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 #include <linux/mount.h>
48 #include <linux/buffer_head.h>
50 #include <obd_class.h>
52 #include <lustre_dlm.h>
53 #include <lustre_fsfilt.h>
54 #include <lprocfs_status.h>
55 #include <lustre_log.h>
56 #include <lustre_commit_confd.h>
57 #include <libcfs/list.h>
58 #include <lustre_disk.h>
59 #include <lustre_quota.h>
60 #include <linux/slab.h>
61 #include <lustre_param.h>
63 #include "filter_internal.h"
65 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
66 #define FILTER_MIN_GROUPS 3
67 static struct lvfs_callback_ops filter_lvfs_ops;
68 cfs_mem_cache_t *ll_fmd_cachep;
70 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
71 void *cb_data, int error)
73 obd_transno_commit_cb(obd, transno, error);
76 /* Assumes caller has already pushed us into the kernel context. */
77 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
78 int rc, int force_sync)
80 struct filter_obd *filter = &exp->exp_obd->u.filter;
81 struct filter_export_data *fed = &exp->exp_filter_data;
82 struct filter_client_data *fcd = fed->fed_fcd;
85 int err, log_pri = D_HA;
87 /* Propagate error code. */
91 if (!exp->exp_obd->obd_replayable || oti == NULL)
94 /* we don't allocate new transnos for replayed requests */
95 if (oti->oti_transno == 0) {
96 spin_lock(&filter->fo_translock);
97 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
98 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
99 spin_unlock(&filter->fo_translock);
100 oti->oti_transno = last_rcvd;
102 spin_lock(&filter->fo_translock);
103 last_rcvd = oti->oti_transno;
104 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
105 filter->fo_fsd->lsd_last_transno =
106 cpu_to_le64(last_rcvd);
107 spin_unlock(&filter->fo_translock);
109 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
111 /* could get xid from oti, if it's ever needed */
112 fcd->fcd_last_xid = 0;
114 off = fed->fed_lr_off;
116 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
117 fed->fed_lr_idx, fed->fed_lr_off);
121 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
127 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
128 fcd, sizeof(*fcd), &off,
129 force_sync | exp->exp_need_sync);
131 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
139 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
140 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
145 void f_dput(struct dentry *dentry)
147 /* Can't go inside filter_ddelete because it can block */
148 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
149 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
150 LASSERT(atomic_read(&dentry->d_count) > 0);
155 static void init_brw_stats(struct brw_stats *brw_stats)
158 for (i = 0; i < BRW_LAST; i++)
159 spin_lock_init(&brw_stats->hist[i].oh_lock);
162 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
163 plus the procfs overhead :( */
164 static int filter_export_stats_init(struct obd_device *obd,
165 struct obd_export *exp)
167 struct filter_export_data *fed = &exp->exp_filter_data;
168 struct proc_dir_entry *brw_entry;
172 init_brw_stats(&fed->fed_brw_stats);
174 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
175 /* Self-export gets no proc entry */
178 rc = lprocfs_exp_setup(exp);
182 /* Create a per export proc entry for brw_stats */
183 brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
184 if (brw_entry == NULL)
186 brw_entry->proc_fops = &filter_per_export_stats_fops;
187 brw_entry->data = fed;
189 /* Create a per export proc entry for ops stats */
190 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
191 LPROC_FILTER_LAST - 1;
192 exp->exp_ops_stats = lprocfs_alloc_stats(num_stats,
193 LPROCFS_STATS_FLAG_NOPERCPU);
194 if (exp->exp_ops_stats == NULL)
196 lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
197 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
198 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
199 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
200 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
201 lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
205 /* Add client data to the FILTER. We use a bitmap to locate a free space
206 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
207 * Otherwise, we have just read the data from the last_rcvd file and
208 * we know its offset. */
209 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
212 struct filter_obd *filter = &obd->u.filter;
213 struct filter_export_data *fed = &exp->exp_filter_data;
214 unsigned long *bitmap = filter->fo_last_rcvd_slots;
215 int new_client = (cl_idx == -1);
219 LASSERT(bitmap != NULL);
220 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
223 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
226 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
227 * there's no need for extra complication here
230 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
232 if (cl_idx >= LR_MAX_CLIENTS) {
233 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
237 if (test_and_set_bit(cl_idx, bitmap)) {
238 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
243 if (test_and_set_bit(cl_idx, bitmap)) {
244 CERROR("FILTER client %d: bit already set in bitmap!\n",
250 fed->fed_lr_idx = cl_idx;
251 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
252 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
253 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
255 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
256 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
259 struct lvfs_run_ctxt saved;
260 loff_t off = fed->fed_lr_off;
264 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
265 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
267 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
268 /* Transaction needed to fix bug 1403 */
269 handle = fsfilt_start(obd,
270 filter->fo_rcvd_filp->f_dentry->d_inode,
271 FSFILT_OP_SETATTR, NULL);
272 if (IS_ERR(handle)) {
273 rc = PTR_ERR(handle);
274 CERROR("unable to start transaction: rc %d\n", rc);
276 rc = fsfilt_add_journal_cb(obd, 0, handle,
277 target_client_add_cb, exp);
279 spin_lock(&exp->exp_lock);
280 exp->exp_need_sync = 1;
281 spin_unlock(&exp->exp_lock);
283 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
285 sizeof(*fed->fed_fcd),
286 &off, rc /* sync if no cb */);
288 filter->fo_rcvd_filp->f_dentry->d_inode,
291 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
294 CERROR("error writing %s client idx %u: rc %d\n",
295 LAST_RCVD, fed->fed_lr_idx, rc);
302 static int filter_client_free(struct obd_export *exp)
304 struct filter_export_data *fed = &exp->exp_filter_data;
305 struct filter_obd *filter = &exp->exp_obd->u.filter;
306 struct obd_device *obd = exp->exp_obd;
307 struct filter_client_data zero_fcd;
308 struct lvfs_run_ctxt saved;
313 if (fed->fed_fcd == NULL)
316 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
317 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
320 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
321 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
323 LASSERT(filter->fo_last_rcvd_slots != NULL);
325 off = fed->fed_lr_off;
327 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
328 * we leak a client slot that will be cleaned on the next recovery. */
330 CERROR("%s: client idx %d has med_off %lld\n",
331 obd->obd_name, fed->fed_lr_idx, off);
332 GOTO(free, rc = -EINVAL);
335 /* Clear the bit _after_ zeroing out the client so we don't
336 race with filter_client_add and zero out new clients.*/
337 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
338 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
343 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
344 memset(&zero_fcd, 0, sizeof zero_fcd);
345 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
346 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
347 sizeof(zero_fcd), &off,
348 (!exp->exp_libclient ||
349 exp->exp_need_sync));
351 /* update server's transno */
352 filter_update_server_data(obd, filter->fo_rcvd_filp,
354 !exp->exp_libclient);
355 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
357 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
358 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
359 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
363 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
364 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
371 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
377 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
378 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
379 struct filter_mod_data *fmd)
381 LASSERT_SPIN_LOCKED(&fed->fed_lock);
382 if (--fmd->fmd_refcount == 0) {
383 /* XXX when we have persistent reservations and the handle
384 * is stored herein we need to drop it here. */
385 fed->fed_mod_count--;
386 list_del(&fmd->fmd_list);
387 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
391 /* drop fmd reference, free it if last ref */
392 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
394 struct filter_export_data *fed;
399 fed = &exp->exp_filter_data;
400 spin_lock(&fed->fed_lock);
401 filter_fmd_put_nolock(fed, fmd); /* caller reference */
402 spin_unlock(&fed->fed_lock);
405 /* expire entries from the end of the list if there are too many
406 * or they are too old */
407 static void filter_fmd_expire_nolock(struct filter_obd *filter,
408 struct filter_export_data *fed,
409 struct filter_mod_data *keep)
411 struct filter_mod_data *fmd, *tmp;
413 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
417 if (time_before(jiffies, fmd->fmd_expire) &&
418 fed->fed_mod_count < filter->fo_fmd_max_num)
421 list_del_init(&fmd->fmd_list);
422 filter_fmd_put_nolock(fed, fmd); /* list reference */
426 void filter_fmd_expire(struct obd_export *exp)
428 spin_lock(&exp->exp_filter_data.fed_lock);
429 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
430 &exp->exp_filter_data, NULL);
431 spin_unlock(&exp->exp_filter_data.fed_lock);
434 /* find specified objid, group in export fmd list.
