1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 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);
167 class_export_cb_get(exp); /* released when the cb is called */
169 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
175 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
176 lcd, sizeof(*lcd), &off,
177 force_sync | exp->exp_need_sync);
179 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
187 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
188 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
193 void f_dput(struct dentry *dentry)
195 /* Can't go inside filter_ddelete because it can block */
196 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
197 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
198 LASSERT(atomic_read(&dentry->d_count) > 0);
203 static void init_brw_stats(struct brw_stats *brw_stats)
206 for (i = 0; i < BRW_LAST; i++)
207 spin_lock_init(&brw_stats->hist[i].oh_lock);
210 static int lprocfs_init_rw_stats(struct obd_device *obd,
211 struct lprocfs_stats **stats)
215 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
216 LPROC_FILTER_LAST - 1;
217 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
221 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
222 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
223 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
224 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
225 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
230 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
231 plus the procfs overhead :( */
232 static int filter_export_stats_init(struct obd_device *obd,
233 struct obd_export *exp,
239 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
240 /* Self-export gets no proc entry */
243 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
245 /* Mask error for already created
253 struct nid_stat *tmp = exp->exp_nid_stats;
254 LASSERT(tmp != NULL);
256 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
257 if (tmp->nid_brw_stats == NULL)
260 init_brw_stats(tmp->nid_brw_stats);
261 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
262 0644, &filter_per_nid_stats_fops,
265 CWARN("Error adding the brw_stats file\n");
267 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
271 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
275 /* Always add in ldlm_stats */
276 tmp->nid_ldlm_stats =
277 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
278 LPROCFS_STATS_FLAG_NOPERCPU);
279 if (tmp->nid_ldlm_stats == NULL)
282 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
283 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
284 tmp->nid_ldlm_stats);
292 /* Add client data to the FILTER. We use a bitmap to locate a free space
293 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
294 * Otherwise, we have just read the data from the last_rcvd file and
295 * we know its offset. */
296 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
299 struct filter_obd *filter = &obd->u.filter;
300 struct filter_export_data *fed = &exp->exp_filter_data;
301 unsigned long *bitmap = filter->fo_last_rcvd_slots;
302 int new_client = (cl_idx == -1);
306 LASSERT(bitmap != NULL);
307 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
310 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
313 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
314 * there's no need for extra complication here
317 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
319 if (cl_idx >= LR_MAX_CLIENTS) {
320 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
324 if (test_and_set_bit(cl_idx, bitmap)) {
325 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
330 if (test_and_set_bit(cl_idx, bitmap)) {
331 CERROR("FILTER client %d: bit already set in bitmap!\n",
337 fed->fed_lr_idx = cl_idx;
338 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
339 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
340 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
342 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
343 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
346 struct lvfs_run_ctxt saved;
347 loff_t off = fed->fed_lr_off;
351 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
352 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
354 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
355 /* Transaction needed to fix bug 1403 */
356 handle = fsfilt_start(obd,
357 filter->fo_rcvd_filp->f_dentry->d_inode,
358 FSFILT_OP_SETATTR, NULL);
359 if (IS_ERR(handle)) {
360 rc = PTR_ERR(handle);
361 CERROR("unable to start transaction: rc %d\n", rc);
363 fed->fed_lcd->lcd_last_epoch =
364 filter->fo_fsd->lsd_start_epoch;
365 exp->exp_last_request_time = cfs_time_current_sec();
366 rc = fsfilt_add_journal_cb(obd, 0, handle,
367 target_client_add_cb, exp);
369 spin_lock(&exp->exp_lock);
370 exp->exp_need_sync = 1;
371 spin_unlock(&exp->exp_lock);
373 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
375 sizeof(*fed->fed_lcd),
376 &off, rc /* sync if no cb */);
378 filter->fo_rcvd_filp->f_dentry->d_inode,
381 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
384 CERROR("error writing %s client idx %u: rc %d\n",
385 LAST_RCVD, fed->fed_lr_idx, rc);
392 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
394 static int filter_client_free(struct obd_export *exp)
396 struct filter_export_data *fed = &exp->exp_filter_data;
397 struct filter_obd *filter = &exp->exp_obd->u.filter;
398 struct obd_device *obd = exp->exp_obd;
399 struct lvfs_run_ctxt saved;
404 if (fed->fed_lcd == NULL)
407 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
408 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
411 LASSERT(filter->fo_last_rcvd_slots != NULL);
413 off = fed->fed_lr_off;
415 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
416 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
418 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
419 * we leak a client slot that will be cleaned on the next recovery. */
421 CERROR("%s: client idx %d has med_off %lld\n",
422 obd->obd_name, fed->fed_lr_idx, off);
423 GOTO(free, rc = -EINVAL);
426 /* Clear the bit _after_ zeroing out the client so we don't
427 race with filter_client_add and zero out new clients.*/
428 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
429 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
434 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
435 /* Don't force sync on disconnect if aborting recovery,
436 * or it does num_clients * num_osts. b=17194 */
437 int need_sync = exp->exp_need_sync &&
438 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
439 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
440 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
441 sizeof(zero_lcd), &off, 0);
443 /* Make sure the server's last_transno is up to date. Do this
444 * after the client is freed so we know all the client's
445 * transactions have been committed. */
447 filter_update_server_data(obd, filter->fo_rcvd_filp,
448 filter->fo_fsd, need_sync);
449 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
451 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
452 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
453 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
454 LAST_RCVD, need_sync ? "" : "a", rc);
457 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
458 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
465 OBD_FREE_PTR(fed->fed_lcd);
471 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
472 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
473 struct filter_mod_data *fmd)
475 LASSERT_SPIN_LOCKED(&fed->fed_lock);
476 if (--fmd->fmd_refcount == 0) {
477 /* XXX when we have persistent reservations and the handle
478 * is stored herein we need to drop it here. */
479 fed->fed_mod_count--;
480 list_del(&fmd->fmd_list);
481 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
485 /* drop fmd reference, free it if last ref */
486 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
488 struct filter_export_data *fed;
493 fed = &exp->exp_filter_data;
494 spin_lock(&fed->fed_lock);
495 filter_fmd_put_nolock(fed, fmd); /* caller reference */
496 spin_unlock(&fed->fed_lock);
499 /* expire entries from the end of the list if there are too many
500 * or they are too old */
501 static void filter_fmd_expire_nolock(struct filter_obd *filter,
502 struct filter_export_data *fed,
503 struct filter_mod_data *keep)
505 struct filter_mod_data *fmd, *tmp;
507 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
511 if (time_before(jiffies, fmd->fmd_expire) &&
512 fed->fed_mod_count < filter->fo_fmd_max_num)
515 list_del_init(&fmd->fmd_list);
516 filter_fmd_put_nolock(fed, fmd); /* list reference */
520 void filter_fmd_expire(struct obd_export *exp)
522 spin_lock(&exp->exp_filter_data.fed_lock);
523 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
524 &exp->exp_filter_data, NULL);
525 spin_unlock(&exp->exp_filter_data.fed_lock);
528 /* find specified objid, group in export fmd list.
529 * caller must hold fed_lock and take fmd reference itself */
530 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
531 struct filter_export_data *fed,
532 obd_id objid, obd_gr group)
534 struct filter_mod_data *found = NULL, *fmd;
536 LASSERT_SPIN_LOCKED(&fed->fed_lock);
538 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
539 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
541 list_del(&fmd->fmd_list);
542 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
543 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
548 filter_fmd_expire_nolock(filter, fed, found);
553 /* Find fmd based on objid and group, or return NULL if not found. */
554 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
555 obd_id objid, obd_gr group)
557 struct filter_mod_data *fmd;
559 spin_lock(&exp->exp_filter_data.fed_lock);
560 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
561 &exp->exp_filter_data, objid, group);
563 fmd->fmd_refcount++; /* caller reference */
564 spin_unlock(&exp->exp_filter_data.fed_lock);
569 /* Find fmd based on objid and group, or create a new one if none is found.
570 * It is possible for this function to return NULL under memory pressure,
571 * or if objid = 0 is passed (which will only cause old entries to expire).
572 * Currently this is not fatal because any fmd state is transient and
573 * may also be freed when it gets sufficiently old. */
574 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
575 obd_id objid, obd_gr group)
577 struct filter_export_data *fed = &exp->exp_filter_data;
578 struct filter_mod_data *found = NULL, *fmd_new = NULL;
580 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
582 spin_lock(&fed->fed_lock);
583 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
586 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
587 fmd_new->fmd_id = objid;
588 fmd_new->fmd_gr = group;
589 fmd_new->fmd_refcount++; /* list reference */
591 fed->fed_mod_count++;
593 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
597 found->fmd_refcount++; /* caller reference */
598 found->fmd_expire = jiffies +
599 exp->exp_obd->u.filter.fo_fmd_max_age;
602 spin_unlock(&fed->fed_lock);
608 /* drop fmd list reference so it will disappear when last reference is put.
609 * This isn't so critical because it would in fact only affect the one client
610 * that is doing the unlink and at worst we have an stale entry referencing
611 * an object that should never be used again. */
612 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
614 struct filter_mod_data *found = NULL;
616 spin_lock(&exp->exp_filter_data.fed_lock);
617 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
619 list_del_init(&found->fmd_list);
620 filter_fmd_put_nolock(&exp->exp_filter_data, found);
622 spin_unlock(&exp->exp_filter_data.fed_lock);
625 #define filter_fmd_drop(exp, objid, group)
628 /* remove all entries from fmd list */
629 static void filter_fmd_cleanup(struct obd_export *exp)
631 struct filter_export_data *fed = &exp->exp_filter_data;
632 struct filter_mod_data *fmd = NULL, *tmp;
634 spin_lock(&fed->fed_lock);
635 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
636 list_del_init(&fmd->fmd_list);
637 filter_fmd_put_nolock(fed, fmd);
639 spin_unlock(&fed->fed_lock);
642 static int filter_init_export(struct obd_export *exp)
644 spin_lock_init(&exp->exp_filter_data.fed_lock);
645 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
647 spin_lock(&exp->exp_lock);
648 exp->exp_connecting = 1;
649 spin_unlock(&exp->exp_lock);
651 return ldlm_init_export(exp);
654 static int filter_free_server_data(struct filter_obd *filter)
656 OBD_FREE_PTR(filter->fo_fsd);
657 filter->fo_fsd = NULL;
658 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
659 filter->fo_last_rcvd_slots = NULL;
663 /* assumes caller is already in kernel ctxt */
664 int filter_update_server_data(struct obd_device *obd, struct file *filp,
665 struct lr_server_data *fsd, int force_sync)
671 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
672 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
673 le64_to_cpu(fsd->lsd_last_transno));
674 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
675 le64_to_cpu(fsd->lsd_mount_count));
677 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
679 CERROR("error writing lr_server_data: rc = %d\n", rc);
684 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
687 struct filter_obd *filter = &obd->u.filter;
693 if (filter->fo_last_objid_files[group] == NULL) {
694 CERROR("Object group "LPU64" not fully setup; not updating "
695 "last_objid\n", group);
699 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
700 obd->obd_name, group, filter->fo_last_objids[group]);
702 tmp = cpu_to_le64(filter->fo_last_objids[group]);
703 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
704 &tmp, sizeof(tmp), &off, force_sync);
706 CERROR("error writing group "LPU64" last objid: rc = %d\n",
710 extern int ost_handle(struct ptlrpc_request *req);
711 /* assumes caller has already in kernel ctxt */
712 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
714 struct filter_obd *filter = &obd->u.filter;
715 struct lr_server_data *fsd;
716 struct lsd_client_data *lcd = NULL;
717 struct inode *inode = filp->f_dentry->d_inode;
718 unsigned long last_rcvd_size = i_size_read(inode);
725 /* ensure padding in the struct is the correct size */
726 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
727 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
728 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
729 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
731 OBD_ALLOC(fsd, sizeof(*fsd));
734 filter->fo_fsd = fsd;
736 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
737 if (filter->fo_last_rcvd_slots == NULL) {
738 OBD_FREE(fsd, sizeof(*fsd));
742 if (last_rcvd_size == 0) {
743 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
745 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
746 fsd->lsd_last_transno = 0;
747 mount_count = fsd->lsd_mount_count = 0;
748 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
749 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
750 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
751 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
752 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
753 /* OBD_COMPAT_OST is set in filter_connect_internal when the
754 * MDS first connects and assigns the OST index number. */
755 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
758 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
760 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
764 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
765 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
766 "using the wrong disk %s. Were the "
767 "/dev/ assignments rearranged?\n",
768 obd->obd_uuid.uuid, fsd->lsd_uuid);
769 GOTO(err_fsd, rc = -EINVAL);
771 mount_count = le64_to_cpu(fsd->lsd_mount_count);
772 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
774 /* Assume old last_rcvd format unless I_C_LR is set */
775 if (!(fsd->lsd_feature_incompat &
776 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
777 fsd->lsd_last_transno = fsd->lsd_compat14;
779 /* OBD_COMPAT_OST is set in filter_connect_internal when the
780 * MDS first connects and assigns the OST index number. */
781 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
785 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
786 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
787 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
788 ~FILTER_INCOMPAT_SUPP);
789 GOTO(err_fsd, rc = -EINVAL);
791 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
792 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
793 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
794 ~FILTER_ROCOMPAT_SUPP);
795 /* Do something like remount filesystem read-only */
796 GOTO(err_fsd, rc = -EINVAL);
799 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
801 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
802 obd->obd_name, start_epoch);
803 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
804 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
805 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
806 obd->obd_name, mount_count + 1);
807 CDEBUG(D_INODE, "%s: server data size: %u\n",
808 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
809 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
810 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
811 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
812 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
813 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
814 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
815 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
816 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
817 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
818 le16_to_cpu(fsd->lsd_client_size));
820 if (!obd->obd_replayable) {
821 CWARN("%s: recovery support OFF\n", obd->obd_name);
825 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
826 off < last_rcvd_size; cl_idx++) {
828 struct obd_export *exp;
829 struct filter_export_data *fed;
834 GOTO(err_client, rc = -ENOMEM);
837 /* Don't assume off is incremented properly by
838 * fsfilt_read_record(), in case sizeof(*lcd)
839 * isn't the same as fsd->lsd_client_size. */
840 off = le32_to_cpu(fsd->lsd_client_start) +
841 cl_idx * le16_to_cpu(fsd->lsd_client_size);
842 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
844 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
845 LAST_RCVD, cl_idx, off, rc);
846 break; /* read error shouldn't cause startup to fail */
849 if (lcd->lcd_uuid[0] == '\0') {
850 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
855 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
857 /* These exports are cleaned up by filter_disconnect(), so they
858 * need to be set up like real exports as filter_connect() does.
