1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct filter_obd *filter = &exp->exp_obd->u.filter;
124 struct filter_export_data *fed = &exp->exp_filter_data;
125 struct lsd_client_data *lcd = fed->fed_lcd;
128 int err, log_pri = D_RPCTRACE;
130 /* Propagate error code. */
134 if (!exp->exp_obd->obd_replayable || oti == NULL)
137 /* we don't allocate new transnos for replayed requests */
138 spin_lock(&filter->fo_translock);
139 if (oti->oti_transno == 0) {
140 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
141 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
143 last_rcvd = oti->oti_transno;
144 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
145 filter->fo_fsd->lsd_last_transno =
146 cpu_to_le64(last_rcvd);
148 oti->oti_transno = last_rcvd;
150 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
151 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
152 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
153 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
154 spin_unlock(&filter->fo_translock);
157 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
159 off = fed->fed_lr_off;
161 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
162 fed->fed_lr_idx, fed->fed_lr_off);
165 class_export_cb_get(exp); /* released when the cb is called */
167 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
173 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
174 lcd, sizeof(*lcd), &off,
175 force_sync | exp->exp_need_sync);
177 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
185 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
186 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
191 void f_dput(struct dentry *dentry)
193 /* Can't go inside filter_ddelete because it can block */
194 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
195 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
196 LASSERT(atomic_read(&dentry->d_count) > 0);
201 static void init_brw_stats(struct brw_stats *brw_stats)
204 for (i = 0; i < BRW_LAST; i++)
205 spin_lock_init(&brw_stats->hist[i].oh_lock);
208 static int lprocfs_init_rw_stats(struct obd_device *obd,
209 struct lprocfs_stats **stats)
213 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
214 LPROC_FILTER_LAST - 1;
215 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
219 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
220 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
221 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
222 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
223 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
228 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
229 plus the procfs overhead :( */
230 static int filter_export_stats_init(struct obd_device *obd,
231 struct obd_export *exp,
237 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
238 /* Self-export gets no proc entry */
241 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
243 /* Mask error for already created
251 struct nid_stat *tmp = exp->exp_nid_stats;
252 LASSERT(tmp != NULL);
254 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
255 if (tmp->nid_brw_stats == NULL)
258 init_brw_stats(tmp->nid_brw_stats);
259 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
260 0644, &filter_per_nid_stats_fops,
263 CWARN("Error adding the brw_stats file\n");
265 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
269 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
273 /* Always add in ldlm_stats */
274 tmp->nid_ldlm_stats =
275 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
276 LPROCFS_STATS_FLAG_NOPERCPU);
277 if (tmp->nid_ldlm_stats == NULL)
280 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
282 tmp->nid_ldlm_stats);
290 /* Add client data to the FILTER. We use a bitmap to locate a free space
291 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
292 * Otherwise, we have just read the data from the last_rcvd file and
293 * we know its offset. */
294 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
297 struct filter_obd *filter = &obd->u.filter;
298 struct filter_export_data *fed = &exp->exp_filter_data;
299 unsigned long *bitmap = filter->fo_last_rcvd_slots;
300 int new_client = (cl_idx == -1);
304 LASSERT(bitmap != NULL);
305 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
308 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
311 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
312 * there's no need for extra complication here
315 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
317 if (cl_idx >= LR_MAX_CLIENTS) {
318 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
322 if (test_and_set_bit(cl_idx, bitmap)) {
323 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
328 if (test_and_set_bit(cl_idx, bitmap)) {
329 CERROR("FILTER client %d: bit already set in bitmap!\n",
335 fed->fed_lr_idx = cl_idx;
336 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
337 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
338 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
340 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
341 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
344 struct lvfs_run_ctxt saved;
345 loff_t off = fed->fed_lr_off;
349 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
350 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
352 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
353 /* Transaction needed to fix bug 1403 */
354 handle = fsfilt_start(obd,
355 filter->fo_rcvd_filp->f_dentry->d_inode,
356 FSFILT_OP_SETATTR, NULL);
357 if (IS_ERR(handle)) {
358 rc = PTR_ERR(handle);
359 CERROR("unable to start transaction: rc %d\n", rc);
361 fed->fed_lcd->lcd_last_epoch =
362 filter->fo_fsd->lsd_start_epoch;
363 exp->exp_last_request_time = cfs_time_current_sec();
364 rc = fsfilt_add_journal_cb(obd, 0, handle,
365 target_client_add_cb, exp);
367 spin_lock(&exp->exp_lock);
368 exp->exp_need_sync = 1;
369 spin_unlock(&exp->exp_lock);
371 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
373 sizeof(*fed->fed_lcd),
374 &off, rc /* sync if no cb */);
376 filter->fo_rcvd_filp->f_dentry->d_inode,
379 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
382 CERROR("error writing %s client idx %u: rc %d\n",
383 LAST_RCVD, fed->fed_lr_idx, rc);
390 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
392 static int filter_client_free(struct obd_export *exp)
394 struct filter_export_data *fed = &exp->exp_filter_data;
395 struct filter_obd *filter = &exp->exp_obd->u.filter;
396 struct obd_device *obd = exp->exp_obd;
397 struct lvfs_run_ctxt saved;
402 if (fed->fed_lcd == NULL)
405 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
406 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
409 LASSERT(filter->fo_last_rcvd_slots != NULL);
411 off = fed->fed_lr_off;
413 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
414 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
416 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
417 * we leak a client slot that will be cleaned on the next recovery. */
419 CERROR("%s: client idx %d has med_off %lld\n",
420 obd->obd_name, fed->fed_lr_idx, off);
421 GOTO(free, rc = -EINVAL);
424 /* Clear the bit _after_ zeroing out the client so we don't
425 race with filter_client_add and zero out new clients.*/
426 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
427 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
432 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
433 /* Don't force sync on disconnect if aborting recovery,
434 * or it does num_clients * num_osts. b=17194 */
435 int need_sync = exp->exp_need_sync &&
436 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
437 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
438 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
439 sizeof(zero_lcd), &off, 0);
441 /* Make sure the server's last_transno is up to date. Do this
442 * after the client is freed so we know all the client's
443 * transactions have been committed. */
445 filter_update_server_data(obd, filter->fo_rcvd_filp,
446 filter->fo_fsd, need_sync);
447 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
449 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
450 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
451 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
452 LAST_RCVD, need_sync ? "" : "a", rc);
455 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
456 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
463 OBD_FREE_PTR(fed->fed_lcd);
469 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
470 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
471 struct filter_mod_data *fmd)
473 LASSERT_SPIN_LOCKED(&fed->fed_lock);
474 if (--fmd->fmd_refcount == 0) {
475 /* XXX when we have persistent reservations and the handle
476 * is stored herein we need to drop it here. */
477 fed->fed_mod_count--;
478 list_del(&fmd->fmd_list);
479 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
483 /* drop fmd reference, free it if last ref */
484 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
486 struct filter_export_data *fed;
491 fed = &exp->exp_filter_data;
492 spin_lock(&fed->fed_lock);
493 filter_fmd_put_nolock(fed, fmd); /* caller reference */
494 spin_unlock(&fed->fed_lock);
497 /* expire entries from the end of the list if there are too many
498 * or they are too old */
499 static void filter_fmd_expire_nolock(struct filter_obd *filter,
500 struct filter_export_data *fed,
501 struct filter_mod_data *keep)
503 struct filter_mod_data *fmd, *tmp;
505 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
509 if (time_before(jiffies, fmd->fmd_expire) &&
510 fed->fed_mod_count < filter->fo_fmd_max_num)
513 list_del_init(&fmd->fmd_list);
514 filter_fmd_put_nolock(fed, fmd); /* list reference */
518 void filter_fmd_expire(struct obd_export *exp)
520 spin_lock(&exp->exp_filter_data.fed_lock);
521 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
522 &exp->exp_filter_data, NULL);
523 spin_unlock(&exp->exp_filter_data.fed_lock);
526 /* find specified objid, group in export fmd list.
527 * caller must hold fed_lock and take fmd reference itself */
528 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
529 struct filter_export_data *fed,
530 obd_id objid, obd_gr group)
532 struct filter_mod_data *found = NULL, *fmd;
534 LASSERT_SPIN_LOCKED(&fed->fed_lock);
536 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
537 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
539 list_del(&fmd->fmd_list);
540 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
541 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
546 filter_fmd_expire_nolock(filter, fed, found);
551 /* Find fmd based on objid and group, or return NULL if not found. */
552 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
553 obd_id objid, obd_gr group)
555 struct filter_mod_data *fmd;
557 spin_lock(&exp->exp_filter_data.fed_lock);
558 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
559 &exp->exp_filter_data, objid, group);
561 fmd->fmd_refcount++; /* caller reference */
562 spin_unlock(&exp->exp_filter_data.fed_lock);
567 /* Find fmd based on objid and group, or create a new one if none is found.
568 * It is possible for this function to return NULL under memory pressure,
569 * or if objid = 0 is passed (which will only cause old entries to expire).
570 * Currently this is not fatal because any fmd state is transient and
571 * may also be freed when it gets sufficiently old. */
572 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
573 obd_id objid, obd_gr group)
575 struct filter_export_data *fed = &exp->exp_filter_data;
576 struct filter_mod_data *found = NULL, *fmd_new = NULL;
578 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
580 spin_lock(&fed->fed_lock);
581 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
584 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
585 fmd_new->fmd_id = objid;
586 fmd_new->fmd_gr = group;
587 fmd_new->fmd_refcount++; /* list reference */
589 fed->fed_mod_count++;
591 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
595 found->fmd_refcount++; /* caller reference */
596 found->fmd_expire = jiffies +
597 exp->exp_obd->u.filter.fo_fmd_max_age;
600 spin_unlock(&fed->fed_lock);
606 /* drop fmd list reference so it will disappear when last reference is put.
607 * This isn't so critical because it would in fact only affect the one client
608 * that is doing the unlink and at worst we have an stale entry referencing
609 * an object that should never be used again. */
610 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
612 struct filter_mod_data *found = NULL;
614 spin_lock(&exp->exp_filter_data.fed_lock);
615 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
617 list_del_init(&found->fmd_list);
618 filter_fmd_put_nolock(&exp->exp_filter_data, found);
620 spin_unlock(&exp->exp_filter_data.fed_lock);
623 #define filter_fmd_drop(exp, objid, group)
626 /* remove all entries from fmd list */
627 static void filter_fmd_cleanup(struct obd_export *exp)
629 struct filter_export_data *fed = &exp->exp_filter_data;
630 struct filter_mod_data *fmd = NULL, *tmp;
632 spin_lock(&fed->fed_lock);
633 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
634 list_del_init(&fmd->fmd_list);
635 filter_fmd_put_nolock(fed, fmd);
637 spin_unlock(&fed->fed_lock);
640 static int filter_init_export(struct obd_export *exp)
642 spin_lock_init(&exp->exp_filter_data.fed_lock);
643 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
645 spin_lock(&exp->exp_lock);
646 exp->exp_connecting = 1;
647 spin_unlock(&exp->exp_lock);
649 return ldlm_init_export(exp);
652 static int filter_free_server_data(struct filter_obd *filter)
654 OBD_FREE_PTR(filter->fo_fsd);
655 filter->fo_fsd = NULL;
656 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
657 filter->fo_last_rcvd_slots = NULL;
661 /* assumes caller is already in kernel ctxt */
662 int filter_update_server_data(struct obd_device *obd, struct file *filp,
663 struct lr_server_data *fsd, int force_sync)
669 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
670 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
671 le64_to_cpu(fsd->lsd_last_transno));
672 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
673 le64_to_cpu(fsd->lsd_mount_count));
675 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
677 CERROR("error writing lr_server_data: rc = %d\n", rc);
682 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
685 struct filter_obd *filter = &obd->u.filter;
691 if (filter->fo_last_objid_files[group] == NULL) {
692 CERROR("Object group "LPU64" not fully setup; not updating "
693 "last_objid\n", group);
697 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
698 obd->obd_name, group, filter->fo_last_objids[group]);
700 tmp = cpu_to_le64(filter->fo_last_objids[group]);
701 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
702 &tmp, sizeof(tmp), &off, force_sync);
704 CERROR("error writing group "LPU64" last objid: rc = %d\n",
708 extern int ost_handle(struct ptlrpc_request *req);
709 /* assumes caller has already in kernel ctxt */
710 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
712 struct filter_obd *filter = &obd->u.filter;
713 struct lr_server_data *fsd;
714 struct lsd_client_data *lcd = NULL;
715 struct inode *inode = filp->f_dentry->d_inode;
716 unsigned long last_rcvd_size = i_size_read(inode);
723 /* ensure padding in the struct is the correct size */
724 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
725 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
726 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
727 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
729 OBD_ALLOC(fsd, sizeof(*fsd));
732 filter->fo_fsd = fsd;
734 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
735 if (filter->fo_last_rcvd_slots == NULL) {
736 OBD_FREE(fsd, sizeof(*fsd));
740 if (last_rcvd_size == 0) {
741 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
743 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
744 fsd->lsd_last_transno = 0;
745 mount_count = fsd->lsd_mount_count = 0;
746 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
747 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
748 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
749 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
750 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
751 /* OBD_COMPAT_OST is set in filter_connect_internal when the
752 * MDS first connects and assigns the OST index number. */
753 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
756 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
758 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
762 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
763 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
764 "using the wrong disk %s. Were the "
765 "/dev/ assignments rearranged?\n",
766 obd->obd_uuid.uuid, fsd->lsd_uuid);
767 GOTO(err_fsd, rc = -EINVAL);
769 mount_count = le64_to_cpu(fsd->lsd_mount_count);
770 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
772 /* Assume old last_rcvd format unless I_C_LR is set */
773 if (!(fsd->lsd_feature_incompat &
774 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
775 fsd->lsd_last_transno = fsd->lsd_compat14;
777 /* OBD_COMPAT_OST is set in filter_connect_internal when the
778 * MDS first connects and assigns the OST index number. */
779 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
783 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
784 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
785 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
786 ~FILTER_INCOMPAT_SUPP);
787 GOTO(err_fsd, rc = -EINVAL);
789 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
790 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
791 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
792 ~FILTER_ROCOMPAT_SUPP);
793 /* Do something like remount filesystem read-only */
794 GOTO(err_fsd, rc = -EINVAL);
797 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
799 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
800 obd->obd_name, start_epoch);
801 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
802 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
803 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
804 obd->obd_name, mount_count + 1);
805 CDEBUG(D_INODE, "%s: server data size: %u\n",
806 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
807 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
808 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
809 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
810 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
811 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
812 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
813 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
814 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
815 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
816 le16_to_cpu(fsd->lsd_client_size));
818 if (!obd->obd_replayable) {
819 CWARN("%s: recovery support OFF\n", obd->obd_name);
823 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
824 off < last_rcvd_size; cl_idx++) {
826 struct obd_export *exp;
827 struct filter_export_data *fed;
832 GOTO(err_client, rc = -ENOMEM);
835 /* Don't assume off is incremented properly by
836 * fsfilt_read_record(), in case sizeof(*lcd)
837 * isn't the same as fsd->lsd_client_size. */
838 off = le32_to_cpu(fsd->lsd_client_start) +
839 cl_idx * le16_to_cpu(fsd->lsd_client_size);
840 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
842 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
843 LAST_RCVD, cl_idx, off, rc);
844 break; /* read error shouldn't cause startup to fail */
847 if (lcd->lcd_uuid[0] == '\0') {
848 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
853 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
855 /* These exports are cleaned up by filter_disconnect(), so they
856 * need to be set up like real exports as filter_connect() does.