435 * caller must hold fed_lock and take fmd reference itself */
436 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
437 struct filter_export_data *fed,
438 obd_id objid, obd_gr group)
440 struct filter_mod_data *found = NULL, *fmd;
442 LASSERT_SPIN_LOCKED(&fed->fed_lock);
444 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
445 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
447 list_del(&fmd->fmd_list);
448 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
449 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
454 filter_fmd_expire_nolock(filter, fed, found);
459 /* Find fmd based on objid and group, or return NULL if not found. */
460 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
461 obd_id objid, obd_gr group)
463 struct filter_mod_data *fmd;
465 spin_lock(&exp->exp_filter_data.fed_lock);
466 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
467 &exp->exp_filter_data, objid, group);
469 fmd->fmd_refcount++; /* caller reference */
470 spin_unlock(&exp->exp_filter_data.fed_lock);
475 /* Find fmd based on objid and group, or create a new one if none is found.
476 * It is possible for this function to return NULL under memory pressure,
477 * or if objid = 0 is passed (which will only cause old entries to expire).
478 * Currently this is not fatal because any fmd state is transient and
479 * may also be freed when it gets sufficiently old. */
480 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
481 obd_id objid, obd_gr group)
483 struct filter_export_data *fed = &exp->exp_filter_data;
484 struct filter_mod_data *found = NULL, *fmd_new = NULL;
486 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
488 spin_lock(&fed->fed_lock);
489 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
492 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
493 fmd_new->fmd_id = objid;
494 fmd_new->fmd_gr = group;
495 fmd_new->fmd_refcount++; /* list reference */
497 fed->fed_mod_count++;
499 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
503 found->fmd_refcount++; /* caller reference */
504 found->fmd_expire = jiffies +
505 exp->exp_obd->u.filter.fo_fmd_max_age;
508 spin_unlock(&fed->fed_lock);
514 /* drop fmd list reference so it will disappear when last reference is put.
515 * This isn't so critical because it would in fact only affect the one client
516 * that is doing the unlink and at worst we have an stale entry referencing
517 * an object that should never be used again. */
518 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
520 struct filter_mod_data *found = NULL;
522 spin_lock(&exp->exp_filter_data.fed_lock);
523 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
525 list_del_init(&found->fmd_list);
526 filter_fmd_put_nolock(&exp->exp_filter_data, found);
528 spin_unlock(&exp->exp_filter_data.fed_lock);
531 #define filter_fmd_drop(exp, objid, group)
534 /* remove all entries from fmd list */
535 static void filter_fmd_cleanup(struct obd_export *exp)
537 struct filter_export_data *fed = &exp->exp_filter_data;
538 struct filter_mod_data *fmd = NULL, *tmp;
540 spin_lock(&fed->fed_lock);
541 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
542 list_del_init(&fmd->fmd_list);
543 filter_fmd_put_nolock(fed, fmd);
545 spin_unlock(&fed->fed_lock);
548 static int filter_init_export(struct obd_export *exp)
550 spin_lock_init(&exp->exp_filter_data.fed_lock);
551 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
553 spin_lock(&exp->exp_lock);
554 exp->exp_connecting = 1;
555 spin_unlock(&exp->exp_lock);
560 static int filter_free_server_data(struct filter_obd *filter)
562 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
563 filter->fo_fsd = NULL;
564 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
565 filter->fo_last_rcvd_slots = NULL;
569 /* assumes caller is already in kernel ctxt */
570 int filter_update_server_data(struct obd_device *obd, struct file *filp,
571 struct lr_server_data *fsd, int force_sync)
577 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
578 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
579 le64_to_cpu(fsd->lsd_last_transno));
580 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
581 le64_to_cpu(fsd->lsd_mount_count));
583 fsd->lsd_compat14 = fsd->lsd_last_transno;
584 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
586 CERROR("error writing lr_server_data: rc = %d\n", rc);
591 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
594 struct filter_obd *filter = &obd->u.filter;
600 if (filter->fo_last_objid_files[group] == NULL) {
601 CERROR("Object group "LPU64" not fully setup; not updating "
602 "last_objid\n", group);
606 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
607 obd->obd_name, group, filter->fo_last_objids[group]);
609 tmp = cpu_to_le64(filter->fo_last_objids[group]);
610 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
611 &tmp, sizeof(tmp), &off, force_sync);
613 CERROR("error writing group "LPU64" last objid: rc = %d\n",
617 extern int ost_handle(struct ptlrpc_request *req);
618 /* assumes caller has already in kernel ctxt */
619 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
621 struct filter_obd *filter = &obd->u.filter;
622 struct lr_server_data *fsd;
623 struct filter_client_data *fcd = NULL;
624 struct inode *inode = filp->f_dentry->d_inode;
625 unsigned long last_rcvd_size = i_size_read(inode);
631 /* ensure padding in the struct is the correct size */
632 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
633 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
634 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
635 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
637 OBD_ALLOC(fsd, sizeof(*fsd));
640 filter->fo_fsd = fsd;
642 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
643 if (filter->fo_last_rcvd_slots == NULL) {
644 OBD_FREE(fsd, sizeof(*fsd));
648 if (last_rcvd_size == 0) {
649 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
651 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
652 fsd->lsd_last_transno = 0;
653 mount_count = fsd->lsd_mount_count = 0;
654 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
655 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
656 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
657 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
658 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
659 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
661 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
663 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
667 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
668 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
669 "using the wrong disk %s. Were the "
670 "/dev/ assignments rearranged?\n",
671 obd->obd_uuid.uuid, fsd->lsd_uuid);
672 GOTO(err_fsd, rc = -EINVAL);
674 mount_count = le64_to_cpu(fsd->lsd_mount_count);
675 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
677 /* Assume old last_rcvd format unless I_C_LR is set */
678 if (!(fsd->lsd_feature_incompat &
679 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
680 fsd->lsd_last_transno = fsd->lsd_compat14;
684 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
685 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
686 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
687 ~FILTER_INCOMPAT_SUPP);
688 GOTO(err_fsd, rc = -EINVAL);
690 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
691 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
692 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
693 ~FILTER_ROCOMPAT_SUPP);
694 /* Do something like remount filesystem read-only */
695 GOTO(err_fsd, rc = -EINVAL);
698 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
699 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
700 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
701 obd->obd_name, mount_count + 1);
702 CDEBUG(D_INODE, "%s: server data size: %u\n",
703 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
704 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
705 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
706 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
707 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
708 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
709 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
710 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
711 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
712 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
713 le16_to_cpu(fsd->lsd_client_size));
715 if (!obd->obd_replayable) {
716 CWARN("%s: recovery support OFF\n", obd->obd_name);
720 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
721 off < last_rcvd_size; cl_idx++) {
723 struct obd_export *exp;
724 struct filter_export_data *fed;
727 OBD_ALLOC(fcd, sizeof(*fcd));
729 GOTO(err_client, rc = -ENOMEM);
732 /* Don't assume off is incremented properly by
733 * fsfilt_read_record(), in case sizeof(*fcd)
734 * isn't the same as fsd->lsd_client_size. */
735 off = le32_to_cpu(fsd->lsd_client_start) +
736 cl_idx * le16_to_cpu(fsd->lsd_client_size);
737 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
739 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
740 LAST_RCVD, cl_idx, off, rc);
741 break; /* read error shouldn't cause startup to fail */
744 if (fcd->fcd_uuid[0] == '\0') {
745 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
750 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
752 /* These exports are cleaned up by filter_disconnect(), so they
753 * need to be set up like real exports as filter_connect() does.
