1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 obd_transno_commit_cb(obd, transno, error);
93 /* Assumes caller has already pushed us into the kernel context. */
94 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
95 int rc, int force_sync)
97 struct filter_obd *filter = &exp->exp_obd->u.filter;
98 struct filter_export_data *fed = &exp->exp_filter_data;
99 struct lsd_client_data *lcd = fed->fed_lcd;
102 int err, log_pri = D_RPCTRACE;
104 /* Propagate error code. */
108 if (!exp->exp_obd->obd_replayable || oti == NULL)
111 /* we don't allocate new transnos for replayed requests */
112 if (oti->oti_transno == 0) {
113 spin_lock(&filter->fo_translock);
114 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
115 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
116 spin_unlock(&filter->fo_translock);
117 oti->oti_transno = last_rcvd;
119 spin_lock(&filter->fo_translock);
120 last_rcvd = oti->oti_transno;
121 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
122 filter->fo_fsd->lsd_last_transno =
123 cpu_to_le64(last_rcvd);
124 spin_unlock(&filter->fo_translock);
126 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
128 /* could get xid from oti, if it's ever needed */
129 lcd->lcd_last_xid = 0;
131 off = fed->fed_lr_off;
133 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
134 fed->fed_lr_idx, fed->fed_lr_off);
138 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
144 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
145 lcd, sizeof(*lcd), &off,
146 force_sync | exp->exp_need_sync);
148 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
156 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
157 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
162 void f_dput(struct dentry *dentry)
164 /* Can't go inside filter_ddelete because it can block */
165 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
166 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
167 LASSERT(atomic_read(&dentry->d_count) > 0);
172 static void init_brw_stats(struct brw_stats *brw_stats)
175 for (i = 0; i < BRW_LAST; i++)
176 spin_lock_init(&brw_stats->hist[i].oh_lock);
179 static int lprocfs_init_rw_stats(struct obd_device *obd,
180 struct lprocfs_stats **stats)
184 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
185 LPROC_FILTER_LAST - 1;
186 *stats = lprocfs_alloc_stats(num_stats, 0);
190 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
191 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
192 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
193 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
194 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
199 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
200 plus the procfs overhead :( */
201 static int filter_export_stats_init(struct obd_device *obd,
202 struct obd_export *exp,
205 struct filter_export_data *fed = &exp->exp_filter_data;
209 init_brw_stats(&fed->fed_brw_stats);
211 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
212 /* Self-export gets no proc entry */
215 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
217 /* Mask error for already created
225 struct nid_stat *tmp = exp->exp_nid_stats;
226 LASSERT(tmp != NULL);
228 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
229 if (tmp->nid_brw_stats == NULL)
232 init_brw_stats(tmp->nid_brw_stats);
233 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
234 0644, &filter_per_nid_stats_fops,
237 CWARN("Error adding the brw_stats file\n");
239 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
243 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
247 /* Always add in ldlm_stats */
248 tmp->nid_ldlm_stats = lprocfs_alloc_stats(LDLM_LAST_OPC -
250 if (tmp->nid_ldlm_stats == NULL)
253 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
254 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
255 tmp->nid_ldlm_stats);
263 /* Add client data to the FILTER. We use a bitmap to locate a free space
264 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
265 * Otherwise, we have just read the data from the last_rcvd file and
266 * we know its offset. */
267 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
270 struct filter_obd *filter = &obd->u.filter;
271 struct filter_export_data *fed = &exp->exp_filter_data;
272 unsigned long *bitmap = filter->fo_last_rcvd_slots;
273 int new_client = (cl_idx == -1);
277 LASSERT(bitmap != NULL);
278 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
281 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
284 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
285 * there's no need for extra complication here
288 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
290 if (cl_idx >= LR_MAX_CLIENTS) {
291 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
295 if (test_and_set_bit(cl_idx, bitmap)) {
296 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
301 if (test_and_set_bit(cl_idx, bitmap)) {
302 CERROR("FILTER client %d: bit already set in bitmap!\n",
308 fed->fed_lr_idx = cl_idx;
309 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
310 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
311 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
313 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
314 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
317 struct lvfs_run_ctxt saved;
318 loff_t off = fed->fed_lr_off;
322 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
323 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
325 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
326 /* Transaction needed to fix bug 1403 */
327 handle = fsfilt_start(obd,
328 filter->fo_rcvd_filp->f_dentry->d_inode,
329 FSFILT_OP_SETATTR, NULL);
330 if (IS_ERR(handle)) {
331 rc = PTR_ERR(handle);
332 CERROR("unable to start transaction: rc %d\n", rc);
334 rc = fsfilt_add_journal_cb(obd, 0, handle,
335 target_client_add_cb, exp);
337 spin_lock(&exp->exp_lock);
338 exp->exp_need_sync = 1;
339 spin_unlock(&exp->exp_lock);
341 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
343 sizeof(*fed->fed_lcd),
344 &off, rc /* sync if no cb */);
346 filter->fo_rcvd_filp->f_dentry->d_inode,
349 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
352 CERROR("error writing %s client idx %u: rc %d\n",
353 LAST_RCVD, fed->fed_lr_idx, rc);
360 static int filter_client_free(struct obd_export *exp)
362 struct filter_export_data *fed = &exp->exp_filter_data;
363 struct filter_obd *filter = &exp->exp_obd->u.filter;
364 struct obd_device *obd = exp->exp_obd;
365 struct lsd_client_data zero_lcd;
366 struct lvfs_run_ctxt saved;
371 if (fed->fed_lcd == NULL)
374 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
375 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
378 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
379 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
381 LASSERT(filter->fo_last_rcvd_slots != NULL);
383 off = fed->fed_lr_off;
385 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
386 * we leak a client slot that will be cleaned on the next recovery. */
388 CERROR("%s: client idx %d has med_off %lld\n",
389 obd->obd_name, fed->fed_lr_idx, off);
390 GOTO(free, rc = -EINVAL);
393 /* Clear the bit _after_ zeroing out the client so we don't
394 race with filter_client_add and zero out new clients.*/
395 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
396 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
401 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
402 memset(&zero_lcd, 0, sizeof zero_lcd);
403 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
404 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
405 sizeof(zero_lcd), &off,
406 (!exp->exp_libclient ||
407 exp->exp_need_sync));
409 /* update server's transno */
410 filter_update_server_data(obd, filter->fo_rcvd_filp,
412 !exp->exp_libclient);
413 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
415 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
416 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
417 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
421 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
422 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
429 OBD_FREE_PTR(fed->fed_lcd);
435 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
436 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
437 struct filter_mod_data *fmd)
439 LASSERT_SPIN_LOCKED(&fed->fed_lock);
440 if (--fmd->fmd_refcount == 0) {
441 /* XXX when we have persistent reservations and the handle
442 * is stored herein we need to drop it here. */
443 fed->fed_mod_count--;
444 list_del(&fmd->fmd_list);
445 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
449 /* drop fmd reference, free it if last ref */
450 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
452 struct filter_export_data *fed;
457 fed = &exp->exp_filter_data;
458 spin_lock(&fed->fed_lock);
459 filter_fmd_put_nolock(fed, fmd); /* caller reference */
460 spin_unlock(&fed->fed_lock);
463 /* expire entries from the end of the list if there are too many
464 * or they are too old */
465 static void filter_fmd_expire_nolock(struct filter_obd *filter,
466 struct filter_export_data *fed,
467 struct filter_mod_data *keep)
469 struct filter_mod_data *fmd, *tmp;
471 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
475 if (time_before(jiffies, fmd->fmd_expire) &&
476 fed->fed_mod_count < filter->fo_fmd_max_num)
479 list_del_init(&fmd->fmd_list);
480 filter_fmd_put_nolock(fed, fmd); /* list reference */
484 void filter_fmd_expire(struct obd_export *exp)
486 spin_lock(&exp->exp_filter_data.fed_lock);
487 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
488 &exp->exp_filter_data, NULL);
489 spin_unlock(&exp->exp_filter_data.fed_lock);
492 /* find specified objid, group in export fmd list.
493 * caller must hold fed_lock and take fmd reference itself */
494 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
495 struct filter_export_data *fed,
496 obd_id objid, obd_gr group)
498 struct filter_mod_data *found = NULL, *fmd;
500 LASSERT_SPIN_LOCKED(&fed->fed_lock);
502 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
503 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
505 list_del(&fmd->fmd_list);
506 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
507 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
512 filter_fmd_expire_nolock(filter, fed, found);
517 /* Find fmd based on objid and group, or return NULL if not found. */
518 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
519 obd_id objid, obd_gr group)
521 struct filter_mod_data *fmd;
523 spin_lock(&exp->exp_filter_data.fed_lock);
524 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
525 &exp->exp_filter_data, objid, group);
527 fmd->fmd_refcount++; /* caller reference */
528 spin_unlock(&exp->exp_filter_data.fed_lock);
533 /* Find fmd based on objid and group, or create a new one if none is found.
534 * It is possible for this function to return NULL under memory pressure,
535 * or if objid = 0 is passed (which will only cause old entries to expire).
536 * Currently this is not fatal because any fmd state is transient and
537 * may also be freed when it gets sufficiently old. */
538 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
539 obd_id objid, obd_gr group)
541 struct filter_export_data *fed = &exp->exp_filter_data;
542 struct filter_mod_data *found = NULL, *fmd_new = NULL;
544 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
546 spin_lock(&fed->fed_lock);
547 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
550 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
551 fmd_new->fmd_id = objid;
552 fmd_new->fmd_gr = group;
553 fmd_new->fmd_refcount++; /* list reference */
555 fed->fed_mod_count++;
557 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
561 found->fmd_refcount++; /* caller reference */
562 found->fmd_expire = jiffies +
563 exp->exp_obd->u.filter.fo_fmd_max_age;
566 spin_unlock(&fed->fed_lock);
572 /* drop fmd list reference so it will disappear when last reference is put.
