1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
28 * (which need to get journal_lock, may block if journal full).
30 * Invariant: Call filter_start_transno() before any journal ops to avoid the
31 * same deadlock problem. We can (and want) to get rid of the
32 * transno sem in favour of the dir/inode i_mutex to avoid single
33 * threaded operation on the OST.
36 #define DEBUG_SUBSYSTEM S_FILTER
38 #ifndef AUTOCONF_INCLUDED
39 #include <linux/config.h>
41 #include <linux/module.h>
43 #include <linux/dcache.h>
44 #include <linux/init.h>
45 #include <linux/version.h>
46 #include <linux/sched.h>
47 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
48 # include <linux/mount.h>
49 # include <linux/buffer_head.h>
52 #include <obd_class.h>
54 #include <lustre_dlm.h>
55 #include <lustre_fsfilt.h>
56 #include <lprocfs_status.h>
57 #include <lustre_log.h>
58 #include <lustre_commit_confd.h>
59 #include <libcfs/list.h>
60 #include <lustre_disk.h>
61 #include <lustre_quota.h>
62 #include <linux/slab.h>
63 #include <lustre_param.h>
65 #include "filter_internal.h"
67 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
68 #define FILTER_MIN_GROUPS 3
69 static struct lvfs_callback_ops filter_lvfs_ops;
70 cfs_mem_cache_t *ll_fmd_cachep;
72 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
73 void *cb_data, int error)
75 obd_transno_commit_cb(obd, transno, error);
78 /* Assumes caller has already pushed us into the kernel context. */
79 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
80 int rc, int force_sync)
82 struct filter_obd *filter = &exp->exp_obd->u.filter;
83 struct filter_export_data *fed = &exp->exp_filter_data;
84 struct filter_client_data *fcd = fed->fed_fcd;
87 int err, log_pri = D_HA;
89 /* Propagate error code. */
93 if (!exp->exp_obd->obd_replayable || oti == NULL)
96 /* we don't allocate new transnos for replayed requests */
97 if (oti->oti_transno == 0) {
98 spin_lock(&filter->fo_translock);
99 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
100 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
101 spin_unlock(&filter->fo_translock);
102 oti->oti_transno = last_rcvd;
104 spin_lock(&filter->fo_translock);
105 last_rcvd = oti->oti_transno;
106 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
107 filter->fo_fsd->lsd_last_transno =
108 cpu_to_le64(last_rcvd);
109 spin_unlock(&filter->fo_translock);
111 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
113 /* could get xid from oti, if it's ever needed */
114 fcd->fcd_last_xid = 0;
116 off = fed->fed_lr_off;
118 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
119 fed->fed_lr_idx, fed->fed_lr_off);
123 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
129 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
130 fcd, sizeof(*fcd), &off,
131 force_sync | exp->exp_need_sync);
133 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
141 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
142 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
147 void f_dput(struct dentry *dentry)
149 /* Can't go inside filter_ddelete because it can block */
150 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
151 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
152 LASSERT(atomic_read(&dentry->d_count) > 0);
157 static void init_brw_stats(struct brw_stats *brw_stats)
160 for (i = 0; i < BRW_LAST; i++)
161 spin_lock_init(&brw_stats->hist[i].oh_lock);
164 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
165 plus the procfs overhead :( */
166 static int filter_export_stats_init(struct obd_device *obd,
167 struct obd_export *exp)
169 struct filter_export_data *fed = &exp->exp_filter_data;
170 struct proc_dir_entry *brw_entry;
174 init_brw_stats(&fed->fed_brw_stats);
176 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
177 /* Self-export gets no proc entry */
180 rc = lprocfs_exp_setup(exp);
184 /* Create a per export proc entry for brw_stats */
185 brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
186 if (brw_entry == NULL)
188 brw_entry->proc_fops = &filter_per_export_stats_fops;
189 brw_entry->data = fed;
191 /* Create a per export proc entry for ops stats */
192 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
193 LPROC_FILTER_LAST - 1;
194 exp->exp_ops_stats = lprocfs_alloc_stats(num_stats,
195 LPROCFS_STATS_FLAG_NOPERCPU);
196 if (exp->exp_ops_stats == NULL)
198 lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
199 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
200 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
201 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
202 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
203 lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
207 /* Add client data to the FILTER. We use a bitmap to locate a free space
208 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
209 * Otherwise, we have just read the data from the last_rcvd file and
210 * we know its offset. */
211 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
214 struct filter_obd *filter = &obd->u.filter;
215 struct filter_export_data *fed = &exp->exp_filter_data;
216 unsigned long *bitmap = filter->fo_last_rcvd_slots;
217 int new_client = (cl_idx == -1);
221 LASSERT(bitmap != NULL);
222 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
225 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
228 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
229 * there's no need for extra complication here
232 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
234 if (cl_idx >= LR_MAX_CLIENTS) {
235 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
239 if (test_and_set_bit(cl_idx, bitmap)) {
240 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
245 if (test_and_set_bit(cl_idx, bitmap)) {
246 CERROR("FILTER client %d: bit already set in bitmap!\n",
252 fed->fed_lr_idx = cl_idx;
253 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
254 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
255 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
257 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
258 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
261 struct lvfs_run_ctxt saved;
262 loff_t off = fed->fed_lr_off;
266 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
267 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
269 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
270 /* Transaction needed to fix bug 1403 */
271 handle = fsfilt_start(obd,
272 filter->fo_rcvd_filp->f_dentry->d_inode,
273 FSFILT_OP_SETATTR, NULL);
274 if (IS_ERR(handle)) {
275 rc = PTR_ERR(handle);
276 CERROR("unable to start transaction: rc %d\n", rc);
278 rc = fsfilt_add_journal_cb(obd, 0, handle,
279 target_client_add_cb, exp);
281 spin_lock(&exp->exp_lock);
282 exp->exp_need_sync = 1;
283 spin_unlock(&exp->exp_lock);
285 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
287 sizeof(*fed->fed_fcd),
288 &off, rc /* sync if no cb */);
290 filter->fo_rcvd_filp->f_dentry->d_inode,
293 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
296 CERROR("error writing %s client idx %u: rc %d\n",
297 LAST_RCVD, fed->fed_lr_idx, rc);
304 static int filter_client_free(struct obd_export *exp)
306 struct filter_export_data *fed = &exp->exp_filter_data;
307 struct filter_obd *filter = &exp->exp_obd->u.filter;
308 struct obd_device *obd = exp->exp_obd;
309 struct filter_client_data zero_fcd;
310 struct lvfs_run_ctxt saved;
315 if (fed->fed_fcd == NULL)
318 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
319 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
322 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
323 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
325 LASSERT(filter->fo_last_rcvd_slots != NULL);
327 off = fed->fed_lr_off;
329 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
330 * we leak a client slot that will be cleaned on the next recovery. */
332 CERROR("%s: client idx %d has med_off %lld\n",
333 obd->obd_name, fed->fed_lr_idx, off);
334 GOTO(free, rc = -EINVAL);
337 /* Clear the bit _after_ zeroing out the client so we don't
338 race with filter_client_add and zero out new clients.*/
339 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
340 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
345 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
346 memset(&zero_fcd, 0, sizeof zero_fcd);
347 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
348 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
349 sizeof(zero_fcd), &off,
350 (!exp->exp_libclient ||
351 exp->exp_need_sync));
353 /* update server's transno */
354 filter_update_server_data(obd, filter->fo_rcvd_filp,
356 !exp->exp_libclient);
357 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
359 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
360 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
361 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
365 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
366 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
373 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
379 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
380 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
381 struct filter_mod_data *fmd)
383 LASSERT_SPIN_LOCKED(&fed->fed_lock);
384 if (--fmd->fmd_refcount == 0) {
385 /* XXX when we have persistent reservations and the handle
386 * is stored herein we need to drop it here. */
387 fed->fed_mod_count--;
388 list_del(&fmd->fmd_list);
389 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
393 /* drop fmd reference, free it if last ref */
394 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
396 struct filter_export_data *fed;
401 fed = &exp->exp_filter_data;
402 spin_lock(&fed->fed_lock);
403 filter_fmd_put_nolock(fed, fmd); /* caller reference */
404 spin_unlock(&fed->fed_lock);
407 /* expire entries from the end of the list if there are too many
408 * or they are too old */
409 static void filter_fmd_expire_nolock(struct filter_obd *filter,
410 struct filter_export_data *fed,
411 struct filter_mod_data *keep)
413 struct filter_mod_data *fmd, *tmp;
415 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
419 if (time_before(jiffies, fmd->fmd_expire) &&
420 fed->fed_mod_count < filter->fo_fmd_max_num)
423 list_del_init(&fmd->fmd_list);
424 filter_fmd_put_nolock(fed, fmd); /* list reference */
428 void filter_fmd_expire(struct obd_export *exp)
430 spin_lock(&exp->exp_filter_data.fed_lock);
431 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
432 &exp->exp_filter_data, NULL);
433 spin_unlock(&exp->exp_filter_data.fed_lock);
436 /* find specified objid, group in export fmd list.
437 * caller must hold fed_lock and take fmd reference itself */
438 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
439 struct filter_export_data *fed,
440 obd_id objid, obd_gr group)
442 struct filter_mod_data *found = NULL, *fmd;
444 LASSERT_SPIN_LOCKED(&fed->fed_lock);
446 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
447 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
449 list_del(&fmd->fmd_list);
450 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
451 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
456 filter_fmd_expire_nolock(filter, fed, found);
461 /* Find fmd based on objid and group, or return NULL if not found. */
462 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
463 obd_id objid, obd_gr group)
465 struct filter_mod_data *fmd;
467 spin_lock(&exp->exp_filter_data.fed_lock);
468 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
469 &exp->exp_filter_data, objid, group);
471 fmd->fmd_refcount++; /* caller reference */
472 spin_unlock(&exp->exp_filter_data.fed_lock);
477 /* Find fmd based on objid and group, or create a new one if none is found.
478 * It is possible for this function to return NULL under memory pressure,
479 * or if objid = 0 is passed (which will only cause old entries to expire).
480 * Currently this is not fatal because any fmd state is transient and
481 * may also be freed when it gets sufficiently old. */
482 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
483 obd_id objid, obd_gr group)
485 struct filter_export_data *fed = &exp->exp_filter_data;
486 struct filter_mod_data *found = NULL, *fmd_new = NULL;
488 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
490 spin_lock(&fed->fed_lock);
491 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
494 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
495 fmd_new->fmd_id = objid;
496 fmd_new->fmd_gr = group;
497 fmd_new->fmd_refcount++; /* list reference */
499 fed->fed_mod_count++;
501 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
505 found->fmd_refcount++; /* caller reference */
506 found->fmd_expire = jiffies +
507 exp->exp_obd->u.filter.fo_fmd_max_age;
510 spin_unlock(&fed->fed_lock);
516 /* drop fmd list reference so it will disappear when last reference is put.