858 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
860 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
861 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
862 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
864 if (PTR_ERR(exp) == -EALREADY) {
865 /* export already exists, zero out this one */
866 CERROR("Zeroing out duplicate export due to "
868 lcd->lcd_uuid[0] = '\0';
870 GOTO(err_client, rc = PTR_ERR(exp));
873 fed = &exp->exp_filter_data;
875 fed->fed_group = 0; /* will be assigned at connect */
876 filter_export_stats_init(obd, exp, NULL);
877 rc = filter_client_add(obd, exp, cl_idx);
878 /* can't fail for existing client */
879 LASSERTF(rc == 0, "rc = %d\n", rc);
881 /* VBR: set export last committed */
882 exp->exp_last_committed = last_rcvd;
883 spin_lock(&exp->exp_lock);
884 exp->exp_connecting = 0;
885 exp->exp_in_recovery = 0;
886 spin_unlock(&exp->exp_lock);
887 spin_lock_bh(&obd->obd_processing_task_lock);
888 obd->obd_max_recoverable_clients++;
889 spin_unlock_bh(&obd->obd_processing_task_lock);
891 class_export_put(exp);
894 /* Need to check last_rcvd even for duplicated exports. */
895 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
898 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
899 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
905 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
907 filter->fo_mount_count = mount_count + 1;
908 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
910 /* save it, so mount count and last_transno is current */
911 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
913 GOTO(err_client, rc);
918 class_disconnect_exports(obd);
920 filter_free_server_data(filter);
924 static int filter_cleanup_groups(struct obd_device *obd)
926 struct filter_obd *filter = &obd->u.filter;
928 struct dentry *dentry;
932 if (filter->fo_dentry_O_groups != NULL) {
933 for (i = 0; i < filter->fo_group_count; i++) {
934 dentry = filter->fo_dentry_O_groups[i];
938 OBD_FREE(filter->fo_dentry_O_groups,
939 filter->fo_group_count *
940 sizeof(*filter->fo_dentry_O_groups));
941 filter->fo_dentry_O_groups = NULL;
943 if (filter->fo_last_objid_files != NULL) {
944 for (i = 0; i < filter->fo_group_count; i++) {
945 filp = filter->fo_last_objid_files[i];
949 OBD_FREE(filter->fo_last_objid_files,
950 filter->fo_group_count *
951 sizeof(*filter->fo_last_objid_files));
952 filter->fo_last_objid_files = NULL;
954 if (filter->fo_dentry_O_sub != NULL) {
955 for (i = 0; i < filter->fo_group_count; i++) {
956 for (j = 0; j < filter->fo_subdir_count; j++) {
957 dentry = filter->fo_dentry_O_sub[i].dentry[j];
962 OBD_FREE(filter->fo_dentry_O_sub,
963 filter->fo_group_count *
964 sizeof(*filter->fo_dentry_O_sub));
965 filter->fo_dentry_O_sub = NULL;
967 if (filter->fo_last_objids != NULL) {
968 OBD_FREE(filter->fo_last_objids,
969 filter->fo_group_count *
970 sizeof(*filter->fo_last_objids));
971 filter->fo_last_objids = NULL;
973 if (filter->fo_dentry_O != NULL) {
974 f_dput(filter->fo_dentry_O);
975 filter->fo_dentry_O = NULL;
980 static int filter_update_last_group(struct obd_device *obd, int group)
982 struct filter_obd *filter = &obd->u.filter;
983 struct file *filp = NULL;
984 int last_group = 0, rc;
988 if (group <= filter->fo_committed_group)
991 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
995 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
999 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1001 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1005 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1006 obd->obd_name, last_group, group);
1010 /* must be sync: bXXXX */
1011 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1013 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1017 filter->fo_committed_group = group;
1020 filp_close(filp, 0);
1024 static int filter_read_group_internal(struct obd_device *obd, int group,
1027 struct filter_obd *filter = &obd->u.filter;
1028 __u64 *new_objids = NULL;
1029 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1030 struct dentry **new_groups = NULL;
1031 struct file **new_files = NULL;
1032 struct dentry *dentry;
1034 int old_count = filter->fo_group_count, rc, stage = 0, i;
1038 int len = group + 1;
1040 snprintf(name, 24, "%d", group);
1044 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1046 if (IS_ERR(dentry)) {
1047 CERROR("Cannot lookup expected object group %d: %ld\n",
1048 group, PTR_ERR(dentry));
1049 RETURN(PTR_ERR(dentry));
1052 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1054 if (IS_ERR(dentry)) {
1055 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1057 RETURN(PTR_ERR(dentry));
1062 snprintf(name, 24, "O/%d/LAST_ID", group);
1064 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1066 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1067 GOTO(cleanup, rc = PTR_ERR(filp));
1071 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1073 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1077 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1078 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1079 if (tmp_subdirs == NULL)
1080 GOTO(cleanup, rc = -ENOMEM);
1083 for (i = 0; i < filter->fo_subdir_count; i++) {
1085 snprintf(dir, sizeof(dir), "d%u", i);
1087 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1090 if (IS_ERR(tmp_subdirs->dentry[i])) {
1091 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1092 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1097 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1098 tmp_subdirs->dentry[i]);
1102 /* 'group' is an index; we need an array of length 'group + 1' */
1103 if (group + 1 > old_count) {
1104 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1105 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1106 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1107 OBD_ALLOC(new_files, len * sizeof(*new_files));
1109 if (new_objids == NULL || new_subdirs == NULL ||
1110 new_groups == NULL || new_files == NULL)
1111 GOTO(cleanup, rc = -ENOMEM);
1114 memcpy(new_objids, filter->fo_last_objids,
1115 old_count * sizeof(*new_objids));
1116 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1117 old_count * sizeof(*new_subdirs));
1118 memcpy(new_groups, filter->fo_dentry_O_groups,
1119 old_count * sizeof(*new_groups));
1120 memcpy(new_files, filter->fo_last_objid_files,
1121 old_count * sizeof(*new_files));
1123 OBD_FREE(filter->fo_last_objids,
1124 old_count * sizeof(*new_objids));
1125 OBD_FREE(filter->fo_dentry_O_sub,
1126 old_count * sizeof(*new_subdirs));
1127 OBD_FREE(filter->fo_dentry_O_groups,
1128 old_count * sizeof(*new_groups));
1129 OBD_FREE(filter->fo_last_objid_files,
1130 old_count * sizeof(*new_files));
1132 filter->fo_last_objids = new_objids;
1133 filter->fo_dentry_O_sub = new_subdirs;
1134 filter->fo_dentry_O_groups = new_groups;
1135 filter->fo_last_objid_files = new_files;
1136 filter->fo_group_count = len;
1139 filter->fo_dentry_O_groups[group] = dentry;
1140 filter->fo_last_objid_files[group] = filp;
1141 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1142 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1143 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1146 filter_update_last_group(obd, group);
1148 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1149 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1150 rc = filter_update_last_objid(obd, group, 1);
1154 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1155 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1156 obd->obd_name, group, last_objid);
1161 if (new_objids != NULL)
1162 OBD_FREE(new_objids, len * sizeof(*new_objids));
1163 if (new_subdirs != NULL)
1164 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1165 if (new_groups != NULL)
1166 OBD_FREE(new_groups, len * sizeof(*new_groups));
1167 if (new_files != NULL)
1168 OBD_FREE(new_files, len * sizeof(*new_files));
1170 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1171 for (i = 0; i < filter->fo_subdir_count; i++) {
1172 if (tmp_subdirs->dentry[i] != NULL)
1173 dput(tmp_subdirs->dentry[i]);
1175 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1178 filp_close(filp, 0);
1185 static int filter_read_groups(struct obd_device *obd, int last_group,
1188 struct filter_obd *filter = &obd->u.filter;
1189 int old_count, group, rc = 0;
1191 down(&filter->fo_init_lock);
1192 old_count = filter->fo_group_count;
1193 for (group = old_count; group <= last_group; group++) {
1194 rc = filter_read_group_internal(obd, group, create);
1198 up(&filter->fo_init_lock);
1202 /* FIXME: object groups */
1203 static int filter_prep_groups(struct obd_device *obd)
1205 struct filter_obd *filter = &obd->u.filter;
1206 struct dentry *O_dentry;
1208 int last_group, rc = 0, cleanup_phase = 0;
1212 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1214 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1215 if (IS_ERR(O_dentry)) {
1216 rc = PTR_ERR(O_dentry);
1217 CERROR("cannot open/create O: rc = %d\n", rc);
1220 filter->fo_dentry_O = O_dentry;
1221 cleanup_phase = 1; /* O_dentry */
1223 /* we have to initialize all groups before first connections from
1224 * clients because they may send create/destroy for any group -bzzz */
1225 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1227 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1228 GOTO(cleanup, rc = PTR_ERR(filp));
1230 cleanup_phase = 2; /* filp */
1232 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1234 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1239 last_group = FILTER_GROUP_MDS0;
1241 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1242 FILTER_GROUP_MDS0, last_group);
1243 filter->fo_committed_group = last_group;
1244 rc = filter_read_groups(obd, last_group, 1);
1248 filp_close(filp, 0);
1252 switch (cleanup_phase) {
1254 filp_close(filp, 0);
1256 filter_cleanup_groups(obd);
1257 f_dput(filter->fo_dentry_O);
1258 filter->fo_dentry_O = NULL;
1266 /* setup the object store with correct subdirectories */
1267 static int filter_prep(struct obd_device *obd)
1269 struct lvfs_run_ctxt saved;
1270 struct filter_obd *filter = &obd->u.filter;
1272 struct inode *inode;
1276 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1277 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1278 if (!file || IS_ERR(file)) {
1280 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1284 filter->fo_rcvd_filp = file;
1285 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1286 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1287 file->f_dentry->d_inode->i_mode);
1288 GOTO(err_filp, rc = -ENOENT);
1291 inode = file->f_dentry->d_parent->d_inode;
1292 /* We use i_op->unlink directly in filter_vfs_unlink() */
1293 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1294 CERROR("%s: filesystem does not support create/unlink ops\n",
1296 GOTO(err_filp, rc = -EOPNOTSUPP);
1299 /** lu_target has very limited use in filter now */
1300 lut_init(NULL, &filter->fo_lut, obd, NULL);
1302 rc = filter_init_server_data(obd, file);
1304 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1308 target_recovery_init(&filter->fo_lut, ost_handle);
1310 /* open and create health check io file*/
1311 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1314 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1316 GOTO(err_server_data, rc);
1318 filter->fo_obt.obt_health_check_filp = file;
1319 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1320 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1321 file->f_dentry->d_inode->i_mode);
1322 GOTO(err_health_check, rc = -ENOENT);
1324 rc = lvfs_check_io_health(obd, file);
1326 GOTO(err_health_check, rc);
1328 rc = filter_prep_groups(obd);
1330 GOTO(err_health_check, rc);
1332 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1337 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1338 CERROR("can't close %s after error\n", HEALTH_CHECK);
1339 filter->fo_obt.obt_health_check_filp = NULL;
1341 target_recovery_fini(obd);
1342 filter_free_server_data(filter);
1344 if (filp_close(filter->fo_rcvd_filp, 0))
1345 CERROR("can't close %s after error\n", LAST_RCVD);
1346 filter->fo_rcvd_filp = NULL;
1350 /* cleanup the filter: write last used object id to status file */
1351 static void filter_post(struct obd_device *obd)
1353 struct lvfs_run_ctxt saved;
1354 struct filter_obd *filter = &obd->u.filter;
1357 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1358 * best to start a transaction with h_sync, because we removed this
1361 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1362 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1365 CERROR("error writing server data: rc = %d\n", rc);
1367 for (i = 0; i < filter->fo_group_count; i++) {
1368 rc = filter_update_last_objid(obd, i,
1369 (i == filter->fo_group_count - 1));
1371 CERROR("error writing group %d lastobjid: rc = %d\n",
1375 rc = filp_close(filter->fo_rcvd_filp, 0);
1376 filter->fo_rcvd_filp = NULL;
1378 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1380 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1381 filter->fo_obt.obt_health_check_filp = NULL;
1383 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1385 filter_cleanup_groups(obd);
1386 filter_free_server_data(filter);
1387 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1389 filter_free_capa_keys(filter);
1390 cleanup_capa_hash(filter->fo_capa_hash);
1393 static void filter_set_last_id(struct filter_obd *filter,
1394 obd_id id, obd_gr group)
1396 LASSERT(filter->fo_fsd != NULL);
1397 LASSERT(group <= filter->fo_group_count);
1399 spin_lock(&filter->fo_objidlock);
1400 filter->fo_last_objids[group] = id;
1401 spin_unlock(&filter->fo_objidlock);
1404 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1407 LASSERT(filter->fo_fsd != NULL);
1408 LASSERT(group <= filter->fo_group_count);
1410 /* FIXME: object groups */
1411 spin_lock(&filter->fo_objidlock);
1412 id = filter->fo_last_objids[group];
1413 spin_unlock(&filter->fo_objidlock);
1417 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1419 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1423 /* We never dget the object parent, so DON'T dput it either */
1424 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1426 struct filter_obd *filter = &obd->u.filter;
1427 struct filter_subdirs *subdirs;
1428 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1430 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1431 return filter->fo_dentry_O_groups[group];
1433 subdirs = &filter->fo_dentry_O_sub[group];
1434 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1437 /* We never dget the object parent, so DON'T dput it either */
1438 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1441 unsigned long now = jiffies;
1442 struct dentry *dparent = filter_parent(obd, group, objid);
1445 if (IS_ERR(dparent))
1447 if (dparent == NULL)
1448 return ERR_PTR(-ENOENT);
1450 rc = filter_lock_dentry(obd, dparent);
1451 fsfilt_check_slow(obd, now, "parent lock");
1452 return rc ? ERR_PTR(rc) : dparent;
1455 /* We never dget the object parent, so DON'T dput it either */
1456 static void filter_parent_unlock(struct dentry *dparent)
1458 UNLOCK_INODE_MUTEX(dparent->d_inode);
1461 /* How to get files, dentries, inodes from object id's.