573 * This isn't so critical because it would in fact only affect the one client
574 * that is doing the unlink and at worst we have an stale entry referencing
575 * an object that should never be used again. */
576 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
578 struct filter_mod_data *found = NULL;
580 spin_lock(&exp->exp_filter_data.fed_lock);
581 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
583 list_del_init(&found->fmd_list);
584 filter_fmd_put_nolock(&exp->exp_filter_data, found);
586 spin_unlock(&exp->exp_filter_data.fed_lock);
589 #define filter_fmd_drop(exp, objid, group)
592 /* remove all entries from fmd list */
593 static void filter_fmd_cleanup(struct obd_export *exp)
595 struct filter_export_data *fed = &exp->exp_filter_data;
596 struct filter_mod_data *fmd = NULL, *tmp;
598 spin_lock(&fed->fed_lock);
599 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
600 list_del_init(&fmd->fmd_list);
601 filter_fmd_put_nolock(fed, fmd);
603 spin_unlock(&fed->fed_lock);
606 static int filter_init_export(struct obd_export *exp)
608 spin_lock_init(&exp->exp_filter_data.fed_lock);
609 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
611 spin_lock(&exp->exp_lock);
612 exp->exp_connecting = 1;
613 spin_unlock(&exp->exp_lock);
615 return ldlm_init_export(exp);
618 static int filter_free_server_data(struct filter_obd *filter)
620 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
621 filter->fo_fsd = NULL;
622 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
623 filter->fo_last_rcvd_slots = NULL;
627 /* assumes caller is already in kernel ctxt */
628 int filter_update_server_data(struct obd_device *obd, struct file *filp,
629 struct lr_server_data *fsd, int force_sync)
635 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
636 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
637 le64_to_cpu(fsd->lsd_last_transno));
638 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
639 le64_to_cpu(fsd->lsd_mount_count));
641 fsd->lsd_compat14 = fsd->lsd_last_transno;
642 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
644 CERROR("error writing lr_server_data: rc = %d\n", rc);
649 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
652 struct filter_obd *filter = &obd->u.filter;
658 if (filter->fo_last_objid_files[group] == NULL) {
659 CERROR("Object group "LPU64" not fully setup; not updating "
660 "last_objid\n", group);
664 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
665 obd->obd_name, group, filter->fo_last_objids[group]);
667 tmp = cpu_to_le64(filter->fo_last_objids[group]);
668 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
669 &tmp, sizeof(tmp), &off, force_sync);
671 CERROR("error writing group "LPU64" last objid: rc = %d\n",
675 extern int ost_handle(struct ptlrpc_request *req);
676 /* assumes caller has already in kernel ctxt */
677 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
679 struct filter_obd *filter = &obd->u.filter;
680 struct lr_server_data *fsd;
681 struct lsd_client_data *lcd = NULL;
682 struct inode *inode = filp->f_dentry->d_inode;
683 unsigned long last_rcvd_size = i_size_read(inode);
689 /* ensure padding in the struct is the correct size */
690 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
691 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
692 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
693 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
695 OBD_ALLOC(fsd, sizeof(*fsd));
698 filter->fo_fsd = fsd;
700 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
701 if (filter->fo_last_rcvd_slots == NULL) {
702 OBD_FREE(fsd, sizeof(*fsd));
706 if (last_rcvd_size == 0) {
707 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
709 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
710 fsd->lsd_last_transno = 0;
711 mount_count = fsd->lsd_mount_count = 0;
712 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
713 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
714 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
715 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
716 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
717 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
719 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
721 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
725 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
726 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
727 "using the wrong disk %s. Were the "
728 "/dev/ assignments rearranged?\n",
729 obd->obd_uuid.uuid, fsd->lsd_uuid);
730 GOTO(err_fsd, rc = -EINVAL);
732 mount_count = le64_to_cpu(fsd->lsd_mount_count);
733 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
735 /* Assume old last_rcvd format unless I_C_LR is set */
736 if (!(fsd->lsd_feature_incompat &
737 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
738 fsd->lsd_last_transno = fsd->lsd_compat14;
742 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
743 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
744 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
745 ~FILTER_INCOMPAT_SUPP);
746 GOTO(err_fsd, rc = -EINVAL);
748 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
749 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
750 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
751 ~FILTER_ROCOMPAT_SUPP);
752 /* Do something like remount filesystem read-only */
753 GOTO(err_fsd, rc = -EINVAL);
756 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
757 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
758 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
759 obd->obd_name, mount_count + 1);
760 CDEBUG(D_INODE, "%s: server data size: %u\n",
761 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
762 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
763 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
764 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
765 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
766 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
767 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
768 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
769 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
770 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
771 le16_to_cpu(fsd->lsd_client_size));
773 if (!obd->obd_replayable) {
774 CWARN("%s: recovery support OFF\n", obd->obd_name);
778 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
779 off < last_rcvd_size; cl_idx++) {
781 struct obd_export *exp;
782 struct filter_export_data *fed;
787 GOTO(err_client, rc = -ENOMEM);
790 /* Don't assume off is incremented properly by
791 * fsfilt_read_record(), in case sizeof(*lcd)
792 * isn't the same as fsd->lsd_client_size. */
793 off = le32_to_cpu(fsd->lsd_client_start) +
794 cl_idx * le16_to_cpu(fsd->lsd_client_size);
795 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
797 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
798 LAST_RCVD, cl_idx, off, rc);
799 break; /* read error shouldn't cause startup to fail */
802 if (lcd->lcd_uuid[0] == '\0') {
803 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
808 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
810 /* These exports are cleaned up by filter_disconnect(), so they
811 * need to be set up like real exports as filter_connect() does.
813 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
815 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
816 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
817 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
819 if (PTR_ERR(exp) == -EALREADY) {
820 /* export already exists, zero out this one */
821 CERROR("Zeroing out duplicate export due to "
823 lcd->lcd_uuid[0] = '\0';
825 GOTO(err_client, rc = PTR_ERR(exp));
828 fed = &exp->exp_filter_data;
830 fed->fed_group = 0; /* will be assigned at connect */
831 filter_export_stats_init(obd, exp, NULL);
832 rc = filter_client_add(obd, exp, cl_idx);
833 /* can't fail for existing client */
834 LASSERTF(rc == 0, "rc = %d\n", rc);
837 spin_lock(&exp->exp_lock);
838 exp->exp_connecting = 0;
839 exp->exp_in_recovery = 0;
840 spin_unlock(&exp->exp_lock);
841 obd->obd_max_recoverable_clients++;
842 class_export_put(exp);
845 /* Need to check last_rcvd even for duplicated exports. */
846 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
849 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
850 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
856 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
858 target_recovery_init(obd, ost_handle);
861 filter->fo_mount_count = mount_count + 1;
862 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
864 /* save it, so mount count and last_transno is current */
865 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
867 GOTO(err_client, rc);
872 target_recovery_fini(obd);
874 filter_free_server_data(filter);
878 static int filter_cleanup_groups(struct obd_device *obd)
880 struct filter_obd *filter = &obd->u.filter;
882 struct dentry *dentry;
886 if (filter->fo_dentry_O_groups != NULL) {
887 for (i = 0; i < filter->fo_group_count; i++) {
888 dentry = filter->fo_dentry_O_groups[i];
892 OBD_FREE(filter->fo_dentry_O_groups,
893 filter->fo_group_count *
894 sizeof(*filter->fo_dentry_O_groups));
895 filter->fo_dentry_O_groups = NULL;
897 if (filter->fo_last_objid_files != NULL) {
898 for (i = 0; i < filter->fo_group_count; i++) {
899 filp = filter->fo_last_objid_files[i];
903 OBD_FREE(filter->fo_last_objid_files,
904 filter->fo_group_count *
905 sizeof(*filter->fo_last_objid_files));
906 filter->fo_last_objid_files = NULL;
908 if (filter->fo_dentry_O_sub != NULL) {
909 for (i = 0; i < filter->fo_group_count; i++) {
910 for (j = 0; j < filter->fo_subdir_count; j++) {
911 dentry = filter->fo_dentry_O_sub[i].dentry[j];
916 OBD_FREE(filter->fo_dentry_O_sub,
917 filter->fo_group_count *
918 sizeof(*filter->fo_dentry_O_sub));
919 filter->fo_dentry_O_sub = NULL;
921 if (filter->fo_last_objids != NULL) {
922 OBD_FREE(filter->fo_last_objids,
923 filter->fo_group_count *
924 sizeof(*filter->fo_last_objids));
925 filter->fo_last_objids = NULL;
927 if (filter->fo_dentry_O != NULL) {
928 f_dput(filter->fo_dentry_O);
929 filter->fo_dentry_O = NULL;
934 static int filter_update_last_group(struct obd_device *obd, int group)
936 struct filter_obd *filter = &obd->u.filter;
937 struct file *filp = NULL;
938 int last_group = 0, rc;
942 if (group <= filter->fo_committed_group)
945 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
949 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
953 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
955 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
958 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
959 "off = %llu and last_group = %d\n", off, last_group);
961 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
962 obd->obd_name, last_group, group);
966 /* must be sync: bXXXX */
967 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
969 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
973 filter->fo_committed_group = group;
980 static int filter_read_group_internal(struct obd_device *obd, int group,
983 struct filter_obd *filter = &obd->u.filter;
984 __u64 *new_objids = NULL;
985 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
986 struct dentry **new_groups = NULL;
987 struct file **new_files = NULL;
988 struct dentry *dentry;
990 int old_count = filter->fo_group_count, rc, stage = 0, i;
996 snprintf(name, 24, "%d", group);
1000 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1002 if (IS_ERR(dentry)) {
1003 CERROR("Cannot lookup expected object group %d: %ld\n",
1004 group, PTR_ERR(dentry));
1005 RETURN(PTR_ERR(dentry));
1008 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1010 if (IS_ERR(dentry)) {
1011 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1013 RETURN(PTR_ERR(dentry));
1018 snprintf(name, 24, "O/%d/LAST_ID", group);
1020 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1022 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1023 GOTO(cleanup, rc = PTR_ERR(filp));
1027 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1029 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1033 if (filter->fo_subdir_count) {
1034 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1035 if (tmp_subdirs == NULL)
1036 GOTO(cleanup, rc = -ENOMEM);
1039 for (i = 0; i < filter->fo_subdir_count; i++) {
1041 snprintf(dir, sizeof(dir), "d%u", i);
1043 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1046 if (IS_ERR(tmp_subdirs->dentry[i])) {
1047 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1048 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1053 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1054 tmp_subdirs->dentry[i]);
1058 /* 'group' is an index; we need an array of length 'group + 1' */
1059 if (group + 1 > old_count) {
1060 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1061 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1062 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1063 OBD_ALLOC(new_files, len * sizeof(*new_files));
1065 if (new_objids == NULL || new_subdirs == NULL ||
1066 new_groups == NULL || new_files == NULL)
1067 GOTO(cleanup, rc = -ENOMEM);
1070 memcpy(new_objids, filter->fo_last_objids,
1071 old_count * sizeof(*new_objids));
1072 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1073 old_count * sizeof(*new_subdirs));
1074 memcpy(new_groups, filter->fo_dentry_O_groups,
1075 old_count * sizeof(*new_groups));
1076 memcpy(new_files, filter->fo_last_objid_files,
1077 old_count * sizeof(*new_files));
1079 OBD_FREE(filter->fo_last_objids,
1080 old_count * sizeof(*new_objids));
1081 OBD_FREE(filter->fo_dentry_O_sub,
1082 old_count * sizeof(*new_subdirs));
1083 OBD_FREE(filter->fo_dentry_O_groups,
1084 old_count * sizeof(*new_groups));
1085 OBD_FREE(filter->fo_last_objid_files,
1086 old_count * sizeof(*new_files));
1088 filter->fo_last_objids = new_objids;
1089 filter->fo_dentry_O_sub = new_subdirs;
1090 filter->fo_dentry_O_groups = new_groups;
1091 filter->fo_last_objid_files = new_files;
1092 filter->fo_group_count = len;
1095 filter->fo_dentry_O_groups[group] = dentry;
1096 filter->fo_last_objid_files[group] = filp;
1097 if (filter->fo_subdir_count) {
1098 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1099 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1102 filter_update_last_group(obd, group);
1104 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1105 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1106 rc = filter_update_last_objid(obd, group, 1);
1110 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1111 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1112 obd->obd_name, group, last_objid);
1117 if (new_objids != NULL)
1118 OBD_FREE(new_objids, len * sizeof(*new_objids));
1119 if (new_subdirs != NULL)
1120 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1121 if (new_groups != NULL)
1122 OBD_FREE(new_groups, len * sizeof(*new_groups));
1123 if (new_files != NULL)
1124 OBD_FREE(new_files, len * sizeof(*new_files));
1126 if (filter->fo_subdir_count) {
1127 for (i = 0; i < filter->fo_subdir_count; i++) {
1128 if (tmp_subdirs->dentry[i] != NULL)
1129 dput(tmp_subdirs->dentry[i]);
1131 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1134 filp_close(filp, 0);
1141 static int filter_read_groups(struct obd_device *obd, int last_group,