755 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
757 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
758 " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
759 last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
760 le32_to_cpu(fcd->fcd_group));
762 if (PTR_ERR(exp) == -EALREADY) {
763 /* export already exists, zero out this one */
764 CERROR("Zeroing out duplicate export due to "
766 fcd->fcd_uuid[0] = '\0';
768 GOTO(err_client, rc = PTR_ERR(exp));
771 fed = &exp->exp_filter_data;
773 fed->fed_group = le32_to_cpu(fcd->fcd_group);
774 filter_export_stats_init(obd, exp);
775 rc = filter_client_add(obd, exp, cl_idx);
776 /* can't fail for existing client */
777 LASSERTF(rc == 0, "rc = %d\n", rc);
780 spin_lock(&exp->exp_lock);
781 exp->exp_connecting = 0;
782 exp->exp_in_recovery = 0;
783 spin_unlock(&exp->exp_lock);
784 obd->obd_max_recoverable_clients++;
785 class_export_put(exp);
788 /* Need to check last_rcvd even for duplicated exports. */
789 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
792 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
793 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
797 OBD_FREE(fcd, sizeof(*fcd));
799 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
801 target_recovery_init(obd, ost_handle);
804 filter->fo_mount_count = mount_count + 1;
805 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
807 /* save it, so mount count and last_transno is current */
808 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
810 GOTO(err_client, rc);
815 target_recovery_fini(obd);
817 filter_free_server_data(filter);
821 static int filter_cleanup_groups(struct obd_device *obd)
823 struct filter_obd *filter = &obd->u.filter;
825 struct dentry *dentry;
829 if (filter->fo_dentry_O_groups != NULL) {
830 for (i = 0; i < filter->fo_group_count; i++) {
831 dentry = filter->fo_dentry_O_groups[i];
835 OBD_FREE(filter->fo_dentry_O_groups,
836 filter->fo_group_count *
837 sizeof(*filter->fo_dentry_O_groups));
838 filter->fo_dentry_O_groups = NULL;
840 if (filter->fo_last_objid_files != NULL) {
841 for (i = 0; i < filter->fo_group_count; i++) {
842 filp = filter->fo_last_objid_files[i];
846 OBD_FREE(filter->fo_last_objid_files,
847 filter->fo_group_count *
848 sizeof(*filter->fo_last_objid_files));
849 filter->fo_last_objid_files = NULL;
851 if (filter->fo_dentry_O_sub != NULL) {
852 for (i = 0; i < filter->fo_group_count; i++) {
853 for (j = 0; j < filter->fo_subdir_count; j++) {
854 dentry = filter->fo_dentry_O_sub[i].dentry[j];
859 OBD_FREE(filter->fo_dentry_O_sub,
860 filter->fo_group_count *
861 sizeof(*filter->fo_dentry_O_sub));
862 filter->fo_dentry_O_sub = NULL;
864 if (filter->fo_last_objids != NULL) {
865 OBD_FREE(filter->fo_last_objids,
866 filter->fo_group_count *
867 sizeof(*filter->fo_last_objids));
868 filter->fo_last_objids = NULL;
870 if (filter->fo_dentry_O != NULL) {
871 f_dput(filter->fo_dentry_O);
872 filter->fo_dentry_O = NULL;
877 static int filter_update_last_group(struct obd_device *obd, int group)
879 struct filter_obd *filter = &obd->u.filter;
880 struct file *filp = NULL;
881 int last_group = 0, rc;
885 if (group <= filter->fo_committed_group)
888 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
892 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
896 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
898 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
901 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
902 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
903 obd->obd_name, last_group, group);
907 /* must be sync: bXXXX */
908 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
910 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
914 filter->fo_committed_group = group;
921 static int filter_read_group_internal(struct obd_device *obd, int group,
924 struct filter_obd *filter = &obd->u.filter;
925 __u64 *new_objids = NULL;
926 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
927 struct dentry **new_groups = NULL;
928 struct file **new_files = NULL;
929 struct dentry *dentry;
931 int old_count = filter->fo_group_count, rc, stage = 0, i;
937 snprintf(name, 24, "%d", group);
941 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
943 if (IS_ERR(dentry)) {
944 CERROR("Cannot lookup expected object group %d: %ld\n",
945 group, PTR_ERR(dentry));
946 RETURN(PTR_ERR(dentry));
949 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
950 if (IS_ERR(dentry)) {
951 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
953 RETURN(PTR_ERR(dentry));
958 snprintf(name, 24, "O/%d/LAST_ID", group);
960 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
962 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
963 GOTO(cleanup, rc = PTR_ERR(filp));
967 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
969 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
973 if (filter->fo_subdir_count) {
974 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
975 if (tmp_subdirs == NULL)
976 GOTO(cleanup, rc = -ENOMEM);
979 for (i = 0; i < filter->fo_subdir_count; i++) {
981 snprintf(dir, sizeof(dir), "d%u", i);
983 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
984 if (IS_ERR(tmp_subdirs->dentry[i])) {
985 rc = PTR_ERR(tmp_subdirs->dentry[i]);
986 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
991 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
992 tmp_subdirs->dentry[i]);
996 /* 'group' is an index; we need an array of length 'group + 1' */
997 if (group + 1 > old_count) {
998 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
999 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1000 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1001 OBD_ALLOC(new_files, len * sizeof(*new_files));
1003 if (new_objids == NULL || new_subdirs == NULL ||
1004 new_groups == NULL || new_files == NULL)
1005 GOTO(cleanup, rc = -ENOMEM);
1008 memcpy(new_objids, filter->fo_last_objids,
1009 old_count * sizeof(*new_objids));
1010 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1011 old_count * sizeof(*new_subdirs));
1012 memcpy(new_groups, filter->fo_dentry_O_groups,
1013 old_count * sizeof(*new_groups));
1014 memcpy(new_files, filter->fo_last_objid_files,
1015 old_count * sizeof(*new_files));
1017 OBD_FREE(filter->fo_last_objids,
1018 old_count * sizeof(*new_objids));
1019 OBD_FREE(filter->fo_dentry_O_sub,
1020 old_count * sizeof(*new_subdirs));
1021 OBD_FREE(filter->fo_dentry_O_groups,
1022 old_count * sizeof(*new_groups));
1023 OBD_FREE(filter->fo_last_objid_files,
1024 old_count * sizeof(*new_files));
1026 filter->fo_last_objids = new_objids;
1027 filter->fo_dentry_O_sub = new_subdirs;
1028 filter->fo_dentry_O_groups = new_groups;
1029 filter->fo_last_objid_files = new_files;
1030 filter->fo_group_count = len;
1033 filter->fo_dentry_O_groups[group] = dentry;
1034 filter->fo_last_objid_files[group] = filp;
1035 if (filter->fo_subdir_count) {
1036 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1037 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1040 filter_update_last_group(obd, group);
1042 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1043 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1044 rc = filter_update_last_objid(obd, group, 1);
1048 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1049 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1050 obd->obd_name, group, last_objid);
1055 if (new_objids != NULL)
1056 OBD_FREE(new_objids, len * sizeof(*new_objids));
1057 if (new_subdirs != NULL)
1058 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1059 if (new_groups != NULL)
1060 OBD_FREE(new_groups, len * sizeof(*new_groups));
1061 if (new_files != NULL)
1062 OBD_FREE(new_files, len * sizeof(*new_files));
1064 if (filter->fo_subdir_count) {
1065 for (i = 0; i < filter->fo_subdir_count; i++) {
1066 if (tmp_subdirs->dentry[i] != NULL)
1067 dput(tmp_subdirs->dentry[i]);
1069 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1072 filp_close(filp, 0);
1079 static int filter_read_groups(struct obd_device *obd, int last_group,
1082 struct filter_obd *filter = &obd->u.filter;
1083 int old_count, group, rc = 0;
1085 down(&filter->fo_init_lock);
1086 old_count = filter->fo_group_count;
1087 for (group = old_count; group <= last_group; group++) {
1089 continue; /* no group zero */
1091 rc = filter_read_group_internal(obd, group, create);
1095 up(&filter->fo_init_lock);
1099 /* FIXME: object groups */
1100 static int filter_prep_groups(struct obd_device *obd)
1102 struct filter_obd *filter = &obd->u.filter;
1103 struct dentry *dentry, *O_dentry;
1105 int last_group, rc = 0, cleanup_phase = 0;
1109 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1110 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1111 if (IS_ERR(O_dentry)) {
1112 rc = PTR_ERR(O_dentry);
1113 CERROR("cannot open/create O: rc = %d\n", rc);
1116 filter->fo_dentry_O = O_dentry;
1117 cleanup_phase = 1; /* O_dentry */
1119 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1120 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1121 * some time post 1.0 when all old-style OSTs have converted along
1122 * with the init_objid hack. */
1123 dentry = ll_lookup_one_len("R", O_dentry, 1);
1125 GOTO(cleanup, rc = PTR_ERR(dentry));
1126 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1127 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1130 CWARN("converting OST to new object layout\n");
1131 if (IS_ERR(O0_dentry)) {
1132 rc = PTR_ERR(O0_dentry);
1133 CERROR("error looking up O/0: rc %d\n", rc);
1134 GOTO(cleanup_R, rc);
1137 if (O0_dentry->d_inode) {
1138 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1139 GOTO(cleanup_O0, rc = -EEXIST);
1142 LOCK_INODE_MUTEX(O_dentry->d_inode);
1143 rc = vfs_rename(O_dentry->d_inode, dentry,
1144 O_dentry->d_inode, O0_dentry);
1145 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1148 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1149 GOTO(cleanup_O0, rc);
1151 filter->fo_fsd->lsd_feature_incompat |=