1463 * If dir_dentry is passed, the caller has already locked the parent
1464 * appropriately for this operation (normally a write lock). If
1465 * dir_dentry is NULL, we do a read lock while we do the lookup to
1466 * avoid races with create/destroy and such changing the directory
1467 * internal to the filesystem code. */
1468 struct dentry *filter_fid2dentry(struct obd_device *obd,
1469 struct dentry *dir_dentry,
1470 obd_gr group, obd_id id)
1472 struct dentry *dparent = dir_dentry;
1473 struct dentry *dchild;
1478 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1479 obd->u.filter.fo_destroys_in_progress == 0) {
1480 /* don't fail lookups for orphan recovery, it causes
1481 * later LBUGs when objects still exist during precreate */
1482 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1483 RETURN(ERR_PTR(-ENOENT));
1486 CERROR("fatal: invalid object id 0\n");
1487 RETURN(ERR_PTR(-ESTALE));
1490 len = sprintf(name, LPU64, id);
1491 if (dir_dentry == NULL) {
1492 dparent = filter_parent_lock(obd, group, id);
1493 if (IS_ERR(dparent)) {
1494 CERROR("%s: error getting object "LPU64":"LPU64
1495 " parent: rc %ld\n", obd->obd_name,
1496 id, group, PTR_ERR(dparent));
1500 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1501 dparent->d_name.len, dparent->d_name.name, name);
1502 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1503 if (dir_dentry == NULL)
1504 filter_parent_unlock(dparent);
1505 if (IS_ERR(dchild)) {
1506 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1511 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1512 CERROR("%s: got bad object "LPU64" inode %lu\n",
1513 obd->obd_name, id, dchild->d_inode->i_ino);
1515 RETURN(ERR_PTR(-ENOENT));
1518 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1519 name, dchild, atomic_read(&dchild->d_count));
1521 LASSERT(atomic_read(&dchild->d_count) > 0);
1526 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1527 obd_id group, struct lustre_handle *lockh)
1529 int flags = LDLM_AST_DISCARD_DATA, rc;
1530 struct ldlm_res_id res_id;
1531 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1534 osc_build_res_name(objid, group, &res_id);
1535 /* Tell the clients that the object is gone now and that they should
1536 * throw away any cached pages. */
1537 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1538 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1539 ldlm_completion_ast, NULL, NULL, 0, NULL,
1546 static void filter_fini_destroy(struct obd_device *obd,
1547 struct lustre_handle *lockh)
1550 ldlm_lock_decref(lockh, LCK_PW);
1553 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1554 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1555 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1556 * reverse. Caller must take i_sem before starting the transaction and we
1557 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1558 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1559 struct vfsmount *mnt)
1564 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1565 * itself we already hold dir->i_mutex for child create/unlink ops */
1566 LASSERT(dentry->d_inode != NULL);
1567 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1568 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1572 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1573 GOTO(out, rc = -ENOENT);
1575 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1580 GOTO(out, rc = -EPERM);
1582 /* check_sticky() */
1583 if ((dentry->d_inode->i_uid != current->fsuid &&
1584 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1585 IS_IMMUTABLE(dentry->d_inode))
1586 GOTO(out, rc = -EPERM);
1588 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1589 * that was done because of journal_start (which is already done
1590 * here) or some other ordering issue. */
1593 rc = ll_security_inode_unlink(dir, dentry, mnt);
1597 rc = dir->i_op->unlink(dir, dentry);
1599 /* need to drop i_mutex before we lose inode reference */
1600 UNLOCK_INODE_MUTEX(dentry->d_inode);
1607 /* Caller must hold LCK_PW on parent and push us into kernel context.
1608 * Caller must hold child i_mutex, we drop it always.
1609 * Caller is also required to ensure that dchild->d_inode exists. */
1610 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1611 obd_gr group, struct dentry *dparent,
1612 struct dentry *dchild)
1614 struct inode *inode = dchild->d_inode;
1615 struct filter_obd *filter = &obd->u.filter;
1618 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1619 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1620 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1621 (unsigned long)inode->i_nlink,
1622 atomic_read(&inode->i_count));
1625 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1627 CERROR("error unlinking objid %.*s: rc %d\n",
1628 dchild->d_name.len, dchild->d_name.name, rc);
1632 struct filter_intent_args {
1633 struct ldlm_lock **victim;
1638 static enum interval_iter filter_intent_cb(struct interval_node *n,
1641 struct ldlm_interval *node = (struct ldlm_interval *)n;
1642 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1643 __u64 size = arg->size;
1644 struct ldlm_lock **v = arg->victim;
1645 struct ldlm_lock *lck;
1647 /* If the interval is lower than the current file size,
1649 if (interval_high(n) <= size)
1650 return INTERVAL_ITER_STOP;
1652 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1653 /* Don't send glimpse ASTs to liblustre clients.
1654 * They aren't listening for them, and they do
1655 * entirely synchronous I/O anyways. */
1656 if (lck->l_export == NULL ||
1657 lck->l_export->exp_libclient == 1)
1660 if (*arg->liblustre)
1661 *arg->liblustre = 0;
1664 *v = LDLM_LOCK_GET(lck);
1665 } else if ((*v)->l_policy_data.l_extent.start <
1666 lck->l_policy_data.l_extent.start) {
1667 LDLM_LOCK_RELEASE(*v);
1668 *v = LDLM_LOCK_GET(lck);
1671 /* the same policy group - every lock has the
1672 * same extent, so needn't do it any more */
1676 return INTERVAL_ITER_CONT;
1679 static int filter_intent_policy(struct ldlm_namespace *ns,
1680 struct ldlm_lock **lockp, void *req_cookie,
1681 ldlm_mode_t mode, int flags, void *data)
1683 CFS_LIST_HEAD(rpc_list);
1684 struct ptlrpc_request *req = req_cookie;
1685 struct ldlm_lock *lock = *lockp, *l = NULL;
1686 struct ldlm_resource *res = lock->l_resource;
1687 ldlm_processing_policy policy;
1688 struct ost_lvb *res_lvb, *reply_lvb;
1689 struct ldlm_reply *rep;
1691 int idx, rc, tmpflags = 0, only_liblustre = 1;
1692 struct ldlm_interval_tree *tree;
1693 struct filter_intent_args arg;
1694 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1695 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1696 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1699 policy = ldlm_get_processing_policy(res);
1700 LASSERT(policy != NULL);
1701 LASSERT(req != NULL);
1703 rc = lustre_pack_reply(req, 3, repsize, NULL);
1705 RETURN(req->rq_status = rc);
1707 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1708 LASSERT(rep != NULL);
1710 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1711 sizeof(*reply_lvb));
1712 LASSERT(reply_lvb != NULL);
1714 //fixup_handle_for_resent_req(req, lock, &lockh);
1716 /* Call the extent policy function to see if our request can be
1717 * granted, or is blocked.
1718 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1719 * lock, and should not be granted if the lock will be blocked.
1722 LASSERT(ns == res->lr_namespace);
1724 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1725 check_res_locked(res);
1727 /* FIXME: we should change the policy function slightly, to not make
1728 * this list at all, since we just turn around and free it */
1729 while (!list_empty(&rpc_list)) {
1730 struct ldlm_lock *wlock =
1731 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1732 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1733 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1734 lock->l_flags &= ~LDLM_FL_CP_REQD;
1735 list_del_init(&wlock->l_cp_ast);
1736 LDLM_LOCK_RELEASE(wlock);
1739 /* The lock met with no resistance; we're finished. */
1740 if (rc == LDLM_ITER_CONTINUE) {
1741 /* do not grant locks to the liblustre clients: they cannot
1742 * handle ASTs robustly. We need to do this while still
1743 * holding ns_lock to avoid the lock remaining on the res_link
1744 * list (and potentially being added to l_pending_list by an
1745 * AST) when we are going to drop this lock ASAP. */
1746 if (lock->l_export->exp_libclient ||
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1748 ldlm_resource_unlink_lock(lock);
1749 err = ELDLM_LOCK_ABORTED;
1751 err = ELDLM_LOCK_REPLACED;
1757 /* Do not grant any lock, but instead send GL callbacks. The extent
1758 * policy nicely created a list of all PW locks for us. We will choose
1759 * the highest of those which are larger than the size in the LVB, if
1760 * any, and perform a glimpse callback. */
1761 res_lvb = res->lr_lvb_data;
1762 LASSERT(res_lvb != NULL);
1763 *reply_lvb = *res_lvb;
1766 * ->ns_lock guarantees that no new locks are granted, and,
1767 * therefore, that res->lr_lvb_data cannot increase beyond the
1768 * end of already granted lock. As a result, it is safe to
1769 * check against "stale" reply_lvb->lvb_size value without
1772 arg.size = reply_lvb->lvb_size;
1774 arg.liblustre = &only_liblustre;
1775 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1776 tree = &res->lr_itree[idx];
1777 if (tree->lit_mode == LCK_PR)
1780 interval_iterate_reverse(tree->lit_root,
1781 filter_intent_cb, &arg);
1785 /* There were no PW locks beyond the size in the LVB; finished. */
1787 if (only_liblustre) {
1788 /* If we discovered a liblustre client with a PW lock,
1789 * however, the LVB may be out of date! The LVB is
1790 * updated only on glimpse (which we don't do for
1791 * liblustre clients) and cancel (which the client
1792 * obviously has not yet done). So if it has written
1793 * data but kept the lock, the LVB is stale and needs
1794 * to be updated from disk.
1796 * Of course, this will all disappear when we switch to
1797 * taking liblustre locks on the OST. */
1798 ldlm_res_lvbo_update(res, NULL, 0, 1);
1800 RETURN(ELDLM_LOCK_ABORTED);
1804 * This check is for lock taken in filter_prepare_destroy() that does
1805 * not have l_glimpse_ast set. So the logic is: if there is a lock
1806 * with no l_glimpse_ast set, this object is being destroyed already.