1144 struct filter_obd *filter = &obd->u.filter;
1145 int old_count, group, rc = 0;
1147 down(&filter->fo_init_lock);
1148 old_count = filter->fo_group_count;
1149 for (group = old_count; group <= last_group; group++) {
1151 rc = filter_read_group_internal(obd, group, create);
1155 up(&filter->fo_init_lock);
1159 /* FIXME: object groups */
1160 static int filter_prep_groups(struct obd_device *obd)
1162 struct filter_obd *filter = &obd->u.filter;
1163 struct dentry *dentry, *O_dentry;
1165 int last_group, rc = 0, cleanup_phase = 0;
1169 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1171 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1172 if (IS_ERR(O_dentry)) {
1173 rc = PTR_ERR(O_dentry);
1174 CERROR("cannot open/create O: rc = %d\n", rc);
1177 filter->fo_dentry_O = O_dentry;
1178 cleanup_phase = 1; /* O_dentry */
1180 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1181 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1182 * some time post 1.0 when all old-style OSTs have converted along
1183 * with the init_objid hack. */
1184 dentry = ll_lookup_one_len("R", O_dentry, 1);
1186 GOTO(cleanup, rc = PTR_ERR(dentry));
1187 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1188 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1191 CWARN("converting OST to new object layout\n");
1192 if (IS_ERR(O0_dentry)) {
1193 rc = PTR_ERR(O0_dentry);
1194 CERROR("error looking up O/0: rc %d\n", rc);
1195 GOTO(cleanup_R, rc);
1198 if (O0_dentry->d_inode) {
1199 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1200 GOTO(cleanup_O0, rc = -EEXIST);
1203 LOCK_INODE_MUTEX(O_dentry->d_inode);
1204 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1205 O_dentry->d_inode, O0_dentry,
1207 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1210 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1211 GOTO(cleanup_O0, rc);
1213 filter->fo_fsd->lsd_feature_incompat |=
1214 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1215 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1217 GOTO(cleanup_O0, rc);
1229 cleanup_phase = 2; /* groups */
1231 /* we have to initialize all groups before first connections from
1232 * clients because they may send create/destroy for any group -bzzz */
1233 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1235 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1236 GOTO(cleanup, rc = PTR_ERR(filp));
1238 cleanup_phase = 3; /* filp */
1240 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1242 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1246 last_group = FILTER_MIN_GROUPS;
1248 LASSERT_MDS_GROUP(last_group);
1251 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1252 FILTER_MIN_GROUPS, last_group);
1253 filter->fo_committed_group = last_group;
1254 rc = filter_read_groups(obd, last_group, 1);
1258 filp_close(filp, 0);
1262 switch (cleanup_phase) {
1264 filp_close(filp, 0);
1266 filter_cleanup_groups(obd);
1268 f_dput(filter->fo_dentry_O);
1269 filter->fo_dentry_O = NULL;
1277 /* setup the object store with correct subdirectories */
1278 static int filter_prep(struct obd_device *obd)
1280 struct lvfs_run_ctxt saved;
1281 struct filter_obd *filter = &obd->u.filter;
1283 struct inode *inode;
1287 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1288 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1289 if (!file || IS_ERR(file)) {
1291 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1295 filter->fo_rcvd_filp = file;
1296 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1297 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1298 file->f_dentry->d_inode->i_mode);
1299 GOTO(err_filp, rc = -ENOENT);
1302 inode = file->f_dentry->d_parent->d_inode;
1303 /* We use i_op->unlink directly in filter_vfs_unlink() */
1304 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1305 CERROR("%s: filesystem does not support create/unlink ops\n",
1307 GOTO(err_filp, rc = -EOPNOTSUPP);
1310 rc = filter_init_server_data(obd, file);
1312 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1315 /* open and create health check io file*/
1316 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1319 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1321 GOTO(err_server_data, rc);
1323 filter->fo_health_check_filp = file;
1324 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1325 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1326 file->f_dentry->d_inode->i_mode);
1327 GOTO(err_health_check, rc = -ENOENT);
1329 rc = lvfs_check_io_health(obd, file);
1331 GOTO(err_health_check, rc);
1333 rc = filter_prep_groups(obd);
1335 GOTO(err_health_check, rc);
1337 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1342 if (filp_close(filter->fo_health_check_filp, 0))
1343 CERROR("can't close %s after error\n", HEALTH_CHECK);
1344 filter->fo_health_check_filp = NULL;
1346 target_recovery_fini(obd);
1347 filter_free_server_data(filter);
1349 if (filp_close(filter->fo_rcvd_filp, 0))
1350 CERROR("can't close %s after error\n", LAST_RCVD);
1351 filter->fo_rcvd_filp = NULL;
1355 /* cleanup the filter: write last used object id to status file */
1356 static void filter_post(struct obd_device *obd)
1358 struct lvfs_run_ctxt saved;
1359 struct filter_obd *filter = &obd->u.filter;
1362 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1363 * best to start a transaction with h_sync, because we removed this
1366 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1367 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1370 CERROR("error writing server data: rc = %d\n", rc);
1372 for (i = 0; i < filter->fo_group_count; i++) {
1373 rc = filter_update_last_objid(obd, i,
1374 (i == filter->fo_group_count - 1));
1376 CERROR("error writing group %d lastobjid: rc = %d\n",
1380 rc = filp_close(filter->fo_rcvd_filp, 0);
1381 filter->fo_rcvd_filp = NULL;
1383 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1385 rc = filp_close(filter->fo_health_check_filp, 0);
1386 filter->fo_health_check_filp = NULL;
1388 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1390 filter_cleanup_groups(obd);
1391 filter_free_server_data(filter);
1392 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1394 filter_free_capa_keys(filter);
1395 cleanup_capa_hash(filter->fo_capa_hash);
1398 static void filter_set_last_id(struct filter_obd *filter,
1399 obd_id id, obd_gr group)
1401 LASSERT(filter->fo_fsd != NULL);
1402 LASSERT(group <= filter->fo_group_count);
1404 spin_lock(&filter->fo_objidlock);
1405 filter->fo_last_objids[group] = id;
1406 spin_unlock(&filter->fo_objidlock);
1409 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1412 LASSERT(filter->fo_fsd != NULL);
1413 LASSERT(group <= filter->fo_group_count);
1415 /* FIXME: object groups */
1416 spin_lock(&filter->fo_objidlock);
1417 id = filter->fo_last_objids[group];
1418 spin_unlock(&filter->fo_objidlock);
1422 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1424 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1428 /* We never dget the object parent, so DON'T dput it either */
1429 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1431 struct filter_obd *filter = &obd->u.filter;
1432 struct filter_subdirs *subdirs;
1433 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1435 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1436 filter->fo_subdir_count == 0)
1437 return filter->fo_dentry_O_groups[group];
1439 subdirs = &filter->fo_dentry_O_sub[group];
1440 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1443 /* We never dget the object parent, so DON'T dput it either */
1444 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1447 unsigned long now = jiffies;
1448 struct dentry *dparent = filter_parent(obd, group, objid);
1451 if (IS_ERR(dparent))
1453 if (dparent == NULL)
1454 return ERR_PTR(-ENOENT);
1456 rc = filter_lock_dentry(obd, dparent);
1457 fsfilt_check_slow(obd, now, "parent lock");
1458 return rc ? ERR_PTR(rc) : dparent;
1461 /* We never dget the object parent, so DON'T dput it either */
1462 static void filter_parent_unlock(struct dentry *dparent)
1464 UNLOCK_INODE_MUTEX(dparent->d_inode);
1467 /* How to get files, dentries, inodes from object id's.
1469 * If dir_dentry is passed, the caller has already locked the parent
1470 * appropriately for this operation (normally a write lock). If
1471 * dir_dentry is NULL, we do a read lock while we do the lookup to
1472 * avoid races with create/destroy and such changing the directory
1473 * internal to the filesystem code. */
1474 struct dentry *filter_fid2dentry(struct obd_device *obd,
1475 struct dentry *dir_dentry,
1476 obd_gr group, obd_id id)
1478 struct dentry *dparent = dir_dentry;
1479 struct dentry *dchild;
1484 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1485 obd->u.filter.fo_destroys_in_progress == 0) {
1486 /* don't fail lookups for orphan recovery, it causes
1487 * later LBUGs when objects still exist during precreate */
1488 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1489 RETURN(ERR_PTR(-ENOENT));
1492 CERROR("fatal: invalid object id 0\n");
1493 RETURN(ERR_PTR(-ESTALE));
1496 len = sprintf(name, LPU64, id);
1497 if (dir_dentry == NULL) {
1498 dparent = filter_parent_lock(obd, group, id);
1499 if (IS_ERR(dparent)) {
1500 CERROR("%s: error getting object "LPU64":"LPU64
1501 " parent: rc %ld\n", obd->obd_name,
1502 id, group, PTR_ERR(dparent));
1506 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1507 dparent->d_name.len, dparent->d_name.name, name);
1508 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1509 if (dir_dentry == NULL)
1510 filter_parent_unlock(dparent);
1511 if (IS_ERR(dchild)) {
1512 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1517 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1518 CERROR("%s: got bad object "LPU64" inode %lu\n",
1519 obd->obd_name, id, dchild->d_inode->i_ino);
1521 RETURN(ERR_PTR(-ENOENT));
1524 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1525 name, dchild, atomic_read(&dchild->d_count));
1527 LASSERT(atomic_read(&dchild->d_count) > 0);
1532 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1535 struct lustre_handle lockh;
1536 int flags = LDLM_AST_DISCARD_DATA, rc;
1537 struct ldlm_res_id res_id;
1538 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1541 osc_build_res_name(objid, group, &res_id);
1542 /* Tell the clients that the object is gone now and that they should
1543 * throw away any cached pages. */
1544 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1545 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1546 ldlm_completion_ast, NULL, NULL, 0, NULL,
1549 /* We only care about the side-effects, just drop the lock. */
1551 ldlm_lock_decref(&lockh, LCK_PW);
1556 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1557 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1558 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1559 * reverse. Caller must take i_sem before starting the transaction and we
1560 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1561 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1562 struct vfsmount *mnt)
1567 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1568 * itself we already hold dir->i_mutex for child create/unlink ops */
1569 LASSERT(dentry->d_inode != NULL);
1570 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1571 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1575 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1576 GOTO(out, rc = -ENOENT);
1578 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1583 GOTO(out, rc = -EPERM);
1585 /* check_sticky() */
1586 if ((dentry->d_inode->i_uid != current->fsuid &&
1587 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1588 IS_IMMUTABLE(dentry->d_inode))
1589 GOTO(out, rc = -EPERM);
1591 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1592 * that was done because of journal_start (which is already done
1593 * here) or some other ordering issue. */
1596 rc = ll_security_inode_unlink(dir, dentry, mnt);
1600 rc = dir->i_op->unlink(dir, dentry);
1602 /* need to drop i_mutex before we lose inode reference */
1603 UNLOCK_INODE_MUTEX(dentry->d_inode);
1610 /* Caller must hold LCK_PW on parent and push us into kernel context.
1611 * Caller must hold child i_mutex, we drop it always.
1612 * Caller is also required to ensure that dchild->d_inode exists. */
1613 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1614 obd_gr group, struct dentry *dparent,
1615 struct dentry *dchild)
1617 struct inode *inode = dchild->d_inode;
1618 struct filter_obd *filter = &obd->u.filter;
1621 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1622 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1623 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1624 (unsigned long)inode->i_nlink,
1625 atomic_read(&inode->i_count));
1628 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1630 CERROR("error unlinking objid %.*s: rc %d\n",
1631 dchild->d_name.len, dchild->d_name.name, rc);
1635 struct filter_intent_args {
1636 struct ldlm_lock **victim;
1641 static enum interval_iter filter_intent_cb(struct interval_node *n,
1644 struct ldlm_interval *node = (struct ldlm_interval *)n;
1645 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1646 __u64 size = arg->size;
1647 struct ldlm_lock **v = arg->victim;
1648 struct ldlm_lock *lck;
1650 /* If the interval is lower than the current file size,
1652 if (interval_high(n) <= size)
1653 return INTERVAL_ITER_STOP;
1655 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1656 /* Don't send glimpse ASTs to liblustre clients.
1657 * They aren't listening for them, and they do
1658 * entirely synchronous I/O anyways. */
1659 if (lck->l_export == NULL ||
1660 lck->l_export->exp_libclient == 1)
1663 if (*arg->liblustre)
1664 *arg->liblustre = 0;
1667 *v = LDLM_LOCK_GET(lck);
1668 } else if ((*v)->l_policy_data.l_extent.start <
1669 lck->l_policy_data.l_extent.start) {
1670 LDLM_LOCK_RELEASE(*v);
1671 *v = LDLM_LOCK_GET(lck);
1674 /* the same policy group - every lock has the
1675 * same extent, so needn't do it any more */
1679 return INTERVAL_ITER_CONT;
1682 static int filter_intent_policy(struct ldlm_namespace *ns,
1683 struct ldlm_lock **lockp, void *req_cookie,
1684 ldlm_mode_t mode, int flags, void *data)
1686 CFS_LIST_HEAD(rpc_list);
1687 struct ptlrpc_request *req = req_cookie;
1688 struct ldlm_lock *lock = *lockp, *l = NULL;
1689 struct ldlm_resource *res = lock->l_resource;
1690 ldlm_processing_policy policy;
1691 struct ost_lvb *res_lvb, *reply_lvb;
1692 struct ldlm_reply *rep;
1694 int idx, rc, tmpflags = 0, only_liblustre = 1;
1695 struct ldlm_interval_tree *tree;
1696 struct filter_intent_args arg;
1697 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1698 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1699 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1702 policy = ldlm_get_processing_policy(res);
1703 LASSERT(policy != NULL);
1704 LASSERT(req != NULL);
1706 rc = lustre_pack_reply(req, 3, repsize, NULL);
1708 RETURN(req->rq_status = rc);
1710 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1711 LASSERT(rep != NULL);
1713 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1714 sizeof(*reply_lvb));
1715 LASSERT(reply_lvb != NULL);
1717 //fixup_handle_for_resent_req(req, lock, &lockh);
1719 /* Call the extent policy function to see if our request can be
1720 * granted, or is blocked.