860 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
862 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
863 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
864 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
866 if (PTR_ERR(exp) == -EALREADY) {
867 /* export already exists, zero out this one */
868 CERROR("Zeroing out duplicate export due to "
870 lcd->lcd_uuid[0] = '\0';
872 GOTO(err_client, rc = PTR_ERR(exp));
875 fed = &exp->exp_filter_data;
877 fed->fed_group = 0; /* will be assigned at connect */
878 filter_export_stats_init(obd, exp, NULL);
879 rc = filter_client_add(obd, exp, cl_idx);
880 /* can't fail for existing client */
881 LASSERTF(rc == 0, "rc = %d\n", rc);
883 /* VBR: set export last committed */
884 exp->exp_last_committed = last_rcvd;
885 spin_lock(&exp->exp_lock);
886 exp->exp_connecting = 0;
887 exp->exp_in_recovery = 0;
888 spin_unlock(&exp->exp_lock);
889 spin_lock_bh(&obd->obd_processing_task_lock);
890 obd->obd_max_recoverable_clients++;
891 spin_unlock_bh(&obd->obd_processing_task_lock);
893 class_export_put(exp);
896 /* Need to check last_rcvd even for duplicated exports. */
897 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
900 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
901 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
907 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
909 filter->fo_mount_count = mount_count + 1;
910 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
912 /* save it, so mount count and last_transno is current */
913 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
915 GOTO(err_client, rc);
920 class_disconnect_exports(obd);
922 filter_free_server_data(filter);
926 static int filter_cleanup_groups(struct obd_device *obd)
928 struct filter_obd *filter = &obd->u.filter;
930 struct dentry *dentry;
934 if (filter->fo_dentry_O_groups != NULL) {
935 for (i = 0; i < filter->fo_group_count; i++) {
936 dentry = filter->fo_dentry_O_groups[i];
940 OBD_FREE(filter->fo_dentry_O_groups,
941 filter->fo_group_count *
942 sizeof(*filter->fo_dentry_O_groups));
943 filter->fo_dentry_O_groups = NULL;
945 if (filter->fo_last_objid_files != NULL) {
946 for (i = 0; i < filter->fo_group_count; i++) {
947 filp = filter->fo_last_objid_files[i];
951 OBD_FREE(filter->fo_last_objid_files,
952 filter->fo_group_count *
953 sizeof(*filter->fo_last_objid_files));
954 filter->fo_last_objid_files = NULL;
956 if (filter->fo_dentry_O_sub != NULL) {
957 for (i = 0; i < filter->fo_group_count; i++) {
958 for (j = 0; j < filter->fo_subdir_count; j++) {
959 dentry = filter->fo_dentry_O_sub[i].dentry[j];
964 OBD_FREE(filter->fo_dentry_O_sub,
965 filter->fo_group_count *
966 sizeof(*filter->fo_dentry_O_sub));
967 filter->fo_dentry_O_sub = NULL;
969 if (filter->fo_last_objids != NULL) {
970 OBD_FREE(filter->fo_last_objids,
971 filter->fo_group_count *
972 sizeof(*filter->fo_last_objids));
973 filter->fo_last_objids = NULL;
975 if (filter->fo_dentry_O != NULL) {
976 f_dput(filter->fo_dentry_O);
977 filter->fo_dentry_O = NULL;
982 static int filter_update_last_group(struct obd_device *obd, int group)
984 struct filter_obd *filter = &obd->u.filter;
985 struct file *filp = NULL;
986 int last_group = 0, rc;
990 if (group <= filter->fo_committed_group)
993 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
997 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1001 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1003 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1007 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1008 obd->obd_name, last_group, group);
1012 /* must be sync: bXXXX */
1013 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1015 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1019 filter->fo_committed_group = group;
1022 filp_close(filp, 0);
1026 static int filter_read_group_internal(struct obd_device *obd, int group,
1029 struct filter_obd *filter = &obd->u.filter;
1030 __u64 *new_objids = NULL;
1031 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1032 struct dentry **new_groups = NULL;
1033 struct file **new_files = NULL;
1034 struct dentry *dentry;
1036 int old_count = filter->fo_group_count, rc, stage = 0, i;
1040 int len = group + 1;
1042 snprintf(name, 24, "%d", group);
1046 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1048 if (IS_ERR(dentry)) {
1049 CERROR("Cannot lookup expected object group %d: %ld\n",
1050 group, PTR_ERR(dentry));
1051 RETURN(PTR_ERR(dentry));
1054 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1056 if (IS_ERR(dentry)) {
1057 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1059 RETURN(PTR_ERR(dentry));
1064 snprintf(name, 24, "O/%d/LAST_ID", group);
1066 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1068 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1069 GOTO(cleanup, rc = PTR_ERR(filp));
1073 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1075 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1079 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1080 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1081 if (tmp_subdirs == NULL)
1082 GOTO(cleanup, rc = -ENOMEM);
1085 for (i = 0; i < filter->fo_subdir_count; i++) {
1087 snprintf(dir, sizeof(dir), "d%u", i);
1089 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1092 if (IS_ERR(tmp_subdirs->dentry[i])) {
1093 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1094 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1099 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1100 tmp_subdirs->dentry[i]);
1104 /* 'group' is an index; we need an array of length 'group + 1' */
1105 if (group + 1 > old_count) {
1106 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1107 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1108 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1109 OBD_ALLOC(new_files, len * sizeof(*new_files));
1111 if (new_objids == NULL || new_subdirs == NULL ||
1112 new_groups == NULL || new_files == NULL)
1113 GOTO(cleanup, rc = -ENOMEM);
1116 memcpy(new_objids, filter->fo_last_objids,
1117 old_count * sizeof(*new_objids));
1118 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1119 old_count * sizeof(*new_subdirs));
1120 memcpy(new_groups, filter->fo_dentry_O_groups,
1121 old_count * sizeof(*new_groups));
1122 memcpy(new_files, filter->fo_last_objid_files,
1123 old_count * sizeof(*new_files));
1125 OBD_FREE(filter->fo_last_objids,
1126 old_count * sizeof(*new_objids));
1127 OBD_FREE(filter->fo_dentry_O_sub,
1128 old_count * sizeof(*new_subdirs));
1129 OBD_FREE(filter->fo_dentry_O_groups,
1130 old_count * sizeof(*new_groups));
1131 OBD_FREE(filter->fo_last_objid_files,
1132 old_count * sizeof(*new_files));
1134 filter->fo_last_objids = new_objids;
1135 filter->fo_dentry_O_sub = new_subdirs;
1136 filter->fo_dentry_O_groups = new_groups;
1137 filter->fo_last_objid_files = new_files;
1138 filter->fo_group_count = len;
1141 filter->fo_dentry_O_groups[group] = dentry;
1142 filter->fo_last_objid_files[group] = filp;
1143 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1144 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1145 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1148 filter_update_last_group(obd, group);
1150 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1151 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1152 rc = filter_update_last_objid(obd, group, 1);
1156 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1157 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1158 obd->obd_name, group, last_objid);
1163 if (new_objids != NULL)
1164 OBD_FREE(new_objids, len * sizeof(*new_objids));
1165 if (new_subdirs != NULL)
1166 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1167 if (new_groups != NULL)
1168 OBD_FREE(new_groups, len * sizeof(*new_groups));
1169 if (new_files != NULL)
1170 OBD_FREE(new_files, len * sizeof(*new_files));
1172 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1173 for (i = 0; i < filter->fo_subdir_count; i++) {
1174 if (tmp_subdirs->dentry[i] != NULL)
1175 dput(tmp_subdirs->dentry[i]);
1177 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1180 filp_close(filp, 0);
1187 static int filter_read_groups(struct obd_device *obd, int last_group,
1190 struct filter_obd *filter = &obd->u.filter;
1191 int old_count, group, rc = 0;
1193 down(&filter->fo_init_lock);
1194 old_count = filter->fo_group_count;
1195 for (group = old_count; group <= last_group; group++) {
1196 rc = filter_read_group_internal(obd, group, create);
1200 up(&filter->fo_init_lock);
1204 /* FIXME: object groups */
1205 static int filter_prep_groups(struct obd_device *obd)
1207 struct filter_obd *filter = &obd->u.filter;
1208 struct dentry *O_dentry;
1210 int last_group, rc = 0, cleanup_phase = 0;
1214 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1216 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1217 if (IS_ERR(O_dentry)) {
1218 rc = PTR_ERR(O_dentry);
1219 CERROR("cannot open/create O: rc = %d\n", rc);
1222 filter->fo_dentry_O = O_dentry;
1223 cleanup_phase = 1; /* O_dentry */
1225 /* we have to initialize all groups before first connections from
1226 * clients because they may send create/destroy for any group -bzzz */
1227 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1229 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1230 GOTO(cleanup, rc = PTR_ERR(filp));
1232 cleanup_phase = 2; /* filp */
1234 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1236 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1241 last_group = FILTER_GROUP_MDS0;
1243 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1244 FILTER_GROUP_MDS0, last_group);
1245 filter->fo_committed_group = last_group;
1246 rc = filter_read_groups(obd, last_group, 1);
1250 filp_close(filp, 0);
1254 switch (cleanup_phase) {
1256 filp_close(filp, 0);
1258 filter_cleanup_groups(obd);
1259 f_dput(filter->fo_dentry_O);
1260 filter->fo_dentry_O = NULL;
1268 /* setup the object store with correct subdirectories */
1269 static int filter_prep(struct obd_device *obd)
1271 struct lvfs_run_ctxt saved;
1272 struct filter_obd *filter = &obd->u.filter;
1274 struct inode *inode;
1278 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1279 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1280 if (!file || IS_ERR(file)) {
1282 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1286 filter->fo_rcvd_filp = file;
1287 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1288 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1289 file->f_dentry->d_inode->i_mode);
1290 GOTO(err_filp, rc = -ENOENT);
1293 inode = file->f_dentry->d_parent->d_inode;
1294 /* We use i_op->unlink directly in filter_vfs_unlink() */
1295 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1296 CERROR("%s: filesystem does not support create/unlink ops\n",
1298 GOTO(err_filp, rc = -EOPNOTSUPP);
1301 /** lu_target has very limited use in filter now */
1302 lut_init(NULL, &filter->fo_lut, obd, NULL);
1304 rc = filter_init_server_data(obd, file);
1306 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1310 target_recovery_init(&filter->fo_lut, ost_handle);
1312 /* open and create health check io file*/
1313 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1316 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1318 GOTO(err_server_data, rc);
1320 filter->fo_obt.obt_health_check_filp = file;
1321 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1322 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1323 file->f_dentry->d_inode->i_mode);
1324 GOTO(err_health_check, rc = -ENOENT);
1326 rc = lvfs_check_io_health(obd, file);
1328 GOTO(err_health_check, rc);
1330 rc = filter_prep_groups(obd);
1332 GOTO(err_health_check, rc);
1334 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1339 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1340 CERROR("can't close %s after error\n", HEALTH_CHECK);
1341 filter->fo_obt.obt_health_check_filp = NULL;
1343 target_recovery_fini(obd);
1344 filter_free_server_data(filter);
1346 if (filp_close(filter->fo_rcvd_filp, 0))
1347 CERROR("can't close %s after error\n", LAST_RCVD);
1348 filter->fo_rcvd_filp = NULL;
1352 /* cleanup the filter: write last used object id to status file */
1353 static void filter_post(struct obd_device *obd)
1355 struct lvfs_run_ctxt saved;
1356 struct filter_obd *filter = &obd->u.filter;
1359 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1360 * best to start a transaction with h_sync, because we removed this
1363 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1364 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1367 CERROR("error writing server data: rc = %d\n", rc);
1369 for (i = 0; i < filter->fo_group_count; i++) {
1370 rc = filter_update_last_objid(obd, i,
1371 (i == filter->fo_group_count - 1));
1373 CERROR("error writing group %d lastobjid: rc = %d\n",
1377 rc = filp_close(filter->fo_rcvd_filp, 0);
1378 filter->fo_rcvd_filp = NULL;
1380 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1382 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1383 filter->fo_obt.obt_health_check_filp = NULL;
1385 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1387 filter_cleanup_groups(obd);
1388 filter_free_server_data(filter);
1389 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1391 filter_free_capa_keys(filter);
1392 cleanup_capa_hash(filter->fo_capa_hash);
1395 static void filter_set_last_id(struct filter_obd *filter,
1396 obd_id id, obd_gr group)
1398 LASSERT(filter->fo_fsd != NULL);
1399 LASSERT(group <= filter->fo_group_count);
1401 spin_lock(&filter->fo_objidlock);
1402 filter->fo_last_objids[group] = id;
1403 spin_unlock(&filter->fo_objidlock);
1406 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1409 LASSERT(filter->fo_fsd != NULL);
1410 LASSERT(group <= filter->fo_group_count);
1412 /* FIXME: object groups */
1413 spin_lock(&filter->fo_objidlock);
1414 id = filter->fo_last_objids[group];
1415 spin_unlock(&filter->fo_objidlock);
1419 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1421 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1425 /* We never dget the object parent, so DON'T dput it either */
1426 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1428 struct filter_obd *filter = &obd->u.filter;
1429 struct filter_subdirs *subdirs;
1430 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1432 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1433 return filter->fo_dentry_O_groups[group];
1435 subdirs = &filter->fo_dentry_O_sub[group];
1436 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1439 /* We never dget the object parent, so DON'T dput it either */
1440 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1443 unsigned long now = jiffies;
1444 struct dentry *dparent = filter_parent(obd, group, objid);
1447 if (IS_ERR(dparent))
1449 if (dparent == NULL)
1450 return ERR_PTR(-ENOENT);
1452 rc = filter_lock_dentry(obd, dparent);
1453 fsfilt_check_slow(obd, now, "parent lock");
1454 return rc ? ERR_PTR(rc) : dparent;
1457 /* We never dget the object parent, so DON'T dput it either */
1458 static void filter_parent_unlock(struct dentry *dparent)
1460 UNLOCK_INODE_MUTEX(dparent->d_inode);
1463 /* How to get files, dentries, inodes from object id's.