1808 * Hence, if you are grabbing DLM locks on the server, always set
1809 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1811 if (l->l_glimpse_ast == NULL) {
1812 /* We are racing with unlink(); just return -ENOENT */
1813 rep->lock_policy_res1 = -ENOENT;
1817 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1818 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1819 /* Update the LVB from disk if the AST failed (this is a legal race) */
1821 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1822 * sending ast is not handled. This can result in lost client writes.
1825 ldlm_res_lvbo_update(res, NULL, 0, 1);
1828 *reply_lvb = *res_lvb;
1832 LDLM_LOCK_RELEASE(l);
1834 RETURN(ELDLM_LOCK_ABORTED);
1838 * per-obd_device iobuf pool.
1840 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1841 * allocations in write-path has to be minimized (see bug 5137).
1843 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1844 * OST threads (see ost_thread_{init,done}()).
1846 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1847 * at the OST layer there are only (potentially) multiple obd_device of type
1848 * unknown at the time of OST thread creation.
1850 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1851 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1852 * it's very own iobuf.
1856 * filter_kiobuf_pool_init()
1858 * filter_kiobuf_pool_done()
1860 * filter_iobuf_get()
1862 * operate on this array. They are "generic" in a sense that they don't depend
1863 * on actual type of iobuf's (the latter depending on Linux kernel version).
1867 * destroy pool created by filter_iobuf_pool_init
1869 static void filter_iobuf_pool_done(struct filter_obd *filter)
1871 struct filter_iobuf **pool;
1876 pool = filter->fo_iobuf_pool;
1878 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1879 if (pool[i] != NULL)
1880 filter_free_iobuf(pool[i]);
1882 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1883 filter->fo_iobuf_pool = NULL;
1888 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1890 struct filter_obd *filter = &obd->u.filter;
1891 struct sptlrpc_rule_set tmp_rset;
1894 sptlrpc_rule_set_init(&tmp_rset);
1895 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1897 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1902 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1904 write_lock(&filter->fo_sptlrpc_lock);
1905 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1906 filter->fo_sptlrpc_rset = tmp_rset;
1907 write_unlock(&filter->fo_sptlrpc_lock);
1913 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1915 static int filter_iobuf_pool_init(struct filter_obd *filter)
1922 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1924 if (filter->fo_iobuf_pool == NULL)
1927 filter->fo_iobuf_count = OSS_THREADS_MAX;
1932 /* Return iobuf allocated for @thread_id. We don't know in advance how
1933 * many threads there will be so we allocate a large empty array and only
1934 * fill in those slots that are actually in use.
1935 * If we haven't allocated a pool entry for this thread before, do so now. */
1936 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1938 int thread_id = (oti && oti->oti_thread) ?
1939 oti->oti_thread->t_id : -1;
1940 struct filter_iobuf *pool = NULL;
1941 struct filter_iobuf **pool_place = NULL;
1943 if (thread_id >= 0) {
1944 LASSERT(thread_id < filter->fo_iobuf_count);
1945 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1948 if (unlikely(pool == NULL)) {
1949 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1950 PTLRPC_MAX_BRW_PAGES);
1951 if (pool_place != NULL)
1958 /* mount the file system (secretly). lustre_cfg parameters are:
1961 * 3 = flags: failover=f, failout=n
1964 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1967 struct filter_obd *filter = &obd->u.filter;
1968 struct vfsmount *mnt;
1969 struct lustre_mount_info *lmi;
1970 struct obd_uuid uuid;
1978 if (lcfg->lcfg_bufcount < 3 ||
1979 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1980 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1983 lmi = server_get_mount(obd->obd_name);
1985 /* We already mounted in lustre_fill_super.
1986 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1987 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1989 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1991 /* old path - used by lctl */
1992 CERROR("Using old MDS mount method\n");
1993 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1994 MS_NOATIME|MS_NODIRATIME,
1995 lustre_cfg_string(lcfg, 1), option);
1998 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1999 lustre_cfg_string(lcfg, 1), rc);
2003 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2005 if (IS_ERR(obd->obd_fsops))
2006 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2008 rc = filter_iobuf_pool_init(filter);
2012 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2013 CERROR("%s: Underlying device is marked as read-only. "
2014 "Setup failed\n", obd->obd_name);
2015 GOTO(err_ops, rc = -EROFS);
2018 /* failover is the default */
2019 obd->obd_replayable = 1;
2021 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2022 str = lustre_cfg_string(lcfg, 3);
2023 if (strchr(str, 'n')) {
2024 CWARN("%s: recovery disabled\n", obd->obd_name);
2025 obd->obd_replayable = 0;
2029 filter->fo_vfsmnt = mnt;
2030 obd->u.obt.obt_sb = mnt->mnt_sb;
2031 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2032 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2034 fsfilt_setup(obd, obd->u.obt.obt_sb);
2036 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2037 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2038 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2039 obd->obd_lvfs_ctxt.fs = get_ds();
2040 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2042 init_mutex(&filter->fo_init_lock);
2043 filter->fo_committed_group = 0;
2044 filter->fo_destroys_in_progress = 0;
2045 for (i = 0; i < 32; i++)
2046 sema_init(&filter->fo_create_locks[i], 1);
2048 spin_lock_init(&filter->fo_translock);
2049 spin_lock_init(&filter->fo_objidlock);
2050 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2051 sema_init(&filter->fo_alloc_lock, 1);
2052 init_brw_stats(&filter->fo_filter_stats);
2053 filter->fo_read_cache = 1; /* enable read-only cache by default */
2054 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2055 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2056 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2057 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2059 rc = filter_prep(obd);
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;
2068 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2069 filter->fo_capa_hash = init_capa_hash();
2070 if (filter->fo_capa_hash == NULL)
2071 GOTO(err_post, rc = -ENOMEM);
2073 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2074 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2075 LDLM_NAMESPACE_GREEDY);
2076 if (obd->obd_namespace == NULL)
2077 GOTO(err_post, rc = -ENOMEM);
2078 obd->obd_namespace->ns_lvbp = obd;
2079 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2080 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2082 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2083 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2085 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2087 CERROR("failed to setup llogging subsystems\n");
2091 rwlock_init(&filter->fo_sptlrpc_lock);
2092 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2093 /* do this after llog being initialized */
2094 filter_adapt_sptlrpc_conf(obd, 1);
2096 rc = lquota_setup(filter_quota_interface_ref, obd);
2100 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2101 if (q->max_sectors < q->max_hw_sectors &&
2102 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2103 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2104 "for larger I/O requests: max_sectors = %u "
2105 "could be up to max_hw_sectors=%u\n",
2106 obd->obd_name, mnt->mnt_sb->s_id,
2107 q->max_sectors, q->max_hw_sectors);
2109 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2110 if (uuid_ptr != NULL) {
2111 class_uuid_unparse(uuid_ptr, &uuid);
2117 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2119 if (obd->obd_recovering) {
2120 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2121 "recovery for at least %d:%.02d, or until %d "
2122 "client%s reconnect%s.\n",
2123 obd->obd_name, lustre_cfg_string(lcfg, 1),
2124 label ?: "", label ? "/" : "", str,
2125 obd->obd_recovery_timeout / 60,
2126 obd->obd_recovery_timeout % 60,
2127 obd->obd_max_recoverable_clients,
2128 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2129 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2131 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2132 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2133 label ?: "", label ? "/" : "", str,
2134 obd->obd_replayable ? "enabled" : "disabled");
2142 fsfilt_put_ops(obd->obd_fsops);
2143 filter_iobuf_pool_done(filter);
2145 server_put_mount(obd->obd_name, mnt);
2146 obd->u.obt.obt_sb = 0;
2150 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2152 struct lprocfs_static_vars lvars;
2157 CLASSERT(offsetof(struct obd_device, u.obt) ==
2158 offsetof(struct obd_device, u.filter.fo_obt));
2160 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2163 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2164 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2167 addr = (unsigned long)cfs_page_address(page);
2168 clear_page((void *)addr);
2170 /* lprocfs must be setup before the filter so state can be safely added
2171 * to /proc incrementally as the filter is setup */
2172 lprocfs_filter_init_vars(&lvars);
2173 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2174 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2175 /* Init obdfilter private stats here */
2176 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2177 LPROCFS_CNTR_AVGMINMAX,
2178 "read_bytes", "bytes");
2179 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2180 LPROCFS_CNTR_AVGMINMAX,
2181 "write_bytes", "bytes");
2182 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2183 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2184 "get_page", "usec");
2185 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2186 LPROCFS_CNTR_AVGMINMAX,
2187 "get_page_failures", "num");
2188 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2189 LPROCFS_CNTR_AVGMINMAX,
2190 "cache_access", "pages");
2191 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2192 LPROCFS_CNTR_AVGMINMAX,
2193 "cache_hit", "pages");
2194 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2195 LPROCFS_CNTR_AVGMINMAX,
2196 "cache_miss", "pages");
2198 lproc_filter_attach_seqstat(obd);
2199 obd->obd_proc_exports_entry = lprocfs_register("exports",
2200 obd->obd_proc_entry,
2202 if (IS_ERR(obd->obd_proc_exports_entry)) {
2203 rc = PTR_ERR(obd->obd_proc_exports_entry);
2204 CERROR("error %d setting up lprocfs for %s\n",
2206 obd->obd_proc_exports_entry = NULL;
2209 if (obd->obd_proc_exports_entry)
2210 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2211 lprocfs_nid_stats_clear_read,
2212 lprocfs_nid_stats_clear_write, obd, NULL);
2214 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2215 LUSTRE_CFG_BUFLEN(lcfg, 4));
2216 rc = filter_common_setup(obd, lcfg, (void *)addr);
2217 OBD_PAGE_FREE(page);
2220 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2221 lprocfs_free_per_client_stats(obd);
2222 lprocfs_free_obd_stats(obd);
2223 lprocfs_obd_cleanup(obd);
2229 static struct llog_operations filter_mds_ost_repl_logops;
2231 static struct llog_operations filter_size_orig_logops = {
2232 .lop_setup = llog_obd_origin_setup,
2233 .lop_cleanup = llog_obd_origin_cleanup,
2234 .lop_add = llog_obd_origin_add
2237 static int filter_olg_fini(struct obd_llog_group *olg)
2239 struct llog_ctxt *ctxt;
2240 int rc = 0, rc2 = 0;
2243 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2245 rc = llog_cleanup(ctxt);
2247 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2249 rc2 = llog_cleanup(ctxt);
2254 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2256 rc2 = llog_cleanup(ctxt);
2265 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2266 struct obd_device *tgt)
2271 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2272 &filter_mds_ost_repl_logops);
2276 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2277 &filter_size_orig_logops);
2283 filter_olg_fini(olg);
2288 * Init the default olg, which is embeded in the obd_device, for filter.
2291 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2292 struct obd_device *tgt)
2294 struct filter_obd *filter = &obd->u.filter;
2295 struct llog_ctxt *ctxt;
2299 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2300 if (!filter->fo_lcm)
2303 filter_mds_ost_repl_logops = llog_client_ops;
2304 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2305 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2306 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2308 rc = filter_olg_init(obd, olg, tgt);
2310 GOTO(cleanup_lcm, rc);
2312 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2315 GOTO(cleanup_olg, rc);
2317 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2319 CERROR("Can't get ctxt for %p:%x\n", olg,
2320 LLOG_MDS_OST_REPL_CTXT);
2321 GOTO(cleanup_olg, rc = -ENODEV);
2323 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2324 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2325 llog_ctxt_put(ctxt);
2329 filter_olg_fini(olg);
2331 llog_recov_thread_fini(filter->fo_lcm, 1);
2332 filter->fo_lcm = NULL;
2337 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2338 struct obd_device *tgt, int *index)
2340 struct filter_obd *filter = &obd->u.filter;
2341 struct llog_ctxt *ctxt;
2345 LASSERT(olg != NULL);
2346 if (olg == &obd->obd_olg)
2347 return filter_default_olg_init(obd, olg, tgt);
2349 LASSERT(filter->fo_lcm != NULL);
2350 rc = filter_olg_init(obd, olg, tgt);
2353 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2355 CERROR("Can't get ctxt for %p:%x\n", olg,
2356 LLOG_MDS_OST_REPL_CTXT);
2357 filter_olg_fini(olg);
2360 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2361 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2362 llog_ctxt_put(ctxt);
2366 static int filter_llog_finish(struct obd_device *obd, int count)
2368 struct filter_obd *filter = &obd->u.filter;
2369 struct llog_ctxt *ctxt;
2372 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2375 * Make sure that no cached llcds left in recov_thread.
2376 * We actually do sync in disconnect time, but disconnect
2377 * may not come being marked rq_no_resend = 1.
2379 llog_sync(ctxt, NULL);
2382 * Balance class_import_get() in llog_receptor_accept().
2383 * This is safe to do, as llog is already synchronized
2384 * and its import may go.