1721 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1722 * lock, and should not be granted if the lock will be blocked.
1725 LASSERT(ns == res->lr_namespace);
1727 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1728 check_res_locked(res);
1730 /* FIXME: we should change the policy function slightly, to not make
1731 * this list at all, since we just turn around and free it */
1732 while (!list_empty(&rpc_list)) {
1733 struct ldlm_lock *wlock =
1734 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1735 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1736 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1737 lock->l_flags &= ~LDLM_FL_CP_REQD;
1738 list_del_init(&wlock->l_cp_ast);
1739 LDLM_LOCK_RELEASE(wlock);
1742 /* The lock met with no resistance; we're finished. */
1743 if (rc == LDLM_ITER_CONTINUE) {
1744 /* do not grant locks to the liblustre clients: they cannot
1745 * handle ASTs robustly. We need to do this while still
1746 * holding ns_lock to avoid the lock remaining on the res_link
1747 * list (and potentially being added to l_pending_list by an
1748 * AST) when we are going to drop this lock ASAP. */
1749 if (lock->l_export->exp_libclient ||
1750 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1751 ldlm_resource_unlink_lock(lock);
1752 err = ELDLM_LOCK_ABORTED;
1754 err = ELDLM_LOCK_REPLACED;
1760 /* Do not grant any lock, but instead send GL callbacks. The extent
1761 * policy nicely created a list of all PW locks for us. We will choose
1762 * the highest of those which are larger than the size in the LVB, if
1763 * any, and perform a glimpse callback. */
1764 res_lvb = res->lr_lvb_data;
1765 LASSERT(res_lvb != NULL);
1766 *reply_lvb = *res_lvb;
1769 * ->ns_lock guarantees that no new locks are granted, and,
1770 * therefore, that res->lr_lvb_data cannot increase beyond the
1771 * end of already granted lock. As a result, it is safe to
1772 * check against "stale" reply_lvb->lvb_size value without
1775 arg.size = reply_lvb->lvb_size;
1777 arg.liblustre = &only_liblustre;
1778 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1779 tree = &res->lr_itree[idx];
1780 if (tree->lit_mode == LCK_PR)
1783 interval_iterate_reverse(tree->lit_root,
1784 filter_intent_cb, &arg);
1788 /* There were no PW locks beyond the size in the LVB; finished. */
1790 if (only_liblustre) {
1791 /* If we discovered a liblustre client with a PW lock,
1792 * however, the LVB may be out of date! The LVB is
1793 * updated only on glimpse (which we don't do for
1794 * liblustre clients) and cancel (which the client
1795 * obviously has not yet done). So if it has written
1796 * data but kept the lock, the LVB is stale and needs
1797 * to be updated from disk.
1799 * Of course, this will all disappear when we switch to
1800 * taking liblustre locks on the OST. */
1801 ldlm_res_lvbo_update(res, NULL, 0, 1);
1803 RETURN(ELDLM_LOCK_ABORTED);
1807 * This check is for lock taken in filter_prepare_destroy() that does
1808 * not have l_glimpse_ast set. So the logic is: if there is a lock
1809 * with no l_glimpse_ast set, this object is being destroyed already.
1811 * Hence, if you are grabbing DLM locks on the server, always set
1812 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1814 if (l->l_glimpse_ast == NULL) {
1815 /* We are racing with unlink(); just return -ENOENT */
1816 rep->lock_policy_res1 = -ENOENT;
1820 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1821 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1822 /* Update the LVB from disk if the AST failed (this is a legal race) */
1824 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1825 * sending ast is not handled. This can result in lost client writes.
1828 ldlm_res_lvbo_update(res, NULL, 0, 1);
1831 *reply_lvb = *res_lvb;
1835 LDLM_LOCK_RELEASE(l);
1837 RETURN(ELDLM_LOCK_ABORTED);
1841 * per-obd_device iobuf pool.
1843 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1844 * allocations in write-path has to be minimized (see bug 5137).
1846 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1847 * OST threads (see ost_thread_{init,done}()).
1849 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1850 * at the OST layer there are only (potentially) multiple obd_device of type
1851 * unknown at the time of OST thread creation.
1853 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1854 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1855 * it's very own iobuf.
1859 * filter_kiobuf_pool_init()
1861 * filter_kiobuf_pool_done()
1863 * filter_iobuf_get()
1865 * operate on this array. They are "generic" in a sense that they don't depend
1866 * on actual type of iobuf's (the latter depending on Linux kernel version).
1870 * destroy pool created by filter_iobuf_pool_init
1872 static void filter_iobuf_pool_done(struct filter_obd *filter)
1874 struct filter_iobuf **pool;
1879 pool = filter->fo_iobuf_pool;
1881 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1882 if (pool[i] != NULL)
1883 filter_free_iobuf(pool[i]);
1885 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1886 filter->fo_iobuf_pool = NULL;
1891 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1893 struct filter_obd *filter = &obd->u.filter;
1894 struct sptlrpc_rule_set tmp_rset;
1897 sptlrpc_rule_set_init(&tmp_rset);
1898 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1900 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1905 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1907 write_lock(&filter->fo_sptlrpc_lock);
1908 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1909 filter->fo_sptlrpc_rset = tmp_rset;
1910 write_unlock(&filter->fo_sptlrpc_lock);
1916 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1918 static int filter_iobuf_pool_init(struct filter_obd *filter)
1925 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1927 if (filter->fo_iobuf_pool == NULL)
1930 filter->fo_iobuf_count = OSS_THREADS_MAX;
1935 /* Return iobuf allocated for @thread_id. We don't know in advance how
1936 * many threads there will be so we allocate a large empty array and only
1937 * fill in those slots that are actually in use.
1938 * If we haven't allocated a pool entry for this thread before, do so now. */
1939 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1941 int thread_id = (oti && oti->oti_thread) ?
1942 oti->oti_thread->t_id : -1;
1943 struct filter_iobuf *pool = NULL;
1944 struct filter_iobuf **pool_place = NULL;
1946 if (thread_id >= 0) {
1947 LASSERT(thread_id < filter->fo_iobuf_count);
1948 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1951 if (unlikely(pool == NULL)) {
1952 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1953 PTLRPC_MAX_BRW_PAGES);
1954 if (pool_place != NULL)
1961 /* mount the file system (secretly). lustre_cfg parameters are:
1964 * 3 = flags: failover=f, failout=n
1967 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1970 struct filter_obd *filter = &obd->u.filter;
1971 struct vfsmount *mnt;
1972 struct lustre_mount_info *lmi;
1973 struct obd_uuid uuid;
1981 if (lcfg->lcfg_bufcount < 3 ||
1982 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1983 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1986 lmi = server_get_mount(obd->obd_name);
1988 /* We already mounted in lustre_fill_super.
1989 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1990 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1992 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1994 /* old path - used by lctl */
1995 CERROR("Using old MDS mount method\n");
1996 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1997 MS_NOATIME|MS_NODIRATIME,
1998 lustre_cfg_string(lcfg, 1), option);
2001 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2002 lustre_cfg_string(lcfg, 1), rc);
2006 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2008 if (IS_ERR(obd->obd_fsops))
2009 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2011 rc = filter_iobuf_pool_init(filter);
2015 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
2017 /* failover is the default */
2018 obd->obd_replayable = 1;
2020 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2021 str = lustre_cfg_string(lcfg, 3);
2022 if (strchr(str, 'n')) {
2023 CWARN("%s: recovery disabled\n", obd->obd_name);
2024 obd->obd_replayable = 0;
2028 filter->fo_vfsmnt = mnt;
2029 obd->u.obt.obt_sb = mnt->mnt_sb;
2030 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2031 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2033 fsfilt_setup(obd, obd->u.obt.obt_sb);
2035 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2036 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2037 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2038 obd->obd_lvfs_ctxt.fs = get_ds();
2039 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2041 init_mutex(&filter->fo_init_lock);
2042 filter->fo_committed_group = 0;
2044 rc = filter_prep(obd);
2048 filter->fo_destroys_in_progress = 0;
2049 for (i = 0; i < 32; i++)
2050 sema_init(&filter->fo_create_locks[i], 1);
2052 spin_lock_init(&filter->fo_translock);
2053 spin_lock_init(&filter->fo_objidlock);
2054 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2055 sema_init(&filter->fo_alloc_lock, 1);
2056 init_brw_stats(&filter->fo_filter_stats);
2057 filter->fo_read_cache = 1; /* enable read-only cache by default */
2058 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2059 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2060 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2061 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2063 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2064 spin_lock_init(&filter->fo_llog_list_lock);
2066 filter->fo_fl_oss_capa = 1;
2067 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2068 filter->fo_capa_hash = init_capa_hash();
2069 if (filter->fo_capa_hash == NULL)
2070 GOTO(err_ops, rc = -ENOMEM);
2072 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2073 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2074 LDLM_NAMESPACE_GREEDY);
2075 if (obd->obd_namespace == NULL)
2076 GOTO(err_post, rc = -ENOMEM);
2077 obd->obd_namespace->ns_lvbp = obd;
2078 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2079 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2081 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2082 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2084 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2086 CERROR("failed to setup llogging subsystems\n");
2090 rwlock_init(&filter->fo_sptlrpc_lock);
2091 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2092 /* do this after llog being initialized */
2093 filter_adapt_sptlrpc_conf(obd, 1);
2095 rc = lquota_setup(filter_quota_interface_ref, obd);
2099 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2100 if (q->max_sectors < q->max_hw_sectors &&
2101 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2102 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2103 "for larger I/O requests: max_sectors = %u "
2104 "could be up to max_hw_sectors=%u\n",
2105 obd->obd_name, mnt->mnt_sb->s_id,
2106 q->max_sectors, q->max_hw_sectors);
2108 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2109 if (uuid_ptr != NULL) {
2110 class_uuid_unparse(uuid_ptr, &uuid);
2116 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2118 if (obd->obd_recovering) {
2119 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2120 "recovery for at least %d:%.02d, or until %d "
2121 "client%s reconnect. During this time new clients"
2122 " will not be allowed to connect. "
2123 "Recovery progress can be monitored by watching "
2124 "/proc/fs/lustre/obdfilter/%s/recovery_status.\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",
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 = 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 count, struct llog_catid *catid,
2341 struct obd_uuid *uuid)
2343 struct filter_obd *filter = &obd->u.filter;
2344 struct llog_ctxt *ctxt;
2348 LASSERT(olg != NULL);
2349 if (olg == &obd->obd_olg)
2350 return filter_default_olg_init(obd, olg, tgt);
2352 LASSERT(filter->fo_lcm != NULL);
2353 rc = filter_olg_init(obd, olg, tgt);
2356 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2358 CERROR("Can't get ctxt for %p:%x\n", olg,
2359 LLOG_MDS_OST_REPL_CTXT);
2360 filter_olg_fini(olg);
2363 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2364 ctxt->loc_lcm = filter->fo_lcm;
2365 llog_ctxt_put(ctxt);
2369 static int filter_llog_finish(struct obd_device *obd, int count)
2371 struct filter_obd *filter = &obd->u.filter;
2372 struct llog_ctxt *ctxt;
2375 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2378 * Make sure that no cached llcds left in recov_thread.
2379 * We actually do sync in disconnect time, but disconnect
2380 * may not come being marked rq_no_resend = 1.
2382 llog_sync(ctxt, NULL);
2385 * Balance class_import_get() in llog_receptor_accept().
2386 * This is safe to do, as llog is already synchronized
2387 * and its import may go.