1152 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1153 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1155 GOTO(cleanup_O0, rc);
1167 cleanup_phase = 2; /* groups */
1169 /* we have to initialize all groups before first connections from
1170 * clients because they may send create/destroy for any group -bzzz */
1171 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1173 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1174 GOTO(cleanup, rc = PTR_ERR(filp));
1176 cleanup_phase = 3; /* filp */
1178 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1180 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1184 last_group = FILTER_MIN_GROUPS;
1186 LASSERT(last_group >= FILTER_MIN_GROUPS);
1189 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1190 FILTER_MIN_GROUPS, last_group);
1191 filter->fo_committed_group = last_group;
1192 rc = filter_read_groups(obd, last_group, 1);
1196 filp_close(filp, 0);
1200 switch (cleanup_phase) {
1202 filp_close(filp, 0);
1204 filter_cleanup_groups(obd);
1206 f_dput(filter->fo_dentry_O);
1207 filter->fo_dentry_O = NULL;
1215 /* setup the object store with correct subdirectories */
1216 static int filter_prep(struct obd_device *obd)
1218 struct lvfs_run_ctxt saved;
1219 struct filter_obd *filter = &obd->u.filter;
1221 struct inode *inode;
1225 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1226 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1227 if (!file || IS_ERR(file)) {
1229 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1233 filter->fo_rcvd_filp = file;
1234 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1235 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1236 file->f_dentry->d_inode->i_mode);
1237 GOTO(err_filp, rc = -ENOENT);
1240 inode = file->f_dentry->d_parent->d_inode;
1241 /* We use i_op->unlink directly in filter_vfs_unlink() */
1242 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1243 CERROR("%s: filesystem does not support create/unlink ops\n",
1245 GOTO(err_filp, rc = -EOPNOTSUPP);
1248 rc = filter_init_server_data(obd, file);
1250 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1253 /* open and create health check io file*/
1254 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1257 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1259 GOTO(err_server_data, rc);
1261 filter->fo_health_check_filp = file;
1262 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1263 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1264 file->f_dentry->d_inode->i_mode);
1265 GOTO(err_health_check, rc = -ENOENT);
1267 rc = lvfs_check_io_health(obd, file);
1269 GOTO(err_health_check, rc);
1271 rc = filter_prep_groups(obd);
1273 GOTO(err_health_check, rc);
1275 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1280 if (filp_close(filter->fo_health_check_filp, 0))
1281 CERROR("can't close %s after error\n", HEALTH_CHECK);
1282 filter->fo_health_check_filp = NULL;
1284 target_recovery_fini(obd);
1285 filter_free_server_data(filter);
1287 if (filp_close(filter->fo_rcvd_filp, 0))
1288 CERROR("can't close %s after error\n", LAST_RCVD);
1289 filter->fo_rcvd_filp = NULL;
1293 /* cleanup the filter: write last used object id to status file */
1294 static void filter_post(struct obd_device *obd)
1296 struct lvfs_run_ctxt saved;
1297 struct filter_obd *filter = &obd->u.filter;
1300 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1301 * best to start a transaction with h_sync, because we removed this
1304 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1305 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1308 CERROR("error writing server data: rc = %d\n", rc);
1310 for (i = 1; i < filter->fo_group_count; i++) {
1311 rc = filter_update_last_objid(obd, i,
1312 (i == filter->fo_group_count - 1));
1314 CERROR("error writing group %d lastobjid: rc = %d\n",
1318 rc = filp_close(filter->fo_rcvd_filp, 0);
1319 filter->fo_rcvd_filp = NULL;
1321 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1323 rc = filp_close(filter->fo_health_check_filp, 0);
1324 filter->fo_health_check_filp = NULL;
1326 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1328 filter_cleanup_groups(obd);
1329 filter_free_server_data(filter);
1330 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1332 filter_free_capa_keys(filter);
1333 cleanup_capa_hash(filter->fo_capa_hash);
1336 static void filter_set_last_id(struct filter_obd *filter,
1337 obd_id id, obd_gr group)
1339 LASSERT(filter->fo_fsd != NULL);
1340 LASSERT(group <= filter->fo_group_count);
1342 spin_lock(&filter->fo_objidlock);
1343 filter->fo_last_objids[group] = id;
1344 spin_unlock(&filter->fo_objidlock);
1347 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1350 LASSERT(filter->fo_fsd != NULL);
1351 LASSERT(group <= filter->fo_group_count);
1353 /* FIXME: object groups */
1354 spin_lock(&filter->fo_objidlock);
1355 id = filter->fo_last_objids[group];
1356 spin_unlock(&filter->fo_objidlock);
1361 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1363 LOCK_INODE_MUTEX(dparent->d_inode);
1367 /* We never dget the object parent, so DON'T dput it either */
1368 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1370 struct filter_obd *filter = &obd->u.filter;
1371 struct filter_subdirs *subdirs;
1372 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1374 if (group > 0 || filter->fo_subdir_count == 0)
1375 return filter->fo_dentry_O_groups[group];
1377 subdirs = &filter->fo_dentry_O_sub[group];
1378 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1381 /* We never dget the object parent, so DON'T dput it either */
1382 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1385 unsigned long now = jiffies;
1386 struct dentry *dparent = filter_parent(obd, group, objid);
1389 if (IS_ERR(dparent))
1392 rc = filter_lock_dentry(obd, dparent);
1393 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1394 return rc ? ERR_PTR(rc) : dparent;
1397 /* We never dget the object parent, so DON'T dput it either */
1398 static void filter_parent_unlock(struct dentry *dparent)
1400 UNLOCK_INODE_MUTEX(dparent->d_inode);
1403 /* How to get files, dentries, inodes from object id's.
1405 * If dir_dentry is passed, the caller has already locked the parent
1406 * appropriately for this operation (normally a write lock). If
1407 * dir_dentry is NULL, we do a read lock while we do the lookup to
1408 * avoid races with create/destroy and such changing the directory
1409 * internal to the filesystem code. */
1410 struct dentry *filter_fid2dentry(struct obd_device *obd,
1411 struct dentry *dir_dentry,
1412 obd_gr group, obd_id id)
1414 struct dentry *dparent = dir_dentry;
1415 struct dentry *dchild;
1420 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1421 obd->u.filter.fo_destroys_in_progress == 0) {
1422 /* don't fail lookups for orphan recovery, it causes
1423 * later LBUGs when objects still exist during precreate */
1424 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1425 RETURN(ERR_PTR(-ENOENT));
1428 CERROR("fatal: invalid object id 0\n");
1429 RETURN(ERR_PTR(-ESTALE));
1432 len = sprintf(name, LPU64, id);
1433 if (dir_dentry == NULL) {
1434 dparent = filter_parent_lock(obd, group, id);
1435 if (IS_ERR(dparent)) {
1436 CERROR("%s: error getting object "LPU64":"LPU64
1437 " parent: rc %ld\n", obd->obd_name,
1438 id, group, PTR_ERR(dparent));
1442 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1443 dparent->d_name.len, dparent->d_name.name, name);
1444 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1445 if (dir_dentry == NULL)
1446 filter_parent_unlock(dparent);
1447 if (IS_ERR(dchild)) {
1448 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1453 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1454 CERROR("%s: got bad object "LPU64" inode %lu\n",
1455 obd->obd_name, id, dchild->d_inode->i_ino);
1457 RETURN(ERR_PTR(-ENOENT));
1460 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1461 name, dchild, atomic_read(&dchild->d_count));
1463 LASSERT(atomic_read(&dchild->d_count) > 0);
1468 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1471 struct lustre_handle lockh;
1472 int flags = LDLM_AST_DISCARD_DATA, rc;
1473 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1474 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1477 /* Tell the clients that the object is gone now and that they should
1478 * throw away any cached pages. */
1479 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1480 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1481 ldlm_completion_ast, NULL, NULL, 0, NULL,
1484 /* We only care about the side-effects, just drop the lock. */
1486 ldlm_lock_decref(&lockh, LCK_PW);
1491 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1492 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1493 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1494 * reverse. Caller must take i_sem before starting the transaction and we
1495 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1496 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1501 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1502 * itself we already hold dir->i_mutex for child create/unlink ops */
1503 LASSERT(dentry->d_inode != NULL);
1504 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1505 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1509 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1510 GOTO(out, rc = -ENOENT);
1512 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1517 GOTO(out, rc = -EPERM);
1519 /* check_sticky() */
1520 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1521 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1522 GOTO(out, rc = -EPERM);
1524 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1525 * that was done because of journal_start (which is already done
1526 * here) or some other ordering issue. */
1529 rc = security_inode_unlink(dir, dentry);
1533 rc = dir->i_op->unlink(dir, dentry);
1535 /* need to drop i_mutex before we lose inode reference */
1536 UNLOCK_INODE_MUTEX(dentry->d_inode);
1543 /* Caller must hold LCK_PW on parent and push us into kernel context.