517 * This isn't so critical because it would in fact only affect the one client
518 * that is doing the unlink and at worst we have an stale entry referencing
519 * an object that should never be used again. */
520 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
522 struct filter_mod_data *found = NULL;
524 spin_lock(&exp->exp_filter_data.fed_lock);
525 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
527 list_del_init(&found->fmd_list);
528 filter_fmd_put_nolock(&exp->exp_filter_data, found);
530 spin_unlock(&exp->exp_filter_data.fed_lock);
533 #define filter_fmd_drop(exp, objid, group)
536 /* remove all entries from fmd list */
537 static void filter_fmd_cleanup(struct obd_export *exp)
539 struct filter_export_data *fed = &exp->exp_filter_data;
540 struct filter_mod_data *fmd = NULL, *tmp;
542 spin_lock(&fed->fed_lock);
543 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
544 list_del_init(&fmd->fmd_list);
545 filter_fmd_put_nolock(fed, fmd);
547 spin_unlock(&fed->fed_lock);
550 static int filter_init_export(struct obd_export *exp)
552 spin_lock_init(&exp->exp_filter_data.fed_lock);
553 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
555 spin_lock(&exp->exp_lock);
556 exp->exp_connecting = 1;
557 spin_unlock(&exp->exp_lock);
562 static int filter_free_server_data(struct filter_obd *filter)
564 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
565 filter->fo_fsd = NULL;
566 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
567 filter->fo_last_rcvd_slots = NULL;
571 /* assumes caller is already in kernel ctxt */
572 int filter_update_server_data(struct obd_device *obd, struct file *filp,
573 struct lr_server_data *fsd, int force_sync)
579 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
580 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
581 le64_to_cpu(fsd->lsd_last_transno));
582 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
583 le64_to_cpu(fsd->lsd_mount_count));
585 fsd->lsd_compat14 = fsd->lsd_last_transno;
586 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
588 CERROR("error writing lr_server_data: rc = %d\n", rc);
593 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
596 struct filter_obd *filter = &obd->u.filter;
602 if (filter->fo_last_objid_files[group] == NULL) {
603 CERROR("Object group "LPU64" not fully setup; not updating "
604 "last_objid\n", group);
608 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
609 obd->obd_name, group, filter->fo_last_objids[group]);
611 tmp = cpu_to_le64(filter->fo_last_objids[group]);
612 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
613 &tmp, sizeof(tmp), &off, force_sync);
615 CERROR("error writing group "LPU64" last objid: rc = %d\n",
619 extern int ost_handle(struct ptlrpc_request *req);
620 /* assumes caller has already in kernel ctxt */
621 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
623 struct filter_obd *filter = &obd->u.filter;
624 struct lr_server_data *fsd;
625 struct filter_client_data *fcd = NULL;
626 struct inode *inode = filp->f_dentry->d_inode;
627 unsigned long last_rcvd_size = i_size_read(inode);
633 /* ensure padding in the struct is the correct size */
634 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
635 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
636 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
637 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
639 OBD_ALLOC(fsd, sizeof(*fsd));
642 filter->fo_fsd = fsd;
644 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
645 if (filter->fo_last_rcvd_slots == NULL) {
646 OBD_FREE(fsd, sizeof(*fsd));
650 if (last_rcvd_size == 0) {
651 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
653 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
654 fsd->lsd_last_transno = 0;
655 mount_count = fsd->lsd_mount_count = 0;
656 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
657 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
658 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
659 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
660 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
661 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
663 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
665 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
669 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
670 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
671 "using the wrong disk %s. Were the "
672 "/dev/ assignments rearranged?\n",
673 obd->obd_uuid.uuid, fsd->lsd_uuid);
674 GOTO(err_fsd, rc = -EINVAL);
676 mount_count = le64_to_cpu(fsd->lsd_mount_count);
677 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
679 /* Assume old last_rcvd format unless I_C_LR is set */
680 if (!(fsd->lsd_feature_incompat &
681 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
682 fsd->lsd_last_transno = fsd->lsd_compat14;
686 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
687 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
688 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
689 ~FILTER_INCOMPAT_SUPP);
690 GOTO(err_fsd, rc = -EINVAL);
692 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
693 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
694 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
695 ~FILTER_ROCOMPAT_SUPP);
696 /* Do something like remount filesystem read-only */
697 GOTO(err_fsd, rc = -EINVAL);
700 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
701 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
702 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
703 obd->obd_name, mount_count + 1);
704 CDEBUG(D_INODE, "%s: server data size: %u\n",
705 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
706 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
707 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
708 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
709 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
710 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
711 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
712 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
713 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
714 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
715 le16_to_cpu(fsd->lsd_client_size));
717 if (!obd->obd_replayable) {
718 CWARN("%s: recovery support OFF\n", obd->obd_name);
722 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
723 off < last_rcvd_size; cl_idx++) {
725 struct obd_export *exp;
726 struct filter_export_data *fed;
729 OBD_ALLOC(fcd, sizeof(*fcd));
731 GOTO(err_client, rc = -ENOMEM);
734 /* Don't assume off is incremented properly by
735 * fsfilt_read_record(), in case sizeof(*fcd)
736 * isn't the same as fsd->lsd_client_size. */
737 off = le32_to_cpu(fsd->lsd_client_start) +
738 cl_idx * le16_to_cpu(fsd->lsd_client_size);
739 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
741 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
742 LAST_RCVD, cl_idx, off, rc);
743 break; /* read error shouldn't cause startup to fail */
746 if (fcd->fcd_uuid[0] == '\0') {
747 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
752 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
754 /* These exports are cleaned up by filter_disconnect(), so they
755 * need to be set up like real exports as filter_connect() does.
757 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
759 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
760 " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
761 last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
762 le32_to_cpu(fcd->fcd_group));
764 if (PTR_ERR(exp) == -EALREADY) {
765 /* export already exists, zero out this one */
766 CERROR("Zeroing out duplicate export due to "
768 fcd->fcd_uuid[0] = '\0';
770 GOTO(err_client, rc = PTR_ERR(exp));
773 fed = &exp->exp_filter_data;
775 fed->fed_group = le32_to_cpu(fcd->fcd_group);
776 filter_export_stats_init(obd, exp);
777 rc = filter_client_add(obd, exp, cl_idx);
778 /* can't fail for existing client */
779 LASSERTF(rc == 0, "rc = %d\n", rc);
782 spin_lock(&exp->exp_lock);
783 exp->exp_connecting = 0;
784 exp->exp_in_recovery = 0;
785 spin_unlock(&exp->exp_lock);
786 obd->obd_max_recoverable_clients++;
787 class_export_put(exp);
790 /* Need to check last_rcvd even for duplicated exports. */
791 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
794 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
795 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
799 OBD_FREE(fcd, sizeof(*fcd));
801 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
803 target_recovery_init(obd, ost_handle);
806 filter->fo_mount_count = mount_count + 1;
807 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
809 /* save it, so mount count and last_transno is current */
810 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
812 GOTO(err_client, rc);
817 target_recovery_fini(obd);
819 filter_free_server_data(filter);
823 static int filter_cleanup_groups(struct obd_device *obd)
825 struct filter_obd *filter = &obd->u.filter;
827 struct dentry *dentry;
831 if (filter->fo_dentry_O_groups != NULL) {
832 for (i = 0; i < filter->fo_group_count; i++) {
833 dentry = filter->fo_dentry_O_groups[i];
837 OBD_FREE(filter->fo_dentry_O_groups,
838 filter->fo_group_count *
839 sizeof(*filter->fo_dentry_O_groups));
840 filter->fo_dentry_O_groups = NULL;
842 if (filter->fo_last_objid_files != NULL) {
843 for (i = 0; i < filter->fo_group_count; i++) {
844 filp = filter->fo_last_objid_files[i];
848 OBD_FREE(filter->fo_last_objid_files,
849 filter->fo_group_count *
850 sizeof(*filter->fo_last_objid_files));
851 filter->fo_last_objid_files = NULL;
853 if (filter->fo_dentry_O_sub != NULL) {
854 for (i = 0; i < filter->fo_group_count; i++) {
855 for (j = 0; j < filter->fo_subdir_count; j++) {
856 dentry = filter->fo_dentry_O_sub[i].dentry[j];
861 OBD_FREE(filter->fo_dentry_O_sub,
862 filter->fo_group_count *
863 sizeof(*filter->fo_dentry_O_sub));
864 filter->fo_dentry_O_sub = NULL;
866 if (filter->fo_last_objids != NULL) {
867 OBD_FREE(filter->fo_last_objids,
868 filter->fo_group_count *
869 sizeof(*filter->fo_last_objids));
870 filter->fo_last_objids = NULL;
872 if (filter->fo_dentry_O != NULL) {
873 f_dput(filter->fo_dentry_O);
874 filter->fo_dentry_O = NULL;
879 static int filter_update_last_group(struct obd_device *obd, int group)
881 struct filter_obd *filter = &obd->u.filter;
882 struct file *filp = NULL;
883 int last_group = 0, rc;
887 if (group <= filter->fo_committed_group)
890 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
894 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
898 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
900 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
903 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
904 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
905 obd->obd_name, last_group, group);
909 /* must be sync: bXXXX */
910 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
912 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
916 filter->fo_committed_group = group;
923 static int filter_read_group_internal(struct obd_device *obd, int group,
926 struct filter_obd *filter = &obd->u.filter;
927 __u64 *new_objids = NULL;
928 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
929 struct dentry **new_groups = NULL;
930 struct file **new_files = NULL;
931 struct dentry *dentry;
933 int old_count = filter->fo_group_count, rc, stage = 0, i;
939 snprintf(name, 24, "%d", group);
943 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
945 if (IS_ERR(dentry)) {
946 CERROR("Cannot lookup expected object group %d: %ld\n",
947 group, PTR_ERR(dentry));
948 RETURN(PTR_ERR(dentry));
951 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
952 if (IS_ERR(dentry)) {
953 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
955 RETURN(PTR_ERR(dentry));
960 snprintf(name, 24, "O/%d/LAST_ID", group);
962 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
964 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
965 GOTO(cleanup, rc = PTR_ERR(filp));
969 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
971 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
975 if (filter->fo_subdir_count) {
976 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
977 if (tmp_subdirs == NULL)
978 GOTO(cleanup, rc = -ENOMEM);
981 for (i = 0; i < filter->fo_subdir_count; i++) {
983 snprintf(dir, sizeof(dir), "d%u", i);
985 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
986 if (IS_ERR(tmp_subdirs->dentry[i])) {
987 rc = PTR_ERR(tmp_subdirs->dentry[i]);
988 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
993 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
994 tmp_subdirs->dentry[i]);
998 /* 'group' is an index; we need an array of length 'group + 1' */
999 if (group + 1 > old_count) {
1000 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1001 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1002 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1003 OBD_ALLOC(new_files, len * sizeof(*new_files));
1005 if (new_objids == NULL || new_subdirs == NULL ||
1006 new_groups == NULL || new_files == NULL)
1007 GOTO(cleanup, rc = -ENOMEM);
1010 memcpy(new_objids, filter->fo_last_objids,
1011 old_count * sizeof(*new_objids));
1012 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1013 old_count * sizeof(*new_subdirs));
1014 memcpy(new_groups, filter->fo_dentry_O_groups,
1015 old_count * sizeof(*new_groups));
1016 memcpy(new_files, filter->fo_last_objid_files,
1017 old_count * sizeof(*new_files));
1019 OBD_FREE(filter->fo_last_objids,
1020 old_count * sizeof(*new_objids));
1021 OBD_FREE(filter->fo_dentry_O_sub,
1022 old_count * sizeof(*new_subdirs));
1023 OBD_FREE(filter->fo_dentry_O_groups,
1024 old_count * sizeof(*new_groups));
1025 OBD_FREE(filter->fo_last_objid_files,
1026 old_count * sizeof(*new_files));
1028 filter->fo_last_objids = new_objids;
1029 filter->fo_dentry_O_sub = new_subdirs;
1030 filter->fo_dentry_O_groups = new_groups;
1031 filter->fo_last_objid_files = new_files;
1032 filter->fo_group_count = len;
1035 filter->fo_dentry_O_groups[group] = dentry;
1036 filter->fo_last_objid_files[group] = filp;
1037 if (filter->fo_subdir_count) {
1038 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1039 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1042 filter_update_last_group(obd, group);
1044 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1045 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1046 rc = filter_update_last_objid(obd, group, 1);
1050 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1051 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1052 obd->obd_name, group, last_objid);