2386 mutex_down(&ctxt->loc_sem);
2387 if (ctxt->loc_imp) {
2388 class_import_put(ctxt->loc_imp);
2389 ctxt->loc_imp = NULL;
2391 mutex_up(&ctxt->loc_sem);
2392 llog_ctxt_put(ctxt);
2395 if (filter->fo_lcm) {
2396 mutex_down(&ctxt->loc_sem);
2397 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2398 filter->fo_lcm = NULL;
2399 mutex_up(&ctxt->loc_sem);
2401 RETURN(filter_olg_fini(&obd->obd_olg));
2405 * Find the group llog according to group index in the llog group list.
2407 static struct obd_llog_group *
2408 filter_find_olg_internal(struct filter_obd *filter, int group)
2410 struct obd_llog_group *olg;
2412 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2413 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2414 if (olg->olg_group == group)
2421 * Find the group llog according to group index on the filter
2423 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2425 struct obd_llog_group *olg = NULL;
2426 struct filter_obd *filter;
2428 filter = &obd->u.filter;
2430 if (group == FILTER_GROUP_LLOG)
2431 RETURN(&obd->obd_olg);
2433 spin_lock(&filter->fo_llog_list_lock);
2434 olg = filter_find_olg_internal(filter, group);
2435 spin_unlock(&filter->fo_llog_list_lock);
2440 * Find the llog_group of the filter according to the group. If it can not
2441 * find, create the llog_group, which only happens when mds is being synced
2444 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2446 struct obd_llog_group *olg = NULL;
2447 struct filter_obd *filter;
2450 filter = &obd->u.filter;
2452 if (group == FILTER_GROUP_LLOG)
2453 RETURN(&obd->obd_olg);
2455 spin_lock(&filter->fo_llog_list_lock);
2456 olg = filter_find_olg_internal(filter, group);
2458 if (olg->olg_initializing) {
2459 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2461 GOTO(out_unlock, olg);
2466 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2468 llog_group_init(olg, group);
2469 list_add(&olg->olg_list, &filter->fo_llog_list);
2470 olg->olg_initializing = 1;
2471 spin_unlock(&filter->fo_llog_list_lock);
2473 rc = obd_llog_init(obd, olg, obd, NULL);
2475 spin_lock(&filter->fo_llog_list_lock);
2476 list_del(&olg->olg_list);
2477 spin_unlock(&filter->fo_llog_list_lock);
2479 GOTO(out, olg = ERR_PTR(-ENOMEM));
2481 spin_lock(&filter->fo_llog_list_lock);
2482 olg->olg_initializing = 0;
2483 spin_unlock(&filter->fo_llog_list_lock);
2484 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2485 obd->obd_name, group, olg);
2490 spin_unlock(&filter->fo_llog_list_lock);
2494 static int filter_llog_connect(struct obd_export *exp,
2495 struct llogd_conn_body *body)
2497 struct obd_device *obd = exp->exp_obd;
2498 struct llog_ctxt *ctxt;
2499 struct obd_llog_group *olg;
2503 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2504 obd->obd_name, body->lgdc_logid.lgl_oid,
2505 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2507 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2509 CERROR(" %s: can not find olg of group %d\n",
2510 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2513 llog_group_set_export(olg, exp);
2515 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2516 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2517 body->lgdc_ctxt_idx);
2519 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2520 obd->obd_name, body->lgdc_logid.lgl_oid,
2521 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2523 spin_lock_bh(&obd->obd_processing_task_lock);
2524 obd->u.filter.fo_mds_ost_sync = 1;
2525 spin_unlock_bh(&obd->obd_processing_task_lock);
2526 rc = llog_connect(ctxt, &body->lgdc_logid,
2527 &body->lgdc_gen, NULL);
2528 llog_ctxt_put(ctxt);
2530 CERROR("failed to connect rc %d idx %d\n", rc,
2531 body->lgdc_ctxt_idx);
2536 static int filter_llog_preclean(struct obd_device *obd)
2538 struct obd_llog_group *olg, *tmp;
2539 struct filter_obd *filter;
2540 struct list_head remove_list;
2544 rc = obd_llog_finish(obd, 0);
2546 CERROR("failed to cleanup llogging subsystem\n");
2548 filter = &obd->u.filter;
2549 CFS_INIT_LIST_HEAD(&remove_list);
2551 spin_lock(&filter->fo_llog_list_lock);
2552 while (!list_empty(&filter->fo_llog_list)) {
2553 olg = list_entry(filter->fo_llog_list.next,
2554 struct obd_llog_group, olg_list);
2555 list_del(&olg->olg_list);
2556 list_add(&olg->olg_list, &remove_list);
2558 spin_unlock(&filter->fo_llog_list_lock);
2560 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2561 list_del_init(&olg->olg_list);
2562 rc = filter_olg_fini(olg);
2564 CERROR("failed to cleanup llogging subsystem for %u\n",
2572 static int filter_precleanup(struct obd_device *obd,
2573 enum obd_cleanup_stage stage)
2579 case OBD_CLEANUP_EARLY:
2581 case OBD_CLEANUP_EXPORTS:
2582 /* Stop recovery before namespace cleanup. */
2583 target_recovery_fini(obd);
2584 rc = filter_llog_preclean(obd);
2590 static int filter_cleanup(struct obd_device *obd)
2592 struct filter_obd *filter = &obd->u.filter;
2596 LCONSOLE_WARN("%s: shutting down for failover; client state "
2597 "will be preserved.\n", obd->obd_name);
2599 obd_exports_barrier(obd);
2600 obd_zombie_barrier();
2602 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2603 lprocfs_free_per_client_stats(obd);
2604 lprocfs_free_obd_stats(obd);
2605 lprocfs_obd_cleanup(obd);
2606 lquota_cleanup(filter_quota_interface_ref, obd);
2608 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2609 obd->obd_namespace = NULL;
2611 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2613 if (obd->u.obt.obt_sb == NULL)
2618 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2619 shrink_dcache_sb(obd->u.obt.obt_sb);
2621 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2622 obd->u.obt.obt_sb = NULL;
2624 fsfilt_put_ops(obd->obd_fsops);
2626 filter_iobuf_pool_done(filter);
2628 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2633 static int filter_connect_internal(struct obd_export *exp,
2634 struct obd_connect_data *data,
2637 struct filter_export_data *fed = &exp->exp_filter_data;
2642 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2643 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2644 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2645 data->ocd_connect_flags, data->ocd_version,
2646 data->ocd_grant, data->ocd_index);
2648 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2649 CWARN("!!! This export (nid %s) used object group %d "
2650 "earlier; now it's trying to use group %d! This could "
2651 "be a bug in the MDS. Please report to "
2652 "http://bugzilla.lustre.org/\n",
2653 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2656 fed->fed_group = data->ocd_group;
2658 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2659 exp->exp_connect_flags = data->ocd_connect_flags;
2660 data->ocd_version = LUSTRE_VERSION_CODE;
2662 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2663 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2664 data->ocd_connect_flags & OBD_CONNECT_MDS))
2667 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2668 struct filter_obd *filter = &exp->exp_obd->u.filter;
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, (reconnect == 0));
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);
2683 filter->fo_tot_granted_clients ++;
2686 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2687 struct filter_obd *filter = &exp->exp_obd->u.filter;
2688 struct lr_server_data *lsd = filter->fo_fsd;
2689 int index = le32_to_cpu(lsd->lsd_ost_index);
2691 if (!(lsd->lsd_feature_compat &
2692 cpu_to_le32(OBD_COMPAT_OST))) {
2693 /* this will only happen on the first connect */
2694 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2695 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2696 filter_update_server_data(exp->exp_obd,
2697 filter->fo_rcvd_filp, lsd, 1);
2698 } else if (index != data->ocd_index) {
2699 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2700 " %u doesn't match actual OST index"
2701 " %u in last_rcvd file, bad "
2703 obd_export_nid2str(exp), index,
2707 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2708 * FIXME: We don't strictly need the COMPAT flag for that,
2709 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2710 * FIXME: We needed it for the index, as index 0 is valid. */
2713 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2714 data->ocd_brw_size = 65536;
2715 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2716 data->ocd_brw_size = min(data->ocd_brw_size,
2717 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2718 LASSERT(data->ocd_brw_size);
2721 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2722 __u32 cksum_types = data->ocd_cksum_types;
2724 /* The client set in ocd_cksum_types the checksum types it
2725 * supports. We have to mask off the algorithms that we don't
2727 if (cksum_types & OBD_CKSUM_ALL)
2728 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2730 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2732 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2733 "%x\n", exp->exp_obd->obd_name,
2734 obd_export_nid2str(exp), cksum_types,
2735 data->ocd_cksum_types);
2737 /* This client does not support OBD_CONNECT_CKSUM
2738 * fall back to CRC32 */
2739 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2740 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2741 exp->exp_obd->obd_name,
2742 obd_export_nid2str(exp));
2748 static int filter_reconnect(const struct lu_env *env,
2749 struct obd_export *exp, struct obd_device *obd,
2750 struct obd_uuid *cluuid,
2751 struct obd_connect_data *data,
2757 if (exp == NULL || obd == NULL || cluuid == NULL)
2760 rc = filter_connect_internal(exp, data, 1);
2762 filter_export_stats_init(obd, exp, localdata);
2767 /* nearly identical to mds_connect */
2768 static int filter_connect(const struct lu_env *env,
2769 struct obd_export **exp, struct obd_device *obd,
2770 struct obd_uuid *cluuid,
2771 struct obd_connect_data *data, void *localdata)
2773 struct lvfs_run_ctxt saved;
2774 struct lustre_handle conn = { 0 };
2775 struct obd_export *lexp;
2776 struct filter_export_data *fed;
2777 struct lsd_client_data *lcd = NULL;
2782 if (exp == NULL || obd == NULL || cluuid == NULL)
2785 rc = class_connect(&conn, obd, cluuid);
2788 lexp = class_conn2export(&conn);
2789 LASSERT(lexp != NULL);
2791 fed = &lexp->exp_filter_data;
2793 rc = filter_connect_internal(lexp, data, 0);
2797 filter_export_stats_init(obd, lexp, localdata);
2798 if (obd->obd_replayable) {
2799 OBD_ALLOC(lcd, sizeof(*lcd));
2801 CERROR("filter: out of memory for client data\n");
2802 GOTO(cleanup, rc = -ENOMEM);
2805 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2807 rc = filter_client_add(obd, lexp, -1);
2812 group = data->ocd_group;
2814 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2815 obd->obd_name, lexp->exp_handle.h_cookie, group);
2817 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2818 rc = filter_read_groups(obd, group, 1);
2819 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2821 CERROR("can't read group %u\n", group);
2831 fed->fed_lcd = NULL;
2833 class_disconnect(lexp);
2842 /* Do extra sanity checks for grant accounting. We do this at connect,
2843 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2844 * always get rid of it or turn it off when we know accounting is good. */
2845 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2847 struct filter_export_data *fed;
2848 struct obd_export *exp;
2849 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2850 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2851 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2853 if (list_empty(&obd->obd_exports))
2856 /* We don't want to do this for large machines that do lots of
2857 mounts or unmounts. It burns... */
2858 if (obd->obd_num_exports > 100)
2861 spin_lock(&obd->obd_osfs_lock);
2862 spin_lock(&obd->obd_dev_lock);
2863 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2865 fed = &exp->exp_filter_data;
2866 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2869 if (maxsize > 0) { /* we may not have done a statfs yet */
2870 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2871 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2872 exp->exp_client_uuid.uuid, exp,
2873 fed->fed_grant, fed->fed_pending, maxsize);
2874 LASSERTF(fed->fed_dirty <= maxsize,
2875 "%s: cli %s/%p %ld > "LPU64"\n", func,
2876 exp->exp_client_uuid.uuid, exp,
2877 fed->fed_dirty, maxsize);
2880 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2881 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2882 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2884 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2885 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2886 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2887 tot_granted += fed->fed_grant + fed->fed_pending;
2888 tot_pending += fed->fed_pending;
2889 tot_dirty += fed->fed_dirty;
2891 fo_tot_granted = obd->u.filter.fo_tot_granted;
2892 fo_tot_pending = obd->u.filter.fo_tot_pending;
2893 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2894 spin_unlock(&obd->obd_dev_lock);
2895 spin_unlock(&obd->obd_osfs_lock);
2897 /* Do these assertions outside the spinlocks so we don't kill system */
2898 if (tot_granted != fo_tot_granted)
2899 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2900 func, tot_granted, fo_tot_granted);
2901 if (tot_pending != fo_tot_pending)
2902 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2903 func, tot_pending, fo_tot_pending);
2904 if (tot_dirty != fo_tot_dirty)
2905 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2906 func, tot_dirty, fo_tot_dirty);
2907 if (tot_pending > tot_granted)
2908 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2909 func, tot_pending, tot_granted);
2910 if (tot_granted > maxsize)
2911 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2912 func, tot_granted, maxsize);
2913 if (tot_dirty > maxsize)
2914 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2915 func, tot_dirty, maxsize);
2918 /* Remove this client from the grant accounting totals. We also remove
2919 * the export from the obd device under the osfs and dev locks to ensure
2920 * that the filter_grant_sanity_check() calculations are always valid.