1544 * Caller must hold child i_mutex, we drop it always.
1545 * Caller is also required to ensure that dchild->d_inode exists. */
1546 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1547 obd_gr group, struct dentry *dparent,
1548 struct dentry *dchild)
1550 struct inode *inode = dchild->d_inode;
1553 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1554 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1555 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1556 (unsigned long)inode->i_nlink,
1557 atomic_read(&inode->i_count));
1560 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1562 CERROR("error unlinking objid %.*s: rc %d\n",
1563 dchild->d_name.len, dchild->d_name.name, rc);
1567 static int filter_intent_policy(struct ldlm_namespace *ns,
1568 struct ldlm_lock **lockp, void *req_cookie,
1569 ldlm_mode_t mode, int flags, void *data)
1571 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1572 struct ptlrpc_request *req = req_cookie;
1573 struct ldlm_lock *lock = *lockp, *l = NULL;
1574 struct ldlm_resource *res = lock->l_resource;
1575 ldlm_processing_policy policy;
1576 struct ost_lvb *res_lvb, *reply_lvb;
1577 struct ldlm_reply *rep;
1578 struct list_head *tmp;
1580 int rc, tmpflags = 0, only_liblustre = 0;
1581 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1582 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1583 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1586 policy = ldlm_get_processing_policy(res);
1587 LASSERT(policy != NULL);
1588 LASSERT(req != NULL);
1590 rc = lustre_pack_reply(req, 3, repsize, NULL);
1592 RETURN(req->rq_status = rc);
1594 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1595 LASSERT(rep != NULL);
1597 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1598 sizeof(*reply_lvb));
1599 LASSERT(reply_lvb != NULL);
1601 //fixup_handle_for_resent_req(req, lock, &lockh);
1603 /* If we grant any lock at all, it will be a whole-file read lock.
1604 * Call the extent policy function to see if our request can be
1605 * granted, or is blocked. */
1606 lock->l_policy_data.l_extent.start = 0;
1607 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1608 lock->l_req_mode = LCK_PR;
1610 LASSERT(ns == res->lr_namespace);
1612 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1613 check_res_locked(res);
1615 /* FIXME: we should change the policy function slightly, to not make
1616 * this list at all, since we just turn around and free it */
1617 while (!list_empty(&rpc_list)) {
1618 struct ldlm_lock *wlock =
1619 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1620 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1621 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1622 lock->l_flags &= ~LDLM_FL_CP_REQD;
1623 list_del_init(&wlock->l_cp_ast);
1624 LDLM_LOCK_PUT(wlock);
1627 /* The lock met with no resistance; we're finished. */
1628 if (rc == LDLM_ITER_CONTINUE) {
1629 /* do not grant locks to the liblustre clients: they cannot
1630 * handle ASTs robustly. We need to do this while still
1631 * holding ns_lock to avoid the lock remaining on the res_link
1632 * list (and potentially being added to l_pending_list by an
1633 * AST) when we are going to drop this lock ASAP. */
1634 if (lock->l_export->exp_libclient ||
1635 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1636 ldlm_resource_unlink_lock(lock);
1637 err = ELDLM_LOCK_ABORTED;
1639 err = ELDLM_LOCK_REPLACED;
1645 /* Do not grant any lock, but instead send GL callbacks. The extent
1646 * policy nicely created a list of all PW locks for us. We will choose
1647 * the highest of those which are larger than the size in the LVB, if
1648 * any, and perform a glimpse callback. */
1649 res_lvb = res->lr_lvb_data;
1650 LASSERT(res_lvb != NULL);
1651 *reply_lvb = *res_lvb;
1653 list_for_each(tmp, &res->lr_granted) {
1654 struct ldlm_lock *tmplock =
1655 list_entry(tmp, struct ldlm_lock, l_res_link);
1657 if (tmplock->l_granted_mode == LCK_PR)
1660 * ->ns_lock guarantees that no new locks are granted, and,
1661 * therefore, that res->lr_lvb_data cannot increase beyond the
1662 * end of already granted lock. As a result, it is safe to
1663 * check against "stale" reply_lvb->lvb_size value without
1666 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1669 /* Don't send glimpse ASTs to liblustre clients. They aren't
1670 * listening for them, and they do entirely synchronous I/O
1672 if (tmplock->l_export == NULL ||
1673 tmplock->l_export->exp_libclient == 1) {
1679 l = LDLM_LOCK_GET(tmplock);
1683 if (l->l_policy_data.l_extent.start >
1684 tmplock->l_policy_data.l_extent.start)
1688 l = LDLM_LOCK_GET(tmplock);
1692 /* There were no PW locks beyond the size in the LVB; finished. */
1694 if (only_liblustre) {
1695 /* If we discovered a liblustre client with a PW lock,
1696 * however, the LVB may be out of date! The LVB is
1697 * updated only on glimpse (which we don't do for
1698 * liblustre clients) and cancel (which the client
1699 * obviously has not yet done). So if it has written
1700 * data but kept the lock, the LVB is stale and needs
1701 * to be updated from disk.
1703 * Of course, this will all disappear when we switch to
1704 * taking liblustre locks on the OST. */
1705 ldlm_res_lvbo_update(res, NULL, 0, 1);
1707 RETURN(ELDLM_LOCK_ABORTED);
1711 * This check is for lock taken in filter_prepare_destroy() that does
1712 * not have l_glimpse_ast set. So the logic is: if there is a lock
1713 * with no l_glimpse_ast set, this object is being destroyed already.
1715 * Hence, if you are grabbing DLM locks on the server, always set
1716 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1718 if (l->l_glimpse_ast == NULL) {
1719 /* We are racing with unlink(); just return -ENOENT */
1720 rep->lock_policy_res1 = -ENOENT;
1724 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1725 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1726 /* Update the LVB from disk if the AST failed (this is a legal race) */
1728 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1729 * sending ast is not handled. This can result in lost client writes.
1732 ldlm_res_lvbo_update(res, NULL, 0, 1);
1735 *reply_lvb = *res_lvb;
1741 RETURN(ELDLM_LOCK_ABORTED);
1745 * per-obd_device iobuf pool.
1747 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1748 * allocations in write-path has to be minimized (see bug 5137).
1750 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1751 * OST threads (see ost_thread_{init,done}()).
1753 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1754 * at the OST layer there are only (potentially) multiple obd_device of type
1755 * unknown at the time of OST thread creation.
1757 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1758 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1759 * it's very own iobuf.
1763 * filter_kiobuf_pool_init()
1765 * filter_kiobuf_pool_done()
1767 * filter_iobuf_get()
1769 * operate on this array. They are "generic" in a sense that they don't depend
1770 * on actual type of iobuf's (the latter depending on Linux kernel version).
1774 * destroy pool created by filter_iobuf_pool_init
1776 static void filter_iobuf_pool_done(struct filter_obd *filter)
1778 struct filter_iobuf **pool;
1783 pool = filter->fo_iobuf_pool;
1785 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1786 if (pool[i] != NULL)
1787 filter_free_iobuf(pool[i]);
1789 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1790 filter->fo_iobuf_pool = NULL;
1796 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1798 static int filter_iobuf_pool_init(struct filter_obd *filter)
1805 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1807 if (filter->fo_iobuf_pool == NULL)
1810 filter->fo_iobuf_count = OSS_THREADS_MAX;
1815 /* Return iobuf allocated for @thread_id. We don't know in advance how
1816 * many threads there will be so we allocate a large empty array and only
1817 * fill in those slots that are actually in use.
1818 * If we haven't allocated a pool entry for this thread before, do so now. */
1819 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1821 int thread_id = oti ? oti->oti_thread_id : -1;
1822 struct filter_iobuf *pool = NULL;
1823 struct filter_iobuf **pool_place = NULL;
1825 if (thread_id >= 0) {
1826 LASSERT(thread_id < filter->fo_iobuf_count);
1827 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1830 if (unlikely(pool == NULL)) {
1831 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1832 PTLRPC_MAX_BRW_PAGES);
1833 if (pool_place != NULL)
1840 /* mount the file system (secretly). lustre_cfg parameters are:
1843 * 3 = flags: failover=f, failout=n
1846 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1849 struct filter_obd *filter = &obd->u.filter;
1850 struct vfsmount *mnt;
1851 struct lustre_mount_info *lmi;
1852 struct obd_uuid uuid;
1859 if (lcfg->lcfg_bufcount < 3 ||
1860 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1861 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1864 lmi = server_get_mount(obd->obd_name);
1866 /* We already mounted in lustre_fill_super.