1057 if (new_objids != NULL)
1058 OBD_FREE(new_objids, len * sizeof(*new_objids));
1059 if (new_subdirs != NULL)
1060 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1061 if (new_groups != NULL)
1062 OBD_FREE(new_groups, len * sizeof(*new_groups));
1063 if (new_files != NULL)
1064 OBD_FREE(new_files, len * sizeof(*new_files));
1066 if (filter->fo_subdir_count) {
1067 for (i = 0; i < filter->fo_subdir_count; i++) {
1068 if (tmp_subdirs->dentry[i] != NULL)
1069 dput(tmp_subdirs->dentry[i]);
1071 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1074 filp_close(filp, 0);
1081 static int filter_read_groups(struct obd_device *obd, int last_group,
1084 struct filter_obd *filter = &obd->u.filter;
1085 int old_count, group, rc = 0;
1087 down(&filter->fo_init_lock);
1088 old_count = filter->fo_group_count;
1089 for (group = old_count; group <= last_group; group++) {
1091 continue; /* no group zero */
1093 rc = filter_read_group_internal(obd, group, create);
1097 up(&filter->fo_init_lock);
1101 /* FIXME: object groups */
1102 static int filter_prep_groups(struct obd_device *obd)
1104 struct filter_obd *filter = &obd->u.filter;
1105 struct dentry *dentry, *O_dentry;
1107 int last_group, rc = 0, cleanup_phase = 0;
1111 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1112 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1113 if (IS_ERR(O_dentry)) {
1114 rc = PTR_ERR(O_dentry);
1115 CERROR("cannot open/create O: rc = %d\n", rc);
1118 filter->fo_dentry_O = O_dentry;
1119 cleanup_phase = 1; /* O_dentry */
1121 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1122 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1123 * some time post 1.0 when all old-style OSTs have converted along
1124 * with the init_objid hack. */
1125 dentry = ll_lookup_one_len("R", O_dentry, 1);
1127 GOTO(cleanup, rc = PTR_ERR(dentry));
1128 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1129 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1132 CWARN("converting OST to new object layout\n");
1133 if (IS_ERR(O0_dentry)) {
1134 rc = PTR_ERR(O0_dentry);
1135 CERROR("error looking up O/0: rc %d\n", rc);
1136 GOTO(cleanup_R, rc);
1139 if (O0_dentry->d_inode) {
1140 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1141 GOTO(cleanup_O0, rc = -EEXIST);
1144 LOCK_INODE_MUTEX(O_dentry->d_inode);
1145 rc = vfs_rename(O_dentry->d_inode, dentry,
1146 O_dentry->d_inode, O0_dentry);
1147 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1150 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1151 GOTO(cleanup_O0, rc);
1153 filter->fo_fsd->lsd_feature_incompat |=
1154 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1155 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1157 GOTO(cleanup_O0, rc);
1169 cleanup_phase = 2; /* groups */
1171 /* we have to initialize all groups before first connections from
1172 * clients because they may send create/destroy for any group -bzzz */
1173 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1175 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1176 GOTO(cleanup, rc = PTR_ERR(filp));
1178 cleanup_phase = 3; /* filp */
1180 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1182 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1186 last_group = FILTER_MIN_GROUPS;
1188 LASSERT(last_group >= FILTER_MIN_GROUPS);
1191 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1192 FILTER_MIN_GROUPS, last_group);
1193 filter->fo_committed_group = last_group;
1194 rc = filter_read_groups(obd, last_group, 1);
1198 filp_close(filp, 0);
1202 switch (cleanup_phase) {
1204 filp_close(filp, 0);
1206 filter_cleanup_groups(obd);
1208 f_dput(filter->fo_dentry_O);
1209 filter->fo_dentry_O = NULL;
1217 /* setup the object store with correct subdirectories */
1218 static int filter_prep(struct obd_device *obd)
1220 struct lvfs_run_ctxt saved;
1221 struct filter_obd *filter = &obd->u.filter;
1223 struct inode *inode;
1227 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1228 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1229 if (!file || IS_ERR(file)) {
1231 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1235 filter->fo_rcvd_filp = file;
1236 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1237 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1238 file->f_dentry->d_inode->i_mode);
1239 GOTO(err_filp, rc = -ENOENT);
1242 inode = file->f_dentry->d_parent->d_inode;
1243 /* We use i_op->unlink directly in filter_vfs_unlink() */
1244 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1245 CERROR("%s: filesystem does not support create/unlink ops\n",
1247 GOTO(err_filp, rc = -EOPNOTSUPP);
1250 rc = filter_init_server_data(obd, file);
1252 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1255 /* open and create health check io file*/
1256 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1259 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1261 GOTO(err_server_data, rc);
1263 filter->fo_health_check_filp = file;
1264 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1265 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1266 file->f_dentry->d_inode->i_mode);
1267 GOTO(err_health_check, rc = -ENOENT);
1269 rc = lvfs_check_io_health(obd, file);
1271 GOTO(err_health_check, rc);
1273 rc = filter_prep_groups(obd);
1275 GOTO(err_health_check, rc);
1277 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1282 if (filp_close(filter->fo_health_check_filp, 0))
1283 CERROR("can't close %s after error\n", HEALTH_CHECK);
1284 filter->fo_health_check_filp = NULL;
1286 target_recovery_fini(obd);
1287 filter_free_server_data(filter);
1289 if (filp_close(filter->fo_rcvd_filp, 0))
1290 CERROR("can't close %s after error\n", LAST_RCVD);
1291 filter->fo_rcvd_filp = NULL;
1295 /* cleanup the filter: write last used object id to status file */
1296 static void filter_post(struct obd_device *obd)
1298 struct lvfs_run_ctxt saved;
1299 struct filter_obd *filter = &obd->u.filter;
1302 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1303 * best to start a transaction with h_sync, because we removed this
1306 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1307 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1310 CERROR("error writing server data: rc = %d\n", rc);
1312 for (i = 1; i < filter->fo_group_count; i++) {
1313 rc = filter_update_last_objid(obd, i,
1314 (i == filter->fo_group_count - 1));
1316 CERROR("error writing group %d lastobjid: rc = %d\n",
1320 rc = filp_close(filter->fo_rcvd_filp, 0);
1321 filter->fo_rcvd_filp = NULL;
1323 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1325 rc = filp_close(filter->fo_health_check_filp, 0);
1326 filter->fo_health_check_filp = NULL;
1328 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1330 filter_cleanup_groups(obd);
1331 filter_free_server_data(filter);
1332 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1334 filter_free_capa_keys(filter);
1335 cleanup_capa_hash(filter->fo_capa_hash);
1338 static void filter_set_last_id(struct filter_obd *filter,
1339 obd_id id, obd_gr group)
1341 LASSERT(filter->fo_fsd != NULL);
1342 LASSERT(group <= filter->fo_group_count);
1344 spin_lock(&filter->fo_objidlock);
1345 filter->fo_last_objids[group] = id;
1346 spin_unlock(&filter->fo_objidlock);
1349 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1352 LASSERT(filter->fo_fsd != NULL);
1353 LASSERT(group <= filter->fo_group_count);
1355 /* FIXME: object groups */
1356 spin_lock(&filter->fo_objidlock);
1357 id = filter->fo_last_objids[group];
1358 spin_unlock(&filter->fo_objidlock);
1363 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1365 LOCK_INODE_MUTEX(dparent->d_inode);
1369 /* We never dget the object parent, so DON'T dput it either */
1370 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1372 struct filter_obd *filter = &obd->u.filter;
1373 struct filter_subdirs *subdirs;
1374 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1376 if (group > 0 || filter->fo_subdir_count == 0)
1377 return filter->fo_dentry_O_groups[group];
1379 subdirs = &filter->fo_dentry_O_sub[group];
1380 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1383 /* We never dget the object parent, so DON'T dput it either */
1384 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1387 unsigned long now = jiffies;
1388 struct dentry *dparent = filter_parent(obd, group, objid);
1391 if (IS_ERR(dparent))
1394 rc = filter_lock_dentry(obd, dparent);
1395 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1396 return rc ? ERR_PTR(rc) : dparent;
1399 /* We never dget the object parent, so DON'T dput it either */
1400 static void filter_parent_unlock(struct dentry *dparent)
1402 UNLOCK_INODE_MUTEX(dparent->d_inode);
1405 /* How to get files, dentries, inodes from object id's.
1407 * If dir_dentry is passed, the caller has already locked the parent
1408 * appropriately for this operation (normally a write lock). If
1409 * dir_dentry is NULL, we do a read lock while we do the lookup to
1410 * avoid races with create/destroy and such changing the directory
1411 * internal to the filesystem code. */
1412 struct dentry *filter_fid2dentry(struct obd_device *obd,
1413 struct dentry *dir_dentry,
1414 obd_gr group, obd_id id)
1416 struct dentry *dparent = dir_dentry;
1417 struct dentry *dchild;
1422 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1423 obd->u.filter.fo_destroys_in_progress == 0) {
1424 /* don't fail lookups for orphan recovery, it causes
1425 * later LBUGs when objects still exist during precreate */
1426 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1427 RETURN(ERR_PTR(-ENOENT));
1430 CERROR("fatal: invalid object id 0\n");
1431 RETURN(ERR_PTR(-ESTALE));
1434 len = sprintf(name, LPU64, id);
1435 if (dir_dentry == NULL) {
1436 dparent = filter_parent_lock(obd, group, id);
1437 if (IS_ERR(dparent)) {
1438 CERROR("%s: error getting object "LPU64":"LPU64
1439 " parent: rc %ld\n", obd->obd_name,
1440 id, group, PTR_ERR(dparent));
1444 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1445 dparent->d_name.len, dparent->d_name.name, name);
1446 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1447 if (dir_dentry == NULL)
1448 filter_parent_unlock(dparent);
1449 if (IS_ERR(dchild)) {
1450 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1455 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1456 CERROR("%s: got bad object "LPU64" inode %lu\n",
1457 obd->obd_name, id, dchild->d_inode->i_ino);
1459 RETURN(ERR_PTR(-ENOENT));
1462 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1463 name, dchild, atomic_read(&dchild->d_count));
1465 LASSERT(atomic_read(&dchild->d_count) > 0);
1470 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1473 struct lustre_handle lockh;
1474 int flags = LDLM_AST_DISCARD_DATA, rc;
1475 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1476 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1479 /* Tell the clients that the object is gone now and that they should
1480 * throw away any cached pages. */
1481 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1482 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1483 ldlm_completion_ast, NULL, NULL, 0, NULL,
1486 /* We only care about the side-effects, just drop the lock. */
1488 ldlm_lock_decref(&lockh, LCK_PW);
1493 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1494 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1495 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1496 * reverse. Caller must take i_sem before starting the transaction and we
1497 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1498 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1503 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1504 * itself we already hold dir->i_mutex for child create/unlink ops */
1505 LASSERT(dentry->d_inode != NULL);
1506 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1507 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1511 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1512 GOTO(out, rc = -ENOENT);
1514 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1519 GOTO(out, rc = -EPERM);
1521 /* check_sticky() */
1522 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1523 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1524 GOTO(out, rc = -EPERM);
1526 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1527 * that was done because of journal_start (which is already done
1528 * here) or some other ordering issue. */
1531 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1532 rc = security_inode_unlink(dir, dentry);
1537 rc = dir->i_op->unlink(dir, dentry);
1539 /* need to drop i_mutex before we lose inode reference */
1540 UNLOCK_INODE_MUTEX(dentry->d_inode);
1547 /* Caller must hold LCK_PW on parent and push us into kernel context.
1548 * Caller must hold child i_mutex, we drop it always.
1549 * Caller is also required to ensure that dchild->d_inode exists. */
1550 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1551 obd_gr group, struct dentry *dparent,
1552 struct dentry *dchild)
1554 struct inode *inode = dchild->d_inode;
1557 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1558 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1559 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1560 (unsigned long)inode->i_nlink,
1561 atomic_read(&inode->i_count));
1564 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1566 CERROR("error unlinking objid %.*s: rc %d\n",
1567 dchild->d_name.len, dchild->d_name.name, rc);
1571 static int filter_intent_policy(struct ldlm_namespace *ns,
1572 struct ldlm_lock **lockp, void *req_cookie,
1573 ldlm_mode_t mode, int flags, void *data)
1575 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1576 struct ptlrpc_request *req = req_cookie;
1577 struct ldlm_lock *lock = *lockp, *l = NULL;
1578 struct ldlm_resource *res = lock->l_resource;
1579 ldlm_processing_policy policy;
1580 struct ost_lvb *res_lvb, *reply_lvb;
1581 struct ldlm_reply *rep;
1582 struct list_head *tmp;
1584 int rc, tmpflags = 0, only_liblustre = 0;
1585 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1586 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1587 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1590 policy = ldlm_get_processing_policy(res);
1591 LASSERT(policy != NULL);
1592 LASSERT(req != NULL);
1594 rc = lustre_pack_reply(req, 3, repsize, NULL);
1596 RETURN(req->rq_status = rc);
1598 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1599 LASSERT(rep != NULL);
1601 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1602 sizeof(*reply_lvb));
1603 LASSERT(reply_lvb != NULL);
1605 //fixup_handle_for_resent_req(req, lock, &lockh);
1607 /* If we grant any lock at all, it will be a whole-file read lock.