2921 * The client should do something similar when it invalidates its import. */
2922 static void filter_grant_discard(struct obd_export *exp)
2924 struct obd_device *obd = exp->exp_obd;
2925 struct filter_obd *filter = &obd->u.filter;
2926 struct filter_export_data *fed = &exp->exp_filter_data;
2928 spin_lock(&obd->obd_osfs_lock);
2929 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2930 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2931 obd->obd_name, filter->fo_tot_granted,
2932 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2933 filter->fo_tot_granted -= fed->fed_grant;
2934 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2935 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2936 obd->obd_name, filter->fo_tot_pending,
2937 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2938 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2939 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2940 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2941 obd->obd_name, filter->fo_tot_dirty,
2942 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2943 filter->fo_tot_dirty -= fed->fed_dirty;
2947 spin_unlock(&obd->obd_osfs_lock);
2950 static int filter_destroy_export(struct obd_export *exp)
2954 if (exp->exp_filter_data.fed_pending)
2955 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2956 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2957 exp, exp->exp_filter_data.fed_pending);
2959 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2961 target_destroy_export(exp);
2962 ldlm_destroy_export(exp);
2964 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2968 if (exp->exp_obd->obd_replayable)
2969 filter_client_free(exp);
2971 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2973 filter_grant_discard(exp);
2974 filter_fmd_cleanup(exp);
2976 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2977 struct filter_obd *filter = &exp->exp_obd->u.filter;
2978 if (filter->fo_tot_granted_clients > 0)
2979 filter->fo_tot_granted_clients --;
2982 if (!(exp->exp_flags & OBD_OPT_FORCE))
2983 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2988 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2990 struct obd_llog_group *olg_min, *olg;
2991 struct filter_obd *filter;
2992 int worked = -1, group;
2993 struct llog_ctxt *ctxt;
2996 filter = &obd->u.filter;
2998 /* we can't sync log holding spinlock. also, we do not want to get
2999 * into livelock. so we do following: loop over MDS's exports in
3000 * group order and skip already synced llogs -bzzz */
3002 /* look for group with min. number, but > worked */
3005 spin_lock(&filter->fo_llog_list_lock);
3006 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3007 if (olg->olg_group <= worked) {
3008 /* this group is already synced */
3011 if (group < olg->olg_group) {
3012 /* we have group with smaller number to sync */
3015 /* store current minimal group */
3017 group = olg->olg_group;
3019 spin_unlock(&filter->fo_llog_list_lock);
3021 if (olg_min == NULL)
3024 worked = olg_min->olg_group;
3025 if (olg_min->olg_exp &&
3026 (dexp == olg_min->olg_exp || dexp == NULL)) {
3028 ctxt = llog_group_get_ctxt(olg_min,
3029 LLOG_MDS_OST_REPL_CTXT);
3031 err = llog_sync(ctxt, olg_min->olg_exp);
3032 llog_ctxt_put(ctxt);
3034 CERROR("error flushing logs to MDS: "
3039 } while (olg_min != NULL);
3042 /* Also incredibly similar to mds_disconnect */
3043 static int filter_disconnect(struct obd_export *exp)
3045 struct obd_device *obd = exp->exp_obd;
3050 class_export_get(exp);
3052 if (!(exp->exp_flags & OBD_OPT_FORCE))
3053 filter_grant_sanity_check(obd, __FUNCTION__);
3054 filter_grant_discard(exp);
3056 /* Flush any remaining cancel messages out to the target */
3057 filter_sync_llogs(obd, exp);
3059 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3061 /* Disconnect early so that clients can't keep using export */
3062 rc = class_disconnect(exp);
3063 if (exp->exp_obd->obd_namespace != NULL)
3064 ldlm_cancel_locks_for_export(exp);
3066 fsfilt_sync(obd, obd->u.obt.obt_sb);
3068 lprocfs_exp_cleanup(exp);
3070 class_export_put(exp);
3074 /* reverse import is changed, sync all cancels */
3075 static void filter_revimp_update(struct obd_export *exp)
3080 class_export_get(exp);
3082 /* flush any remaining cancel messages out to the target */
3083 filter_sync_llogs(exp->exp_obd, exp);
3084 class_export_put(exp);
3088 static int filter_ping(struct obd_export *exp)
3090 filter_fmd_expire(exp);
3094 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3095 const char *what, int quiet)
3097 struct dentry *dchild = NULL;
3100 if (oa->o_valid & OBD_MD_FLGROUP)
3103 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3105 if (IS_ERR(dchild)) {
3106 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3107 what, group, oa->o_id);
3111 if (dchild->d_inode == NULL) {
3113 CERROR("%s: %s on non-existent object: "LPU64"\n",
3114 obd->obd_name, what, oa->o_id);
3116 RETURN(ERR_PTR(-ENOENT));
3122 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3124 struct dentry *dentry = NULL;
3125 struct obd_device *obd;
3129 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3130 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3131 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3135 obd = class_exp2obd(exp);
3137 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3141 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3143 RETURN(PTR_ERR(dentry));
3145 /* Limit the valid bits in the return data to what we actually use */
3146 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3147 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3153 /* this should be enabled/disabled in condition to enabled/disabled large
3154 * inodes (fast EAs) in backing store FS. */
3155 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3156 void *handle, struct obdo *oa)
3158 struct obd_device *obd = exp->exp_obd;
3162 if (oa->o_valid & OBD_MD_FLFID) {
3163 struct filter_fid ff;
3165 if (!(oa->o_valid & OBD_MD_FLGROUP))
3167 /* packing fid and converting it to LE for storing into EA.
3168 * Here ->o_stripe_idx should be filled by LOV and rest of
3169 * fields - by client. */
3170 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3171 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3172 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3173 ff.ff_objid = cpu_to_le64(oa->o_id);
3174 ff.ff_group = cpu_to_le64(oa->o_gr);
3176 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3177 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3178 oa->o_generation, oa->o_id, oa->o_gr);
3180 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3182 CERROR("store fid in object failed! rc: %d\n", rc);
3184 CDEBUG(D_HA, "OSS object without fid info!\n");
3190 /* this is called from filter_truncate() until we have filter_punch() */
3191 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3192 struct obdo *oa, struct obd_trans_info *oti)
3194 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3195 struct llog_cookie *fcc = NULL;
3196 struct filter_obd *filter;
3197 int rc, err, sync = 0;
3198 loff_t old_size = 0;
3199 unsigned int ia_valid;
3200 struct inode *inode;
3201 struct page *page = NULL;
3206 LASSERT(dentry != NULL);
3207 LASSERT(!IS_ERR(dentry));
3209 inode = dentry->d_inode;
3210 LASSERT(inode != NULL);
3212 filter = &exp->exp_obd->u.filter;
3213 iattr_from_obdo(&iattr, oa, oa->o_valid);
3214 ia_valid = iattr.ia_valid;
3216 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3217 OBD_ALLOC(fcc, sizeof(*fcc));
3219 *fcc = oa->o_lcookie;
3221 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3223 /* Filter truncates and writes are serialized by
3224 * i_alloc_sem, see the comment in
3225 * filter_preprw_write.*/
3226 if (ia_valid & ATTR_SIZE)
3227 down_write(&inode->i_alloc_sem);
3228 LOCK_INODE_MUTEX(inode);
3229 old_size = i_size_read(inode);
3232 /* VBR: version recovery check */
3233 rc = filter_version_get_check(exp, oti, inode);
3235 GOTO(out_unlock, rc);
3237 /* Let's pin the last page so that ldiskfs_truncate
3238 * should not start GFP_FS allocation. */
3239 if (ia_valid & ATTR_SIZE) {
3240 page = grab_cache_page(inode->i_mapping,
3241 iattr.ia_size >> PAGE_CACHE_SHIFT);
3243 GOTO(out_unlock, rc = -ENOMEM);
3248 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3249 * then we will accept the UID+GID sent by the client during write for
3250 * initializing the ownership of this inode. We only allow this to
3251 * happen once so clear these bits in setattr. In 2.6 kernels it is
3252 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3253 * the flags that are actually being set. */
3254 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3255 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3256 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3258 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3259 if (!(ia_valid & ATTR_MODE)) {
3260 iattr.ia_mode = inode->i_mode;
3261 iattr.ia_valid |= ATTR_MODE;
3263 iattr.ia_mode &= ~S_ISUID;
3265 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3266 if (!(iattr.ia_valid & ATTR_MODE)) {
3267 iattr.ia_mode = inode->i_mode;
3268 iattr.ia_valid |= ATTR_MODE;
3270 iattr.ia_mode &= ~S_ISGID;
3273 orig_ids[USRQUOTA] = inode->i_uid;
3274 orig_ids[GRPQUOTA] = inode->i_gid;
3275 handle = fsfilt_start_log(exp->exp_obd, inode,
3276 FSFILT_OP_SETATTR, oti, 1);
3278 GOTO(out_unlock, rc = PTR_ERR(handle));
3280 /* update inode EA only once when inode is suid bit marked. As
3281 * on 2.6.x UID and GID may be set separately, we check here
3282 * only one of them to avoid double setting. */
3283 if (inode->i_mode & S_ISUID)
3284 filter_update_fidea(exp, inode, handle, oa);
3286 handle = fsfilt_start(exp->exp_obd, inode,
3287 FSFILT_OP_SETATTR, oti);
3289 GOTO(out_unlock, rc = PTR_ERR(handle));
3291 if (oa->o_valid & OBD_MD_FLFLAGS) {
3292 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3293 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3295 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3297 /* set cancel cookie callback function */
3298 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3299 filter_cancel_cookies_cb,
3303 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3304 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3306 /* The truncate might have used up our transaction credits. Make
3307 * sure we have one left for the last_rcvd update. */
3308 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3310 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3312 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3316 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3318 CERROR("error on commit, err = %d\n", err);
3325 /* For a partial-page truncate flush the page to disk immediately
3326 * to avoid data corruption during direct disk write. b=17397 */
3327 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3328 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3329 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3339 page_cache_release(page);
3341 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3342 UNLOCK_INODE_MUTEX(inode);
3343 if (ia_valid & ATTR_SIZE)
3344 up_write(&inode->i_alloc_sem);
3346 OBD_FREE(fcc, sizeof(*fcc));
3348 /* trigger quota release */
3349 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3350 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3351 int rc2 = lquota_adjust(filter_quota_interface_ref,
3352 exp->exp_obd, cur_ids,
3353 orig_ids, rc, FSFILT_OP_SETATTR);
3354 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3355 "filter adjust qunit. (rc:%d)\n", rc2);
3360 /* this is called from filter_truncate() until we have filter_punch() */
3361 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3362 struct obd_trans_info *oti)
3364 struct obdo *oa = oinfo->oi_oa;
3365 struct lustre_capa *capa = oinfo_capa(oinfo);
3366 struct ldlm_res_id res_id;
3367 struct filter_mod_data *fmd;
3368 struct lvfs_run_ctxt saved;
3369 struct filter_obd *filter;
3370 struct ldlm_resource *res;
3371 struct dentry *dentry;
3372 __u64 opc = CAPA_OPC_META_WRITE;
3376 if (oa->o_valid & OBD_FL_TRUNC)
3377 opc |= CAPA_OPC_OSS_TRUNC;
3379 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3380 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3384 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3385 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3390 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3391 /* This would be very bad - accidentally truncating a file when
3392 * changing the time or similar - bug 12203. */
3393 if (oa->o_valid & OBD_MD_FLSIZE &&
3394 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3395 static char mdsinum[48];
3397 if (oa->o_valid & OBD_MD_FLFID)
3398 snprintf(mdsinum, sizeof(mdsinum) - 1,
3399 " of inode "LPU64"/%u", oa->o_fid,
3404 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3406 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3411 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3413 RETURN(PTR_ERR(dentry));
3415 filter = &exp->exp_obd->u.filter;
3416 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3419 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3420 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3421 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3422 fmd->fmd_mactime_xid = oti->oti_xid;
3423 filter_fmd_put(exp, fmd);
3426 /* setting objects attributes (including owner/group) */
3427 rc = filter_setattr_internal(exp, dentry, oa, oti);
3429 GOTO(out_unlock, rc);
3431 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3432 &res_id, LDLM_EXTENT, 0);
3435 LDLM_RESOURCE_ADDREF(res);
3436 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3437 LDLM_RESOURCE_DELREF(res);
3438 ldlm_resource_putref(res);
3441 oa->o_valid = OBD_MD_FLID;
3443 /* Quota release need uid/gid info */
3444 obdo_from_inode(oa, dentry->d_inode,
3445 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3450 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3454 /* XXX identical to osc_unpackmd */
3455 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3456 struct lov_mds_md *lmm, int lmm_bytes)
3462 if (lmm_bytes < sizeof (*lmm)) {
3463 CERROR("lov_mds_md too small: %d, need %d\n",
3464 lmm_bytes, (int)sizeof(*lmm));
3467 /* XXX LOV_MAGIC etc check? */
3469 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3470 CERROR("lov_mds_md: zero lmm_object_id\n");
3475 lsm_size = lov_stripe_md_size(1);
3479 if (*lsmp != NULL && lmm == NULL) {
3480 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3481 OBD_FREE(*lsmp, lsm_size);
3486 if (*lsmp == NULL) {
3487 OBD_ALLOC(*lsmp, lsm_size);
3491 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3492 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3493 OBD_FREE(*lsmp, lsm_size);
3496 loi_init((*lsmp)->lsm_oinfo[0]);
3500 /* XXX zero *lsmp? */
3501 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3502 LASSERT((*lsmp)->lsm_object_id);