1867 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1868 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1870 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1872 /* old path - used by lctl */
1873 CERROR("Using old MDS mount method\n");
1874 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1875 MS_NOATIME|MS_NODIRATIME,
1876 lustre_cfg_string(lcfg, 1), option);
1879 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1880 lustre_cfg_string(lcfg, 1), rc);
1884 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1886 if (IS_ERR(obd->obd_fsops))
1887 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1889 rc = filter_iobuf_pool_init(filter);
1893 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1895 /* failover is the default */
1896 obd->obd_replayable = 1;
1898 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1899 str = lustre_cfg_string(lcfg, 3);
1900 if (strchr(str, 'n')) {
1901 CWARN("%s: recovery disabled\n", obd->obd_name);
1902 obd->obd_replayable = 0;
1906 filter->fo_vfsmnt = mnt;
1907 obd->u.obt.obt_sb = mnt->mnt_sb;
1908 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1909 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1911 fsfilt_setup(obd, obd->u.obt.obt_sb);
1913 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1914 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1915 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1916 obd->obd_lvfs_ctxt.fs = get_ds();
1917 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1919 sema_init(&filter->fo_init_lock, 1);
1920 filter->fo_committed_group = 0;
1922 rc = filter_prep(obd);
1926 filter->fo_destroys_in_progress = 0;
1927 for (i = 0; i < 32; i++)
1928 sema_init(&filter->fo_create_locks[i], 1);
1930 spin_lock_init(&filter->fo_translock);
1931 spin_lock_init(&filter->fo_objidlock);
1932 INIT_LIST_HEAD(&filter->fo_export_list);
1933 sema_init(&filter->fo_alloc_lock, 1);
1934 init_brw_stats(&filter->fo_filter_stats);
1935 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1936 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1937 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1939 INIT_LIST_HEAD(&filter->fo_llog_list);
1940 spin_lock_init(&filter->fo_llog_list_lock);
1942 filter->fo_fl_oss_capa = 0;
1943 INIT_LIST_HEAD(&filter->fo_capa_keys);
1944 filter->fo_capa_hash = init_capa_hash();
1945 if (filter->fo_capa_hash == NULL)
1946 GOTO(err_ops, rc = -ENOMEM);
1948 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1949 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1950 LDLM_NAMESPACE_GREEDY);
1951 if (obd->obd_namespace == NULL)
1952 GOTO(err_post, rc = -ENOMEM);
1953 obd->obd_namespace->ns_lvbp = obd;
1954 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1955 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1957 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1958 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1960 rc = llog_cat_initialize(obd, NULL, 1, NULL);
1962 CERROR("failed to setup llogging subsystems\n");
1966 rc = lquota_setup(filter_quota_interface_ref, obd);
1970 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1971 if (uuid_ptr != NULL) {
1972 class_uuid_unparse(uuid_ptr, &uuid);
1978 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1980 if (obd->obd_recovering) {
1981 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1982 "recovery until %d %s reconnect, or if no clients"
1983 " reconnect for %d:%.02d; during that time new "
1984 "clients will not be allowed to connect. "
1985 "Recovery progress can be monitored by watching "
1986 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1987 obd->obd_name, lustre_cfg_string(lcfg, 1),
1988 label ?: "", label ? "/" : "", str,
1989 obd->obd_max_recoverable_clients,
1990 (obd->obd_max_recoverable_clients == 1)
1991 ? "client" : "clients",
1992 (int)(OBD_RECOVERY_TIMEOUT) / 60,
1993 (int)(OBD_RECOVERY_TIMEOUT) % 60,
1996 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1997 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
1998 label ?: "", label ? "/" : "", str,
1999 obd->obd_replayable ? "enabled" : "disabled");
2007 fsfilt_put_ops(obd->obd_fsops);
2008 filter_iobuf_pool_done(filter);
2010 server_put_mount(obd->obd_name, mnt);
2011 obd->u.obt.obt_sb = 0;
2015 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2017 struct lprocfs_static_vars lvars;
2022 CLASSERT(offsetof(struct obd_device, u.obt) ==
2023 offsetof(struct obd_device, u.filter.fo_obt));
2025 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2028 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2029 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2032 addr = (unsigned long)cfs_page_address(page);
2033 clear_page((void *)addr);
2035 /* lprocfs must be setup before the filter so state can be safely added
2036 * to /proc incrementally as the filter is setup */
2037 lprocfs_init_vars(filter, &lvars);
2038 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2039 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2040 /* Init obdfilter private stats here */
2041 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2042 LPROCFS_CNTR_AVGMINMAX,
2043 "read_bytes", "bytes");
2044 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2045 LPROCFS_CNTR_AVGMINMAX,
2046 "write_bytes", "bytes");
2048 lproc_filter_attach_seqstat(obd);
2049 obd->obd_proc_exports = proc_mkdir("exports",
2050 obd->obd_proc_entry);
2053 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2054 LUSTRE_CFG_BUFLEN(lcfg, 4));
2055 rc = filter_common_setup(obd, lcfg, (void *)addr);
2056 OBD_PAGE_FREE(page);
2059 lprocfs_obd_cleanup(obd);
2060 lprocfs_free_obd_stats(obd);
2066 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2067 static struct llog_operations filter_size_orig_logops = {
2068 lop_setup: llog_obd_origin_setup,
2069 lop_cleanup: llog_obd_origin_cleanup,
2070 lop_add: llog_obd_origin_add
2073 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2074 struct obd_device *tgt, int count,
2075 struct llog_catid *catid,
2076 struct obd_uuid *uuid)
2078 struct llog_ctxt *ctxt;
2082 filter_mds_ost_repl_logops = llog_client_ops;
2083 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2084 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2085 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2087 rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2088 &filter_mds_ost_repl_logops);
2092 /* FIXME - assign unlink_cb for filter's recovery */
2094 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2096 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2098 LASSERT(ctxt != NULL);
2099 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2101 rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2102 &filter_size_orig_logops);
2106 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2111 if (CTXTP(ctxt, cleanup))
2112 rc = CTXTP(ctxt, cleanup)(ctxt);
2115 class_export_put(ctxt->loc_exp);
2116 OBD_FREE(ctxt, sizeof(*ctxt));
2121 static int filter_group_llog_finish(struct obd_llogs *llogs)
2123 struct llog_ctxt *ctxt;
2124 int rc = 0, rc2 = 0;
2127 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2129 rc = filter_group_llog_cleanup(ctxt);
2131 ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2133 rc2 = filter_group_llog_cleanup(ctxt);
2140 static int filter_llog_finish(struct obd_device *obd, int count)
2142 struct llog_ctxt *ctxt;
2143 int rc = 0, rc2 = 0;
2146 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2148 rc = llog_cleanup(ctxt);
2150 ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2152 rc2 = llog_cleanup(ctxt);
2159 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2160 struct obd_export *export)
2162 struct filter_group_llog *fglog, *nlog;
2163 struct filter_obd *filter;
2164 struct llog_ctxt *ctxt;
2165 struct list_head *cur;
2168 filter = &obd->u.filter;
2170 spin_lock(&filter->fo_llog_list_lock);
2171 list_for_each(cur, &filter->fo_llog_list) {
2172 fglog = list_entry(cur, struct filter_group_llog, list);
2173 if (fglog->group == group) {
2174 if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2175 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2176 obd->obd_name, group, fglog->exp, export);
2177 spin_unlock(&filter->fo_llog_list_lock);
2181 spin_unlock(&filter->fo_llog_list_lock);
2186 OBD_ALLOC_PTR(fglog);
2189 fglog->group = group;
2191 OBD_ALLOC_PTR(fglog->llogs);
2192 if (fglog->llogs == NULL) {
2193 OBD_FREE_PTR(fglog);
2197 spin_lock(&filter->fo_llog_list_lock);
2198 list_for_each(cur, &filter->fo_llog_list) {
2199 nlog = list_entry(cur, struct filter_group_llog, list);
2200 LASSERT(nlog->group != group);
2202 list_add(&fglog->list, &filter->fo_llog_list);
2203 spin_unlock(&filter->fo_llog_list_lock);
2205 rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2207 OBD_FREE_PTR(fglog->llogs);
2208 OBD_FREE_PTR(fglog);
2214 fglog->exp = export;
2215 ctxt = llog_get_context_from_llogs(fglog->llogs,
2216 LLOG_MDS_OST_REPL_CTXT);
2217 LASSERT(ctxt != NULL);
2219 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2221 CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2222 obd->obd_name, fglog->llogs, group);
2224 RETURN(fglog->llogs);
2227 static int filter_llog_connect(struct obd_export *exp,
2228 struct llogd_conn_body *body)
2230 struct obd_device *obd = exp->exp_obd;
2231 struct llog_ctxt *ctxt;
2232 struct obd_llogs *llog;
2236 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2237 (unsigned) body->lgdc_logid.lgl_ogr,
2238 (unsigned) body->lgdc_logid.lgl_oid,
2239 (unsigned) body->lgdc_logid.lgl_ogen);
2241 llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2242 LASSERT(llog != NULL);
2243 ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2244 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2245 body->lgdc_ctxt_idx);
2246 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2247 &body->lgdc_gen, NULL);