2389 mutex_down(&ctxt->loc_sem);
2390 if (ctxt->loc_imp) {
2391 class_import_put(ctxt->loc_imp);
2392 ctxt->loc_imp = NULL;
2394 mutex_up(&ctxt->loc_sem);
2395 llog_ctxt_put(ctxt);
2398 if (filter->fo_lcm) {
2399 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2400 filter->fo_lcm = NULL;
2402 RETURN(filter_olg_fini(&obd->obd_olg));
2406 * Find the group llog according to group index in the llog group list.
2408 static struct obd_llog_group *
2409 filter_find_olg_internal(struct filter_obd *filter, int group)
2411 struct obd_llog_group *olg;
2413 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2414 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2415 if (olg->olg_group == group)
2422 * Find the group llog according to group index on the filter
2424 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2426 struct obd_llog_group *olg = NULL;
2427 struct filter_obd *filter;
2429 filter = &obd->u.filter;
2431 if (group == FILTER_GROUP_LLOG)
2432 RETURN(&obd->obd_olg);
2434 spin_lock(&filter->fo_llog_list_lock);
2435 olg = filter_find_olg_internal(filter, group);
2436 spin_unlock(&filter->fo_llog_list_lock);
2441 * Find the llog_group of the filter according to the group. If it can not
2442 * find, create the llog_group, which only happens when mds is being synced
2445 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2447 struct obd_llog_group *olg = NULL;
2448 struct filter_obd *filter;
2451 filter = &obd->u.filter;
2453 if (group == FILTER_GROUP_LLOG)
2454 RETURN(&obd->obd_olg);
2456 spin_lock(&filter->fo_llog_list_lock);
2457 olg = filter_find_olg_internal(filter, group);
2459 if (olg->olg_initializing) {
2460 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2462 GOTO(out_unlock, olg);
2467 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2469 llog_group_init(olg, group);
2470 list_add(&olg->olg_list, &filter->fo_llog_list);
2471 olg->olg_initializing = 1;
2472 spin_unlock(&filter->fo_llog_list_lock);
2474 rc = llog_cat_initialize(obd, olg, 1, NULL);
2476 spin_lock(&filter->fo_llog_list_lock);
2477 list_del(&olg->olg_list);
2478 spin_unlock(&filter->fo_llog_list_lock);
2480 GOTO(out, olg = ERR_PTR(-ENOMEM));
2482 spin_lock(&filter->fo_llog_list_lock);
2483 olg->olg_initializing = 0;
2484 spin_unlock(&filter->fo_llog_list_lock);
2485 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2486 obd->obd_name, group, olg);
2491 spin_unlock(&filter->fo_llog_list_lock);
2495 static int filter_llog_connect(struct obd_export *exp,
2496 struct llogd_conn_body *body)
2498 struct obd_device *obd = exp->exp_obd;
2499 struct llog_ctxt *ctxt;
2500 struct obd_llog_group *olg;
2504 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2505 obd->obd_name, body->lgdc_logid.lgl_oid,
2506 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2508 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2510 CERROR(" %s: can not find olg of group %d\n",
2511 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2514 llog_group_set_export(olg, exp);
2516 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2517 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2518 body->lgdc_ctxt_idx);
2520 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2521 obd->obd_name, body->lgdc_logid.lgl_oid,
2522 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2524 rc = llog_connect(ctxt, &body->lgdc_logid,
2525 &body->lgdc_gen, NULL);
2526 llog_ctxt_put(ctxt);
2528 CERROR("failed to connect rc %d idx %d\n", rc,
2529 body->lgdc_ctxt_idx);
2534 static int filter_llog_preclean(struct obd_device *obd)
2536 struct obd_llog_group *olg, *tmp;
2537 struct filter_obd *filter;
2538 struct list_head remove_list;
2542 rc = obd_llog_finish(obd, 0);
2544 CERROR("failed to cleanup llogging subsystem\n");
2546 filter = &obd->u.filter;
2547 CFS_INIT_LIST_HEAD(&remove_list);
2549 spin_lock(&filter->fo_llog_list_lock);
2550 while (!list_empty(&filter->fo_llog_list)) {
2551 olg = list_entry(filter->fo_llog_list.next,
2552 struct obd_llog_group, olg_list);
2553 list_del(&olg->olg_list);
2554 list_add(&olg->olg_list, &remove_list);
2556 spin_unlock(&filter->fo_llog_list_lock);
2558 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2559 list_del_init(&olg->olg_list);
2560 rc = filter_olg_fini(olg);
2562 CERROR("failed to cleanup llogging subsystem for %u\n",
2570 static int filter_precleanup(struct obd_device *obd,
2571 enum obd_cleanup_stage stage)
2577 case OBD_CLEANUP_EARLY:
2579 case OBD_CLEANUP_EXPORTS:
2580 target_cleanup_recovery(obd);
2581 rc = filter_llog_preclean(obd);
2587 static int filter_cleanup(struct obd_device *obd)
2589 struct filter_obd *filter = &obd->u.filter;
2593 LCONSOLE_WARN("%s: shutting down for failover; client state "
2594 "will be preserved.\n", obd->obd_name);
2596 if (!list_empty(&obd->obd_exports)) {
2597 CERROR("%s: still has clients!\n", obd->obd_name);
2598 class_disconnect_exports(obd);
2599 if (!list_empty(&obd->obd_exports)) {
2600 CERROR("still has exports after forced cleanup?\n");
2605 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2606 lprocfs_free_per_client_stats(obd);
2607 lprocfs_free_obd_stats(obd);
2608 lprocfs_obd_cleanup(obd);
2609 lquota_cleanup(filter_quota_interface_ref, obd);
2611 /* Stop recovery before namespace cleanup. */
2612 target_stop_recovery_thread(obd);
2613 target_cleanup_recovery(obd);
2615 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2616 obd->obd_namespace = NULL;
2618 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2620 if (obd->u.obt.obt_sb == NULL)
2625 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2626 shrink_dcache_sb(obd->u.obt.obt_sb);
2628 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2629 obd->u.obt.obt_sb = NULL;
2631 fsfilt_put_ops(obd->obd_fsops);
2633 filter_iobuf_pool_done(filter);
2635 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2640 static int filter_connect_internal(struct obd_export *exp,
2641 struct obd_connect_data *data)
2643 struct filter_export_data *fed = &exp->exp_filter_data;
2648 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2649 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2650 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2651 data->ocd_connect_flags, data->ocd_version,
2652 data->ocd_grant, data->ocd_index);
2654 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2655 CWARN("!!! This export (nid %s) used object group %d "
2656 "earlier; now it's trying to use group %d! This could "
2657 "be a bug in the MDS. Please report to "
2658 "http://bugzilla.lustre.org/\n",
2659 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2662 fed->fed_group = data->ocd_group;
2664 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2665 exp->exp_connect_flags = data->ocd_connect_flags;
2666 data->ocd_version = LUSTRE_VERSION_CODE;
2668 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2669 obd_size left, want;
2671 spin_lock(&exp->exp_obd->obd_osfs_lock);
2672 left = filter_grant_space_left(exp);
2673 want = data->ocd_grant;
2674 filter_grant(exp, fed->fed_grant, want, left);
2675 data->ocd_grant = fed->fed_grant;
2676 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2678 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2679 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2680 exp->exp_client_uuid.uuid, exp,
2681 data->ocd_grant, want, left);
2684 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2685 struct filter_obd *filter = &exp->exp_obd->u.filter;
2686 struct lr_server_data *lsd = filter->fo_fsd;
2687 int index = le32_to_cpu(lsd->lsd_ost_index);
2689 if (!(lsd->lsd_feature_compat &
2690 cpu_to_le32(OBD_COMPAT_OST))) {
2691 /* this will only happen on the first connect */
2692 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2693 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2694 filter_update_server_data(exp->exp_obd,
2695 filter->fo_rcvd_filp, lsd, 1);
2696 } else if (index != data->ocd_index) {
2697 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2698 " %u doesn't match actual OST index"
2699 " %u in last_rcvd file, bad "
2701 obd_export_nid2str(exp), index,
2707 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2708 data->ocd_brw_size = 65536;
2709 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2710 data->ocd_brw_size = min(data->ocd_brw_size,
2711 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2712 LASSERT(data->ocd_brw_size);
2715 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2716 __u32 cksum_types = data->ocd_cksum_types;
2718 /* The client set in ocd_cksum_types the checksum types it
2719 * supports. We have to mask off the algorithms that we don't
2721 if (cksum_types & OBD_CKSUM_ALL)
2722 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2724 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2726 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2727 "%x\n", exp->exp_obd->obd_name,
2728 obd_export_nid2str(exp), cksum_types,
2729 data->ocd_cksum_types);
2731 /* This client does not support OBD_CONNECT_CKSUM
2732 * fall back to CRC32 */
2733 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2734 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2735 exp->exp_obd->obd_name,
2736 obd_export_nid2str(exp));
2739 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2740 * FIXME: We don't strictly need the COMPAT flag for that,
2741 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2742 * FIXME: We needed it for the index, as index 0 is valid. */
2747 static int filter_reconnect(const struct lu_env *env,
2748 struct obd_export *exp, struct obd_device *obd,
2749 struct obd_uuid *cluuid,
2750 struct obd_connect_data *data,
2756 if (exp == NULL || obd == NULL || cluuid == NULL)
2759 rc = filter_connect_internal(exp, data);
2761 filter_export_stats_init(obd, exp, localdata);
2766 /* nearly identical to mds_connect */
2767 static int filter_connect(const struct lu_env *env,
2768 struct lustre_handle *conn, struct obd_device *obd,
2769 struct obd_uuid *cluuid,
2770 struct obd_connect_data *data, void *localdata)
2772 struct lvfs_run_ctxt saved;
2773 struct obd_export *exp;
2774 struct filter_export_data *fed;
2775 struct lsd_client_data *lcd = NULL;
2780 if (conn == NULL || obd == NULL || cluuid == NULL)
2783 rc = class_connect(conn, obd, cluuid);
2786 exp = class_conn2export(conn);
2787 LASSERT(exp != NULL);
2789 fed = &exp->exp_filter_data;
2791 rc = filter_connect_internal(exp, data);
2795 filter_export_stats_init(obd, exp, localdata);
2796 if (obd->obd_replayable) {
2797 OBD_ALLOC(lcd, sizeof(*lcd));
2799 CERROR("filter: out of memory for client data\n");
2800 GOTO(cleanup, rc = -ENOMEM);
2803 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2805 rc = filter_client_add(obd, exp, -1);
2810 group = data->ocd_group;
2812 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2813 obd->obd_name, exp->exp_handle.h_cookie, group);
2815 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2816 rc = filter_read_groups(obd, group, 1);
2817 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2819 CERROR("can't read group %u\n", group);
2829 fed->fed_lcd = NULL;
2831 class_disconnect(exp);
2833 class_export_put(exp);
2839 /* Do extra sanity checks for grant accounting. We do this at connect,
2840 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2841 * always get rid of it or turn it off when we know accounting is good. */
2842 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2844 struct filter_export_data *fed;
2845 struct obd_export *exp;
2846 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2847 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2848 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2850 if (list_empty(&obd->obd_exports))
2853 /* We don't want to do this for large machines that do lots of
2854 mounts or unmounts. It burns... */
2855 if (obd->obd_num_exports > 100)
2858 spin_lock(&obd->obd_osfs_lock);
2859 spin_lock(&obd->obd_dev_lock);
2860 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2862 fed = &exp->exp_filter_data;
2863 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2866 if (maxsize > 0) { /* we may not have done a statfs yet */
2867 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2868 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2869 exp->exp_client_uuid.uuid, exp,
2870 fed->fed_grant, fed->fed_pending, maxsize);
2871 LASSERTF(fed->fed_dirty <= maxsize,
2872 "%s: cli %s/%p %ld > "LPU64"\n", func,
2873 exp->exp_client_uuid.uuid, exp,
2874 fed->fed_dirty, maxsize);
2877 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2878 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2879 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2881 CDEBUG(D_CACHE, "%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);
2884 tot_granted += fed->fed_grant + fed->fed_pending;
2885 tot_pending += fed->fed_pending;
2886 tot_dirty += fed->fed_dirty;
2888 fo_tot_granted = obd->u.filter.fo_tot_granted;
2889 fo_tot_pending = obd->u.filter.fo_tot_pending;
2890 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2891 spin_unlock(&obd->obd_dev_lock);
2892 spin_unlock(&obd->obd_osfs_lock);
2894 /* Do these assertions outside the spinlocks so we don't kill system */
2895 if (tot_granted != fo_tot_granted)
2896 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2897 func, tot_granted, fo_tot_granted);
2898 if (tot_pending != fo_tot_pending)
2899 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2900 func, tot_pending, fo_tot_pending);
2901 if (tot_dirty != fo_tot_dirty)
2902 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2903 func, tot_dirty, fo_tot_dirty);
2904 if (tot_pending > tot_granted)
2905 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2906 func, tot_pending, tot_granted);
2907 if (tot_granted > maxsize)
2908 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2909 func, tot_granted, maxsize);
2910 if (tot_dirty > maxsize)
2911 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2912 func, tot_dirty, maxsize);
2915 /* Remove this client from the grant accounting totals. We also remove
2916 * the export from the obd device under the osfs and dev locks to ensure
2917 * that the filter_grant_sanity_check() calculations are always valid.