3505 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3510 /* caller must hold fo_create_locks[oa->o_gr] */
3511 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3512 struct filter_obd *filter)
3514 struct obdo doa = { 0 }; /* XXX obdo on stack */
3521 LASSERT_MDS_GROUP(oa->o_gr);
3522 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3523 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3525 memset(&doa, 0, sizeof(doa));
3527 doa.o_valid |= OBD_MD_FLGROUP;
3528 doa.o_gr = oa->o_gr;
3529 doa.o_mode = S_IFREG;
3531 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3532 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3533 exp->exp_obd->obd_name, doa.o_gr);
3537 last = filter_last_id(filter, doa.o_gr);
3539 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3541 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3542 exp->exp_obd->obd_name, oa->o_id + 1, last,
3543 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3545 for (id = last; id > oa->o_id; id--) {
3547 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3548 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3549 CEMERG("error destroying precreate objid "LPU64": %d\n",
3552 /* update last_id on disk periodically so that if we restart
3553 * we don't need to re-scan all of the just-deleted objects. */
3554 if ((id & 511) == 0 && !skip_orphan) {
3555 filter_set_last_id(filter, id - 1, doa.o_gr);
3556 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3560 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3561 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3564 filter_set_last_id(filter, id, doa.o_gr);
3565 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3567 /* don't reuse orphan object, return last used objid */
3571 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3576 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3577 obd_gr group, int *num);
3578 /* returns a negative error or a nonnegative number of files to create */
3579 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3580 obd_gr group, struct obd_trans_info *oti)
3582 struct obd_device *obd = exp->exp_obd;
3583 struct filter_obd *filter = &obd->u.filter;
3587 /* delete orphans request */
3588 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3589 obd_id last = filter_last_id(filter, group);
3591 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3592 CERROR("%s: dropping old orphan cleanup request\n",
3596 /* This causes inflight precreates to abort and drop lock */
3597 set_bit(group, &filter->fo_destroys_in_progress);
3598 down(&filter->fo_create_locks[group]);
3599 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3600 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3601 exp->exp_obd->obd_name, group);
3602 up(&filter->fo_create_locks[group]);
3605 diff = oa->o_id - last;
3606 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3609 if (-diff > OST_MAX_PRECREATE) {
3610 CERROR("%s: ignoring bogus orphan destroy request: "
3611 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3613 /* FIXME: should reset precreate_next_id on MDS */
3614 GOTO(out, rc = -EINVAL);
3617 rc = filter_destroy_precreated(exp, oa, filter);
3619 CERROR("%s: unable to write lastobjid, but "
3620 "orphans were deleted\n", obd->obd_name);
3623 /* XXX: Used by MDS for the first time! */
3624 clear_bit(group, &filter->fo_destroys_in_progress);
3627 down(&filter->fo_create_locks[group]);
3628 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3629 CERROR("%s: dropping old precreate request\n",
3633 /* only precreate if group == 0 and o_id is specfied */
3634 if (!filter_group_is_mds(group) || oa->o_id == 0)
3637 diff = oa->o_id - filter_last_id(filter, group);
3638 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3639 filter_last_id(filter, group), diff);
3641 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3642 oa->o_id, filter_last_id(filter, group), diff);
3646 oa->o_id = filter_last_id(&obd->u.filter, group);
3647 rc = filter_precreate(obd, oa, group, &diff);
3648 oa->o_id = filter_last_id(&obd->u.filter, group);
3650 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3653 /* else diff == 0 */
3656 up(&filter->fo_create_locks[group]);
3660 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3661 __u64 max_age, __u32 flags)
3663 struct filter_obd *filter = &obd->u.filter;
3664 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3668 /* at least try to account for cached pages. its still racey and
3669 * might be under-reporting if clients haven't announced their
3670 * caches with brw recently */
3671 spin_lock(&obd->obd_osfs_lock);
3672 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3673 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3674 spin_unlock(&obd->obd_osfs_lock);
3676 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3677 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3678 filter->fo_tot_dirty, filter->fo_tot_granted,
3679 filter->fo_tot_pending,
3680 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3682 filter_grant_sanity_check(obd, __FUNCTION__);
3684 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3685 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3686 osfs->os_bsize - 1) >> blockbits));
3688 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3689 * stop creating files on MDS if OST is not good shape to create
3693 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3694 osfs->os_state = OS_STATE_READONLY;
3696 if (filter->fo_raid_degraded)
3697 osfs->os_state |= OS_STATE_DEGRADED;
3701 static int filter_use_existing_obj(struct obd_device *obd,
3702 struct dentry *dchild, void **handle,
3705 struct inode *inode = dchild->d_inode;
3709 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3712 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3713 if (IS_ERR(*handle))
3714 return PTR_ERR(*handle);
3716 iattr.ia_valid = ATTR_MODE;
3717 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3718 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3725 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3728 __u64 os_ffree = -1;
3730 spin_lock(&obd->obd_osfs_lock);
3731 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3733 os_ffree = obd->obd_osfs.os_ffree;
3734 spin_unlock(&obd->obd_osfs_lock);
3739 /* We rely on the fact that only one thread will be creating files in a given
3740 * group at a time, which is why we don't need an atomic filter_get_new_id.
3741 * Even if we had that atomic function, the following race would exist:
3743 * thread 1: gets id x from filter_next_id
3744 * thread 2: gets id (x + 1) from filter_next_id
3745 * thread 2: creates object (x + 1)
3746 * thread 1: tries to create object x, gets -ENOSPC
3748 * Caller must hold fo_create_locks[group]
3750 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3751 obd_gr group, int *num)
3753 struct dentry *dchild = NULL, *dparent = NULL;
3754 struct filter_obd *filter;
3755 struct obd_statfs *osfs;
3756 int err = 0, rc = 0, recreate_obj = 0, i;
3757 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3760 void *handle = NULL;
3763 filter = &obd->u.filter;
3765 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3767 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3769 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3770 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3773 OBD_ALLOC(osfs, sizeof(*osfs));
3776 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3777 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3778 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3779 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3780 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3784 OBD_FREE(osfs, sizeof(*osfs));
3789 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3790 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3792 for (i = 0; i < *num && err == 0; i++) {
3793 int cleanup_phase = 0;
3795 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3796 CWARN("%s: create aborted by destroy\n",
3805 last_id = filter_last_id(filter, group);
3806 if (next_id > last_id) {
3807 CERROR("Error: Trying to recreate obj greater"
3808 "than last id "LPD64" > "LPD64"\n",
3810 GOTO(cleanup, rc = -EINVAL);
3813 next_id = filter_last_id(filter, group) + 1;
3815 dparent = filter_parent_lock(obd, group, next_id);
3816 if (IS_ERR(dparent))
3817 GOTO(cleanup, rc = PTR_ERR(dparent));
3818 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3820 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3822 GOTO(cleanup, rc = PTR_ERR(dchild));
3823 cleanup_phase = 2; /* f_dput(dchild) */
3825 if (dchild->d_inode != NULL) {
3826 /* This would only happen if lastobjid was bad on disk*/
3827 /* Could also happen if recreating missing obj but it
3828 * already exists. */
3830 CERROR("%s: recreating existing object %.*s?\n",
3831 obd->obd_name, dchild->d_name.len,
3832 dchild->d_name.name);
3834 /* Use these existing objects if they are
3836 if (dchild->d_inode->i_size == 0) {
3837 rc = filter_use_existing_obj(obd,dchild,
3838 &handle, &cleanup_phase);
3845 CERROR("%s: Serious error: objid %.*s already "
3846 "exists; is this filesystem corrupt?\n",
3847 obd->obd_name, dchild->d_name.len,
3848 dchild->d_name.name);
3851 GOTO(cleanup, rc = -EEXIST);
3854 handle = fsfilt_start_log(obd, dparent->d_inode,
3855 FSFILT_OP_CREATE, NULL, 1);
3857 GOTO(cleanup, rc = PTR_ERR(handle));
3860 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3861 ",od->o_id="LPU64")\n", obd->obd_name, group,
3864 /* We mark object SUID+SGID to flag it for accepting UID+GID
3865 * from client on first write. Currently the permission bits
3866 * on the OST are never used, so this is OK. */
3867 rc = ll_vfs_create(dparent->d_inode, dchild,
3868 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3870 CERROR("create failed rc = %d\n", rc);
3871 if (rc == -ENOSPC) {
3872 os_ffree = filter_calc_free_inodes(obd);
3874 CERROR("%s: free inode "LPU64"\n",
3875 obd->obd_name, os_ffree);
3880 if (dchild->d_inode)
3881 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3882 dchild->d_inode->i_ino);
3885 if (!recreate_obj) {
3886 filter_set_last_id(filter, next_id, group);
3887 err = filter_update_last_objid(obd, group, 0);
3889 CERROR("unable to write lastobjid "
3890 "but file created\n");
3894 switch(cleanup_phase) {
3896 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3898 CERROR("error on commit, err = %d\n", err);
3905 filter_parent_unlock(dparent);
3912 if (time_after(jiffies, enough_time)) {
3914 "%s: precreate slow - want %d got %d \n",
3915 obd->obd_name, *num, i);
3922 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3923 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3928 static int filter_create(struct obd_export *exp, struct obdo *oa,
3929 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3931 struct obd_device *obd = exp->exp_obd;
3932 struct filter_export_data *fed;
3933 struct filter_obd *filter;
3934 struct lvfs_run_ctxt saved;
3935 struct lov_stripe_md *lsm = NULL;
3936 int rc = 0, diff, group = oa->o_gr;
3939 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3940 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3942 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3943 CERROR("!!! nid %s sent invalid object group %d\n",
3944 obd_export_nid2str(exp), group);
3948 fed = &exp->exp_filter_data;
3949 filter = &obd->u.filter;
3951 if (fed->fed_group != group) {
3952 CERROR("!!! this export (nid %s) used object group %d "
3953 "earlier; now it's trying to use group %d! This could "
3954 "be a bug in the MDS. Please report to "
3955 "http://bugzilla.lustre.org/\n",
3956 obd_export_nid2str(exp), fed->fed_group, group);
3963 rc = obd_alloc_memmd(exp, &lsm);
3970 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3972 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3973 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3974 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3975 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3976 oa->o_id, filter_last_id(filter,
3981 down(&filter->fo_create_locks[oa->o_gr]);
3982 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3983 up(&filter->fo_create_locks[oa->o_gr]);
3986 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3989 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3990 if (rc && ea != NULL && *ea != lsm) {
3991 obd_free_memmd(exp, &lsm);
3992 } else if (rc == 0 && ea != NULL) {
3993 /* XXX LOV STACKING: the lsm that is passed to us from
3994 * LOV does not have valid lsm_oinfo data structs, so
3995 * don't go touching that. This needs to be fixed in a
3997 lsm->lsm_object_id = oa->o_id;
4004 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4005 struct lov_stripe_md *md, struct obd_trans_info *oti,
4006 struct obd_export *md_exp, void *capa)
4008 unsigned int qcids[MAXQUOTAS] = {0, 0};
4009 struct obd_device *obd;
4010 struct filter_obd *filter;
4011 struct dentry *dchild = NULL, *dparent = NULL;
4012 struct lustre_handle lockh = { 0 };
4013 struct lvfs_run_ctxt saved;
4014 void *handle = NULL;
4015 struct llog_cookie *fcc = NULL;
4016 int rc, rc2, cleanup_phase = 0, sync = 0;
4020 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4021 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4022 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4027 filter = &obd->u.filter;
4029 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4032 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4033 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4035 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4037 GOTO(cleanup, rc = PTR_ERR(dchild));
4040 if (dchild->d_inode == NULL) {
4041 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4043 /* If object already gone, cancel cookie right now */
4044 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4045 struct llog_ctxt *ctxt;
4046 struct obd_llog_group *olg;
4048 olg = filter_find_olg(obd, oa->o_gr);
4050 CERROR(" %s: can not find olg of group %d\n",
4051 obd->obd_name, (int)oa->o_gr);
4052 GOTO(cleanup, rc = PTR_ERR(olg));
4054 fcc = &oa->o_lcookie;
4055 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4056 llog_cancel(ctxt, NULL, 1, fcc, 0);
4057 llog_ctxt_put(ctxt);
4058 fcc = NULL; /* we didn't allocate fcc, don't free it */
4060 GOTO(cleanup, rc = -ENOENT);
4063 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4065 /* Our MDC connection is established by the MDS to us */
4066 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4067 OBD_ALLOC(fcc, sizeof(*fcc));
4069 *fcc = oa->o_lcookie;
4071 DQUOT_INIT(dchild->d_inode);
4073 /* we're gonna truncate it first in order to avoid possible deadlock:
4075 * open trasaction open transaction
4076 * down(i_zombie) down(i_zombie)
4077 * restart transaction
4078 * (see BUG 4180) -bzzz
4080 * take i_alloc_sem too to prevent other threads from writing to the
4081 * file while we are truncating it. This can cause lock ordering issue
4082 * between page lock, i_mutex & starting new journal handle.