1465 * If dir_dentry is passed, the caller has already locked the parent
1466 * appropriately for this operation (normally a write lock). If
1467 * dir_dentry is NULL, we do a read lock while we do the lookup to
1468 * avoid races with create/destroy and such changing the directory
1469 * internal to the filesystem code. */
1470 struct dentry *filter_fid2dentry(struct obd_device *obd,
1471 struct dentry *dir_dentry,
1472 obd_gr group, obd_id id)
1474 struct dentry *dparent = dir_dentry;
1475 struct dentry *dchild;
1480 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1481 obd->u.filter.fo_destroys_in_progress == 0) {
1482 /* don't fail lookups for orphan recovery, it causes
1483 * later LBUGs when objects still exist during precreate */
1484 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1485 RETURN(ERR_PTR(-ENOENT));
1488 CERROR("fatal: invalid object id 0\n");
1489 RETURN(ERR_PTR(-ESTALE));
1492 len = sprintf(name, LPU64, id);
1493 if (dir_dentry == NULL) {
1494 dparent = filter_parent_lock(obd, group, id);
1495 if (IS_ERR(dparent)) {
1496 CERROR("%s: error getting object "LPU64":"LPU64
1497 " parent: rc %ld\n", obd->obd_name,
1498 id, group, PTR_ERR(dparent));
1502 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1503 dparent->d_name.len, dparent->d_name.name, name);
1504 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1505 if (dir_dentry == NULL)
1506 filter_parent_unlock(dparent);
1507 if (IS_ERR(dchild)) {
1508 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1513 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1514 CERROR("%s: got bad object "LPU64" inode %lu\n",
1515 obd->obd_name, id, dchild->d_inode->i_ino);
1517 RETURN(ERR_PTR(-ENOENT));
1520 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1521 name, dchild, atomic_read(&dchild->d_count));
1523 LASSERT(atomic_read(&dchild->d_count) > 0);
1528 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1529 obd_id group, struct lustre_handle *lockh)
1531 int flags = LDLM_AST_DISCARD_DATA, rc;
1532 struct ldlm_res_id res_id;
1533 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1536 osc_build_res_name(objid, group, &res_id);
1537 /* Tell the clients that the object is gone now and that they should
1538 * throw away any cached pages. */
1539 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1540 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1541 ldlm_completion_ast, NULL, NULL, 0, NULL,
1548 static void filter_fini_destroy(struct obd_device *obd,
1549 struct lustre_handle *lockh)
1552 ldlm_lock_decref(lockh, LCK_PW);
1555 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1556 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1557 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1558 * reverse. Caller must take i_sem before starting the transaction and we
1559 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1560 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1561 struct vfsmount *mnt)
1566 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1567 * itself we already hold dir->i_mutex for child create/unlink ops */
1568 LASSERT(dentry->d_inode != NULL);
1569 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1570 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1574 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1575 GOTO(out, rc = -ENOENT);
1577 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1582 GOTO(out, rc = -EPERM);
1584 /* check_sticky() */
1585 if ((dentry->d_inode->i_uid != current->fsuid &&
1586 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1587 IS_IMMUTABLE(dentry->d_inode))
1588 GOTO(out, rc = -EPERM);
1590 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1591 * that was done because of journal_start (which is already done
1592 * here) or some other ordering issue. */
1595 rc = ll_security_inode_unlink(dir, dentry, mnt);
1599 rc = dir->i_op->unlink(dir, dentry);
1601 /* need to drop i_mutex before we lose inode reference */
1602 UNLOCK_INODE_MUTEX(dentry->d_inode);
1609 /* Caller must hold LCK_PW on parent and push us into kernel context.
1610 * Caller must hold child i_mutex, we drop it always.
1611 * Caller is also required to ensure that dchild->d_inode exists. */
1612 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1613 obd_gr group, struct dentry *dparent,
1614 struct dentry *dchild)
1616 struct inode *inode = dchild->d_inode;
1617 struct filter_obd *filter = &obd->u.filter;
1620 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1621 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1622 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1623 (unsigned long)inode->i_nlink,
1624 atomic_read(&inode->i_count));
1627 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1629 CERROR("error unlinking objid %.*s: rc %d\n",
1630 dchild->d_name.len, dchild->d_name.name, rc);
1634 struct filter_intent_args {
1635 struct ldlm_lock **victim;
1640 static enum interval_iter filter_intent_cb(struct interval_node *n,
1643 struct ldlm_interval *node = (struct ldlm_interval *)n;
1644 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1645 __u64 size = arg->size;
1646 struct ldlm_lock **v = arg->victim;
1647 struct ldlm_lock *lck;
1649 /* If the interval is lower than the current file size,
1651 if (interval_high(n) <= size)
1652 return INTERVAL_ITER_STOP;
1654 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1655 /* Don't send glimpse ASTs to liblustre clients.
1656 * They aren't listening for them, and they do
1657 * entirely synchronous I/O anyways. */
1658 if (lck->l_export == NULL ||
1659 lck->l_export->exp_libclient == 1)
1662 if (*arg->liblustre)
1663 *arg->liblustre = 0;
1666 *v = LDLM_LOCK_GET(lck);
1667 } else if ((*v)->l_policy_data.l_extent.start <
1668 lck->l_policy_data.l_extent.start) {
1669 LDLM_LOCK_RELEASE(*v);
1670 *v = LDLM_LOCK_GET(lck);
1673 /* the same policy group - every lock has the
1674 * same extent, so needn't do it any more */
1678 return INTERVAL_ITER_CONT;
1681 static int filter_intent_policy(struct ldlm_namespace *ns,
1682 struct ldlm_lock **lockp, void *req_cookie,
1683 ldlm_mode_t mode, int flags, void *data)
1685 CFS_LIST_HEAD(rpc_list);
1686 struct ptlrpc_request *req = req_cookie;
1687 struct ldlm_lock *lock = *lockp, *l = NULL;
1688 struct ldlm_resource *res = lock->l_resource;
1689 ldlm_processing_policy policy;
1690 struct ost_lvb *res_lvb, *reply_lvb;
1691 struct ldlm_reply *rep;
1693 int idx, rc, tmpflags = 0, only_liblustre = 1;
1694 struct ldlm_interval_tree *tree;
1695 struct filter_intent_args arg;
1696 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1697 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1698 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1701 policy = ldlm_get_processing_policy(res);
1702 LASSERT(policy != NULL);
1703 LASSERT(req != NULL);
1705 rc = lustre_pack_reply(req, 3, repsize, NULL);
1707 RETURN(req->rq_status = rc);
1709 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1710 LASSERT(rep != NULL);
1712 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1713 sizeof(*reply_lvb));
1714 LASSERT(reply_lvb != NULL);
1716 //fixup_handle_for_resent_req(req, lock, &lockh);
1718 /* Call the extent policy function to see if our request can be
1719 * granted, or is blocked.
1720 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1721 * lock, and should not be granted if the lock will be blocked.
1724 LASSERT(ns == res->lr_namespace);
1726 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1727 check_res_locked(res);
1729 /* FIXME: we should change the policy function slightly, to not make
1730 * this list at all, since we just turn around and free it */
1731 while (!list_empty(&rpc_list)) {
1732 struct ldlm_lock *wlock =
1733 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1734 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1735 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1736 lock->l_flags &= ~LDLM_FL_CP_REQD;
1737 list_del_init(&wlock->l_cp_ast);
1738 LDLM_LOCK_RELEASE(wlock);
1741 /* The lock met with no resistance; we're finished. */
1742 if (rc == LDLM_ITER_CONTINUE) {
1743 /* do not grant locks to the liblustre clients: they cannot
1744 * handle ASTs robustly. We need to do this while still
1745 * holding ns_lock to avoid the lock remaining on the res_link
1746 * list (and potentially being added to l_pending_list by an
1747 * AST) when we are going to drop this lock ASAP. */
1748 if (lock->l_export->exp_libclient ||
1749 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1750 ldlm_resource_unlink_lock(lock);
1751 err = ELDLM_LOCK_ABORTED;
1753 err = ELDLM_LOCK_REPLACED;
1759 /* Do not grant any lock, but instead send GL callbacks. The extent
1760 * policy nicely created a list of all PW locks for us. We will choose
1761 * the highest of those which are larger than the size in the LVB, if
1762 * any, and perform a glimpse callback. */
1763 res_lvb = res->lr_lvb_data;
1764 LASSERT(res_lvb != NULL);
1765 *reply_lvb = *res_lvb;
1768 * ->ns_lock guarantees that no new locks are granted, and,
1769 * therefore, that res->lr_lvb_data cannot increase beyond the
1770 * end of already granted lock. As a result, it is safe to
1771 * check against "stale" reply_lvb->lvb_size value without
1774 arg.size = reply_lvb->lvb_size;
1776 arg.liblustre = &only_liblustre;
1777 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1778 tree = &res->lr_itree[idx];
1779 if (tree->lit_mode == LCK_PR)
1782 interval_iterate_reverse(tree->lit_root,
1783 filter_intent_cb, &arg);
1787 /* There were no PW locks beyond the size in the LVB; finished. */
1789 if (only_liblustre) {
1790 /* If we discovered a liblustre client with a PW lock,
1791 * however, the LVB may be out of date! The LVB is
1792 * updated only on glimpse (which we don't do for
1793 * liblustre clients) and cancel (which the client
1794 * obviously has not yet done). So if it has written
1795 * data but kept the lock, the LVB is stale and needs
1796 * to be updated from disk.
1798 * Of course, this will all disappear when we switch to
1799 * taking liblustre locks on the OST. */
1800 ldlm_res_lvbo_update(res, NULL, 0, 1);
1802 RETURN(ELDLM_LOCK_ABORTED);
1806 * This check is for lock taken in filter_prepare_destroy() that does
1807 * not have l_glimpse_ast set. So the logic is: if there is a lock
1808 * with no l_glimpse_ast set, this object is being destroyed already.
1810 * Hence, if you are grabbing DLM locks on the server, always set
1811 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1813 if (l->l_glimpse_ast == NULL) {
1814 /* We are racing with unlink(); just return -ENOENT */
1815 rep->lock_policy_res1 = -ENOENT;
1819 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1820 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1821 /* Update the LVB from disk if the AST failed (this is a legal race) */
1823 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1824 * sending ast is not handled. This can result in lost client writes.