1608 * Call the extent policy function to see if our request can be
1609 * granted, or is blocked. */
1610 lock->l_policy_data.l_extent.start = 0;
1611 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1612 lock->l_req_mode = LCK_PR;
1614 LASSERT(ns == res->lr_namespace);
1616 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1617 check_res_locked(res);
1619 /* FIXME: we should change the policy function slightly, to not make
1620 * this list at all, since we just turn around and free it */
1621 while (!list_empty(&rpc_list)) {
1622 struct ldlm_lock *wlock =
1623 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1624 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1625 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1626 lock->l_flags &= ~LDLM_FL_CP_REQD;
1627 list_del_init(&wlock->l_cp_ast);
1628 LDLM_LOCK_PUT(wlock);
1631 /* The lock met with no resistance; we're finished. */
1632 if (rc == LDLM_ITER_CONTINUE) {
1633 /* do not grant locks to the liblustre clients: they cannot
1634 * handle ASTs robustly. We need to do this while still
1635 * holding ns_lock to avoid the lock remaining on the res_link
1636 * list (and potentially being added to l_pending_list by an
1637 * AST) when we are going to drop this lock ASAP. */
1638 if (lock->l_export->exp_libclient ||
1639 OBD_FAIL_CHECK(OBD_FAIL_LDLM_GLIMPSE)) {
1640 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1641 ldlm_resource_unlink_lock(lock);
1642 err = ELDLM_LOCK_ABORTED;
1644 err = ELDLM_LOCK_REPLACED;
1650 /* Do not grant any lock, but instead send GL callbacks. The extent
1651 * policy nicely created a list of all PW locks for us. We will choose
1652 * the highest of those which are larger than the size in the LVB, if
1653 * any, and perform a glimpse callback. */
1654 res_lvb = res->lr_lvb_data;
1655 LASSERT(res_lvb != NULL);
1656 *reply_lvb = *res_lvb;
1658 list_for_each(tmp, &res->lr_granted) {
1659 struct ldlm_lock *tmplock =
1660 list_entry(tmp, struct ldlm_lock, l_res_link);
1662 if (tmplock->l_granted_mode == LCK_PR)
1665 * ->ns_lock guarantees that no new locks are granted, and,
1666 * therefore, that res->lr_lvb_data cannot increase beyond the
1667 * end of already granted lock. As a result, it is safe to
1668 * check against "stale" reply_lvb->lvb_size value without
1671 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1674 /* Don't send glimpse ASTs to liblustre clients. They aren't
1675 * listening for them, and they do entirely synchronous I/O
1677 if (tmplock->l_export == NULL ||
1678 tmplock->l_export->exp_libclient == 1) {
1684 l = LDLM_LOCK_GET(tmplock);
1688 if (l->l_policy_data.l_extent.start >
1689 tmplock->l_policy_data.l_extent.start)
1693 l = LDLM_LOCK_GET(tmplock);
1697 /* There were no PW locks beyond the size in the LVB; finished. */
1699 if (only_liblustre) {
1700 /* If we discovered a liblustre client with a PW lock,
1701 * however, the LVB may be out of date! The LVB is
1702 * updated only on glimpse (which we don't do for
1703 * liblustre clients) and cancel (which the client
1704 * obviously has not yet done). So if it has written
1705 * data but kept the lock, the LVB is stale and needs
1706 * to be updated from disk.
1708 * Of course, this will all disappear when we switch to
1709 * taking liblustre locks on the OST. */
1710 ldlm_res_lvbo_update(res, NULL, 0, 1);
1712 RETURN(ELDLM_LOCK_ABORTED);
1716 * This check is for lock taken in filter_prepare_destroy() that does
1717 * not have l_glimpse_ast set. So the logic is: if there is a lock
1718 * with no l_glimpse_ast set, this object is being destroyed already.
1720 * Hence, if you are grabbing DLM locks on the server, always set
1721 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1723 if (l->l_glimpse_ast == NULL) {
1724 /* We are racing with unlink(); just return -ENOENT */
1725 rep->lock_policy_res1 = -ENOENT;
1729 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1730 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1731 /* Update the LVB from disk if the AST failed (this is a legal race) */
1733 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1734 * sending ast is not handled. This can result in lost client writes.
1737 ldlm_res_lvbo_update(res, NULL, 0, 1);
1740 *reply_lvb = *res_lvb;
1746 RETURN(ELDLM_LOCK_ABORTED);
1750 * per-obd_device iobuf pool.
1752 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1753 * allocations in write-path has to be minimized (see bug 5137).
1755 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1756 * OST threads (see ost_thread_{init,done}()).
1758 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1759 * at the OST layer there are only (potentially) multiple obd_device of type
1760 * unknown at the time of OST thread creation.
1762 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1763 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1764 * it's very own iobuf.
1768 * filter_kiobuf_pool_init()
1770 * filter_kiobuf_pool_done()
1772 * filter_iobuf_get()
1774 * operate on this array. They are "generic" in a sense that they don't depend
1775 * on actual type of iobuf's (the latter depending on Linux kernel version).
1779 * destroy pool created by filter_iobuf_pool_init
1781 static void filter_iobuf_pool_done(struct filter_obd *filter)
1783 struct filter_iobuf **pool;
1788 pool = filter->fo_iobuf_pool;
1790 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1791 if (pool[i] != NULL)
1792 filter_free_iobuf(pool[i]);
1794 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1795 filter->fo_iobuf_pool = NULL;
1801 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1803 static int filter_iobuf_pool_init(struct filter_obd *filter)
1810 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1812 if (filter->fo_iobuf_pool == NULL)
1815 filter->fo_iobuf_count = OSS_THREADS_MAX;
1820 /* Return iobuf allocated for @thread_id. We don't know in advance how
1821 * many threads there will be so we allocate a large empty array and only
1822 * fill in those slots that are actually in use.
1823 * If we haven't allocated a pool entry for this thread before, do so now. */
1824 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1826 int thread_id = oti ? oti->oti_thread_id : -1;
1827 struct filter_iobuf *pool = NULL;
1828 struct filter_iobuf **pool_place = NULL;
1830 if (thread_id >= 0) {
1831 LASSERT(thread_id < filter->fo_iobuf_count);
1832 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1835 if (unlikely(pool == NULL)) {
1836 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1837 PTLRPC_MAX_BRW_PAGES);
1838 if (pool_place != NULL)
1845 /* mount the file system (secretly). lustre_cfg parameters are:
1848 * 3 = flags: failover=f, failout=n
1851 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1854 struct filter_obd *filter = &obd->u.filter;
1855 struct vfsmount *mnt;
1856 struct lustre_mount_info *lmi;
1857 struct obd_uuid uuid;
1864 if (lcfg->lcfg_bufcount < 3 ||
1865 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1866 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1869 lmi = server_get_mount(obd->obd_name);
1871 /* We already mounted in lustre_fill_super.
1872 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1873 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1875 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1877 /* old path - used by lctl */
1878 CERROR("Using old MDS mount method\n");
1879 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1880 MS_NOATIME|MS_NODIRATIME,
1881 lustre_cfg_string(lcfg, 1), option);
1884 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1885 lustre_cfg_string(lcfg, 1), rc);
1889 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1891 if (IS_ERR(obd->obd_fsops))
1892 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1894 rc = filter_iobuf_pool_init(filter);
1898 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1900 /* failover is the default */
1901 obd->obd_replayable = 1;
1903 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1904 str = lustre_cfg_string(lcfg, 3);
1905 if (strchr(str, 'n')) {
1906 CWARN("%s: recovery disabled\n", obd->obd_name);
1907 obd->obd_replayable = 0;
1911 filter->fo_vfsmnt = mnt;
1912 obd->u.obt.obt_sb = mnt->mnt_sb;
1913 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1914 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1916 fsfilt_setup(obd, obd->u.obt.obt_sb);
1918 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1919 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1920 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1921 obd->obd_lvfs_ctxt.fs = get_ds();
1922 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1924 sema_init(&filter->fo_init_lock, 1);
1925 filter->fo_committed_group = 0;
1927 rc = filter_prep(obd);
1931 filter->fo_destroys_in_progress = 0;
1932 for (i = 0; i < 32; i++)
1933 sema_init(&filter->fo_create_locks[i], 1);
1935 spin_lock_init(&filter->fo_translock);
1936 spin_lock_init(&filter->fo_objidlock);
1937 INIT_LIST_HEAD(&filter->fo_export_list);
1938 sema_init(&filter->fo_alloc_lock, 1);
1939 init_brw_stats(&filter->fo_filter_stats);
1940 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1941 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1942 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1944 INIT_LIST_HEAD(&filter->fo_llog_list);
1945 spin_lock_init(&filter->fo_llog_list_lock);
1947 filter->fo_fl_oss_capa = 0;
1948 INIT_LIST_HEAD(&filter->fo_capa_keys);
1949 filter->fo_capa_hash = init_capa_hash();
1950 if (filter->fo_capa_hash == NULL)
1951 GOTO(err_ops, rc = -ENOMEM);
1953 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1954 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1955 LDLM_NAMESPACE_GREEDY);
1956 if (obd->obd_namespace == NULL)
1957 GOTO(err_post, rc = -ENOMEM);
1958 obd->obd_namespace->ns_lvbp = obd;
1959 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1960 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1962 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1963 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1965 rc = llog_cat_initialize(obd, NULL, 1, NULL);
1967 CERROR("failed to setup llogging subsystems\n");
1971 rc = lquota_setup(filter_quota_interface_ref, obd);
1975 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1976 if (uuid_ptr != NULL) {
1977 class_uuid_unparse(uuid_ptr, &uuid);
1983 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1985 if (obd->obd_recovering) {
1986 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1987 "recovery until %d %s reconnect, or if no clients"
1988 " reconnect for %d:%.02d; during that time new "
1989 "clients will not be allowed to connect. "
1990 "Recovery progress can be monitored by watching "
1991 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1992 obd->obd_name, lustre_cfg_string(lcfg, 1),
1993 label ?: "", label ? "/" : "", str,
1994 obd->obd_max_recoverable_clients,
1995 (obd->obd_max_recoverable_clients == 1)
1996 ? "client" : "clients",
1997 (int)(OBD_RECOVERY_TIMEOUT) / 60,
1998 (int)(OBD_RECOVERY_TIMEOUT) % 60,
2001 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2002 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2003 label ?: "", label ? "/" : "", str,
2004 obd->obd_replayable ? "enabled" : "disabled");
2012 fsfilt_put_ops(obd->obd_fsops);
2013 filter_iobuf_pool_done(filter);
2015 server_put_mount(obd->obd_name, mnt);
2016 obd->u.obt.obt_sb = 0;
2020 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2022 struct lprocfs_static_vars lvars;
2026 CLASSERT(offsetof(struct obd_device, u.obt) ==
2027 offsetof(struct obd_device, u.filter.fo_obt));
2029 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2032 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2033 page = get_zeroed_page(GFP_KERNEL);
2037 /* lprocfs must be setup before the filter so state can be safely added
2038 * to /proc incrementally as the filter is setup */
2039 lprocfs_init_vars(filter, &lvars);
2040 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2041 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2042 /* Init obdfilter private stats here */
2043 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2044 LPROCFS_CNTR_AVGMINMAX,
2045 "read_bytes", "bytes");
2046 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2047 LPROCFS_CNTR_AVGMINMAX,
2048 "write_bytes", "bytes");
2050 lproc_filter_attach_seqstat(obd);
2051 obd->obd_proc_exports = proc_mkdir("exports",
2052 obd->obd_proc_entry);
2055 memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
2056 LUSTRE_CFG_BUFLEN(lcfg, 4));
2057 rc = filter_common_setup(obd, lcfg, (void *)page);
2061 lprocfs_obd_cleanup(obd);
2062 lprocfs_free_obd_stats(obd);
2068 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2069 static struct llog_operations filter_size_orig_logops = {
2070 lop_setup: llog_obd_origin_setup,
2071 lop_cleanup: llog_obd_origin_cleanup,
2072 lop_add: llog_obd_origin_add
2075 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2076 struct obd_device *tgt, int count,
2077 struct llog_catid *catid,
2078 struct obd_uuid *uuid)
2080 struct llog_ctxt *ctxt;
2084 filter_mds_ost_repl_logops = llog_client_ops;
2085 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2086 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2087 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2089 rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2090 &filter_mds_ost_repl_logops);