2918 * The client should do something similar when it invalidates its import. */
2919 static void filter_grant_discard(struct obd_export *exp)
2921 struct obd_device *obd = exp->exp_obd;
2922 struct filter_obd *filter = &obd->u.filter;
2923 struct filter_export_data *fed = &exp->exp_filter_data;
2925 spin_lock(&obd->obd_osfs_lock);
2926 spin_lock(&obd->obd_dev_lock);
2927 list_del_init(&exp->exp_obd_chain);
2928 spin_unlock(&obd->obd_dev_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_flags & OBD_OPT_FORCE))
2978 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2983 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2985 struct obd_llog_group *olg_min, *olg;
2986 struct filter_obd *filter;
2987 int worked = -1, group;
2988 struct llog_ctxt *ctxt;
2991 filter = &obd->u.filter;
2993 /* we can't sync log holding spinlock. also, we do not want to get
2994 * into livelock. so we do following: loop over MDS's exports in
2995 * group order and skip already synced llogs -bzzz */
2997 /* look for group with min. number, but > worked */
3000 spin_lock(&filter->fo_llog_list_lock);
3001 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3002 if (olg->olg_group <= worked) {
3003 /* this group is already synced */
3006 if (group < olg->olg_group) {
3007 /* we have group with smaller number to sync */
3010 /* store current minimal group */
3012 group = olg->olg_group;
3014 spin_unlock(&filter->fo_llog_list_lock);
3016 if (olg_min == NULL)
3019 worked = olg_min->olg_group;
3020 if (olg_min->olg_exp &&
3021 (dexp == olg_min->olg_exp || dexp == NULL)) {
3023 ctxt = llog_group_get_ctxt(olg_min,
3024 LLOG_MDS_OST_REPL_CTXT);
3026 err = llog_sync(ctxt, olg_min->olg_exp);
3027 llog_ctxt_put(ctxt);
3029 CERROR("error flushing logs to MDS: "
3034 } while (olg_min != NULL);
3037 /* Also incredibly similar to mds_disconnect */
3038 static int filter_disconnect(struct obd_export *exp)
3040 struct obd_device *obd = exp->exp_obd;
3045 class_export_get(exp);
3047 if (!(exp->exp_flags & OBD_OPT_FORCE))
3048 filter_grant_sanity_check(obd, __FUNCTION__);
3049 filter_grant_discard(exp);
3051 /* Flush any remaining cancel messages out to the target */
3052 filter_sync_llogs(obd, exp);
3054 /* Disconnect early so that clients can't keep using export */
3055 rc = class_disconnect(exp);
3056 if (exp->exp_obd->obd_namespace != NULL)
3057 ldlm_cancel_locks_for_export(exp);
3059 fsfilt_sync(obd, obd->u.obt.obt_sb);
3061 lprocfs_exp_cleanup(exp);
3063 class_export_put(exp);
3067 /* reverse import is changed, sync all cancels */
3068 static void filter_revimp_update(struct obd_export *exp)
3073 class_export_get(exp);
3075 /* flush any remaining cancel messages out to the target */
3076 filter_sync_llogs(exp->exp_obd, exp);
3077 class_export_put(exp);
3081 static int filter_ping(struct obd_export *exp)
3083 filter_fmd_expire(exp);
3088 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3089 const char *what, int quiet)
3091 struct dentry *dchild = NULL;
3094 if (oa->o_valid & OBD_MD_FLGROUP)
3097 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3099 if (IS_ERR(dchild)) {
3100 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3101 what, group, oa->o_id);
3105 if (dchild->d_inode == NULL) {
3107 CERROR("%s: %s on non-existent object: "LPU64"\n",
3108 obd->obd_name, what, oa->o_id);
3110 RETURN(ERR_PTR(-ENOENT));
3116 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3118 struct dentry *dentry = NULL;
3119 struct obd_device *obd;
3123 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3124 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3128 obd = class_exp2obd(exp);
3130 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3134 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3136 RETURN(PTR_ERR(dentry));
3138 /* Limit the valid bits in the return data to what we actually use */
3139 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3140 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3146 /* this should be enabled/disabled in condition to enabled/disabled large
3147 * inodes (fast EAs) in backing store FS. */
3148 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3149 void *handle, struct obdo *oa)
3151 struct obd_device *obd = exp->exp_obd;
3155 if (oa->o_valid & OBD_MD_FLFID) {
3156 struct filter_fid ff;
3158 if (!(oa->o_valid & OBD_MD_FLGROUP))
3160 /* packing fid and converting it to LE for storing into EA.
3161 * Here ->o_stripe_idx should be filled by LOV and rest of
3162 * fields - by client. */
3163 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3164 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3165 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3166 ff.ff_objid = cpu_to_le64(oa->o_id);
3167 ff.ff_group = cpu_to_le64(oa->o_gr);
3169 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3170 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3171 oa->o_generation, oa->o_id, oa->o_gr);
3173 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3175 CERROR("store fid in object failed! rc: %d\n", rc);
3177 CDEBUG(D_HA, "OSS object without fid info!\n");
3183 /* this is called from filter_truncate() until we have filter_punch() */
3184 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3185 struct obdo *oa, struct obd_trans_info *oti)
3187 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3188 struct llog_cookie *fcc = NULL;
3189 struct filter_obd *filter;
3190 int rc, err, locked = 0, sync = 0;
3191 loff_t old_size = 0;
3192 unsigned int ia_valid;
3193 struct inode *inode;
3198 LASSERT(dentry != NULL);
3199 LASSERT(!IS_ERR(dentry));
3201 inode = dentry->d_inode;
3202 LASSERT(inode != NULL);
3204 filter = &exp->exp_obd->u.filter;
3205 iattr_from_obdo(&iattr, oa, oa->o_valid);
3206 ia_valid = iattr.ia_valid;
3208 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3209 OBD_ALLOC(fcc, sizeof(*fcc));
3211 *fcc = oa->o_lcookie;
3214 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3216 LOCK_INODE_MUTEX(inode);
3217 old_size = i_size_read(inode);
3221 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3222 * then we will accept the UID+GID sent by the client during write for
3223 * initializing the ownership of this inode. We only allow this to
3224 * happen once so clear these bits in setattr. In 2.6 kernels it is
3225 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3226 * the flags that are actually being set. */
3227 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3228 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3229 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3231 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3232 if (!(ia_valid & ATTR_MODE)) {
3233 iattr.ia_mode = inode->i_mode;
3234 iattr.ia_valid |= ATTR_MODE;
3236 iattr.ia_mode &= ~S_ISUID;
3238 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3239 if (!(iattr.ia_valid & ATTR_MODE)) {
3240 iattr.ia_mode = inode->i_mode;
3241 iattr.ia_valid |= ATTR_MODE;
3243 iattr.ia_mode &= ~S_ISGID;
3246 orig_ids[USRQUOTA] = inode->i_uid;
3247 orig_ids[GRPQUOTA] = inode->i_gid;
3248 handle = fsfilt_start_log(exp->exp_obd, inode,
3249 FSFILT_OP_SETATTR, oti, 1);
3251 GOTO(out_unlock, rc = PTR_ERR(handle));
3253 /* update inode EA only once when inode is suid bit marked. As
3254 * on 2.6.x UID and GID may be set separately, we check here
3255 * only one of them to avoid double setting. */
3256 if (inode->i_mode & S_ISUID)
3257 filter_update_fidea(exp, inode, handle, oa);
3259 handle = fsfilt_start(exp->exp_obd, inode,
3260 FSFILT_OP_SETATTR, oti);
3262 GOTO(out_unlock, rc = PTR_ERR(handle));
3264 if (oa->o_valid & OBD_MD_FLFLAGS) {
3265 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3266 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3268 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3270 /* set cancel cookie callback function */
3271 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3272 filter_cancel_cookies_cb,
3276 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3277 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3279 /* The truncate might have used up our transaction credits. Make
3280 * sure we have one left for the last_rcvd update. */
3281 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3283 rc = filter_finish_transno(exp, oti, rc, sync);
3285 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3289 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3291 CERROR("error on commit, err = %d\n", err);
3298 /* For a partial-page truncate flush the page to disk immediately
3299 * to avoid data corruption during direct disk write. b=17397 */
3300 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3301 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3302 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3309 UNLOCK_INODE_MUTEX(inode);
3316 UNLOCK_INODE_MUTEX(inode);
3319 OBD_FREE(fcc, sizeof(*fcc));
3321 /* trigger quota release */
3322 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3323 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3324 int rc2 = lquota_adjust(filter_quota_interface_ref,
3325 exp->exp_obd, cur_ids,
3326 orig_ids, rc, FSFILT_OP_SETATTR);
3327 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3328 "filter adjust qunit. (rc:%d)\n", rc2);
3333 /* this is called from filter_truncate() until we have filter_punch() */
3334 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3335 struct obd_trans_info *oti)
3337 struct obdo *oa = oinfo->oi_oa;
3338 struct lustre_capa *capa = oinfo_capa(oinfo);
3339 struct ldlm_res_id res_id;
3340 struct filter_mod_data *fmd;
3341 struct lvfs_run_ctxt saved;
3342 struct filter_obd *filter;
3343 struct ldlm_resource *res;
3344 struct dentry *dentry;
3345 __u64 opc = CAPA_OPC_META_WRITE;
3349 if (oa->o_valid & OBD_FL_TRUNC)
3350 opc |= CAPA_OPC_OSS_TRUNC;
3351 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3355 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3356 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3361 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3362 /* This would be very bad - accidentally truncating a file when
3363 * changing the time or similar - bug 12203. */
3364 if (oa->o_valid & OBD_MD_FLSIZE &&
3365 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3366 static char mdsinum[48];
3368 if (oa->o_valid & OBD_MD_FLFID)
3369 snprintf(mdsinum, sizeof(mdsinum) - 1,
3370 " of inode "LPU64"/%u", oa->o_fid,
3375 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3377 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3382 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3384 RETURN(PTR_ERR(dentry));
3386 filter = &exp->exp_obd->u.filter;
3387 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3391 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3392 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3393 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3394 fmd->fmd_mactime_xid = oti->oti_xid;
3395 filter_fmd_put(exp, fmd);
3398 /* setting objects attributes (including owner/group) */
3399 rc = filter_setattr_internal(exp, dentry, oa, oti);
3401 GOTO(out_unlock, rc);
3403 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3404 &res_id, LDLM_EXTENT, 0);
3407 LDLM_RESOURCE_ADDREF(res);
3408 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3409 LDLM_RESOURCE_DELREF(res);
3410 ldlm_resource_putref(res);
3413 oa->o_valid = OBD_MD_FLID;
3415 /* Quota release need uid/gid info */
3416 obdo_from_inode(oa, dentry->d_inode,
3417 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3423 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3427 /* XXX identical to osc_unpackmd */
3428 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3429 struct lov_mds_md *lmm, int lmm_bytes)
3435 if (lmm_bytes < sizeof (*lmm)) {
3436 CERROR("lov_mds_md too small: %d, need %d\n",
3437 lmm_bytes, (int)sizeof(*lmm));
3440 /* XXX LOV_MAGIC etc check? */
3442 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3443 CERROR("lov_mds_md: zero lmm_object_id\n");
3448 lsm_size = lov_stripe_md_size(1);
3452 if (*lsmp != NULL && lmm == NULL) {
3453 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3454 OBD_FREE(*lsmp, lsm_size);
3459 if (*lsmp == NULL) {
3460 OBD_ALLOC(*lsmp, lsm_size);
3464 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3465 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3466 OBD_FREE(*lsmp, lsm_size);
3469 loi_init((*lsmp)->lsm_oinfo[0]);
3473 /* XXX zero *lsmp? */
3474 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3475 LASSERT((*lsmp)->lsm_object_id);
3478 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3483 /* caller must hold fo_create_locks[oa->o_gr] */
3484 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3485 struct filter_obd *filter)
3487 struct obdo doa; /* XXX obdo on stack */
3493 LASSERT_MDS_GROUP(oa->o_gr);
3494 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3495 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3497 memset(&doa, 0, sizeof(doa));
3499 doa.o_valid |= OBD_MD_FLGROUP;
3500 doa.o_gr = oa->o_gr;
3501 doa.o_mode = S_IFREG;
3503 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3504 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3505 exp->exp_obd->obd_name, doa.o_gr);