1827 ldlm_res_lvbo_update(res, NULL, 0, 1);
1830 *reply_lvb = *res_lvb;
1834 LDLM_LOCK_RELEASE(l);
1836 RETURN(ELDLM_LOCK_ABORTED);
1840 * per-obd_device iobuf pool.
1842 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1843 * allocations in write-path has to be minimized (see bug 5137).
1845 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1846 * OST threads (see ost_thread_{init,done}()).
1848 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1849 * at the OST layer there are only (potentially) multiple obd_device of type
1850 * unknown at the time of OST thread creation.
1852 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1853 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1854 * it's very own iobuf.
1858 * filter_kiobuf_pool_init()
1860 * filter_kiobuf_pool_done()
1862 * filter_iobuf_get()
1864 * operate on this array. They are "generic" in a sense that they don't depend
1865 * on actual type of iobuf's (the latter depending on Linux kernel version).
1869 * destroy pool created by filter_iobuf_pool_init
1871 static void filter_iobuf_pool_done(struct filter_obd *filter)
1873 struct filter_iobuf **pool;
1878 pool = filter->fo_iobuf_pool;
1880 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1881 if (pool[i] != NULL)
1882 filter_free_iobuf(pool[i]);
1884 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1885 filter->fo_iobuf_pool = NULL;
1890 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1892 struct filter_obd *filter = &obd->u.filter;
1893 struct sptlrpc_rule_set tmp_rset;
1896 sptlrpc_rule_set_init(&tmp_rset);
1897 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1899 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1904 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1906 write_lock(&filter->fo_sptlrpc_lock);
1907 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1908 filter->fo_sptlrpc_rset = tmp_rset;
1909 write_unlock(&filter->fo_sptlrpc_lock);
1915 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1917 static int filter_iobuf_pool_init(struct filter_obd *filter)
1924 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1926 if (filter->fo_iobuf_pool == NULL)
1929 filter->fo_iobuf_count = OSS_THREADS_MAX;
1934 /* Return iobuf allocated for @thread_id. We don't know in advance how
1935 * many threads there will be so we allocate a large empty array and only
1936 * fill in those slots that are actually in use.
1937 * If we haven't allocated a pool entry for this thread before, do so now. */
1938 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1940 int thread_id = (oti && oti->oti_thread) ?
1941 oti->oti_thread->t_id : -1;
1942 struct filter_iobuf *pool = NULL;
1943 struct filter_iobuf **pool_place = NULL;
1945 if (thread_id >= 0) {
1946 LASSERT(thread_id < filter->fo_iobuf_count);
1947 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1950 if (unlikely(pool == NULL)) {
1951 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1952 PTLRPC_MAX_BRW_PAGES);
1953 if (pool_place != NULL)
1960 /* mount the file system (secretly). lustre_cfg parameters are:
1963 * 3 = flags: failover=f, failout=n
1966 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1969 struct filter_obd *filter = &obd->u.filter;
1970 struct vfsmount *mnt;
1971 struct lustre_mount_info *lmi;
1972 struct obd_uuid uuid;
1980 if (lcfg->lcfg_bufcount < 3 ||
1981 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1982 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1985 lmi = server_get_mount(obd->obd_name);
1987 /* We already mounted in lustre_fill_super.
1988 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1989 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1991 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1993 /* old path - used by lctl */
1994 CERROR("Using old MDS mount method\n");
1995 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1996 MS_NOATIME|MS_NODIRATIME,
1997 lustre_cfg_string(lcfg, 1), option);
2000 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2001 lustre_cfg_string(lcfg, 1), rc);
2005 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2007 if (IS_ERR(obd->obd_fsops))
2008 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2010 rc = filter_iobuf_pool_init(filter);
2014 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2015 CERROR("%s: Underlying device is marked as read-only. "
2016 "Setup failed\n", obd->obd_name);
2017 GOTO(err_ops, rc = -EROFS);
2020 /* failover is the default */
2021 obd->obd_replayable = 1;
2023 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2024 str = lustre_cfg_string(lcfg, 3);
2025 if (strchr(str, 'n')) {
2026 CWARN("%s: recovery disabled\n", obd->obd_name);
2027 obd->obd_replayable = 0;
2031 filter->fo_vfsmnt = mnt;
2032 obd->u.obt.obt_sb = mnt->mnt_sb;
2033 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2034 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2036 fsfilt_setup(obd, obd->u.obt.obt_sb);
2038 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2039 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2040 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2041 obd->obd_lvfs_ctxt.fs = get_ds();
2042 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2044 init_mutex(&filter->fo_init_lock);
2045 filter->fo_committed_group = 0;
2046 filter->fo_destroys_in_progress = 0;
2047 for (i = 0; i < 32; i++)
2048 sema_init(&filter->fo_create_locks[i], 1);
2050 spin_lock_init(&filter->fo_translock);
2051 spin_lock_init(&filter->fo_objidlock);
2052 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2053 sema_init(&filter->fo_alloc_lock, 1);
2054 init_brw_stats(&filter->fo_filter_stats);
2055 filter->fo_read_cache = 1; /* enable read-only cache by default */
2056 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2057 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2058 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2059 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2061 rc = filter_prep(obd);
2065 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2066 spin_lock_init(&filter->fo_llog_list_lock);
2068 filter->fo_fl_oss_capa = 1;
2070 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2071 filter->fo_capa_hash = init_capa_hash();
2072 if (filter->fo_capa_hash == NULL)
2073 GOTO(err_post, rc = -ENOMEM);
2075 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2076 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2077 LDLM_NAMESPACE_GREEDY);
2078 if (obd->obd_namespace == NULL)
2079 GOTO(err_post, rc = -ENOMEM);
2080 obd->obd_namespace->ns_lvbp = obd;
2081 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2082 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2084 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2085 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2087 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2089 CERROR("failed to setup llogging subsystems\n");
2093 rwlock_init(&filter->fo_sptlrpc_lock);
2094 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2095 /* do this after llog being initialized */
2096 filter_adapt_sptlrpc_conf(obd, 1);
2098 rc = lquota_setup(filter_quota_interface_ref, obd);
2102 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2103 if (q->max_sectors < q->max_hw_sectors &&
2104 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2105 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2106 "for larger I/O requests: max_sectors = %u "
2107 "could be up to max_hw_sectors=%u\n",
2108 obd->obd_name, mnt->mnt_sb->s_id,
2109 q->max_sectors, q->max_hw_sectors);
2111 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2112 if (uuid_ptr != NULL) {
2113 class_uuid_unparse(uuid_ptr, &uuid);
2119 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2121 if (obd->obd_recovering) {
2122 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2123 "recovery for at least %d:%.02d, or until %d "
2124 "client%s reconnect%s.\n",
2125 obd->obd_name, lustre_cfg_string(lcfg, 1),
2126 label ?: "", label ? "/" : "", str,
2127 obd->obd_recovery_timeout / 60,
2128 obd->obd_recovery_timeout % 60,
2129 obd->obd_max_recoverable_clients,
2130 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2131 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2133 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2134 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2135 label ?: "", label ? "/" : "", str,
2136 obd->obd_replayable ? "enabled" : "disabled");
2144 fsfilt_put_ops(obd->obd_fsops);
2145 filter_iobuf_pool_done(filter);
2147 server_put_mount(obd->obd_name, mnt);
2148 obd->u.obt.obt_sb = 0;
2152 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2154 struct lprocfs_static_vars lvars;
2159 CLASSERT(offsetof(struct obd_device, u.obt) ==
2160 offsetof(struct obd_device, u.filter.fo_obt));
2162 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2165 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2166 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2169 addr = (unsigned long)cfs_page_address(page);
2170 clear_page((void *)addr);
2172 /* lprocfs must be setup before the filter so state can be safely added
2173 * to /proc incrementally as the filter is setup */
2174 lprocfs_filter_init_vars(&lvars);
2175 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2176 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2177 /* Init obdfilter private stats here */
2178 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2179 LPROCFS_CNTR_AVGMINMAX,
2180 "read_bytes", "bytes");
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "write_bytes", "bytes");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2185 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2186 "get_page", "usec");
2187 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2188 LPROCFS_CNTR_AVGMINMAX,
2189 "get_page_failures", "num");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2191 LPROCFS_CNTR_AVGMINMAX,
2192 "cache_access", "pages");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2194 LPROCFS_CNTR_AVGMINMAX,
2195 "cache_hit", "pages");
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2197 LPROCFS_CNTR_AVGMINMAX,
2198 "cache_miss", "pages");
2200 lproc_filter_attach_seqstat(obd);
2201 obd->obd_proc_exports_entry = lprocfs_register("exports",
2202 obd->obd_proc_entry,
2204 if (IS_ERR(obd->obd_proc_exports_entry)) {
2205 rc = PTR_ERR(obd->obd_proc_exports_entry);
2206 CERROR("error %d setting up lprocfs for %s\n",
2208 obd->obd_proc_exports_entry = NULL;
2211 if (obd->obd_proc_exports_entry)
2212 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2213 lprocfs_nid_stats_clear_read,
2214 lprocfs_nid_stats_clear_write, obd, NULL);
2216 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2217 LUSTRE_CFG_BUFLEN(lcfg, 4));
2218 rc = filter_common_setup(obd, lcfg, (void *)addr);
2219 OBD_PAGE_FREE(page);
2222 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2223 lprocfs_free_per_client_stats(obd);
2224 lprocfs_free_obd_stats(obd);
2225 lprocfs_obd_cleanup(obd);
2231 static struct llog_operations filter_mds_ost_repl_logops;
2233 static struct llog_operations filter_size_orig_logops = {
2234 .lop_setup = llog_obd_origin_setup,
2235 .lop_cleanup = llog_obd_origin_cleanup,
2236 .lop_add = llog_obd_origin_add
2239 static int filter_olg_fini(struct obd_llog_group *olg)
2241 struct llog_ctxt *ctxt;
2242 int rc = 0, rc2 = 0;
2245 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2247 rc = llog_cleanup(ctxt);
2249 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2251 rc2 = llog_cleanup(ctxt);
2256 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2258 rc2 = llog_cleanup(ctxt);
2267 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2268 struct obd_device *tgt)
2273 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2274 &filter_mds_ost_repl_logops);
2278 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2279 &filter_size_orig_logops);
2285 filter_olg_fini(olg);
2290 * Init the default olg, which is embeded in the obd_device, for filter.
2293 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2294 struct obd_device *tgt)
2296 struct filter_obd *filter = &obd->u.filter;
2297 struct llog_ctxt *ctxt;
2301 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2302 if (!filter->fo_lcm)
2305 filter_mds_ost_repl_logops = llog_client_ops;
2306 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2307 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2308 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2310 rc = filter_olg_init(obd, olg, tgt);
2312 GOTO(cleanup_lcm, rc);
2314 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2317 GOTO(cleanup_olg, rc);
2319 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2321 CERROR("Can't get ctxt for %p:%x\n", olg,
2322 LLOG_MDS_OST_REPL_CTXT);
2323 GOTO(cleanup_olg, rc = -ENODEV);
2325 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2326 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2327 llog_ctxt_put(ctxt);
2331 filter_olg_fini(olg);
2333 llog_recov_thread_fini(filter->fo_lcm, 1);
2334 filter->fo_lcm = NULL;
2339 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2340 struct obd_device *tgt, int *index)
2342 struct filter_obd *filter = &obd->u.filter;
2343 struct llog_ctxt *ctxt;
2347 LASSERT(olg != NULL);
2348 if (olg == &obd->obd_olg)
2349 return filter_default_olg_init(obd, olg, tgt);
2351 LASSERT(filter->fo_lcm != NULL);
2352 rc = filter_olg_init(obd, olg, tgt);
2355 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2357 CERROR("Can't get ctxt for %p:%x\n", olg,
2358 LLOG_MDS_OST_REPL_CTXT);
2359 filter_olg_fini(olg);
2362 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2363 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2364 llog_ctxt_put(ctxt);
2368 static int filter_llog_finish(struct obd_device *obd, int count)
2370 struct filter_obd *filter = &obd->u.filter;
2371 struct llog_ctxt *ctxt;
2374 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2377 * Make sure that no cached llcds left in recov_thread.
2378 * We actually do sync in disconnect time, but disconnect
2379 * may not come being marked rq_no_resend = 1.
2381 llog_sync(ctxt, NULL);
2384 * Balance class_import_get() in llog_receptor_accept().
2385 * This is safe to do, as llog is already synchronized
2386 * and its import may go.
2388 mutex_down(&ctxt->loc_sem);
2389 if (ctxt->loc_imp) {
2390 class_import_put(ctxt->loc_imp);
2391 ctxt->loc_imp = NULL;
2393 mutex_up(&ctxt->loc_sem);
2394 llog_ctxt_put(ctxt);
2397 if (filter->fo_lcm) {
2398 mutex_down(&ctxt->loc_sem);
2399 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2400 filter->fo_lcm = NULL;
2401 mutex_up(&ctxt->loc_sem);
2403 RETURN(filter_olg_fini(&obd->obd_olg));
2407 * Find the group llog according to group index in the llog group list.