2094 /* FIXME - assign unlink_cb for filter's recovery */
2096 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2098 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2100 LASSERT(ctxt != NULL);
2101 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2103 rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2104 &filter_size_orig_logops);
2108 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2113 if (CTXTP(ctxt, cleanup))
2114 rc = CTXTP(ctxt, cleanup)(ctxt);
2117 class_export_put(ctxt->loc_exp);
2118 OBD_FREE(ctxt, sizeof(*ctxt));
2123 static int filter_group_llog_finish(struct obd_llogs *llogs)
2125 struct llog_ctxt *ctxt;
2126 int rc = 0, rc2 = 0;
2129 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2131 rc = filter_group_llog_cleanup(ctxt);
2133 ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2135 rc2 = filter_group_llog_cleanup(ctxt);
2142 static int filter_llog_finish(struct obd_device *obd, int count)
2144 struct llog_ctxt *ctxt;
2145 int rc = 0, rc2 = 0;
2148 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2150 rc = llog_cleanup(ctxt);
2152 ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2154 rc2 = llog_cleanup(ctxt);
2161 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2162 struct obd_export *export)
2164 struct filter_group_llog *fglog, *nlog;
2165 struct filter_obd *filter;
2166 struct llog_ctxt *ctxt;
2167 struct list_head *cur;
2170 filter = &obd->u.filter;
2172 spin_lock(&filter->fo_llog_list_lock);
2173 list_for_each(cur, &filter->fo_llog_list) {
2174 fglog = list_entry(cur, struct filter_group_llog, list);
2175 if (fglog->group == group) {
2176 if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2177 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2178 obd->obd_name, group, fglog->exp, export);
2179 spin_unlock(&filter->fo_llog_list_lock);
2183 spin_unlock(&filter->fo_llog_list_lock);
2188 OBD_ALLOC_PTR(fglog);
2191 fglog->group = group;
2193 OBD_ALLOC_PTR(fglog->llogs);
2194 if (fglog->llogs == NULL) {
2195 OBD_FREE_PTR(fglog);
2199 spin_lock(&filter->fo_llog_list_lock);
2200 list_for_each(cur, &filter->fo_llog_list) {
2201 nlog = list_entry(cur, struct filter_group_llog, list);
2202 LASSERT(nlog->group != group);
2204 list_add(&fglog->list, &filter->fo_llog_list);
2205 spin_unlock(&filter->fo_llog_list_lock);
2207 rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2209 OBD_FREE_PTR(fglog->llogs);
2210 OBD_FREE_PTR(fglog);
2216 fglog->exp = export;
2217 ctxt = llog_get_context_from_llogs(fglog->llogs,
2218 LLOG_MDS_OST_REPL_CTXT);
2219 LASSERT(ctxt != NULL);
2221 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2223 CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2224 obd->obd_name, fglog->llogs, group);
2226 RETURN(fglog->llogs);
2229 static int filter_llog_connect(struct obd_export *exp,
2230 struct llogd_conn_body *body)
2232 struct obd_device *obd = exp->exp_obd;
2233 struct llog_ctxt *ctxt;
2234 struct obd_llogs *llog;
2238 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2239 (unsigned) body->lgdc_logid.lgl_ogr,
2240 (unsigned) body->lgdc_logid.lgl_oid,
2241 (unsigned) body->lgdc_logid.lgl_ogen);
2243 llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2244 LASSERT(llog != NULL);
2245 ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2246 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2247 body->lgdc_ctxt_idx);
2248 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2249 &body->lgdc_gen, NULL);
2251 CERROR("failed to connect rc %d idx %d\n", rc,
2252 body->lgdc_ctxt_idx);
2257 static int filter_llog_preclean (struct obd_device *obd)
2259 struct filter_group_llog *log;
2260 struct filter_obd *filter;
2264 filter = &obd->u.filter;
2265 spin_lock(&filter->fo_llog_list_lock);
2266 while (!list_empty(&filter->fo_llog_list)) {
2267 log = list_entry(filter->fo_llog_list.next,
2268 struct filter_group_llog, list);
2269 list_del(&log->list);
2270 spin_unlock(&filter->fo_llog_list_lock);
2272 rc = filter_group_llog_finish(log->llogs);
2274 CERROR("failed to cleanup llogging subsystem for %u\n",
2276 OBD_FREE_PTR(log->llogs);
2278 spin_lock(&filter->fo_llog_list_lock);
2280 spin_unlock(&filter->fo_llog_list_lock);
2282 rc = obd_llog_finish(obd, 0);
2284 CERROR("failed to cleanup llogging subsystem\n");
2289 static int filter_precleanup(struct obd_device *obd,
2290 enum obd_cleanup_stage stage)
2296 case OBD_CLEANUP_EARLY:
2298 case OBD_CLEANUP_EXPORTS:
2299 target_cleanup_recovery(obd);
2301 case OBD_CLEANUP_SELF_EXP:
2302 rc = filter_llog_preclean(obd);
2304 case OBD_CLEANUP_OBD:
2310 static int filter_cleanup(struct obd_device *obd)
2312 struct filter_obd *filter = &obd->u.filter;
2316 LCONSOLE_WARN("%s: shutting down for failover; client state "
2317 "will be preserved.\n", obd->obd_name);
2319 if (!list_empty(&obd->obd_exports)) {
2320 CERROR("%s: still has clients!\n", obd->obd_name);
2321 class_disconnect_exports(obd);
2322 if (!list_empty(&obd->obd_exports)) {
2323 CERROR("still has exports after forced cleanup?\n");
2328 lprocfs_obd_cleanup(obd);
2329 lprocfs_free_obd_stats(obd);
2330 lquota_cleanup(filter_quota_interface_ref, obd);
2332 /* Stop recovery before namespace cleanup. */
2333 target_stop_recovery_thread(obd);
2334 target_cleanup_recovery(obd);
2336 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2338 if (obd->u.obt.obt_sb == NULL)
2343 shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2345 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2347 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2348 obd->u.obt.obt_sb = NULL;
2350 fsfilt_put_ops(obd->obd_fsops);
2352 filter_iobuf_pool_done(filter);
2354 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2359 static int filter_connect_internal(struct obd_export *exp,
2360 struct obd_connect_data *data)
2365 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2366 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2367 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2368 data->ocd_connect_flags, data->ocd_version,
2369 data->ocd_grant, data->ocd_index);
2371 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2372 exp->exp_connect_flags = data->ocd_connect_flags;
2373 data->ocd_version = LUSTRE_VERSION_CODE;
2375 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2376 struct filter_export_data *fed = &exp->exp_filter_data;
2377 obd_size left, want;
2379 spin_lock(&exp->exp_obd->obd_osfs_lock);
2380 left = filter_grant_space_left(exp);
2381 want = data->ocd_grant;
2382 filter_grant(exp, fed->fed_grant, want, left);
2383 data->ocd_grant = fed->fed_grant;
2384 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2386 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2387 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2388 exp->exp_client_uuid.uuid, exp,
2389 data->ocd_grant, want, left);
2392 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2393 struct filter_obd *filter = &exp->exp_obd->u.filter;
2394 struct lr_server_data *lsd = filter->fo_fsd;
2395 int index = le32_to_cpu(lsd->lsd_ost_index);
2397 if (!(lsd->lsd_feature_compat &
2398 cpu_to_le32(OBD_COMPAT_OST))) {
2399 /* this will only happen on the first connect */
2400 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2401 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2402 filter_update_server_data(exp->exp_obd,
2403 filter->fo_rcvd_filp, lsd, 1);
2404 } else if (index != data->ocd_index) {
2405 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2406 " %u doesn't match actual OST index"
2407 " %u in last_rcvd file, bad "
2409 obd_export_nid2str(exp), index,
2415 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2416 data->ocd_brw_size = 65536;
2417 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2418 data->ocd_brw_size = min(data->ocd_brw_size,
2419 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2420 LASSERT(data->ocd_brw_size);
2423 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2424 * FIXME: We don't strictly need the COMPAT flag for that,
2425 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2426 * FIXME: We needed it for the index, as index 0 is valid. */
2431 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2432 struct obd_uuid *cluuid,
2433 struct obd_connect_data *data)
2438 if (exp == NULL || obd == NULL || cluuid == NULL)
2441 rc = filter_connect_internal(exp, data);
2446 /* nearly identical to mds_connect */
2447 static int filter_connect(const struct lu_env *env,
2448 struct lustre_handle *conn, struct obd_device *obd,
2449 struct obd_uuid *cluuid,
2450 struct obd_connect_data *data)
2452 struct lvfs_run_ctxt saved;
2453 struct obd_export *exp;
2454 struct filter_export_data *fed;
2455 struct filter_client_data *fcd = NULL;
2460 if (conn == NULL || obd == NULL || cluuid == NULL)
2463 rc = class_connect(conn, obd, cluuid);
2466 exp = class_conn2export(conn);
2467 LASSERT(exp != NULL);
2469 fed = &exp->exp_filter_data;
2471 rc = filter_connect_internal(exp, data);
2475 filter_export_stats_init(obd, exp);
2476 group = data->ocd_group;
2477 if (obd->obd_replayable) {
2478 OBD_ALLOC(fcd, sizeof(*fcd));
2480 CERROR("filter: out of memory for client data\n");
2481 GOTO(cleanup, rc = -ENOMEM);
2484 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2486 fed->fed_fcd->fcd_group = group;
2487 rc = filter_client_add(obd, exp, -1);
2491 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2492 obd->obd_name, exp->exp_handle.h_cookie, group);
2496 if (fed->fed_group != 0 && fed->fed_group != group) {
2497 CERROR("!!! This export (nid %s) used object group %d "
2498 "earlier; now it's trying to use group %d! This could "
2499 "be a bug in the MDS. Tell CFS.\n",
2500 obd_export_nid2str(exp), fed->fed_group, group);
2501 GOTO(cleanup, rc = -EPROTO);
2503 fed->fed_group = group;
2505 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2506 rc = filter_read_groups(obd, group, 1);
2507 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2509 CERROR("can't read group %u\n", group);
2518 OBD_FREE(fcd, sizeof(*fcd));
2519 fed->fed_fcd = NULL;
2521 class_disconnect(exp);
2523 class_export_put(exp);
2529 /* Do extra sanity checks for grant accounting. We do this at connect,
2530 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2531 * always get rid of it or turn it off when we know accounting is good. */
2532 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2534 struct filter_export_data *fed;
2535 struct obd_export *exp;
2536 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2537 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2538 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2540 if (list_empty(&obd->obd_exports))
2543 /* We don't want to do this for large machines that do lots of
2544 mounts or unmounts. It burns... */
2545 if (obd->obd_num_exports > 100)
2548 spin_lock(&obd->obd_osfs_lock);
2549 spin_lock(&obd->obd_dev_lock);
2550 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2552 fed = &exp->exp_filter_data;
2553 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2556 if (maxsize > 0) { /* we may not have done a statfs yet */
2557 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2558 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2559 exp->exp_client_uuid.uuid, exp,
2560 fed->fed_grant, fed->fed_pending, maxsize);
2561 LASSERTF(fed->fed_dirty <= maxsize,
2562 "%s: cli %s/%p %ld > "LPU64"\n", func,
2563 exp->exp_client_uuid.uuid, exp,
2564 fed->fed_dirty, maxsize);
2567 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2568 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2569 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2571 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2572 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2573 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2574 tot_granted += fed->fed_grant + fed->fed_pending;
2575 tot_pending += fed->fed_pending;
2576 tot_dirty += fed->fed_dirty;
2578 fo_tot_granted = obd->u.filter.fo_tot_granted;
2579 fo_tot_pending = obd->u.filter.fo_tot_pending;
2580 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2581 spin_unlock(&obd->obd_dev_lock);
2582 spin_unlock(&obd->obd_osfs_lock);
2584 /* Do these assertions outside the spinlocks so we don't kill system */
2585 if (tot_granted != fo_tot_granted)
2586 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2587 func, tot_granted, fo_tot_granted);
2588 if (tot_pending != fo_tot_pending)
2589 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2590 func, tot_pending, fo_tot_pending);
2591 if (tot_dirty != fo_tot_dirty)
2592 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2593 func, tot_dirty, fo_tot_dirty);
2594 if (tot_pending > tot_granted)
2595 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2596 func, tot_pending, tot_granted);
2597 if (tot_granted > maxsize)
2598 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2599 func, tot_granted, maxsize);
2600 if (tot_dirty > maxsize)
2601 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2602 func, tot_dirty, maxsize);
2605 /* Remove this client from the grant accounting totals. We also remove
2606 * the export from the obd device under the osfs and dev locks to ensure
2607 * that the filter_grant_sanity_check() calculations are always valid.