4083 * (see bug 20321) -johann
4085 down_write(&dchild->d_inode->i_alloc_sem);
4086 LOCK_INODE_MUTEX(dchild->d_inode);
4088 /* VBR: version recovery check */
4089 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4091 UNLOCK_INODE_MUTEX(dchild->d_inode);
4092 up_write(&dchild->d_inode->i_alloc_sem);
4096 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4098 if (IS_ERR(handle)) {
4099 UNLOCK_INODE_MUTEX(dchild->d_inode);
4100 up_write(&dchild->d_inode->i_alloc_sem);
4101 GOTO(cleanup, rc = PTR_ERR(handle));
4104 iattr.ia_valid = ATTR_SIZE;
4106 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4107 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4108 UNLOCK_INODE_MUTEX(dchild->d_inode);
4109 up_write(&dchild->d_inode->i_alloc_sem);
4113 GOTO(cleanup, rc = rc2);
4115 /* We don't actually need to lock the parent until we are unlinking
4116 * here, and not while truncating above. That avoids holding the
4117 * parent lock for a long time during truncate, which can block other
4118 * threads from doing anything to objects in that directory. bug 7171 */
4119 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4120 if (IS_ERR(dparent))
4121 GOTO(cleanup, rc = PTR_ERR(dparent));
4122 cleanup_phase = 3; /* filter_parent_unlock */
4124 LOCK_INODE_MUTEX(dchild->d_inode);
4125 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4126 if (IS_ERR(handle)) {
4127 UNLOCK_INODE_MUTEX(dchild->d_inode);
4128 GOTO(cleanup, rc = PTR_ERR(handle));
4130 cleanup_phase = 4; /* fsfilt_commit */
4132 /* Quota release need uid/gid of inode */
4133 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4135 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4137 /* this drops dchild->d_inode->i_mutex unconditionally */
4138 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4142 switch(cleanup_phase) {
4145 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4146 oti->oti_handle : handle,
4147 filter_cancel_cookies_cb,
4149 /* If add_journal_cb failed, then filter_finish_transno
4150 * will commit the handle and we will do a sync
4151 * on commit. then we call callback directly to free
4154 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4156 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4159 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4161 CERROR("error on commit, err = %d\n", rc2);
4168 filter_parent_unlock(dparent);
4170 filter_fini_destroy(obd, &lockh);
4174 OBD_FREE(fcc, sizeof(*fcc));
4176 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4179 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4183 /* trigger quota release */
4184 qcids[USRQUOTA] = oa->o_uid;
4185 qcids[GRPQUOTA] = oa->o_gid;
4186 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4189 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4193 /* NB start and end are used for punch, but not truncate */
4194 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4195 struct obd_trans_info *oti,
4196 struct ptlrpc_request_set *rqset)
4201 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4202 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4203 oinfo->oi_policy.l_extent.end);
4207 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4208 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4209 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4211 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4212 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4213 rc = filter_setattr(exp, oinfo, oti);
4214 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4218 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4219 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4222 struct lvfs_run_ctxt saved;
4223 struct filter_obd *filter;
4224 struct dentry *dentry;
4228 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4229 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4230 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4234 filter = &exp->exp_obd->u.filter;
4236 /* An objid of zero is taken to mean "sync whole filesystem" */
4237 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4238 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4239 /* Flush any remaining cancel messages out to the target */
4240 filter_sync_llogs(exp->exp_obd, exp);
4244 dentry = filter_oa2dentry(exp->exp_obd, oa);
4246 RETURN(PTR_ERR(dentry));
4248 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4250 LOCK_INODE_MUTEX(dentry->d_inode);
4252 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4254 /* just any file to grab fsync method - "file" arg unused */
4255 struct file *file = filter->fo_rcvd_filp;
4257 if (file->f_op && file->f_op->fsync)
4258 rc = file->f_op->fsync(NULL, dentry, 1);
4260 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4264 UNLOCK_INODE_MUTEX(dentry->d_inode);
4266 oa->o_valid = OBD_MD_FLID;
4267 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4269 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4275 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4276 void *key, __u32 *vallen, void *val,
4277 struct lov_stripe_md *lsm)
4279 struct obd_device *obd;
4282 obd = class_exp2obd(exp);
4284 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4288 if (KEY_IS(KEY_BLOCKSIZE)) {
4289 __u32 *blocksize = val;
4291 if (*vallen < sizeof(*blocksize))
4293 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4295 *vallen = sizeof(*blocksize);
4299 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4300 __u32 *blocksize_bits = val;
4301 if (blocksize_bits) {
4302 if (*vallen < sizeof(*blocksize_bits))
4304 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4306 *vallen = sizeof(*blocksize_bits);
4310 if (KEY_IS(KEY_LAST_ID)) {
4311 obd_id *last_id = val;
4312 /* FIXME: object groups */
4314 if (*vallen < sizeof(*last_id))
4316 *last_id = filter_last_id(&obd->u.filter,
4317 exp->exp_filter_data.fed_group);
4319 *vallen = sizeof(*last_id);
4323 if (KEY_IS(KEY_FIEMAP)) {
4324 struct ll_fiemap_info_key *fm_key = key;
4325 struct dentry *dentry;
4326 struct ll_user_fiemap *fiemap = val;
4327 struct lvfs_run_ctxt saved;
4330 if (fiemap == NULL) {
4331 *vallen = fiemap_count_to_size(
4332 fm_key->fiemap.fm_extent_count);
4336 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4339 RETURN(PTR_ERR(dentry));
4341 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4342 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4343 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4345 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4351 CDEBUG(D_IOCTL, "invalid key\n");
4355 static inline int filter_setup_llog_group(struct obd_export *exp,
4356 struct obd_device *obd,
4359 struct obd_llog_group *olg;
4360 struct llog_ctxt *ctxt;
4363 olg = filter_find_create_olg(obd, group);
4365 RETURN(PTR_ERR(olg));
4367 llog_group_set_export(olg, exp);
4369 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4370 LASSERTF(ctxt != NULL, "ctxt is null\n");
4372 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4373 llog_ctxt_put(ctxt);
4377 static int filter_set_grant_shrink(struct obd_export *exp,
4378 struct ost_body *body)
4380 /* handle shrink grant */
4381 spin_lock(&exp->exp_obd->obd_osfs_lock);
4382 filter_grant_incoming(exp, &body->oa);
4383 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4389 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4391 struct obd_device *obd;
4397 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4401 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4402 obd_export_nid2str(exp));
4403 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4405 /* setup llog imports */
4407 group = (int)(*(__u32 *)val);
4409 group = 0; /* default value */
4411 LASSERT_MDS_GROUP(group);
4412 rc = filter_setup_llog_group(exp, obd, group);
4416 if (group == FILTER_GROUP_MDS0) {
4417 /* setup llog group 1 for interop */
4418 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4421 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4426 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4427 void *key, __u32 vallen, void *val,
4428 struct ptlrpc_request_set *set)
4430 struct obd_device *obd;
4435 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4439 if (KEY_IS(KEY_CAPA_KEY)) {
4441 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4443 CERROR("filter update capability key failed: %d\n", rc);
4447 if (KEY_IS(KEY_REVIMP_UPD)) {
4448 filter_revimp_update(exp);
4449 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4453 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4454 filter_adapt_sptlrpc_conf(obd, 0);
4458 if (KEY_IS(KEY_MDS_CONN))
4459 RETURN(filter_set_mds_conn(exp, val));
4461 if (KEY_IS(KEY_GRANT_SHRINK))
4462 RETURN(filter_set_grant_shrink(exp, val));
4467 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4468 int len, void *karg, void *uarg)
4470 struct obd_device *obd = exp->exp_obd;
4471 struct obd_ioctl_data *data = karg;
4475 case OBD_IOC_ABORT_RECOVERY: {
4476 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4477 target_stop_recovery_thread(obd);
4481 case OBD_IOC_SYNC: {
4482 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4483 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4487 case OBD_IOC_SET_READONLY: {
4489 struct super_block *sb = obd->u.obt.obt_sb;
4490 struct inode *inode = sb->s_root->d_inode;
4491 BDEVNAME_DECLARE_STORAGE(tmp);
4492 CERROR("*** setting device %s read-only ***\n",
4493 ll_bdevname(sb, tmp));
4495 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4496 if (!IS_ERR(handle))
4497 rc = fsfilt_commit(obd, inode, handle, 1);
4499 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4500 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4502 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4506 case OBD_IOC_CATLOGLIST: {
4507 rc = llog_catalog_list(obd, 1, data);
4511 case OBD_IOC_LLOG_CANCEL:
4512 case OBD_IOC_LLOG_REMOVE:
4513 case OBD_IOC_LLOG_INFO:
4514 case OBD_IOC_LLOG_PRINT: {
4515 /* FIXME to be finished */
4516 RETURN(-EOPNOTSUPP);
4518 struct llog_ctxt *ctxt = NULL;
4520 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4521 rc = llog_ioctl(ctxt, cmd, data);
4522 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4535 static int filter_health_check(struct obd_device *obd)
4537 #ifdef USE_HEALTH_CHECK_WRITE
4538 struct filter_obd *filter = &obd->u.filter;
4543 * health_check to return 0 on healthy
4544 * and 1 on unhealthy.
4546 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4549 #ifdef USE_HEALTH_CHECK_WRITE
4550 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4551 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4556 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4559 return filter_fid2dentry(data, NULL, gr, id);
4562 static int filter_process_config(struct obd_device *obd, obd_count len,
4565 struct lustre_cfg *lcfg = buf;
4566 struct lprocfs_static_vars lvars;
4569 switch (lcfg->lcfg_command) {
4571 lprocfs_filter_init_vars(&lvars);
4573 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4583 static struct lvfs_callback_ops filter_lvfs_ops = {
4584 l_fid2dentry: filter_lvfs_fid2dentry,
4587 static struct obd_ops filter_obd_ops = {
4588 .o_owner = THIS_MODULE,
4589 .o_get_info = filter_get_info,
4590 .o_set_info_async = filter_set_info_async,
4591 .o_setup = filter_setup,
4592 .o_precleanup = filter_precleanup,
4593 .o_cleanup = filter_cleanup,
4594 .o_connect = filter_connect,
4595 .o_reconnect = filter_reconnect,
4596 .o_disconnect = filter_disconnect,
4597 .o_ping = filter_ping,
4598 .o_init_export = filter_init_export,
4599 .o_destroy_export = filter_destroy_export,
4600 .o_statfs = filter_statfs,
4601 .o_getattr = filter_getattr,
4602 .o_unpackmd = filter_unpackmd,
4603 .o_create = filter_create,
4604 .o_setattr = filter_setattr,
4605 .o_destroy = filter_destroy,
4606 .o_brw = filter_brw,
4607 .o_punch = filter_truncate,
4608 .o_sync = filter_sync,
4609 .o_preprw = filter_preprw,
4610 .o_commitrw = filter_commitrw,
4611 .o_llog_init = filter_llog_init,
4612 .o_llog_connect = filter_llog_connect,
4613 .o_llog_finish = filter_llog_finish,
4614 .o_iocontrol = filter_iocontrol,
4615 .o_health_check = filter_health_check,
4616 .o_process_config = filter_process_config,
4619 quota_interface_t *filter_quota_interface_ref;
4620 extern quota_interface_t filter_quota_interface;
4622 static int __init obdfilter_init(void)
4624 struct lprocfs_static_vars lvars;
4627 /** sanity check for group<->mdsno conversion */
4628 for (i = 0; i < 32; i++)
4629 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4631 lprocfs_filter_init_vars(&lvars);
4633 request_module("%s", "lquota");
4634 OBD_ALLOC(obdfilter_created_scratchpad,
4635 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4636 sizeof(*obdfilter_created_scratchpad));
4637 if (obdfilter_created_scratchpad == NULL)
4640 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4641 sizeof(struct filter_mod_data),
4644 GOTO(out, rc = -ENOMEM);
4646 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4647 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4649 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4650 LUSTRE_OST_NAME, NULL);
4654 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4655 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4656 ll_fmd_cachep = NULL;
4658 if (filter_quota_interface_ref)
4659 PORTAL_SYMBOL_PUT(filter_quota_interface);
4661 OBD_FREE(obdfilter_created_scratchpad,
4662 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4663 sizeof(*obdfilter_created_scratchpad));
4669 static void __exit obdfilter_exit(void)
4671 if (filter_quota_interface_ref)
4672 PORTAL_SYMBOL_PUT(filter_quota_interface);
4674 if (ll_fmd_cachep) {
4675 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4676 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4677 ll_fmd_cachep = NULL;
4680 class_unregister_type(LUSTRE_OST_NAME);
4681 OBD_FREE(obdfilter_created_scratchpad,
4682 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4683 sizeof(*obdfilter_created_scratchpad));
4686 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4687 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4688 MODULE_LICENSE("GPL");
4690 module_init(obdfilter_init);
4691 module_exit(obdfilter_exit);