2409 static struct obd_llog_group *
2410 filter_find_olg_internal(struct filter_obd *filter, int group)
2412 struct obd_llog_group *olg;
2414 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2415 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2416 if (olg->olg_group == group)
2423 * Find the group llog according to group index on the filter
2425 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2427 struct obd_llog_group *olg = NULL;
2428 struct filter_obd *filter;
2430 filter = &obd->u.filter;
2432 if (group == FILTER_GROUP_LLOG)
2433 RETURN(&obd->obd_olg);
2435 spin_lock(&filter->fo_llog_list_lock);
2436 olg = filter_find_olg_internal(filter, group);
2437 spin_unlock(&filter->fo_llog_list_lock);
2442 * Find the llog_group of the filter according to the group. If it can not
2443 * find, create the llog_group, which only happens when mds is being synced
2446 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2448 struct obd_llog_group *olg = NULL;
2449 struct filter_obd *filter;
2452 filter = &obd->u.filter;
2454 if (group == FILTER_GROUP_LLOG)
2455 RETURN(&obd->obd_olg);
2457 spin_lock(&filter->fo_llog_list_lock);
2458 olg = filter_find_olg_internal(filter, group);
2460 if (olg->olg_initializing) {
2461 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2463 GOTO(out_unlock, olg);
2468 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2470 llog_group_init(olg, group);
2471 list_add(&olg->olg_list, &filter->fo_llog_list);
2472 olg->olg_initializing = 1;
2473 spin_unlock(&filter->fo_llog_list_lock);
2475 rc = obd_llog_init(obd, olg, obd, NULL);
2477 spin_lock(&filter->fo_llog_list_lock);
2478 list_del(&olg->olg_list);
2479 spin_unlock(&filter->fo_llog_list_lock);
2481 GOTO(out, olg = ERR_PTR(-ENOMEM));
2483 spin_lock(&filter->fo_llog_list_lock);
2484 olg->olg_initializing = 0;
2485 spin_unlock(&filter->fo_llog_list_lock);
2486 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2487 obd->obd_name, group, olg);
2492 spin_unlock(&filter->fo_llog_list_lock);
2496 static int filter_llog_connect(struct obd_export *exp,
2497 struct llogd_conn_body *body)
2499 struct obd_device *obd = exp->exp_obd;
2500 struct llog_ctxt *ctxt;
2501 struct obd_llog_group *olg;
2505 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2506 obd->obd_name, body->lgdc_logid.lgl_oid,
2507 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2509 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2511 CERROR(" %s: can not find olg of group %d\n",
2512 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2515 llog_group_set_export(olg, exp);
2517 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2518 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2519 body->lgdc_ctxt_idx);
2521 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2522 obd->obd_name, body->lgdc_logid.lgl_oid,
2523 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2525 spin_lock_bh(&obd->obd_processing_task_lock);
2526 obd->u.filter.fo_mds_ost_sync = 1;
2527 spin_unlock_bh(&obd->obd_processing_task_lock);
2528 rc = llog_connect(ctxt, &body->lgdc_logid,
2529 &body->lgdc_gen, NULL);
2530 llog_ctxt_put(ctxt);
2532 CERROR("failed to connect rc %d idx %d\n", rc,
2533 body->lgdc_ctxt_idx);
2538 static int filter_llog_preclean(struct obd_device *obd)
2540 struct obd_llog_group *olg, *tmp;
2541 struct filter_obd *filter;
2542 struct list_head remove_list;
2546 rc = obd_llog_finish(obd, 0);
2548 CERROR("failed to cleanup llogging subsystem\n");
2550 filter = &obd->u.filter;
2551 CFS_INIT_LIST_HEAD(&remove_list);
2553 spin_lock(&filter->fo_llog_list_lock);
2554 while (!list_empty(&filter->fo_llog_list)) {
2555 olg = list_entry(filter->fo_llog_list.next,
2556 struct obd_llog_group, olg_list);
2557 list_del(&olg->olg_list);
2558 list_add(&olg->olg_list, &remove_list);
2560 spin_unlock(&filter->fo_llog_list_lock);
2562 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2563 list_del_init(&olg->olg_list);
2564 rc = filter_olg_fini(olg);
2566 CERROR("failed to cleanup llogging subsystem for %u\n",
2574 static int filter_precleanup(struct obd_device *obd,
2575 enum obd_cleanup_stage stage)
2581 case OBD_CLEANUP_EARLY:
2583 case OBD_CLEANUP_EXPORTS:
2584 /* Stop recovery before namespace cleanup. */
2585 target_recovery_fini(obd);
2586 rc = filter_llog_preclean(obd);
2592 static int filter_cleanup(struct obd_device *obd)
2594 struct filter_obd *filter = &obd->u.filter;
2598 LCONSOLE_WARN("%s: shutting down for failover; client state "
2599 "will be preserved.\n", obd->obd_name);
2601 obd_exports_barrier(obd);
2602 obd_zombie_barrier();
2604 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2605 lprocfs_free_per_client_stats(obd);
2606 lprocfs_free_obd_stats(obd);
2607 lprocfs_obd_cleanup(obd);
2608 lquota_cleanup(filter_quota_interface_ref, obd);
2610 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2611 obd->obd_namespace = NULL;
2613 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2615 if (obd->u.obt.obt_sb == NULL)
2620 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2621 shrink_dcache_sb(obd->u.obt.obt_sb);
2623 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2624 obd->u.obt.obt_sb = NULL;
2626 fsfilt_put_ops(obd->obd_fsops);
2628 filter_iobuf_pool_done(filter);
2630 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2635 static int filter_connect_internal(struct obd_export *exp,
2636 struct obd_connect_data *data)
2638 struct filter_export_data *fed = &exp->exp_filter_data;
2643 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2644 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2645 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2646 data->ocd_connect_flags, data->ocd_version,
2647 data->ocd_grant, data->ocd_index);
2649 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2650 CWARN("!!! This export (nid %s) used object group %d "
2651 "earlier; now it's trying to use group %d! This could "
2652 "be a bug in the MDS. Please report to "
2653 "http://bugzilla.lustre.org/\n",
2654 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2657 fed->fed_group = data->ocd_group;
2659 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2660 exp->exp_connect_flags = data->ocd_connect_flags;
2661 data->ocd_version = LUSTRE_VERSION_CODE;
2663 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2664 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2665 data->ocd_connect_flags & OBD_CONNECT_MDS))
2668 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2669 struct filter_obd *filter = &exp->exp_obd->u.filter;
2670 obd_size left, want;
2672 spin_lock(&exp->exp_obd->obd_osfs_lock);
2673 left = filter_grant_space_left(exp);
2674 want = data->ocd_grant;
2675 filter_grant(exp, fed->fed_grant, want, left);
2676 data->ocd_grant = fed->fed_grant;
2677 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2679 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2680 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2681 exp->exp_client_uuid.uuid, exp,
2682 data->ocd_grant, want, left);
2684 filter->fo_tot_granted_clients ++;
2687 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2688 struct filter_obd *filter = &exp->exp_obd->u.filter;
2689 struct lr_server_data *lsd = filter->fo_fsd;
2690 int index = le32_to_cpu(lsd->lsd_ost_index);
2692 if (!(lsd->lsd_feature_compat &
2693 cpu_to_le32(OBD_COMPAT_OST))) {
2694 /* this will only happen on the first connect */
2695 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2696 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2697 filter_update_server_data(exp->exp_obd,
2698 filter->fo_rcvd_filp, lsd, 1);
2699 } else if (index != data->ocd_index) {
2700 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2701 " %u doesn't match actual OST index"
2702 " %u in last_rcvd file, bad "
2704 obd_export_nid2str(exp), index,
2708 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2709 * FIXME: We don't strictly need the COMPAT flag for that,
2710 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2711 * FIXME: We needed it for the index, as index 0 is valid. */
2714 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2715 data->ocd_brw_size = 65536;
2716 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2717 data->ocd_brw_size = min(data->ocd_brw_size,
2718 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2719 LASSERT(data->ocd_brw_size);
2722 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2723 __u32 cksum_types = data->ocd_cksum_types;
2725 /* The client set in ocd_cksum_types the checksum types it
2726 * supports. We have to mask off the algorithms that we don't
2728 if (cksum_types & OBD_CKSUM_ALL)
2729 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2731 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2733 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2734 "%x\n", exp->exp_obd->obd_name,
2735 obd_export_nid2str(exp), cksum_types,
2736 data->ocd_cksum_types);
2738 /* This client does not support OBD_CONNECT_CKSUM
2739 * fall back to CRC32 */
2740 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2741 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2742 exp->exp_obd->obd_name,
2743 obd_export_nid2str(exp));
2749 static int filter_reconnect(const struct lu_env *env,
2750 struct obd_export *exp, struct obd_device *obd,
2751 struct obd_uuid *cluuid,
2752 struct obd_connect_data *data,
2758 if (exp == NULL || obd == NULL || cluuid == NULL)
2761 rc = filter_connect_internal(exp, data);
2763 filter_export_stats_init(obd, exp, localdata);
2768 /* nearly identical to mds_connect */
2769 static int filter_connect(const struct lu_env *env,
2770 struct obd_export **exp, struct obd_device *obd,
2771 struct obd_uuid *cluuid,
2772 struct obd_connect_data *data, void *localdata)
2774 struct lvfs_run_ctxt saved;
2775 struct lustre_handle conn = { 0 };
2776 struct obd_export *lexp;
2777 struct filter_export_data *fed;
2778 struct lsd_client_data *lcd = NULL;
2783 if (exp == NULL || obd == NULL || cluuid == NULL)
2786 rc = class_connect(&conn, obd, cluuid);
2789 lexp = class_conn2export(&conn);
2790 LASSERT(lexp != NULL);
2792 fed = &lexp->exp_filter_data;
2794 rc = filter_connect_internal(lexp, data);
2798 filter_export_stats_init(obd, lexp, localdata);
2799 if (obd->obd_replayable) {
2800 OBD_ALLOC(lcd, sizeof(*lcd));
2802 CERROR("filter: out of memory for client data\n");
2803 GOTO(cleanup, rc = -ENOMEM);
2806 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2808 rc = filter_client_add(obd, lexp, -1);
2813 group = data->ocd_group;
2815 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2816 obd->obd_name, lexp->exp_handle.h_cookie, group);
2818 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2819 rc = filter_read_groups(obd, group, 1);
2820 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2822 CERROR("can't read group %u\n", group);
2832 fed->fed_lcd = NULL;
2834 class_disconnect(lexp);
2843 /* Do extra sanity checks for grant accounting. We do this at connect,
2844 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2845 * always get rid of it or turn it off when we know accounting is good. */
2846 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2848 struct filter_export_data *fed;
2849 struct obd_export *exp;
2850 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2851 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2852 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2854 if (list_empty(&obd->obd_exports))
2857 /* We don't want to do this for large machines that do lots of
2858 mounts or unmounts. It burns... */
2859 if (obd->obd_num_exports > 100)
2862 spin_lock(&obd->obd_osfs_lock);
2863 spin_lock(&obd->obd_dev_lock);
2864 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2866 fed = &exp->exp_filter_data;
2867 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2870 if (maxsize > 0) { /* we may not have done a statfs yet */
2871 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2872 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2873 exp->exp_client_uuid.uuid, exp,
2874 fed->fed_grant, fed->fed_pending, maxsize);
2875 LASSERTF(fed->fed_dirty <= maxsize,
2876 "%s: cli %s/%p %ld > "LPU64"\n", func,
2877 exp->exp_client_uuid.uuid, exp,
2878 fed->fed_dirty, maxsize);
2881 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2882 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2883 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2885 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2886 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2887 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2888 tot_granted += fed->fed_grant + fed->fed_pending;
2889 tot_pending += fed->fed_pending;
2890 tot_dirty += fed->fed_dirty;
2892 fo_tot_granted = obd->u.filter.fo_tot_granted;
2893 fo_tot_pending = obd->u.filter.fo_tot_pending;
2894 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2895 spin_unlock(&obd->obd_dev_lock);
2896 spin_unlock(&obd->obd_osfs_lock);
2898 /* Do these assertions outside the spinlocks so we don't kill system */
2899 if (tot_granted != fo_tot_granted)
2900 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2901 func, tot_granted, fo_tot_granted);
2902 if (tot_pending != fo_tot_pending)
2903 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2904 func, tot_pending, fo_tot_pending);
2905 if (tot_dirty != fo_tot_dirty)
2906 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2907 func, tot_dirty, fo_tot_dirty);
2908 if (tot_pending > tot_granted)
2909 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2910 func, tot_pending, tot_granted);
2911 if (tot_granted > maxsize)
2912 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2913 func, tot_granted, maxsize);
2914 if (tot_dirty > maxsize)
2915 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2916 func, tot_dirty, maxsize);
2919 /* Remove this client from the grant accounting totals. We also remove
2920 * the export from the obd device under the osfs and dev locks to ensure
2921 * that the filter_grant_sanity_check() calculations are always valid.