2608 * The client should do something similar when it invalidates its import. */
2609 static void filter_grant_discard(struct obd_export *exp)
2611 struct obd_device *obd = exp->exp_obd;
2612 struct filter_obd *filter = &obd->u.filter;
2613 struct filter_export_data *fed = &exp->exp_filter_data;
2615 spin_lock(&obd->obd_osfs_lock);
2616 spin_lock(&obd->obd_dev_lock);
2617 list_del_init(&exp->exp_obd_chain);
2618 spin_unlock(&obd->obd_dev_lock);
2620 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2621 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2622 obd->obd_name, filter->fo_tot_granted,
2623 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2624 filter->fo_tot_granted -= fed->fed_grant;
2625 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2626 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2627 obd->obd_name, filter->fo_tot_pending,
2628 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2629 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2630 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2631 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2632 obd->obd_name, filter->fo_tot_dirty,
2633 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2634 filter->fo_tot_dirty -= fed->fed_dirty;
2638 spin_unlock(&obd->obd_osfs_lock);
2641 static int filter_destroy_export(struct obd_export *exp)
2645 if (exp->exp_filter_data.fed_pending)
2646 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2647 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2648 exp, exp->exp_filter_data.fed_pending);
2650 target_destroy_export(exp);
2652 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2655 lprocfs_exp_cleanup(exp);
2657 if (exp->exp_obd->obd_replayable)
2658 filter_client_free(exp);
2660 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2662 filter_grant_discard(exp);
2663 filter_fmd_cleanup(exp);
2665 if (!(exp->exp_flags & OBD_OPT_FORCE))
2666 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2671 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2673 struct filter_group_llog *fglog, *nlog;
2674 struct filter_obd *filter;
2675 int worked = 0, group;
2676 struct llog_ctxt *ctxt;
2679 filter = &obd->u.filter;
2681 /* we can't sync log holding spinlock. also, we do not want to get
2682 * into livelock. so we do following: loop over MDS's exports in
2683 * group order and skip already synced llogs -bzzz */
2685 /* look for group with min. number, but > worked */
2688 spin_lock(&filter->fo_llog_list_lock);
2689 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2690 if (nlog->group <= worked) {
2691 /* this group is already synced */
2694 if (group < nlog->group) {
2695 /* we have group with smaller number to sync */
2698 /* store current minimal group */
2700 group = nlog->group;
2702 spin_unlock(&filter->fo_llog_list_lock);
2707 worked = fglog->group;
2708 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2709 ctxt = llog_get_context_from_llogs(fglog->llogs,
2710 LLOG_MDS_OST_REPL_CTXT);
2711 LASSERT(ctxt != NULL);
2712 llog_sync(ctxt, fglog->exp);
2714 } while (fglog != NULL);
2717 /* also incredibly similar to mds_disconnect */
2718 static int filter_disconnect(struct obd_export *exp)
2720 struct obd_device *obd = exp->exp_obd;
2725 class_export_get(exp);
2727 if (!(exp->exp_flags & OBD_OPT_FORCE))
2728 filter_grant_sanity_check(obd, __FUNCTION__);
2729 filter_grant_discard(exp);
2731 /* Disconnect early so that clients can't keep using export */
2732 rc = class_disconnect(exp);
2733 if (exp->exp_obd->obd_namespace != NULL)
2734 ldlm_cancel_locks_for_export(exp);
2736 fsfilt_sync(obd, obd->u.obt.obt_sb);
2738 /* flush any remaining cancel messages out to the target */
2739 filter_sync_llogs(obd, exp);
2740 class_export_put(exp);
2744 /* reverse import is changed, sync all cancels */
2745 static void filter_revimp_update(struct obd_export *exp)
2750 class_export_get(exp);
2752 /* flush any remaining cancel messages out to the target */
2753 filter_sync_llogs(exp->exp_obd, exp);
2754 class_export_put(exp);
2758 static int filter_ping(struct obd_export *exp)
2760 filter_fmd_expire(exp);
2765 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2766 const char *what, int quiet)
2768 struct dentry *dchild = NULL;
2771 if (oa->o_valid & OBD_MD_FLGROUP)
2774 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2776 if (IS_ERR(dchild)) {
2777 CERROR("%s error looking up object: "LPU64"\n",
2782 if (dchild->d_inode == NULL) {
2784 CERROR("%s: %s on non-existent object: "LPU64"\n",
2785 obd->obd_name, what, oa->o_id);
2787 RETURN(ERR_PTR(-ENOENT));
2793 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2795 struct dentry *dentry = NULL;
2796 struct obd_device *obd;
2800 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2801 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2805 obd = class_exp2obd(exp);
2807 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2811 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2813 RETURN(PTR_ERR(dentry));
2815 /* Limit the valid bits in the return data to what we actually use */
2816 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2817 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2823 /* this should be enabled/disabled in condition to enabled/disabled large
2824 * inodes (fast EAs) in backing store FS. */
2825 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2826 void *handle, struct obdo *oa)
2828 struct obd_device *obd = exp->exp_obd;
2832 if (oa->o_valid & OBD_MD_FLFID) {
2833 struct filter_fid ff;
2835 if (!(oa->o_valid & OBD_MD_FLGROUP))
2837 /* packing fid and converting it to LE for storing into EA.
2838 * Here ->o_stripe_idx should be filled by LOV and rest of
2839 * fields - by client. */
2840 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2841 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2842 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2843 ff.ff_objid = cpu_to_le64(oa->o_id);
2844 ff.ff_group = cpu_to_le64(oa->o_gr);
2846 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2847 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2848 oa->o_generation, oa->o_id, oa->o_gr);
2850 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2852 CERROR("store fid in object failed! rc: %d\n", rc);
2854 CDEBUG(D_HA, "OSS object without fid info!\n");
2860 /* this is called from filter_truncate() until we have filter_punch() */
2861 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2862 struct obdo *oa, struct obd_trans_info *oti)
2864 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2865 struct llog_cookie *fcc = NULL;
2866 struct filter_obd *filter;
2867 int rc, err, locked = 0, sync = 0;
2868 unsigned int ia_valid;
2869 struct inode *inode;
2874 LASSERT(dentry != NULL);
2875 LASSERT(!IS_ERR(dentry));
2877 inode = dentry->d_inode;
2878 LASSERT(inode != NULL);
2880 filter = &exp->exp_obd->u.filter;
2881 iattr_from_obdo(&iattr, oa, oa->o_valid);
2882 ia_valid = iattr.ia_valid;
2884 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2885 OBD_ALLOC(fcc, sizeof(*fcc));
2887 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2890 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2892 LOCK_INODE_MUTEX(inode);
2896 /* If the inode still has SUID+SGID bits set (see filter_precreate())
2897 * then we will accept the UID+GID sent by the client during write for
2898 * initializing the ownership of this inode. We only allow this to
2899 * happen once so clear these bits in setattr. In 2.6 kernels it is
2900 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2901 * the flags that are actually being set. */
2902 if (ia_valid & (ATTR_UID | ATTR_GID)) {
2903 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2904 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2906 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2907 if (!(ia_valid & ATTR_MODE)) {
2908 iattr.ia_mode = inode->i_mode;
2909 iattr.ia_valid |= ATTR_MODE;
2911 iattr.ia_mode &= ~S_ISUID;
2913 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2914 if (!(iattr.ia_valid & ATTR_MODE)) {
2915 iattr.ia_mode = inode->i_mode;
2916 iattr.ia_valid |= ATTR_MODE;
2918 iattr.ia_mode &= ~S_ISGID;
2921 orig_ids[USRQUOTA] = inode->i_uid;
2922 orig_ids[GRPQUOTA] = inode->i_gid;
2923 handle = fsfilt_start_log(exp->exp_obd, inode,
2924 FSFILT_OP_SETATTR, oti, 1);
2926 GOTO(out_unlock, rc = PTR_ERR(handle));
2928 /* update inode EA only once when inode is suid bit marked. As
2929 * on 2.6.x UID and GID may be set separately, we check here
2930 * only one of them to avoid double setting. */
2931 if (inode->i_mode & S_ISUID)
2932 filter_update_fidea(exp, inode, handle, oa);
2934 handle = fsfilt_start(exp->exp_obd, inode,
2935 FSFILT_OP_SETATTR, oti);
2937 GOTO(out_unlock, rc = PTR_ERR(handle));
2939 if (oa->o_valid & OBD_MD_FLFLAGS) {
2940 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2941 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2943 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2945 /* set cancel cookie callback function */
2946 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2947 filter_cancel_cookies_cb,
2951 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2952 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2954 /* The truncate might have used up our transaction credits. Make
2955 * sure we have one left for the last_rcvd update. */
2956 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2958 rc = filter_finish_transno(exp, oti, rc, sync);
2960 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2964 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2966 CERROR("error on commit, err = %d\n", err);
2974 /* Let's flush truncated page on disk immediately, then we can
2975 * avoid need to search for page aliases before directio writes
2976 * and this sort of stuff at expense of somewhat slower
2977 * truncates not on a page boundary. I believe this is the only
2978 * place in filter code that can lead to pages getting to
2979 * pagecache so far. */
2980 filter_clear_truncated_page(inode);
2981 UNLOCK_INODE_MUTEX(inode);
2988 UNLOCK_INODE_MUTEX(inode);
2991 OBD_FREE(fcc, sizeof(*fcc));
2993 /* trigger quota release */
2994 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2995 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2996 int rc2 = lquota_adjust(filter_quota_interface_ref,
2997 exp->exp_obd, cur_ids,
2998 orig_ids, rc, FSFILT_OP_SETATTR);
2999 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3000 "filter adjust qunit. (rc:%d)\n", rc2);
3005 /* this is called from filter_truncate() until we have filter_punch() */
3006 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3007 struct obd_trans_info *oti)
3009 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3010 oinfo->oi_oa->o_gr, 0 } };
3011 struct filter_mod_data *fmd;
3012 struct lvfs_run_ctxt saved;
3013 struct filter_obd *filter;
3014 struct ldlm_resource *res;
3015 struct dentry *dentry;
3019 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3020 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3024 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3027 RETURN(PTR_ERR(dentry));
3029 filter = &exp->exp_obd->u.filter;
3030 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3033 if (oinfo->oi_oa->o_valid &
3034 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3035 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3036 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3037 fmd->fmd_mactime_xid = oti->oti_xid;
3038 filter_fmd_put(exp, fmd);
3041 /* setting objects attributes (including owner/group) */
3042 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3044 GOTO(out_unlock, rc);
3046 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3047 &res_id, LDLM_EXTENT, 0);
3050 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3051 ldlm_resource_putref(res);
3054 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3056 /* Quota release need uid/gid info */
3057 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3058 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3064 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3068 /* XXX identical to osc_unpackmd */
3069 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3070 struct lov_mds_md *lmm, int lmm_bytes)
3076 if (lmm_bytes < sizeof (*lmm)) {
3077 CERROR("lov_mds_md too small: %d, need %d\n",
3078 lmm_bytes, (int)sizeof(*lmm));
3081 /* XXX LOV_MAGIC etc check? */
3083 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3084 CERROR("lov_mds_md: zero lmm_object_id\n");
3089 lsm_size = lov_stripe_md_size(1);
