4634 .o_punch = filter_truncate,
4635 .o_sync = filter_sync,
4636 .o_preprw = filter_preprw,
4637 .o_commitrw = filter_commitrw,
4638 .o_llog_init = filter_llog_init,
4639 .o_llog_connect = filter_llog_connect,
4640 .o_llog_finish = filter_llog_finish,
4641 .o_iocontrol = filter_iocontrol,
4642 .o_health_check = filter_health_check,
4643 .o_process_config = filter_process_config,
4646 quota_interface_t *filter_quota_interface_ref;
4647 extern quota_interface_t filter_quota_interface;
4649 static int __init obdfilter_init(void)
4651 struct lprocfs_static_vars lvars;
4654 lprocfs_filter_init_vars(&lvars);
4656 request_module("lquota");
4657 OBD_ALLOC(obdfilter_created_scratchpad,
4658 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4659 sizeof(*obdfilter_created_scratchpad));
4660 if (obdfilter_created_scratchpad == NULL)
4663 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4664 sizeof(struct filter_mod_data),
4667 GOTO(out, rc = -ENOMEM);
4669 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4670 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4672 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4673 LUSTRE_OST_NAME, NULL);
4677 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4678 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4679 ll_fmd_cachep = NULL;
4681 if (filter_quota_interface_ref)
4682 PORTAL_SYMBOL_PUT(filter_quota_interface);
4684 OBD_FREE(obdfilter_created_scratchpad,
4685 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4686 sizeof(*obdfilter_created_scratchpad));
4692 static void __exit obdfilter_exit(void)
4694 if (filter_quota_interface_ref)
4695 PORTAL_SYMBOL_PUT(filter_quota_interface);
4697 if (ll_fmd_cachep) {
4698 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4699 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4700 ll_fmd_cachep = NULL;
4703 class_unregister_type(LUSTRE_OST_NAME);
4704 OBD_FREE(obdfilter_created_scratchpad,
4705 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4706 sizeof(*obdfilter_created_scratchpad));
4709 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4710 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4711 MODULE_LICENSE("GPL");
4713 module_init(obdfilter_init);
4714 module_exit(obdfilter_exit);