3509 last = filter_last_id(filter, doa.o_gr);
3511 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3512 exp->exp_obd->obd_name, oa->o_id + 1, last);
3514 for (id = last; id > oa->o_id; id--) {
3516 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3517 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3518 CEMERG("error destroying precreate objid "LPU64": %d\n",
3520 filter_set_last_id(filter, id - 1, doa.o_gr);
3521 /* update last_id on disk periodically so that if we restart
3522 * we don't need to re-scan all of the just-deleted objects. */
3523 if ((id & 511) == 0)
3524 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3527 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3528 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3530 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3531 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3536 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3537 obd_gr group, int *num);
3538 /* returns a negative error or a nonnegative number of files to create */
3539 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3540 obd_gr group, struct obd_trans_info *oti)
3542 struct obd_device *obd = exp->exp_obd;
3543 struct filter_obd *filter = &obd->u.filter;
3547 /* delete orphans request */
3548 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3549 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3550 CERROR("%s: dropping old orphan cleanup request\n",
3554 /* This causes inflight precreates to abort and drop lock */
3555 set_bit(group, &filter->fo_destroys_in_progress);
3556 down(&filter->fo_create_locks[group]);
3557 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3558 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3559 exp->exp_obd->obd_name, group);
3560 up(&filter->fo_create_locks[group]);
3563 diff = oa->o_id - filter_last_id(filter, group);
3564 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3565 filter_last_id(filter, group), diff);
3567 if (-diff > OST_MAX_PRECREATE) {
3568 CERROR("%s: ignoring bogus orphan destroy request: "
3569 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3570 oa->o_id, filter_last_id(filter, group));
3571 /* FIXME: should reset precreate_next_id on MDS */
3572 GOTO(out, rc = -EINVAL);
3575 rc = filter_destroy_precreated(exp, oa, filter);
3577 CERROR("%s: unable to write lastobjid, but "
3578 "orphans were deleted\n", obd->obd_name);
3581 /* XXX: Used by MDS for the first time! */
3582 clear_bit(group, &filter->fo_destroys_in_progress);
3585 down(&filter->fo_create_locks[group]);
3586 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3587 CERROR("%s: dropping old precreate request\n",
3591 /* only precreate if group == 0 and o_id is specfied */
3592 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3595 diff = oa->o_id - filter_last_id(filter, group);
3596 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3597 filter_last_id(filter, group), diff);
3599 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3600 oa->o_id, filter_last_id(filter, group), diff);
3604 oa->o_id = filter_last_id(&obd->u.filter, group);
3605 rc = filter_precreate(obd, oa, group, &diff);
3606 oa->o_id = filter_last_id(&obd->u.filter, group);
3608 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3611 /* else diff == 0 */
3614 up(&filter->fo_create_locks[group]);
3618 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3619 __u64 max_age, __u32 flags)
3621 struct filter_obd *filter = &obd->u.filter;
3622 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3626 /* at least try to account for cached pages. its still racey and
3627 * might be under-reporting if clients haven't announced their
3628 * caches with brw recently */
3629 spin_lock(&obd->obd_osfs_lock);
3630 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3631 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3632 spin_unlock(&obd->obd_osfs_lock);
3634 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3635 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3636 filter->fo_tot_dirty, filter->fo_tot_granted,
3637 filter->fo_tot_pending,
3638 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3640 filter_grant_sanity_check(obd, __FUNCTION__);
3642 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3643 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3644 osfs->os_bsize - 1) >> blockbits));
3646 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3647 * stop creating files on MDS if OST is not good shape to create
3649 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3654 static int filter_use_existing_obj(struct obd_device *obd,
3655 struct dentry *dchild, void **handle,
3658 struct inode *inode = dchild->d_inode;
3662 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3665 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3666 if (IS_ERR(*handle))
3667 return PTR_ERR(*handle);
3669 iattr.ia_valid = ATTR_MODE;
3670 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3671 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3679 /* We rely on the fact that only one thread will be creating files in a given
3680 * group at a time, which is why we don't need an atomic filter_get_new_id.
3681 * Even if we had that atomic function, the following race would exist:
3683 * thread 1: gets id x from filter_next_id
3684 * thread 2: gets id (x + 1) from filter_next_id
3685 * thread 2: creates object (x + 1)
3686 * thread 1: tries to create object x, gets -ENOSPC
3688 * Caller must hold fo_create_locks[group]
3690 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3691 obd_gr group, int *num)
3693 struct dentry *dchild = NULL, *dparent = NULL;
3694 struct filter_obd *filter;
3695 struct obd_statfs *osfs;
3696 int err = 0, rc = 0, recreate_obj = 0, i;
3697 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3699 void *handle = NULL;
3702 filter = &obd->u.filter;
3704 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3706 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3708 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3709 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3712 OBD_ALLOC(osfs, sizeof(*osfs));
3715 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3716 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3717 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3718 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3719 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3723 OBD_FREE(osfs, sizeof(*osfs));
3728 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3729 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3731 for (i = 0; i < *num && err == 0; i++) {
3732 int cleanup_phase = 0;
3734 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3735 CWARN("%s: create aborted by destroy\n",
3744 last_id = filter_last_id(filter, group);
3745 if (next_id > last_id) {
3746 CERROR("Error: Trying to recreate obj greater"
3747 "than last id "LPD64" > "LPD64"\n",
3749 GOTO(cleanup, rc = -EINVAL);
3752 next_id = filter_last_id(filter, group) + 1;
3754 dparent = filter_parent_lock(obd, group, next_id);
3755 if (IS_ERR(dparent))
3756 GOTO(cleanup, rc = PTR_ERR(dparent));
3757 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3759 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3761 GOTO(cleanup, rc = PTR_ERR(dchild));
3762 cleanup_phase = 2; /* f_dput(dchild) */
3764 if (dchild->d_inode != NULL) {
3765 /* This would only happen if lastobjid was bad on disk*/
3766 /* Could also happen if recreating missing obj but it
3767 * already exists. */
3769 CERROR("%s: recreating existing object %.*s?\n",
3770 obd->obd_name, dchild->d_name.len,
3771 dchild->d_name.name);
3773 /* Use these existing objects if they are
3775 if (dchild->d_inode->i_size == 0) {
3776 rc = filter_use_existing_obj(obd,dchild,
3777 &handle, &cleanup_phase);
3784 CERROR("%s: Serious error: objid %.*s already "
3785 "exists; is this filesystem corrupt?\n",
3786 obd->obd_name, dchild->d_name.len,
3787 dchild->d_name.name);
3790 GOTO(cleanup, rc = -EEXIST);
3793 handle = fsfilt_start_log(obd, dparent->d_inode,
3794 FSFILT_OP_CREATE, NULL, 1);
3796 GOTO(cleanup, rc = PTR_ERR(handle));
3799 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3800 ",od->o_id="LPU64")\n", obd->obd_name, group,
3803 /* We mark object SUID+SGID to flag it for accepting UID+GID
3804 * from client on first write. Currently the permission bits
3805 * on the OST are never used, so this is OK. */
3806 rc = ll_vfs_create(dparent->d_inode, dchild,
3807 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3809 CERROR("create failed rc = %d\n", rc);
3814 if (!recreate_obj) {
3815 filter_set_last_id(filter, next_id, group);
3816 err = filter_update_last_objid(obd, group, 0);
3818 CERROR("unable to write lastobjid "
3819 "but file created\n");
3823 switch(cleanup_phase) {
3825 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3827 CERROR("error on commit, err = %d\n", err);
3834 filter_parent_unlock(dparent);
3841 if (time_after(jiffies, enough_time)) {
3843 "%s: precreate slow - want %d got %d \n",
3844 obd->obd_name, *num, i);
3851 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3852 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3857 static int filter_create(struct obd_export *exp, struct obdo *oa,
3858 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3860 struct obd_device *obd = exp->exp_obd;
3861 struct filter_export_data *fed;
3862 struct filter_obd *filter;
3863 struct lvfs_run_ctxt saved;
3864 struct lov_stripe_md *lsm = NULL;
3865 int rc = 0, diff, group = oa->o_gr;
3868 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3869 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3871 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3872 CERROR("!!! nid %s sent invalid object group %d\n",
3873 obd_export_nid2str(exp), group);
3877 fed = &exp->exp_filter_data;
3878 filter = &obd->u.filter;
3880 if (fed->fed_group != group) {
3881 CERROR("!!! this export (nid %s) used object group %d "
3882 "earlier; now it's trying to use group %d! This could "
3883 "be a bug in the MDS. Please report to "
3884 "http://bugzilla.lustre.org/\n",
3885 obd_export_nid2str(exp), fed->fed_group, group);
3892 rc = obd_alloc_memmd(exp, &lsm);
3899 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3901 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3902 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3903 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3904 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3905 oa->o_id, filter_last_id(filter,
3910 down(&filter->fo_create_locks[oa->o_gr]);
3911 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3912 up(&filter->fo_create_locks[oa->o_gr]);
3915 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3918 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3919 if (rc && ea != NULL && *ea != lsm) {
3920 obd_free_memmd(exp, &lsm);
3921 } else if (rc == 0 && ea != NULL) {
3922 /* XXX LOV STACKING: the lsm that is passed to us from
3923 * LOV does not have valid lsm_oinfo data structs, so
3924 * don't go touching that. This needs to be fixed in a
3926 lsm->lsm_object_id = oa->o_id;
3933 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3934 struct lov_stripe_md *md, struct obd_trans_info *oti,
3935 struct obd_export *md_exp, void *capa)
3937 unsigned int qcids[MAXQUOTAS] = {0, 0};
3938 struct obd_device *obd;
3939 struct filter_obd *filter;
3940 struct dentry *dchild = NULL, *dparent = NULL;
3941 struct lvfs_run_ctxt saved;
3942 void *handle = NULL;
3943 struct llog_cookie *fcc = NULL;
3944 int rc, rc2, cleanup_phase = 0, sync = 0;
3948 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3950 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3951 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
3956 filter = &obd->u.filter;
3958 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3961 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
3962 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3964 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3966 GOTO(cleanup, rc = PTR_ERR(dchild));
3969 if (dchild->d_inode == NULL) {
3970 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3972 /* If object already gone, cancel cookie right now */
3973 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3974 struct llog_ctxt *ctxt;
3975 struct obd_llog_group *olg;
3976 fcc = &oa->o_lcookie;
3977 olg = filter_find_olg(obd, oa->o_gr);
3979 CERROR(" %s: can not find olg of group %d\n",
3980 obd->obd_name, (int)oa->o_gr);
3981 GOTO(cleanup, rc = PTR_ERR(olg));
3983 llog_group_set_export(olg, exp);
3985 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3986 llog_cancel(ctxt, NULL, 1, fcc, 0);
3987 llog_ctxt_put(ctxt);
3988 fcc = NULL; /* we didn't allocate fcc, don't free it */
3990 GOTO(cleanup, rc = -ENOENT);
3993 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3995 /* Our MDC connection is established by the MDS to us */
3996 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3997 OBD_ALLOC(fcc, sizeof(*fcc));
3999 *fcc = oa->o_lcookie;
4001 DQUOT_INIT(dchild->d_inode);
4003 /* we're gonna truncate it first in order to avoid possible deadlock:
4005 * open trasaction open transaction
4006 * down(i_zombie) down(i_zombie)
4007 * restart transaction
4008 * (see BUG 4180) -bzzz
4010 LOCK_INODE_MUTEX(dchild->d_inode);
4011 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4013 if (IS_ERR(handle)) {
4014 UNLOCK_INODE_MUTEX(dchild->d_inode);
4015 GOTO(cleanup, rc = PTR_ERR(handle));
4018 iattr.ia_valid = ATTR_SIZE;
4020 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4021 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4022 UNLOCK_INODE_MUTEX(dchild->d_inode);
4026 GOTO(cleanup, rc = rc2);
4028 /* We don't actually need to lock the parent until we are unlinking
4029 * here, and not while truncating above. That avoids holding the
4030 * parent lock for a long time during truncate, which can block other
4031 * threads from doing anything to objects in that directory. bug 7171 */
4032 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4033 if (IS_ERR(dparent))
4034 GOTO(cleanup, rc = PTR_ERR(dparent));
4035 cleanup_phase = 3; /* filter_parent_unlock */
4037 LOCK_INODE_MUTEX(dchild->d_inode);
4038 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4039 if (IS_ERR(handle)) {
4040 UNLOCK_INODE_MUTEX(dchild->d_inode);
4041 GOTO(cleanup, rc = PTR_ERR(handle));
4043 cleanup_phase = 4; /* fsfilt_commit */
4045 /* Quota release need uid/gid of inode */
4046 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4048 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4050 /* this drops dchild->d_inode->i_mutex unconditionally */
4051 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4055 switch(cleanup_phase) {
4058 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4059 oti->oti_handle : handle,
4060 filter_cancel_cookies_cb,
4062 /* If add_journal_cb failed, then filter_finish_transno
4063 * will commit the handle and we will do a sync
4064 * on commit. then we call callback directly to free
4067 rc = filter_finish_transno(exp, oti, rc, sync);
4069 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4072 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4074 CERROR("error on commit, err = %d\n", rc2);
4081 filter_parent_unlock(dparent);
4085 OBD_FREE(fcc, sizeof(*fcc));
4087 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4090 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4094 /* trigger quota release */
4095 qcids[USRQUOTA] = oa->o_uid;
4096 qcids[GRPQUOTA] = oa->o_gid;
4097 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4100 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4104 /* NB start and end are used for punch, but not truncate */
4105 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4106 struct obd_trans_info *oti,
4107 struct ptlrpc_request_set *rqset)
4112 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4113 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4114 oinfo->oi_policy.l_extent.end);
4118 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4119 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4120 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4122 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4123 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4124 rc = filter_setattr(exp, oinfo, oti);
4125 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4129 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4130 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4133 struct lvfs_run_ctxt saved;
4134 struct filter_obd *filter;
4135 struct dentry *dentry;
4139 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4140 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4144 filter = &exp->exp_obd->u.filter;
4146 /* An objid of zero is taken to mean "sync whole filesystem" */
4147 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4148 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4149 /* Flush any remaining cancel messages out to the target */
4150 filter_sync_llogs(exp->exp_obd, exp);
4154 dentry = filter_oa2dentry(exp->exp_obd, oa);
4156 RETURN(PTR_ERR(dentry));
4158 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4160 LOCK_INODE_MUTEX(dentry->d_inode);
4162 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4164 /* just any file to grab fsync method - "file" arg unused */
4165 struct file *file = filter->fo_rcvd_filp;
4167 if (file->f_op && file->f_op->fsync)
4168 rc = file->f_op->fsync(NULL, dentry, 1);
4170 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4174 UNLOCK_INODE_MUTEX(dentry->d_inode);
4176 oa->o_valid = OBD_MD_FLID;
4177 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4179 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4185 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4186 void *key, __u32 *vallen, void *val,
4187 struct lov_stripe_md *lsm)
4189 struct obd_device *obd;
4192 obd = class_exp2obd(exp);
4194 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4198 if (KEY_IS(KEY_BLOCKSIZE)) {
4199 __u32 *blocksize = val;
4201 if (*vallen < sizeof(*blocksize))
4203 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4205 *vallen = sizeof(*blocksize);
4209 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4210 __u32 *blocksize_bits = val;
4211 if (blocksize_bits) {
4212 if (*vallen < sizeof(*blocksize_bits))
4214 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4216 *vallen = sizeof(*blocksize_bits);
4220 if (KEY_IS(KEY_LAST_ID)) {
4221 obd_id *last_id = val;
4222 /* FIXME: object groups */
4224 if (*vallen < sizeof(*last_id))
4226 *last_id = filter_last_id(&obd->u.filter,
4227 exp->exp_filter_data.fed_group);
4229 *vallen = sizeof(*last_id);
4233 if (KEY_IS(KEY_FIEMAP)) {
4234 struct ll_fiemap_info_key *fm_key = key;
4235 struct dentry *dentry;
4236 struct ll_user_fiemap *fiemap = val;
4237 struct lvfs_run_ctxt saved;
4240 if (fiemap == NULL) {
4241 *vallen = fiemap_count_to_size(
4242 fm_key->fiemap.fm_extent_count);
4246 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4249 RETURN(PTR_ERR(dentry));
4251 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4252 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4253 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4254 EXT3_IOC_FIEMAP, (long)fiemap);
4259 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4265 CDEBUG(D_IOCTL, "invalid key\n");
4269 static inline int filter_setup_llog_group(struct obd_export *exp,
4270 struct obd_device *obd,
4273 struct obd_llog_group *olg;
4274 struct llog_ctxt *ctxt;
4277 olg = filter_find_create_olg(obd, group);
4279 RETURN(PTR_ERR(olg));
4281 llog_group_set_export(olg, exp);
4283 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4284 LASSERTF(ctxt != NULL, "ctxt is null\n");
4286 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4287 llog_ctxt_put(ctxt);
4290 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4291 void *key, __u32 vallen, void *val,
4292 struct ptlrpc_request_set *set)
4294 struct obd_device *obd;
4300 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4304 if (KEY_IS(KEY_CAPA_KEY)) {
4305 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4307 CERROR("filter update capability key failed: %d\n", rc);
4311 if (KEY_IS(KEY_REVIMP_UPD)) {
4312 filter_revimp_update(exp);
4313 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4317 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4318 filter_adapt_sptlrpc_conf(obd, 0);
4322 if (!KEY_IS(KEY_MDS_CONN))
4325 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4326 obd_export_nid2str(exp));
4327 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4329 /* setup llog imports */
4330 LASSERT(val != NULL);
4332 group = (int)(*(__u32 *)val);
4333 LASSERT_MDS_GROUP(group);
4334 rc = filter_setup_llog_group(exp, obd, group);
4338 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4340 if (group == FILTER_GROUP_MDS0) {
4341 /* setup llog group 1 for interop */
4342 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4348 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4349 int len, void *karg, void *uarg)
4351 struct obd_device *obd = exp->exp_obd;
4352 struct obd_ioctl_data *data = karg;
4356 case OBD_IOC_ABORT_RECOVERY: {
4357 CERROR("aborting recovery for device %s\n", obd->obd_name);
4358 target_stop_recovery_thread(obd);
4362 case OBD_IOC_SYNC: {
4363 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4364 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4368 case OBD_IOC_SET_READONLY: {
4370 struct super_block *sb = obd->u.obt.obt_sb;
4371 struct inode *inode = sb->s_root->d_inode;
4372 BDEVNAME_DECLARE_STORAGE(tmp);
4373 CERROR("*** setting device %s read-only ***\n",
4374 ll_bdevname(sb, tmp));
4376 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4377 if (!IS_ERR(handle))
4378 rc = fsfilt_commit(obd, inode, handle, 1);
4380 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4381 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4383 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4387 case OBD_IOC_CATLOGLIST: {
4388 rc = llog_catalog_list(obd, 1, data);
4392 case OBD_IOC_LLOG_CANCEL:
4393 case OBD_IOC_LLOG_REMOVE:
4394 case OBD_IOC_LLOG_INFO:
4395 case OBD_IOC_LLOG_PRINT: {
4396 /* FIXME to be finished */
4397 RETURN(-EOPNOTSUPP);
4399 struct llog_ctxt *ctxt = NULL;
4401 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4402 rc = llog_ioctl(ctxt, cmd, data);
4403 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4416 static int filter_health_check(struct obd_device *obd)
4418 #ifdef USE_HEALTH_CHECK_WRITE
4419 struct filter_obd *filter = &obd->u.filter;
4424 * health_check to return 0 on healthy
4425 * and 1 on unhealthy.
4427 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4430 #ifdef USE_HEALTH_CHECK_WRITE
4431 LASSERT(filter->fo_health_check_filp != NULL);
4432 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4437 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4440 return filter_fid2dentry(data, NULL, gr, id);
4443 static int filter_process_config(struct obd_device *obd, obd_count len,
4446 struct lustre_cfg *lcfg = buf;
4447 struct lprocfs_static_vars lvars;
4450 switch (lcfg->lcfg_command) {
4452 lprocfs_filter_init_vars(&lvars);
4454 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4464 static struct lvfs_callback_ops filter_lvfs_ops = {
4465 l_fid2dentry: filter_lvfs_fid2dentry,
4468 static struct obd_ops filter_obd_ops = {
4469 .o_owner = THIS_MODULE,
4470 .o_get_info = filter_get_info,
4471 .o_set_info_async = filter_set_info_async,
4472 .o_setup = filter_setup,
4473 .o_precleanup = filter_precleanup,
4474 .o_cleanup = filter_cleanup,
4475 .o_connect = filter_connect,
4476 .o_reconnect = filter_reconnect,
4477 .o_disconnect = filter_disconnect,
4478 .o_ping = filter_ping,
4479 .o_init_export = filter_init_export,
4480 .o_destroy_export = filter_destroy_export,
4481 .o_statfs = filter_statfs,
4482 .o_getattr = filter_getattr,
4483 .o_unpackmd = filter_unpackmd,
4484 .o_create = filter_create,
4485 .o_setattr = filter_setattr,
4486 .o_destroy = filter_destroy,
4487 .o_brw = filter_brw,
4488 .o_punch = filter_truncate,
4489 .o_sync = filter_sync,
4490 .o_preprw = filter_preprw,
4491 .o_commitrw = filter_commitrw,
4492 .o_llog_init = filter_llog_init,
4493 .o_llog_connect = filter_llog_connect,
4494 .o_llog_finish = filter_llog_finish,
4495 .o_iocontrol = filter_iocontrol,
4496 .o_health_check = filter_health_check,
4497 .o_process_config = filter_process_config,
4500 quota_interface_t *filter_quota_interface_ref;
4501 extern quota_interface_t filter_quota_interface;
4503 static int __init obdfilter_init(void)
4505 struct lprocfs_static_vars lvars;
4508 lprocfs_filter_init_vars(&lvars);
4510 request_module("lquota");
4511 OBD_ALLOC(obdfilter_created_scratchpad,
4512 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4513 sizeof(*obdfilter_created_scratchpad));
4514 if (obdfilter_created_scratchpad == NULL)
4517 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4518 sizeof(struct filter_mod_data),
4521 GOTO(out, rc = -ENOMEM);
4523 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4524 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4526 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4527 LUSTRE_OST_NAME, NULL);
4531 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4532 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4533 ll_fmd_cachep = NULL;
4535 if (filter_quota_interface_ref)
4536 PORTAL_SYMBOL_PUT(filter_quota_interface);
4538 OBD_FREE(obdfilter_created_scratchpad,
4539 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4540 sizeof(*obdfilter_created_scratchpad));
4546 static void __exit obdfilter_exit(void)
4548 if (filter_quota_interface_ref)
4549 PORTAL_SYMBOL_PUT(filter_quota_interface);
4551 if (ll_fmd_cachep) {
4552 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4553 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4554 ll_fmd_cachep = NULL;
4557 class_unregister_type(LUSTRE_OST_NAME);
4558 OBD_FREE(obdfilter_created_scratchpad,
4559 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4560 sizeof(*obdfilter_created_scratchpad));
4563 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4564 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4565 MODULE_LICENSE("GPL");
4567 module_init(obdfilter_init);
4568 module_exit(obdfilter_exit);