3093 if (*lsmp != NULL && lmm == NULL) {
3094 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3095 OBD_FREE(*lsmp, lsm_size);
3100 if (*lsmp == NULL) {
3101 OBD_ALLOC(*lsmp, lsm_size);
3105 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3106 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3107 OBD_FREE(*lsmp, lsm_size);
3110 loi_init((*lsmp)->lsm_oinfo[0]);
3114 /* XXX zero *lsmp? */
3115 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3116 LASSERT((*lsmp)->lsm_object_id);
3119 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3124 /* caller must hold fo_create_locks[oa->o_gr] */
3125 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3126 struct filter_obd *filter)
3128 struct obdo doa; /* XXX obdo on stack */
3134 LASSERT(oa->o_gr != 0);
3135 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3136 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3138 memset(&doa, 0, sizeof(doa));
3140 doa.o_valid |= OBD_MD_FLGROUP;
3141 doa.o_gr = oa->o_gr;
3142 doa.o_mode = S_IFREG;
3144 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3145 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3146 exp->exp_obd->obd_name, doa.o_gr);
3150 last = filter_last_id(filter, doa.o_gr);
3152 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3153 exp->exp_obd->obd_name, oa->o_id + 1, last);
3155 for (id = last; id > oa->o_id; id--) {
3157 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3158 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3159 CEMERG("error destroying precreate objid "LPU64": %d\n",
3161 filter_set_last_id(filter, id - 1, doa.o_gr);
3162 /* update last_id on disk periodically so that if we restart
3163 * we don't need to re-scan all of the just-deleted objects. */
3164 if ((id & 511) == 0)
3165 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3168 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3169 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3171 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3172 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3177 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3178 obd_gr group, int *num);
3179 /* returns a negative error or a nonnegative number of files to create */
3180 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3181 obd_gr group, struct obd_trans_info *oti)
3183 struct obd_device *obd = exp->exp_obd;
3184 struct filter_obd *filter = &obd->u.filter;
3188 /* delete orphans request */
3189 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3190 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3191 CERROR("%s: dropping old orphan cleanup request\n",
3195 /* This causes inflight precreates to abort and drop lock */
3196 set_bit(group, &filter->fo_destroys_in_progress);
3197 down(&filter->fo_create_locks[group]);
3198 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3199 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3200 exp->exp_obd->obd_name, group);
3201 up(&filter->fo_create_locks[group]);
3204 diff = oa->o_id - filter_last_id(filter, group);
3205 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3206 filter_last_id(filter, group), diff);
3208 if (-diff > OST_MAX_PRECREATE) {
3209 CERROR("%s: ignoring bogus orphan destroy request: "
3210 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3211 oa->o_id, filter_last_id(filter, group));
3212 /* FIXME: should reset precreate_next_id on MDS */
3213 GOTO(out, rc = -EINVAL);
3216 rc = filter_destroy_precreated(exp, oa, filter);
3218 CERROR("%s: unable to write lastobjid, but "
3219 "orphans were deleted\n", obd->obd_name);
3222 /* XXX: Used by MDS for the first time! */
3223 clear_bit(group, &filter->fo_destroys_in_progress);
3226 down(&filter->fo_create_locks[group]);
3227 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3228 CERROR("%s: dropping old precreate request\n",
3232 /* only precreate if group == 0 and o_id is specfied */
3233 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3236 diff = oa->o_id - filter_last_id(filter, group);
3237 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3238 filter_last_id(filter, group), diff);
3240 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3241 oa->o_id, filter_last_id(filter, group), diff);
3245 oa->o_id = filter_last_id(&obd->u.filter, group);
3246 rc = filter_precreate(obd, oa, group, &diff);
3247 oa->o_id = filter_last_id(&obd->u.filter, group);
3249 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3252 /* else diff == 0 */
3255 up(&filter->fo_create_locks[group]);
3259 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3262 struct filter_obd *filter = &obd->u.filter;
3263 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3267 /* at least try to account for cached pages. its still racey and
3268 * might be under-reporting if clients haven't announced their
3269 * caches with brw recently */
3270 spin_lock(&obd->obd_osfs_lock);
3271 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3272 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3273 spin_unlock(&obd->obd_osfs_lock);
3275 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3276 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3277 filter->fo_tot_dirty, filter->fo_tot_granted,
3278 filter->fo_tot_pending,
3279 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3281 filter_grant_sanity_check(obd, __FUNCTION__);
3283 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3284 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3285 osfs->os_bsize - 1) >> blockbits));
3287 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3288 * stop creating files on MDS if OST is not good shape to create
3290 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3295 /* We rely on the fact that only one thread will be creating files in a given
3296 * group at a time, which is why we don't need an atomic filter_get_new_id.
3297 * Even if we had that atomic function, the following race would exist:
3299 * thread 1: gets id x from filter_next_id
3300 * thread 2: gets id (x + 1) from filter_next_id
3301 * thread 2: creates object (x + 1)
3302 * thread 1: tries to create object x, gets -ENOSPC
3304 * Caller must hold fo_create_locks[group]
3306 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3307 obd_gr group, int *num)
3309 struct dentry *dchild = NULL, *dparent = NULL;
3310 struct filter_obd *filter;
3311 struct obd_statfs *osfs;
3312 int err = 0, rc = 0, recreate_obj = 0, i;
3313 unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3315 void *handle = NULL;
3318 filter = &obd->u.filter;
3320 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3322 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3324 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3325 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3328 OBD_ALLOC(osfs, sizeof(*osfs));
3331 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3332 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3333 CDEBUG(D_HA,"%s: not enough space for create "LPU64"\n",
3334 obd->obd_name, osfs->os_bavail <<
3335 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3339 OBD_FREE(osfs, sizeof(*osfs));
3344 CDEBUG(D_HA, "%s: precreating %d objects in group "LPU64" at "LPU64"\n",
3345 obd->obd_name, *num, group, oa->o_id);
3347 for (i = 0; i < *num && err == 0; i++) {
3348 int cleanup_phase = 0;
3350 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3351 CWARN("%s: create aborted by destroy\n",
3360 last_id = filter_last_id(filter, group);
3361 if (next_id > last_id) {
3362 CERROR("Error: Trying to recreate obj greater"
3363 "than last id "LPD64" > "LPD64"\n",
3365 GOTO(cleanup, rc = -EINVAL);
3368 next_id = filter_last_id(filter, group) + 1;
3370 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3372 dparent = filter_parent_lock(obd, group, next_id);
3373 if (IS_ERR(dparent))
3374 GOTO(cleanup, rc = PTR_ERR(dparent));
3375 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3377 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3379 GOTO(cleanup, rc = PTR_ERR(dchild));
3380 cleanup_phase = 2; /* f_dput(dchild) */
3382 if (dchild->d_inode != NULL) {
3383 /* This would only happen if lastobjid was bad on disk*/
3384 /* Could also happen if recreating missing obj but
3388 CERROR("%s: recreating existing object %.*s?\n",
3389 obd->obd_name, dchild->d_name.len,
3390 dchild->d_name.name);
3392 CERROR("%s: Serious error: objid %.*s already "
3393 "exists; is this filesystem corrupt?\n",
3394 obd->obd_name, dchild->d_name.len,
3395 dchild->d_name.name);
3398 GOTO(cleanup, rc = -EEXIST);
3401 handle = fsfilt_start_log(obd, dparent->d_inode,
3402 FSFILT_OP_CREATE, NULL, 1);
3404 GOTO(cleanup, rc = PTR_ERR(handle));
3407 /* We mark object SUID+SGID to flag it for accepting UID+GID
3408 * from client on first write. Currently the permission bits
3409 * on the OST are never used, so this is OK. */
3410 rc = ll_vfs_create(dparent->d_inode, dchild,
3411 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3413 CERROR("create failed rc = %d\n", rc);
3417 if (!recreate_obj) {
3418 filter_set_last_id(filter, next_id, group);
3419 err = filter_update_last_objid(obd, group, 0);
3421 CERROR("unable to write lastobjid "
3422 "but file created\n");
3426 switch(cleanup_phase) {
3428 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3430 CERROR("error on commit, err = %d\n", err);
3437 filter_parent_unlock(dparent);
3444 if (time_after(jiffies, enough_time)) {
3445 CDEBUG(D_HA, "%s: precreate slow - want %d got %d \n",
3446 obd->obd_name, *num, i);
3452 CDEBUG(D_HA, "%s: created %d objects for group "LPU64": "LPU64"\n",
3453 obd->obd_name, i, group, filter->fo_last_objids[group]);
3458 static int filter_create(struct obd_export *exp, struct obdo *oa,
3459 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3461 struct filter_export_data *fed;
3462 struct obd_device *obd = NULL;
3463 struct filter_obd *filter;
3464 struct lvfs_run_ctxt saved;
3465 struct lov_stripe_md *lsm = NULL;
3466 int rc = 0, diff, group = oa->o_gr;
3469 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3470 CERROR("!!! nid %s sent invalid object group %d\n",
3471 obd_export_nid2str(exp), group);
3476 fed = &exp->exp_filter_data;
3477 filter = &obd->u.filter;
3479 if (fed->fed_group != group) {
3480 CERROR("!!! this export (nid %s) used object group %d "
3481 "earlier; now it's trying to use group %d! This could "
3482 "be a bug in the MDS. Tell CFS.\n",
3483 obd_export_nid2str(exp), fed->fed_group, group);
3487 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3488 oa->o_gr, oa->o_id);
3492 rc = obd_alloc_memmd(exp, &lsm);
3499 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3501 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3502 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3503 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3504 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3505 oa->o_id, filter_last_id(filter,
3510 down(&filter->fo_create_locks[oa->o_gr]);
3511 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3512 up(&filter->fo_create_locks[oa->o_gr]);
3515 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3518 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3519 if (rc && ea != NULL && *ea != lsm) {
3520 obd_free_memmd(exp, &lsm);
3521 } else if (rc == 0 && ea != NULL) {
3522 /* XXX LOV STACKING: the lsm that is passed to us from
3523 * LOV does not have valid lsm_oinfo data structs, so
3524 * don't go touching that. This needs to be fixed in a
3526 lsm->lsm_object_id = oa->o_id;
3533 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3534 struct lov_stripe_md *md, struct obd_trans_info *oti,
3535 struct obd_export *md_exp)
3537 unsigned int qcids[MAXQUOTAS] = {0, 0};
3538 struct obd_device *obd;
3539 struct filter_obd *filter;
3540 struct dentry *dchild = NULL, *dparent = NULL;
3541 struct lvfs_run_ctxt saved;
3542 void *handle = NULL;
3543 struct llog_cookie *fcc = NULL;
3544 int rc, rc2, cleanup_phase = 0, sync = 0;
3548 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3551 filter = &obd->u.filter;
3553 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3556 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3558 GOTO(cleanup, rc = PTR_ERR(dchild));
3561 if (dchild->d_inode == NULL) {
3562 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3564 /* If object already gone, cancel cookie right now */
3565 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3566 fcc = obdo_logcookie(oa);
3567 llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3569 fcc = NULL; /* we didn't allocate fcc, don't free it */