2922 * The client should do something similar when it invalidates its import. */
2923 static void filter_grant_discard(struct obd_export *exp)
2925 struct obd_device *obd = exp->exp_obd;
2926 struct filter_obd *filter = &obd->u.filter;
2927 struct filter_export_data *fed = &exp->exp_filter_data;
2929 spin_lock(&obd->obd_osfs_lock);
2930 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2931 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2932 obd->obd_name, filter->fo_tot_granted,
2933 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2934 filter->fo_tot_granted -= fed->fed_grant;
2935 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2936 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2937 obd->obd_name, filter->fo_tot_pending,
2938 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2939 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2940 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2941 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2942 obd->obd_name, filter->fo_tot_dirty,
2943 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2944 filter->fo_tot_dirty -= fed->fed_dirty;
2948 spin_unlock(&obd->obd_osfs_lock);
2951 static int filter_destroy_export(struct obd_export *exp)
2955 if (exp->exp_filter_data.fed_pending)
2956 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2957 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2958 exp, exp->exp_filter_data.fed_pending);
2960 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2962 target_destroy_export(exp);
2963 ldlm_destroy_export(exp);
2965 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2969 if (exp->exp_obd->obd_replayable)
2970 filter_client_free(exp);
2972 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2974 filter_grant_discard(exp);
2975 filter_fmd_cleanup(exp);
2977 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2978 struct filter_obd *filter = &exp->exp_obd->u.filter;
2979 if (filter->fo_tot_granted_clients > 0)
2980 filter->fo_tot_granted_clients --;
2983 if (!(exp->exp_flags & OBD_OPT_FORCE))
2984 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2989 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2991 struct obd_llog_group *olg_min, *olg;
2992 struct filter_obd *filter;
2993 int worked = -1, group;
2994 struct llog_ctxt *ctxt;
2997 filter = &obd->u.filter;
2999 /* we can't sync log holding spinlock. also, we do not want to get
3000 * into livelock. so we do following: loop over MDS's exports in
3001 * group order and skip already synced llogs -bzzz */
3003 /* look for group with min. number, but > worked */
3006 spin_lock(&filter->fo_llog_list_lock);
3007 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3008 if (olg->olg_group <= worked) {
3009 /* this group is already synced */
3012 if (group < olg->olg_group) {
3013 /* we have group with smaller number to sync */
3016 /* store current minimal group */
3018 group = olg->olg_group;
3020 spin_unlock(&filter->fo_llog_list_lock);
3022 if (olg_min == NULL)
3025 worked = olg_min->olg_group;
3026 if (olg_min->olg_exp &&
3027 (dexp == olg_min->olg_exp || dexp == NULL)) {
3029 ctxt = llog_group_get_ctxt(olg_min,
3030 LLOG_MDS_OST_REPL_CTXT);
3032 err = llog_sync(ctxt, olg_min->olg_exp);
3033 llog_ctxt_put(ctxt);
3035 CERROR("error flushing logs to MDS: "
3040 } while (olg_min != NULL);
3043 /* Also incredibly similar to mds_disconnect */
3044 static int filter_disconnect(struct obd_export *exp)
3046 struct obd_device *obd = exp->exp_obd;
3051 class_export_get(exp);
3053 if (!(exp->exp_flags & OBD_OPT_FORCE))
3054 filter_grant_sanity_check(obd, __FUNCTION__);
3055 filter_grant_discard(exp);
3057 /* Flush any remaining cancel messages out to the target */
3058 filter_sync_llogs(obd, exp);
3060 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3062 /* Disconnect early so that clients can't keep using export */
3063 rc = class_disconnect(exp);
3064 if (exp->exp_obd->obd_namespace != NULL)
3065 ldlm_cancel_locks_for_export(exp);
3067 fsfilt_sync(obd, obd->u.obt.obt_sb);
3069 lprocfs_exp_cleanup(exp);
3071 class_export_put(exp);
3075 /* reverse import is changed, sync all cancels */
3076 static void filter_revimp_update(struct obd_export *exp)
3081 class_export_get(exp);
3083 /* flush any remaining cancel messages out to the target */
3084 filter_sync_llogs(exp->exp_obd, exp);
3085 class_export_put(exp);
3089 static int filter_ping(struct obd_export *exp)
3091 filter_fmd_expire(exp);
3095 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3096 const char *what, int quiet)
3098 struct dentry *dchild = NULL;
3101 if (oa->o_valid & OBD_MD_FLGROUP)
3104 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3106 if (IS_ERR(dchild)) {
3107 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3108 what, group, oa->o_id);
3112 if (dchild->d_inode == NULL) {
3114 CERROR("%s: %s on non-existent object: "LPU64"\n",
3115 obd->obd_name, what, oa->o_id);
3117 RETURN(ERR_PTR(-ENOENT));
3123 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3125 struct dentry *dentry = NULL;
3126 struct obd_device *obd;
3130 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3131 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3132 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3136 obd = class_exp2obd(exp);
3138 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3142 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3144 RETURN(PTR_ERR(dentry));
3146 /* Limit the valid bits in the return data to what we actually use */
3147 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3148 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3154 /* this should be enabled/disabled in condition to enabled/disabled large
3155 * inodes (fast EAs) in backing store FS. */
3156 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3157 void *handle, struct obdo *oa)
3159 struct obd_device *obd = exp->exp_obd;
3163 if (oa->o_valid & OBD_MD_FLFID) {
3164 struct filter_fid ff;
3166 if (!(oa->o_valid & OBD_MD_FLGROUP))
3168 /* packing fid and converting it to LE for storing into EA.
3169 * Here ->o_stripe_idx should be filled by LOV and rest of
3170 * fields - by client. */
3171 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3172 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3173 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3174 ff.ff_objid = cpu_to_le64(oa->o_id);
3175 ff.ff_group = cpu_to_le64(oa->o_gr);
3177 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3178 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3179 oa->o_generation, oa->o_id, oa->o_gr);
3181 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3183 CERROR("store fid in object failed! rc: %d\n", rc);
3185 CDEBUG(D_HA, "OSS object without fid info!\n");
3191 /* this is called from filter_truncate() until we have filter_punch() */
3192 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3193 struct obdo *oa, struct obd_trans_info *oti)
3195 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3196 struct llog_cookie *fcc = NULL;
3197 struct filter_obd *filter;
3198 int rc, err, sync = 0;
3199 loff_t old_size = 0;
3200 unsigned int ia_valid;
3201 struct inode *inode;
3202 struct page *page = NULL;
3207 LASSERT(dentry != NULL);
3208 LASSERT(!IS_ERR(dentry));
3210 inode = dentry->d_inode;
3211 LASSERT(inode != NULL);
3213 filter = &exp->exp_obd->u.filter;
3214 iattr_from_obdo(&iattr, oa, oa->o_valid);
3215 ia_valid = iattr.ia_valid;
3217 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3218 OBD_ALLOC(fcc, sizeof(*fcc));
3220 *fcc = oa->o_lcookie;
3222 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3224 /* Filter truncates and writes are serialized by
3225 * i_alloc_sem, see the comment in
3226 * filter_preprw_write.*/
3227 if (ia_valid & ATTR_SIZE)
3228 down_write(&inode->i_alloc_sem);
3229 LOCK_INODE_MUTEX(inode);
3230 old_size = i_size_read(inode);
3233 /* VBR: version recovery check */
3234 rc = filter_version_get_check(exp, oti, inode);
3236 GOTO(out_unlock, rc);
3238 /* Let's pin the last page so that ldiskfs_truncate
3239 * should not start GFP_FS allocation. */
3240 if (ia_valid & ATTR_SIZE) {
3241 page = grab_cache_page(inode->i_mapping,
3242 iattr.ia_size >> PAGE_CACHE_SHIFT);
3244 GOTO(out_unlock, rc = -ENOMEM);
3249 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3250 * then we will accept the UID+GID sent by the client during write for
3251 * initializing the ownership of this inode. We only allow this to
3252 * happen once so clear these bits in setattr. In 2.6 kernels it is
3253 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3254 * the flags that are actually being set. */
3255 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3256 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3257 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3259 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3260 if (!(ia_valid & ATTR_MODE)) {
3261 iattr.ia_mode = inode->i_mode;
3262 iattr.ia_valid |= ATTR_MODE;
3264 iattr.ia_mode &= ~S_ISUID;
3266 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3267 if (!(iattr.ia_valid & ATTR_MODE)) {
3268 iattr.ia_mode = inode->i_mode;
3269 iattr.ia_valid |= ATTR_MODE;
3271 iattr.ia_mode &= ~S_ISGID;
3274 orig_ids[USRQUOTA] = inode->i_uid;
3275 orig_ids[GRPQUOTA] = inode->i_gid;
3276 handle = fsfilt_start_log(exp->exp_obd, inode,
3277 FSFILT_OP_SETATTR, oti, 1);
3279 GOTO(out_unlock, rc = PTR_ERR(handle));
3281 /* update inode EA only once when inode is suid bit marked. As
3282 * on 2.6.x UID and GID may be set separately, we check here
3283 * only one of them to avoid double setting. */
3284 if (inode->i_mode & S_ISUID)
3285 filter_update_fidea(exp, inode, handle, oa);
3287 handle = fsfilt_start(exp->exp_obd, inode,
3288 FSFILT_OP_SETATTR, oti);
3290 GOTO(out_unlock, rc = PTR_ERR(handle));
3292 if (oa->o_valid & OBD_MD_FLFLAGS) {
3293 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3294 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3296 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3298 /* set cancel cookie callback function */
3299 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3300 filter_cancel_cookies_cb,
3304 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3305 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3307 /* The truncate might have used up our transaction credits. Make
3308 * sure we have one left for the last_rcvd update. */
3309 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3311 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3313 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3317 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3319 CERROR("error on commit, err = %d\n", err);
3326 /* For a partial-page truncate flush the page to disk immediately
3327 * to avoid data corruption during direct disk write. b=17397 */
3328 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3329 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3330 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3340 page_cache_release(page);
3342 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3343 UNLOCK_INODE_MUTEX(inode);
3344 if (ia_valid & ATTR_SIZE)
3345 up_write(&inode->i_alloc_sem);
3347 OBD_FREE(fcc, sizeof(*fcc));
3349 /* trigger quota release */
3350 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3351 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3352 int rc2 = lquota_adjust(filter_quota_interface_ref,
3353 exp->exp_obd, cur_ids,
3354 orig_ids, rc, FSFILT_OP_SETATTR);
3355 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3356 "filter adjust qunit. (rc:%d)\n", rc2);
3361 /* this is called from filter_truncate() until we have filter_punch() */
3362 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3363 struct obd_trans_info *oti)
3365 struct obdo *oa = oinfo->oi_oa;
3366 struct lustre_capa *capa = oinfo_capa(oinfo);
3367 struct ldlm_res_id res_id;
3368 struct filter_mod_data *fmd;
3369 struct lvfs_run_ctxt saved;
3370 struct filter_obd *filter;
3371 struct ldlm_resource *res;
3372 struct dentry *dentry;
3373 __u64 opc = CAPA_OPC_META_WRITE;
3377 if (oa->o_valid & OBD_FL_TRUNC)
3378 opc |= CAPA_OPC_OSS_TRUNC;
3380 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3381 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3385 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3386 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3391 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3392 /* This would be very bad - accidentally truncating a file when
3393 * changing the time or similar - bug 12203. */
3394 if (oa->o_valid & OBD_MD_FLSIZE &&
3395 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3396 static char mdsinum[48];
3398 if (oa->o_valid & OBD_MD_FLFID)
3399 snprintf(mdsinum, sizeof(mdsinum) - 1,
3400 " of inode "LPU64"/%u", oa->o_fid,
3405 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3407 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3412 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3414 RETURN(PTR_ERR(dentry));
3416 filter = &exp->exp_obd->u.filter;
3417 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3421 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3422 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3423 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3424 fmd->fmd_mactime_xid = oti->oti_xid;
3425 filter_fmd_put(exp, fmd);
3428 /* setting objects attributes (including owner/group) */
3429 rc = filter_setattr_internal(exp, dentry, oa, oti);
3431 GOTO(out_unlock, rc);
3433 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3434 &res_id, LDLM_EXTENT, 0);
3437 LDLM_RESOURCE_ADDREF(res);
3438 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3439 LDLM_RESOURCE_DELREF(res);
3440 ldlm_resource_putref(res);
3443 oa->o_valid = OBD_MD_FLID;
3445 /* Quota release need uid/gid info */
3446 obdo_from_inode(oa, dentry->d_inode,
3447 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3453 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3457 /* XXX identical to osc_unpackmd */
3458 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3459 struct lov_mds_md *lmm, int lmm_bytes)
3465 if (lmm_bytes < sizeof (*lmm)) {
3466 CERROR("lov_mds_md too small: %d, need %d\n",
3467 lmm_bytes, (int)sizeof(*lmm));
3470 /* XXX LOV_MAGIC etc check? */
3472 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3473 CERROR("lov_mds_md: zero lmm_object_id\n");
3478 lsm_size = lov_stripe_md_size(1);
3482 if (*lsmp != NULL && lmm == NULL) {
3483 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3484 OBD_FREE(*lsmp, lsm_size);
3489 if (*lsmp == NULL) {
3490 OBD_ALLOC(*lsmp, lsm_size);
3494 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3495 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3496 OBD_FREE(*lsmp, lsm_size);
3499 loi_init((*lsmp)->lsm_oinfo[0]);
3503 /* XXX zero *lsmp? */
3504 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3505 LASSERT((*lsmp)->lsm_object_id);