2249 CERROR("failed to connect rc %d idx %d\n", rc,
2250 body->lgdc_ctxt_idx);
2255 static int filter_llog_preclean (struct obd_device *obd)
2257 struct filter_group_llog *log;
2258 struct filter_obd *filter;
2262 filter = &obd->u.filter;
2263 spin_lock(&filter->fo_llog_list_lock);
2264 while (!list_empty(&filter->fo_llog_list)) {
2265 log = list_entry(filter->fo_llog_list.next,
2266 struct filter_group_llog, list);
2267 list_del(&log->list);
2268 spin_unlock(&filter->fo_llog_list_lock);
2270 rc = filter_group_llog_finish(log->llogs);
2272 CERROR("failed to cleanup llogging subsystem for %u\n",
2274 OBD_FREE_PTR(log->llogs);
2276 spin_lock(&filter->fo_llog_list_lock);
2278 spin_unlock(&filter->fo_llog_list_lock);
2280 rc = obd_llog_finish(obd, 0);
2282 CERROR("failed to cleanup llogging subsystem\n");
2287 static int filter_precleanup(struct obd_device *obd,
2288 enum obd_cleanup_stage stage)
2294 case OBD_CLEANUP_EARLY:
2296 case OBD_CLEANUP_EXPORTS:
2297 target_cleanup_recovery(obd);
2299 case OBD_CLEANUP_SELF_EXP:
2300 rc = filter_llog_preclean(obd);
2302 case OBD_CLEANUP_OBD:
2308 static int filter_cleanup(struct obd_device *obd)
2310 struct filter_obd *filter = &obd->u.filter;
2314 LCONSOLE_WARN("%s: shutting down for failover; client state "
2315 "will be preserved.\n", obd->obd_name);
2317 if (!list_empty(&obd->obd_exports)) {
2318 CERROR("%s: still has clients!\n", obd->obd_name);
2319 class_disconnect_exports(obd);
2320 if (!list_empty(&obd->obd_exports)) {
2321 CERROR("still has exports after forced cleanup?\n");
2326 lprocfs_obd_cleanup(obd);
2327 lprocfs_free_obd_stats(obd);
2328 lquota_cleanup(filter_quota_interface_ref, obd);
2330 /* Stop recovery before namespace cleanup. */
2331 target_stop_recovery_thread(obd);
2332 target_cleanup_recovery(obd);
2334 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2336 if (obd->u.obt.obt_sb == NULL)
2341 shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2343 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2345 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2346 obd->u.obt.obt_sb = NULL;
2348 fsfilt_put_ops(obd->obd_fsops);
2350 filter_iobuf_pool_done(filter);
2352 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2357 static int filter_connect_internal(struct obd_export *exp,
2358 struct obd_connect_data *data)
2363 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2364 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2365 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2366 data->ocd_connect_flags, data->ocd_version,
2367 data->ocd_grant, data->ocd_index);
2369 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2370 exp->exp_connect_flags = data->ocd_connect_flags;
2371 data->ocd_version = LUSTRE_VERSION_CODE;
2373 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2374 struct filter_export_data *fed = &exp->exp_filter_data;
2375 obd_size left, want;
2377 spin_lock(&exp->exp_obd->obd_osfs_lock);
2378 left = filter_grant_space_left(exp);
2379 want = data->ocd_grant;
2380 filter_grant(exp, fed->fed_grant, want, left);
2381 data->ocd_grant = fed->fed_grant;
2382 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2384 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2385 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2386 exp->exp_client_uuid.uuid, exp,
2387 data->ocd_grant, want, left);
2390 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2391 struct filter_obd *filter = &exp->exp_obd->u.filter;
2392 struct lr_server_data *lsd = filter->fo_fsd;
2393 int index = le32_to_cpu(lsd->lsd_ost_index);
2395 if (!(lsd->lsd_feature_compat &
2396 cpu_to_le32(OBD_COMPAT_OST))) {
2397 /* this will only happen on the first connect */
2398 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2399 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2400 filter_update_server_data(exp->exp_obd,
2401 filter->fo_rcvd_filp, lsd, 1);
2402 } else if (index != data->ocd_index) {
2403 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2404 " %u doesn't match actual OST index"
2405 " %u in last_rcvd file, bad "
2407 obd_export_nid2str(exp), index,
2413 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2414 data->ocd_brw_size = 65536;
2415 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2416 data->ocd_brw_size = min(data->ocd_brw_size,
2417 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2418 LASSERT(data->ocd_brw_size);
2421 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2422 * FIXME: We don't strictly need the COMPAT flag for that,
2423 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2424 * FIXME: We needed it for the index, as index 0 is valid. */
2429 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2430 struct obd_uuid *cluuid,
2431 struct obd_connect_data *data)
2436 if (exp == NULL || obd == NULL || cluuid == NULL)
2439 rc = filter_connect_internal(exp, data);
2444 /* nearly identical to mds_connect */
2445 static int filter_connect(const struct lu_env *env,
2446 struct lustre_handle *conn, struct obd_device *obd,
2447 struct obd_uuid *cluuid,
2448 struct obd_connect_data *data)
2450 struct lvfs_run_ctxt saved;
2451 struct obd_export *exp;
2452 struct filter_export_data *fed;
2453 struct filter_client_data *fcd = NULL;
2458 if (conn == NULL || obd == NULL || cluuid == NULL)
2461 rc = class_connect(conn, obd, cluuid);
2464 exp = class_conn2export(conn);
2465 LASSERT(exp != NULL);
2467 fed = &exp->exp_filter_data;
2469 rc = filter_connect_internal(exp, data);
2473 filter_export_stats_init(obd, exp);
2474 group = data->ocd_group;
2475 if (obd->obd_replayable) {
2476 OBD_ALLOC(fcd, sizeof(*fcd));
2478 CERROR("filter: out of memory for client data\n");
2479 GOTO(cleanup, rc = -ENOMEM);
2482 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2484 fed->fed_fcd->fcd_group = group;
2485 rc = filter_client_add(obd, exp, -1);
2489 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2490 obd->obd_name, exp->exp_handle.h_cookie, group);
2494 if (fed->fed_group != 0 && fed->fed_group != group) {
2495 CERROR("!!! This export (nid %s) used object group %d "
2496 "earlier; now it's trying to use group %d! This could "
2497 "be a bug in the MDS. Tell CFS.\n",
2498 obd_export_nid2str(exp), fed->fed_group, group);
2499 GOTO(cleanup, rc = -EPROTO);
2501 fed->fed_group = group;
2503 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2504 rc = filter_read_groups(obd, group, 1);
2505 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2507 CERROR("can't read group %u\n", group);
2516 OBD_FREE(fcd, sizeof(*fcd));
2517 fed->fed_fcd = NULL;
2519 class_disconnect(exp);
2521 class_export_put(exp);
2527 /* Do extra sanity checks for grant accounting. We do this at connect,
2528 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2529 * always get rid of it or turn it off when we know accounting is good. */
2530 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2532 struct filter_export_data *fed;
2533 struct obd_export *exp;
2534 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2535 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2536 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2538 if (list_empty(&obd->obd_exports))
2541 /* We don't want to do this for large machines that do lots of
2542 mounts or unmounts. It burns... */
2543 if (obd->obd_num_exports > 100)
2546 spin_lock(&obd->obd_osfs_lock);
2547 spin_lock(&obd->obd_dev_lock);
2548 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2550 fed = &exp->exp_filter_data;
2551 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2554 if (maxsize > 0) { /* we may not have done a statfs yet */
2555 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2556 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2557 exp->exp_client_uuid.uuid, exp,
2558 fed->fed_grant, fed->fed_pending, maxsize);
2559 LASSERTF(fed->fed_dirty <= maxsize,
2560 "%s: cli %s/%p %ld > "LPU64"\n", func,
2561 exp->exp_client_uuid.uuid, exp,
2562 fed->fed_dirty, maxsize);
2565 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2566 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2567 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2569 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2570 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2571 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2572 tot_granted += fed->fed_grant + fed->fed_pending;
2573 tot_pending += fed->fed_pending;
2574 tot_dirty += fed->fed_dirty;
2576 fo_tot_granted = obd->u.filter.fo_tot_granted;
2577 fo_tot_pending = obd->u.filter.fo_tot_pending;
2578 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2579 spin_unlock(&obd->obd_dev_lock);
2580 spin_unlock(&obd->obd_osfs_lock);
2582 /* Do these assertions outside the spinlocks so we don't kill system */
2583 if (tot_granted != fo_tot_granted)
2584 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2585 func, tot_granted, fo_tot_granted);
2586 if (tot_pending != fo_tot_pending)
2587 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2588 func, tot_pending, fo_tot_pending);
2589 if (tot_dirty != fo_tot_dirty)
2590 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2591 func, tot_dirty, fo_tot_dirty);
2592 if (tot_pending > tot_granted)
2593 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2594 func, tot_pending, tot_granted);
2595 if (tot_granted > maxsize)
2596 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2597 func, tot_granted, maxsize);
2598 if (tot_dirty > maxsize)
2599 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2600 func, tot_dirty, maxsize);
2603 /* Remove this client from the grant accounting totals. We also remove
2604 * the export from the obd device under the osfs and dev locks to ensure
2605 * that the filter_grant_sanity_check() calculations are always valid.