3571 GOTO(cleanup, rc = -ENOENT);
3574 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3576 /* Our MDC connection is established by the MDS to us */
3577 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3578 OBD_ALLOC(fcc, sizeof(*fcc));
3580 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3582 DQUOT_INIT(dchild->d_inode);
3584 /* we're gonna truncate it first in order to avoid possible deadlock:
3586 * open trasaction open transaction
3587 * down(i_zombie) down(i_zombie)
3588 * restart transaction
3589 * (see BUG 4180) -bzzz
3591 LOCK_INODE_MUTEX(dchild->d_inode);
3592 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3594 if (IS_ERR(handle)) {
3595 UNLOCK_INODE_MUTEX(dchild->d_inode);
3596 GOTO(cleanup, rc = PTR_ERR(handle));
3599 iattr.ia_valid = ATTR_SIZE;
3601 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3602 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3603 UNLOCK_INODE_MUTEX(dchild->d_inode);
3607 GOTO(cleanup, rc = rc2);
3609 /* We don't actually need to lock the parent until we are unlinking
3610 * here, and not while truncating above. That avoids holding the
3611 * parent lock for a long time during truncate, which can block other
3612 * threads from doing anything to objects in that directory. bug 7171 */
3613 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3614 if (IS_ERR(dparent))
3615 GOTO(cleanup, rc = PTR_ERR(dparent));
3616 cleanup_phase = 3; /* filter_parent_unlock */
3618 LOCK_INODE_MUTEX(dchild->d_inode);
3619 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3620 if (IS_ERR(handle)) {
3621 UNLOCK_INODE_MUTEX(dchild->d_inode);
3622 GOTO(cleanup, rc = PTR_ERR(handle));
3624 cleanup_phase = 4; /* fsfilt_commit */
3626 /* Quota release need uid/gid of inode */
3627 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3629 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3631 /* this drops dchild->d_inode->i_mutex unconditionally */
3632 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3636 switch(cleanup_phase) {
3639 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3640 oti->oti_handle : handle,
3641 filter_cancel_cookies_cb,
3643 /* If add_journal_cb failed, then filter_finish_transno
3644 * will commit the handle and we will do a sync
3645 * on commit. then we call callback directly to free
3648 rc = filter_finish_transno(exp, oti, rc, sync);
3650 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3653 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3655 CERROR("error on commit, err = %d\n", rc2);
3662 filter_parent_unlock(dparent);
3666 OBD_FREE(fcc, sizeof(*fcc));
3668 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3671 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3675 /* trigger quota release */
3676 qcids[USRQUOTA] = oa->o_uid;
3677 qcids[GRPQUOTA] = oa->o_gid;
3678 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3681 CDEBUG(rc ? D_ERROR : D_QUOTA,
3682 "filter adjust qunit! (rc:%d)\n", rc? rc : rc2);
3686 /* NB start and end are used for punch, but not truncate */
3687 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3688 struct obd_trans_info *oti,
3689 struct ptlrpc_request_set *rqset)
3694 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3695 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3696 oinfo->oi_policy.l_extent.end);
3700 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3701 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3702 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3704 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3705 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3709 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3710 rc = filter_setattr(exp, oinfo, oti);
3714 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3715 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3718 struct lvfs_run_ctxt saved;
3719 struct filter_obd *filter;
3720 struct dentry *dentry;
3721 struct llog_ctxt *ctxt;
3725 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3726 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3730 filter = &exp->exp_obd->u.filter;
3732 /* an objid of zero is taken to mean "sync whole filesystem" */
3733 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3734 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3735 /* flush any remaining cancel messages out to the target */
3736 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3737 llog_sync(ctxt, exp);
3741 dentry = filter_oa2dentry(exp->exp_obd, oa);
3743 RETURN(PTR_ERR(dentry));
3745 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3747 LOCK_INODE_MUTEX(dentry->d_inode);
3749 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3751 /* just any file to grab fsync method - "file" arg unused */
3752 struct file *file = filter->fo_rcvd_filp;
3754 if (file->f_op && file->f_op->fsync)
3755 rc = file->f_op->fsync(NULL, dentry, 1);
3757 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3761 UNLOCK_INODE_MUTEX(dentry->d_inode);
3763 oa->o_valid = OBD_MD_FLID;
3764 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3766 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3772 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3773 void *key, __u32 *vallen, void *val)
3775 struct obd_device *obd;
3778 obd = class_exp2obd(exp);
3780 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3784 if (keylen == strlen("blocksize") &&
3785 memcmp(key, "blocksize", keylen) == 0) {
3786 __u32 *blocksize = val;
3787 *vallen = sizeof(*blocksize);
3788 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3792 if (keylen == strlen("blocksize_bits") &&
3793 memcmp(key, "blocksize_bits", keylen) == 0) {
3794 __u32 *blocksize_bits = val;
3795 *vallen = sizeof(*blocksize_bits);
3796 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3800 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3801 obd_id *last_id = val;
3802 /* FIXME: object groups */
3803 *last_id = filter_last_id(&obd->u.filter, 0);
3806 CDEBUG(D_IOCTL, "invalid key\n");
3810 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3811 void *key, __u32 vallen, void *val,
3812 struct ptlrpc_request_set *set)
3814 struct obd_device *obd;
3815 struct obd_llogs *llog;
3816 struct llog_ctxt *ctxt;
3822 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3826 if (KEY_IS(KEY_CAPA_KEY)) {
3827 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3829 CERROR("filter update capability key failed: %d\n", rc);
3833 if (KEY_IS(KEY_REVIMP_UPD)) {
3834 filter_revimp_update(exp);
3838 if (keylen < strlen(KEY_MDS_CONN) ||
3839 memcmp(key, KEY_MDS_CONN, keylen) != 0)
3842 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3843 obd_export_nid2str(exp));
3844 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3846 /* setup llog imports */
3847 LASSERT(val != NULL);
3848 group = (int)(*(__u32 *)val);
3849 LASSERT(group >= FILTER_GROUP_MDS0);
3851 llog = filter_grab_llog_for_group(obd, group, exp);
3852 LASSERT(llog != NULL);
3853 ctxt = llog_get_context_from_llogs(llog, LLOG_MDS_OST_REPL_CTXT);
3854 LASSERTF(ctxt != NULL, "ctxt is not null\n"),
3856 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3858 lquota_setinfo(filter_quota_interface_ref, exp, obd);
3863 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3864 int len, void *karg, void *uarg)
3866 struct obd_device *obd = exp->exp_obd;
3867 struct obd_ioctl_data *data = karg;
3871 case OBD_IOC_ABORT_RECOVERY: {
3872 CERROR("aborting recovery for device %s\n", obd->obd_name);
3873 target_stop_recovery_thread(obd);
3877 case OBD_IOC_SYNC: {
3878 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3879 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3883 case OBD_IOC_SET_READONLY: {
3885 struct super_block *sb = obd->u.obt.obt_sb;
3886 struct inode *inode = sb->s_root->d_inode;
3887 BDEVNAME_DECLARE_STORAGE(tmp);
3888 CERROR("*** setting device %s read-only ***\n",
3889 ll_bdevname(sb, tmp));
3891 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3892 if (!IS_ERR(handle))
3893 rc = fsfilt_commit(obd, inode, handle, 1);
3895 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3896 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3898 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3902 case OBD_IOC_CATLOGLIST: {
3903 rc = llog_catalog_list(obd, 1, data);
3907 case OBD_IOC_LLOG_CANCEL:
3908 case OBD_IOC_LLOG_REMOVE:
3909 case OBD_IOC_LLOG_INFO:
3910 case OBD_IOC_LLOG_PRINT: {
3911 /* FIXME to be finished */
3912 RETURN(-EOPNOTSUPP);
3914 struct llog_ctxt *ctxt = NULL;
3916 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3917 rc = llog_ioctl(ctxt, cmd, data);
3918 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3931 static int filter_health_check(struct obd_device *obd)
3933 #ifdef USE_HEALTH_CHECK_WRITE
3934 struct filter_obd *filter = &obd->u.filter;
3939 * health_check to return 0 on healthy
3940 * and 1 on unhealthy.
3942 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3945 #ifdef USE_HEALTH_CHECK_WRITE
3946 LASSERT(filter->fo_health_check_filp != NULL);
3947 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3952 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3955 return filter_fid2dentry(data, NULL, gr, id);
3958 static int filter_process_config(struct obd_device *obd, obd_count len,
3961 struct lustre_cfg *lcfg = buf;
3962 struct lprocfs_static_vars lvars;
3965 lprocfs_init_vars(filter, &lvars);
3967 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
3971 static struct lvfs_callback_ops filter_lvfs_ops = {
3972 l_fid2dentry: filter_lvfs_fid2dentry,
3975 static struct obd_ops filter_obd_ops = {
3976 .o_owner = THIS_MODULE,
3977 .o_get_info = filter_get_info,
3978 .o_set_info_async = filter_set_info_async,
3979 .o_setup = filter_setup,
3980 .o_precleanup = filter_precleanup,
3981 .o_cleanup = filter_cleanup,
3982 .o_connect = filter_connect,
3983 .o_reconnect = filter_reconnect,
3984 .o_disconnect = filter_disconnect,
3985 .o_ping = filter_ping,
3986 .o_init_export = filter_init_export,
3987 .o_destroy_export = filter_destroy_export,
3988 .o_statfs = filter_statfs,
3989 .o_getattr = filter_getattr,
3990 .o_unpackmd = filter_unpackmd,
3991 .o_create = filter_create,
3992 .o_setattr = filter_setattr,
3993 .o_destroy = filter_destroy,
3994 .o_brw = filter_brw,
3995 .o_punch = filter_truncate,
3996 .o_sync = filter_sync,
3997 .o_preprw = filter_preprw,
3998 .o_commitrw = filter_commitrw,
3999 .o_llog_init = filter_llog_init,
4000 .o_llog_connect = filter_llog_connect,
4001 .o_llog_finish = filter_llog_finish,
4002 .o_iocontrol = filter_iocontrol,
4003 .o_health_check = filter_health_check,
4004 .o_process_config = filter_process_config,
4007 quota_interface_t *filter_quota_interface_ref;
4008 extern quota_interface_t filter_quota_interface;
4010 static int __init obdfilter_init(void)
4012 struct lprocfs_static_vars lvars;
4015 lprocfs_init_vars(filter, &lvars);
4017 request_module("lquota");
4018 OBD_ALLOC(obdfilter_created_scratchpad,
4019 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4020 sizeof(*obdfilter_created_scratchpad));
4021 if (obdfilter_created_scratchpad == NULL)
4024 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4025 sizeof(struct filter_mod_data),
4028 GOTO(out, rc = -ENOMEM);
4030 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4031 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4033 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4034 LUSTRE_OST_NAME, NULL);
4038 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4039 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4040 ll_fmd_cachep = NULL;
4042 if (filter_quota_interface_ref)
4043 PORTAL_SYMBOL_PUT(filter_quota_interface);
4045 OBD_FREE(obdfilter_created_scratchpad,
4046 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4047 sizeof(*obdfilter_created_scratchpad));
4053 static void __exit obdfilter_exit(void)
4055 if (filter_quota_interface_ref)
4056 PORTAL_SYMBOL_PUT(filter_quota_interface);
4058 if (ll_fmd_cachep) {
4059 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4060 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4061 ll_fmd_cachep = NULL;
4064 class_unregister_type(LUSTRE_OST_NAME);
4065 OBD_FREE(obdfilter_created_scratchpad,
4066 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4067 sizeof(*obdfilter_created_scratchpad));
4070 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4071 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4072 MODULE_LICENSE("GPL");
4074 module_init(obdfilter_init);
4075 module_exit(obdfilter_exit);