3508 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3513 /* caller must hold fo_create_locks[oa->o_gr] */
3514 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3515 struct filter_obd *filter)
3517 struct obdo doa = { 0 }; /* XXX obdo on stack */
3524 LASSERT_MDS_GROUP(oa->o_gr);
3525 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3526 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3528 memset(&doa, 0, sizeof(doa));
3530 doa.o_valid |= OBD_MD_FLGROUP;
3531 doa.o_gr = oa->o_gr;
3532 doa.o_mode = S_IFREG;
3534 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3535 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3536 exp->exp_obd->obd_name, doa.o_gr);
3540 last = filter_last_id(filter, doa.o_gr);
3542 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3544 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3545 exp->exp_obd->obd_name, oa->o_id + 1, last,
3546 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3548 for (id = last; id > oa->o_id; id--) {
3550 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3551 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3552 CEMERG("error destroying precreate objid "LPU64": %d\n",
3555 /* update last_id on disk periodically so that if we restart
3556 * we don't need to re-scan all of the just-deleted objects. */
3557 if ((id & 511) == 0 && !skip_orphan) {
3558 filter_set_last_id(filter, id - 1, doa.o_gr);
3559 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3563 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3564 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3567 filter_set_last_id(filter, id, doa.o_gr);
3568 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3570 /* don't reuse orphan object, return last used objid */
3574 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3579 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3580 obd_gr group, int *num);
3581 /* returns a negative error or a nonnegative number of files to create */
3582 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3583 obd_gr group, struct obd_trans_info *oti)
3585 struct obd_device *obd = exp->exp_obd;
3586 struct filter_obd *filter = &obd->u.filter;
3590 /* delete orphans request */
3591 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3592 obd_id last = filter_last_id(filter, group);
3594 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3595 CERROR("%s: dropping old orphan cleanup request\n",
3599 /* This causes inflight precreates to abort and drop lock */
3600 set_bit(group, &filter->fo_destroys_in_progress);
3601 down(&filter->fo_create_locks[group]);
3602 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3603 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3604 exp->exp_obd->obd_name, group);
3605 up(&filter->fo_create_locks[group]);
3608 diff = oa->o_id - last;
3609 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3612 if (-diff > OST_MAX_PRECREATE) {
3613 CERROR("%s: ignoring bogus orphan destroy request: "
3614 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3616 /* FIXME: should reset precreate_next_id on MDS */
3617 GOTO(out, rc = -EINVAL);
3620 rc = filter_destroy_precreated(exp, oa, filter);
3622 CERROR("%s: unable to write lastobjid, but "
3623 "orphans were deleted\n", obd->obd_name);
3626 /* XXX: Used by MDS for the first time! */
3627 clear_bit(group, &filter->fo_destroys_in_progress);
3630 down(&filter->fo_create_locks[group]);
3631 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3632 CERROR("%s: dropping old precreate request\n",
3636 /* only precreate if group == 0 and o_id is specfied */
3637 if (!filter_group_is_mds(group) || oa->o_id == 0)
3640 diff = oa->o_id - filter_last_id(filter, group);
3641 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3642 filter_last_id(filter, group), diff);
3644 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3645 oa->o_id, filter_last_id(filter, group), diff);
3649 oa->o_id = filter_last_id(&obd->u.filter, group);
3650 rc = filter_precreate(obd, oa, group, &diff);
3651 oa->o_id = filter_last_id(&obd->u.filter, group);
3653 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3656 /* else diff == 0 */
3659 up(&filter->fo_create_locks[group]);
3663 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3664 __u64 max_age, __u32 flags)
3666 struct filter_obd *filter = &obd->u.filter;
3667 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3671 /* at least try to account for cached pages. its still racey and
3672 * might be under-reporting if clients haven't announced their
3673 * caches with brw recently */
3674 spin_lock(&obd->obd_osfs_lock);
3675 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3676 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3677 spin_unlock(&obd->obd_osfs_lock);
3679 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3680 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3681 filter->fo_tot_dirty, filter->fo_tot_granted,
3682 filter->fo_tot_pending,
3683 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3685 filter_grant_sanity_check(obd, __FUNCTION__);
3687 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3688 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3689 osfs->os_bsize - 1) >> blockbits));
3691 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3692 * stop creating files on MDS if OST is not good shape to create
3696 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3697 osfs->os_state = OS_STATE_READONLY;
3699 if (filter->fo_raid_degraded)
3700 osfs->os_state |= OS_STATE_DEGRADED;
3704 static int filter_use_existing_obj(struct obd_device *obd,
3705 struct dentry *dchild, void **handle,
3708 struct inode *inode = dchild->d_inode;
3712 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3715 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3716 if (IS_ERR(*handle))
3717 return PTR_ERR(*handle);
3719 iattr.ia_valid = ATTR_MODE;
3720 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3721 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3728 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3731 __u64 os_ffree = -1;
3733 spin_lock(&obd->obd_osfs_lock);
3734 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3736 os_ffree = obd->obd_osfs.os_ffree;
3737 spin_unlock(&obd->obd_osfs_lock);
3742 /* We rely on the fact that only one thread will be creating files in a given
3743 * group at a time, which is why we don't need an atomic filter_get_new_id.
3744 * Even if we had that atomic function, the following race would exist:
3746 * thread 1: gets id x from filter_next_id
3747 * thread 2: gets id (x + 1) from filter_next_id
3748 * thread 2: creates object (x + 1)
3749 * thread 1: tries to create object x, gets -ENOSPC
3751 * Caller must hold fo_create_locks[group]
3753 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3754 obd_gr group, int *num)
3756 struct dentry *dchild = NULL, *dparent = NULL;
3757 struct filter_obd *filter;
3758 struct obd_statfs *osfs;
3759 int err = 0, rc = 0, recreate_obj = 0, i;
3760 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3763 void *handle = NULL;
3766 filter = &obd->u.filter;
3768 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3770 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3772 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3773 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3776 OBD_ALLOC(osfs, sizeof(*osfs));
3779 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3780 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3781 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3782 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3783 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3787 OBD_FREE(osfs, sizeof(*osfs));
3792 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3793 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3795 for (i = 0; i < *num && err == 0; i++) {
3796 int cleanup_phase = 0;
3798 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3799 CWARN("%s: create aborted by destroy\n",
3808 last_id = filter_last_id(filter, group);
3809 if (next_id > last_id) {
3810 CERROR("Error: Trying to recreate obj greater"
3811 "than last id "LPD64" > "LPD64"\n",
3813 GOTO(cleanup, rc = -EINVAL);
3816 next_id = filter_last_id(filter, group) + 1;
3818 dparent = filter_parent_lock(obd, group, next_id);
3819 if (IS_ERR(dparent))
3820 GOTO(cleanup, rc = PTR_ERR(dparent));
3821 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3823 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3825 GOTO(cleanup, rc = PTR_ERR(dchild));
3826 cleanup_phase = 2; /* f_dput(dchild) */
3828 if (dchild->d_inode != NULL) {
3829 /* This would only happen if lastobjid was bad on disk*/
3830 /* Could also happen if recreating missing obj but it
3831 * already exists. */
3833 CERROR("%s: recreating existing object %.*s?\n",
3834 obd->obd_name, dchild->d_name.len,
3835 dchild->d_name.name);
3837 /* Use these existing objects if they are
3839 if (dchild->d_inode->i_size == 0) {
3840 rc = filter_use_existing_obj(obd,dchild,
3841 &handle, &cleanup_phase);
3848 CERROR("%s: Serious error: objid %.*s already "
3849 "exists; is this filesystem corrupt?\n",
3850 obd->obd_name, dchild->d_name.len,
3851 dchild->d_name.name);
3854 GOTO(cleanup, rc = -EEXIST);
3857 handle = fsfilt_start_log(obd, dparent->d_inode,
3858 FSFILT_OP_CREATE, NULL, 1);
3860 GOTO(cleanup, rc = PTR_ERR(handle));
3863 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3864 ",od->o_id="LPU64")\n", obd->obd_name, group,
3867 /* We mark object SUID+SGID to flag it for accepting UID+GID
3868 * from client on first write. Currently the permission bits
3869 * on the OST are never used, so this is OK. */
3870 rc = ll_vfs_create(dparent->d_inode, dchild,
3871 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3873 CERROR("create failed rc = %d\n", rc);
3874 if (rc == -ENOSPC) {
3875 os_ffree = filter_calc_free_inodes(obd);
3877 CERROR("%s: free inode "LPU64"\n",
3878 obd->obd_name, os_ffree);
3883 if (dchild->d_inode)
3884 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3885 dchild->d_inode->i_ino);
3888 if (!recreate_obj) {
3889 filter_set_last_id(filter, next_id, group);
3890 err = filter_update_last_objid(obd, group, 0);
3892 CERROR("unable to write lastobjid "
3893 "but file created\n");
3897 switch(cleanup_phase) {
3899 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3901 CERROR("error on commit, err = %d\n", err);
3908 filter_parent_unlock(dparent);
3915 if (time_after(jiffies, enough_time)) {
3917 "%s: precreate slow - want %d got %d \n",
3918 obd->obd_name, *num, i);
3925 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3926 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3931 static int filter_create(struct obd_export *exp, struct obdo *oa,
3932 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3934 struct obd_device *obd = exp->exp_obd;
3935 struct filter_export_data *fed;
3936 struct filter_obd *filter;
3937 struct lvfs_run_ctxt saved;
3938 struct lov_stripe_md *lsm = NULL;
3939 int rc = 0, diff, group = oa->o_gr;
3942 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3943 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3945 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3946 CERROR("!!! nid %s sent invalid object group %d\n",
3947 obd_export_nid2str(exp), group);
3951 fed = &exp->exp_filter_data;
3952 filter = &obd->u.filter;
3954 if (fed->fed_group != group) {
3955 CERROR("!!! this export (nid %s) used object group %d "
3956 "earlier; now it's trying to use group %d! This could "
3957 "be a bug in the MDS. Please report to "
3958 "http://bugzilla.lustre.org/\n",
3959 obd_export_nid2str(exp), fed->fed_group, group);
3966 rc = obd_alloc_memmd(exp, &lsm);
3973 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3975 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3976 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3977 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3978 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3979 oa->o_id, filter_last_id(filter,
3984 down(&filter->fo_create_locks[oa->o_gr]);
3985 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3986 up(&filter->fo_create_locks[oa->o_gr]);
3989 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3992 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3993 if (rc && ea != NULL && *ea != lsm) {
3994 obd_free_memmd(exp, &lsm);
3995 } else if (rc == 0 && ea != NULL) {
3996 /* XXX LOV STACKING: the lsm that is passed to us from
3997 * LOV does not have valid lsm_oinfo data structs, so
3998 * don't go touching that. This needs to be fixed in a
4000 lsm->lsm_object_id = oa->o_id;
4007 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4008 struct lov_stripe_md *md, struct obd_trans_info *oti,
4009 struct obd_export *md_exp, void *capa)
4011 unsigned int qcids[MAXQUOTAS] = {0, 0};
4012 struct obd_device *obd;
4013 struct filter_obd *filter;
4014 struct dentry *dchild = NULL, *dparent = NULL;
4015 struct lustre_handle lockh = { 0 };
4016 struct lvfs_run_ctxt saved;
4017 void *handle = NULL;
4018 struct llog_cookie *fcc = NULL;
4019 int rc, rc2, cleanup_phase = 0, sync = 0;
4023 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4024 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4025 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4030 filter = &obd->u.filter;
4032 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4035 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4036 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4038 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4040 GOTO(cleanup, rc = PTR_ERR(dchild));
4043 if (dchild->d_inode == NULL) {
4044 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4046 /* If object already gone, cancel cookie right now */
4047 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4048 struct llog_ctxt *ctxt;
4049 struct obd_llog_group *olg;
4051 olg = filter_find_olg(obd, oa->o_gr);
4053 CERROR(" %s: can not find olg of group %d\n",
4054 obd->obd_name, (int)oa->o_gr);
4055 GOTO(cleanup, rc = PTR_ERR(olg));
4057 fcc = &oa->o_lcookie;
4058 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4059 llog_cancel(ctxt, NULL, 1, fcc, 0);
4060 llog_ctxt_put(ctxt);
4061 fcc = NULL; /* we didn't allocate fcc, don't free it */
4063 GOTO(cleanup, rc = -ENOENT);
4066 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4068 /* Our MDC connection is established by the MDS to us */
4069 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4070 OBD_ALLOC(fcc, sizeof(*fcc));
4072 *fcc = oa->o_lcookie;
4074 DQUOT_INIT(dchild->d_inode);
4076 /* we're gonna truncate it first in order to avoid possible deadlock:
4078 * open trasaction open transaction
4079 * down(i_zombie) down(i_zombie)
4080 * restart transaction
4081 * (see BUG 4180) -bzzz
4083 * take i_alloc_sem too to prevent other threads from writing to the
4084 * file while we are truncating it. This can cause lock ordering issue
4085 * between page lock, i_mutex & starting new journal handle.
4086 * (see bug 20321) -johann
4088 down_write(&dchild->d_inode->i_alloc_sem);
4089 LOCK_INODE_MUTEX(dchild->d_inode);
4091 /* VBR: version recovery check */
4092 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4094 UNLOCK_INODE_MUTEX(dchild->d_inode);
4095 up_write(&dchild->d_inode->i_alloc_sem);
4099 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4101 if (IS_ERR(handle)) {
4102 UNLOCK_INODE_MUTEX(dchild->d_inode);
4103 up_write(&dchild->d_inode->i_alloc_sem);
4104 GOTO(cleanup, rc = PTR_ERR(handle));
4107 iattr.ia_valid = ATTR_SIZE;
4109 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4110 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4111 UNLOCK_INODE_MUTEX(dchild->d_inode);
4112 up_write(&dchild->d_inode->i_alloc_sem);
4116 GOTO(cleanup, rc = rc2);
4118 /* We don't actually need to lock the parent until we are unlinking
4119 * here, and not while truncating above. That avoids holding the
4120 * parent lock for a long time during truncate, which can block other
4121 * threads from doing anything to objects in that directory. bug 7171 */
4122 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4123 if (IS_ERR(dparent))
4124 GOTO(cleanup, rc = PTR_ERR(dparent));
4125 cleanup_phase = 3; /* filter_parent_unlock */
4127 LOCK_INODE_MUTEX(dchild->d_inode);
4128 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4129 if (IS_ERR(handle)) {
4130 UNLOCK_INODE_MUTEX(dchild->d_inode);
4131 GOTO(cleanup, rc = PTR_ERR(handle));
4133 cleanup_phase = 4; /* fsfilt_commit */
4135 /* Quota release need uid/gid of inode */
4136 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4138 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4140 /* this drops dchild->d_inode->i_mutex unconditionally */
4141 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4145 switch(cleanup_phase) {
4148 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4149 oti->oti_handle : handle,
4150 filter_cancel_cookies_cb,
4152 /* If add_journal_cb failed, then filter_finish_transno
4153 * will commit the handle and we will do a sync
4154 * on commit. then we call callback directly to free
4157 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4159 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4162 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4164 CERROR("error on commit, err = %d\n", rc2);
4171 filter_parent_unlock(dparent);
4173 filter_fini_destroy(obd, &lockh);
4177 OBD_FREE(fcc, sizeof(*fcc));
4179 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4182 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4186 /* trigger quota release */
4187 qcids[USRQUOTA] = oa->o_uid;
4188 qcids[GRPQUOTA] = oa->o_gid;
4189 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4192 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4196 /* NB start and end are used for punch, but not truncate */
4197 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4198 struct obd_trans_info *oti,
4199 struct ptlrpc_request_set *rqset)
4204 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4205 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4206 oinfo->oi_policy.l_extent.end);
4210 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4211 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4212 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4214 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4215 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4216 rc = filter_setattr(exp, oinfo, oti);
4217 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4221 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4222 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4225 struct lvfs_run_ctxt saved;
4226 struct filter_obd *filter;
4227 struct dentry *dentry;
4231 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4232 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4233 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4237 filter = &exp->exp_obd->u.filter;
4239 /* An objid of zero is taken to mean "sync whole filesystem" */
4240 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4241 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4242 /* Flush any remaining cancel messages out to the target */
4243 filter_sync_llogs(exp->exp_obd, exp);
4247 dentry = filter_oa2dentry(exp->exp_obd, oa);
4249 RETURN(PTR_ERR(dentry));
4251 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4253 LOCK_INODE_MUTEX(dentry->d_inode);
4255 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4257 /* just any file to grab fsync method - "file" arg unused */
4258 struct file *file = filter->fo_rcvd_filp;
4260 if (file->f_op && file->f_op->fsync)
4261 rc = file->f_op->fsync(NULL, dentry, 1);
4263 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4267 UNLOCK_INODE_MUTEX(dentry->d_inode);
4269 oa->o_valid = OBD_MD_FLID;
4270 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4272 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4278 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4279 void *key, __u32 *vallen, void *val,
4280 struct lov_stripe_md *lsm)
4282 struct obd_device *obd;
4285 obd = class_exp2obd(exp);
4287 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4291 if (KEY_IS(KEY_BLOCKSIZE)) {
4292 __u32 *blocksize = val;
4294 if (*vallen < sizeof(*blocksize))
4296 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4298 *vallen = sizeof(*blocksize);
4302 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4303 __u32 *blocksize_bits = val;
4304 if (blocksize_bits) {
4305 if (*vallen < sizeof(*blocksize_bits))
4307 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4309 *vallen = sizeof(*blocksize_bits);
4313 if (KEY_IS(KEY_LAST_ID)) {
4314 obd_id *last_id = val;
4315 /* FIXME: object groups */
4317 if (*vallen < sizeof(*last_id))
4319 *last_id = filter_last_id(&obd->u.filter,
4320 exp->exp_filter_data.fed_group);
4322 *vallen = sizeof(*last_id);
4326 if (KEY_IS(KEY_FIEMAP)) {
4327 struct ll_fiemap_info_key *fm_key = key;
4328 struct dentry *dentry;
4329 struct ll_user_fiemap *fiemap = val;
4330 struct lvfs_run_ctxt saved;
4333 if (fiemap == NULL) {
4334 *vallen = fiemap_count_to_size(
4335 fm_key->fiemap.fm_extent_count);
4339 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4342 RETURN(PTR_ERR(dentry));
4344 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4345 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4346 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4348 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4354 CDEBUG(D_IOCTL, "invalid key\n");
4358 static inline int filter_setup_llog_group(struct obd_export *exp,
4359 struct obd_device *obd,
4362 struct obd_llog_group *olg;
4363 struct llog_ctxt *ctxt;
4366 olg = filter_find_create_olg(obd, group);
4368 RETURN(PTR_ERR(olg));
4370 llog_group_set_export(olg, exp);
4372 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4373 LASSERTF(ctxt != NULL, "ctxt is null\n");
4375 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4376 llog_ctxt_put(ctxt);
4380 static int filter_set_grant_shrink(struct obd_export *exp,
4381 struct ost_body *body)
4383 /* handle shrink grant */
4384 spin_lock(&exp->exp_obd->obd_osfs_lock);
4385 filter_grant_incoming(exp, &body->oa);
4386 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4392 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4394 struct obd_device *obd;
4400 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4404 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4405 obd_export_nid2str(exp));
4406 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4408 /* setup llog imports */
4410 group = (int)(*(__u32 *)val);
4412 group = 0; /* default value */
4414 LASSERT_MDS_GROUP(group);
4415 rc = filter_setup_llog_group(exp, obd, group);
4419 if (group == FILTER_GROUP_MDS0) {
4420 /* setup llog group 1 for interop */
4421 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4424 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4429 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4430 void *key, __u32 vallen, void *val,
4431 struct ptlrpc_request_set *set)
4433 struct obd_device *obd;
4438 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4442 if (KEY_IS(KEY_CAPA_KEY)) {
4444 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4446 CERROR("filter update capability key failed: %d\n", rc);
4450 if (KEY_IS(KEY_REVIMP_UPD)) {
4451 filter_revimp_update(exp);
4452 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4456 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4457 filter_adapt_sptlrpc_conf(obd, 0);
4461 if (KEY_IS(KEY_MDS_CONN))
4462 RETURN(filter_set_mds_conn(exp, val));
4464 if (KEY_IS(KEY_GRANT_SHRINK))
4465 RETURN(filter_set_grant_shrink(exp, val));
4470 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4471 int len, void *karg, void *uarg)
4473 struct obd_device *obd = exp->exp_obd;
4474 struct obd_ioctl_data *data = karg;
4478 case OBD_IOC_ABORT_RECOVERY: {
4479 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4480 target_stop_recovery_thread(obd);
4484 case OBD_IOC_SYNC: {
4485 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4486 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4490 case OBD_IOC_SET_READONLY: {
4492 struct super_block *sb = obd->u.obt.obt_sb;
4493 struct inode *inode = sb->s_root->d_inode;
4494 BDEVNAME_DECLARE_STORAGE(tmp);
4495 CERROR("*** setting device %s read-only ***\n",
4496 ll_bdevname(sb, tmp));
4498 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4499 if (!IS_ERR(handle))
4500 rc = fsfilt_commit(obd, inode, handle, 1);
4502 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4503 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4505 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4509 case OBD_IOC_CATLOGLIST: {
4510 rc = llog_catalog_list(obd, 1, data);
4514 case OBD_IOC_LLOG_CANCEL:
4515 case OBD_IOC_LLOG_REMOVE:
4516 case OBD_IOC_LLOG_INFO:
4517 case OBD_IOC_LLOG_PRINT: {
4518 /* FIXME to be finished */
4519 RETURN(-EOPNOTSUPP);
4521 struct llog_ctxt *ctxt = NULL;
4523 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4524 rc = llog_ioctl(ctxt, cmd, data);
4525 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4538 static int filter_health_check(struct obd_device *obd)
4540 #ifdef USE_HEALTH_CHECK_WRITE
4541 struct filter_obd *filter = &obd->u.filter;
4546 * health_check to return 0 on healthy
4547 * and 1 on unhealthy.
4549 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4552 #ifdef USE_HEALTH_CHECK_WRITE
4553 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4554 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4559 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4562 return filter_fid2dentry(data, NULL, gr, id);
4565 static int filter_process_config(struct obd_device *obd, obd_count len,
4568 struct lustre_cfg *lcfg = buf;
4569 struct lprocfs_static_vars lvars;
4572 switch (lcfg->lcfg_command) {
4574 lprocfs_filter_init_vars(&lvars);
4576 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4586 static struct lvfs_callback_ops filter_lvfs_ops = {
4587 l_fid2dentry: filter_lvfs_fid2dentry,
4590 static struct obd_ops filter_obd_ops = {
4591 .o_owner = THIS_MODULE,
4592 .o_get_info = filter_get_info,
4593 .o_set_info_async = filter_set_info_async,
4594 .o_setup = filter_setup,
4595 .o_precleanup = filter_precleanup,
4596 .o_cleanup = filter_cleanup,
4597 .o_connect = filter_connect,
4598 .o_reconnect = filter_reconnect,
4599 .o_disconnect = filter_disconnect,
4600 .o_ping = filter_ping,
4601 .o_init_export = filter_init_export,
4602 .o_destroy_export = filter_destroy_export,
4603 .o_statfs = filter_statfs,
4604 .o_getattr = filter_getattr,
4605 .o_unpackmd = filter_unpackmd,
4606 .o_create = filter_create,
4607 .o_setattr = filter_setattr,
4608 .o_destroy = filter_destroy,
4609 .o_brw = filter_brw,
4610 .o_punch = filter_truncate,
4611 .o_sync = filter_sync,
4612 .o_preprw = filter_preprw,
4613 .o_commitrw = filter_commitrw,
4614 .o_llog_init = filter_llog_init,
4615 .o_llog_connect = filter_llog_connect,
4616 .o_llog_finish = filter_llog_finish,
4617 .o_iocontrol = filter_iocontrol,
4618 .o_health_check = filter_health_check,
4619 .o_process_config = filter_process_config,
4622 quota_interface_t *filter_quota_interface_ref;
4623 extern quota_interface_t filter_quota_interface;
4625 static int __init obdfilter_init(void)
4627 struct lprocfs_static_vars lvars;
4630 /** sanity check for group<->mdsno conversion */
4631 for (i = 0; i < 32; i++)
4632 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4634 lprocfs_filter_init_vars(&lvars);
4636 request_module("lquota");
4637 OBD_ALLOC(obdfilter_created_scratchpad,
4638 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4639 sizeof(*obdfilter_created_scratchpad));
4640 if (obdfilter_created_scratchpad == NULL)
4643 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4644 sizeof(struct filter_mod_data),
4647 GOTO(out, rc = -ENOMEM);
4649 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4650 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4652 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4653 LUSTRE_OST_NAME, NULL);
4657 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4658 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4659 ll_fmd_cachep = NULL;
4661 if (filter_quota_interface_ref)
4662 PORTAL_SYMBOL_PUT(filter_quota_interface);
4664 OBD_FREE(obdfilter_created_scratchpad,
4665 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4666 sizeof(*obdfilter_created_scratchpad));
4672 static void __exit obdfilter_exit(void)
4674 if (filter_quota_interface_ref)
4675 PORTAL_SYMBOL_PUT(filter_quota_interface);
4677 if (ll_fmd_cachep) {
4678 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4679 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4680 ll_fmd_cachep = NULL;
4683 class_unregister_type(LUSTRE_OST_NAME);
4684 OBD_FREE(obdfilter_created_scratchpad,
4685 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4686 sizeof(*obdfilter_created_scratchpad));
4689 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4690 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4691 MODULE_LICENSE("GPL");
4693 module_init(obdfilter_init);
4694 module_exit(obdfilter_exit);