2606 * The client should do something similar when it invalidates its import. */
2607 static void filter_grant_discard(struct obd_export *exp)
2609 struct obd_device *obd = exp->exp_obd;
2610 struct filter_obd *filter = &obd->u.filter;
2611 struct filter_export_data *fed = &exp->exp_filter_data;
2613 spin_lock(&obd->obd_osfs_lock);
2614 spin_lock(&obd->obd_dev_lock);
2615 list_del_init(&exp->exp_obd_chain);
2616 spin_unlock(&obd->obd_dev_lock);
2618 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2619 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2620 obd->obd_name, filter->fo_tot_granted,
2621 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2622 filter->fo_tot_granted -= fed->fed_grant;
2623 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2624 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2625 obd->obd_name, filter->fo_tot_pending,
2626 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2627 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2628 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2629 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2630 obd->obd_name, filter->fo_tot_dirty,
2631 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2632 filter->fo_tot_dirty -= fed->fed_dirty;
2636 spin_unlock(&obd->obd_osfs_lock);
2639 static int filter_destroy_export(struct obd_export *exp)
2643 if (exp->exp_filter_data.fed_pending)
2644 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2645 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2646 exp, exp->exp_filter_data.fed_pending);
2648 target_destroy_export(exp);
2650 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2653 lprocfs_exp_cleanup(exp);
2655 if (exp->exp_obd->obd_replayable)
2656 filter_client_free(exp);
2658 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2660 filter_grant_discard(exp);
2661 filter_fmd_cleanup(exp);
2663 if (!(exp->exp_flags & OBD_OPT_FORCE))
2664 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2669 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2671 struct filter_group_llog *fglog, *nlog;
2672 struct filter_obd *filter;
2673 int worked = 0, group;
2674 struct llog_ctxt *ctxt;
2677 filter = &obd->u.filter;
2679 /* we can't sync log holding spinlock. also, we do not want to get
2680 * into livelock. so we do following: loop over MDS's exports in
2681 * group order and skip already synced llogs -bzzz */
2683 /* look for group with min. number, but > worked */
2686 spin_lock(&filter->fo_llog_list_lock);
2687 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2688 if (nlog->group <= worked) {
2689 /* this group is already synced */
2692 if (group < nlog->group) {
2693 /* we have group with smaller number to sync */
2696 /* store current minimal group */
2698 group = nlog->group;
2700 spin_unlock(&filter->fo_llog_list_lock);
2705 worked = fglog->group;
2706 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2707 ctxt = llog_get_context_from_llogs(fglog->llogs,
2708 LLOG_MDS_OST_REPL_CTXT);
2709 LASSERT(ctxt != NULL);
2710 llog_sync(ctxt, fglog->exp);
2712 } while (fglog != NULL);
2715 /* also incredibly similar to mds_disconnect */
2716 static int filter_disconnect(struct obd_export *exp)
2718 struct obd_device *obd = exp->exp_obd;
2723 class_export_get(exp);
2725 if (!(exp->exp_flags & OBD_OPT_FORCE))
2726 filter_grant_sanity_check(obd, __FUNCTION__);
2727 filter_grant_discard(exp);
2729 /* Disconnect early so that clients can't keep using export */
2730 rc = class_disconnect(exp);
2731 if (exp->exp_obd->obd_namespace != NULL)
2732 ldlm_cancel_locks_for_export(exp);
2734 fsfilt_sync(obd, obd->u.obt.obt_sb);
2736 /* flush any remaining cancel messages out to the target */
2737 filter_sync_llogs(obd, exp);
2738 class_export_put(exp);
2742 /* reverse import is changed, sync all cancels */
2743 static void filter_revimp_update(struct obd_export *exp)
2748 class_export_get(exp);
2750 /* flush any remaining cancel messages out to the target */
2751 filter_sync_llogs(exp->exp_obd, exp);
2752 class_export_put(exp);
2756 static int filter_ping(struct obd_export *exp)
2758 filter_fmd_expire(exp);
2763 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2764 const char *what, int quiet)
2766 struct dentry *dchild = NULL;
2769 if (oa->o_valid & OBD_MD_FLGROUP)
2772 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2774 if (IS_ERR(dchild)) {
2775 CERROR("%s error looking up object: "LPU64"\n",
2780 if (dchild->d_inode == NULL) {
2782 CERROR("%s: %s on non-existent object: "LPU64"\n",
2783 obd->obd_name, what, oa->o_id);
2785 RETURN(ERR_PTR(-ENOENT));
2791 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2793 struct dentry *dentry = NULL;
2794 struct obd_device *obd;
2798 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2799 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2803 obd = class_exp2obd(exp);
2805 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2809 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2811 RETURN(PTR_ERR(dentry));
2813 /* Limit the valid bits in the return data to what we actually use */
2814 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2815 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2821 /* this should be enabled/disabled in condition to enabled/disabled large
2822 * inodes (fast EAs) in backing store FS. */
2823 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2824 void *handle, struct obdo *oa)
2826 struct obd_device *obd = exp->exp_obd;
2830 if (oa->o_valid & OBD_MD_FLFID) {
2831 struct filter_fid ff;
2833 if (!(oa->o_valid & OBD_MD_FLGROUP))
2835 /* packing fid and converting it to LE for storing into EA.
2836 * Here ->o_stripe_idx should be filled by LOV and rest of
2837 * fields - by client. */
2838 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2839 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2840 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2841 ff.ff_objid = cpu_to_le64(oa->o_id);
2842 ff.ff_group = cpu_to_le64(oa->o_gr);
2844 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2845 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2846 oa->o_generation, oa->o_id, oa->o_gr);
2848 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2850 CERROR("store fid in object failed! rc: %d\n", rc);
2852 CDEBUG(D_HA, "OSS object without fid info!\n");
2858 /* this is called from filter_truncate() until we have filter_punch() */
2859 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2860 struct obdo *oa, struct obd_trans_info *oti)
2862 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2863 struct llog_cookie *fcc = NULL;
2864 struct filter_obd *filter;
2865 int rc, err, locked = 0, sync = 0;
2866 unsigned int ia_valid;
2867 struct inode *inode;
2872 LASSERT(dentry != NULL);
2873 LASSERT(!IS_ERR(dentry));
2875 inode = dentry->d_inode;
2876 LASSERT(inode != NULL);
2878 filter = &exp->exp_obd->u.filter;
2879 iattr_from_obdo(&iattr, oa, oa->o_valid);
2880 ia_valid = iattr.ia_valid;
2882 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2883 OBD_ALLOC(fcc, sizeof(*fcc));
2885 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2888 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2890 LOCK_INODE_MUTEX(inode);
2894 /* If the inode still has SUID+SGID bits set (see filter_precreate())
2895 * then we will accept the UID+GID sent by the client during write for
2896 * initializing the ownership of this inode. We only allow this to
2897 * happen once so clear these bits in setattr. In 2.6 kernels it is
2898 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2899 * the flags that are actually being set. */
2900 if (ia_valid & (ATTR_UID | ATTR_GID)) {
2901 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2902 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2904 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2905 if (!(ia_valid & ATTR_MODE)) {
2906 iattr.ia_mode = inode->i_mode;
2907 iattr.ia_valid |= ATTR_MODE;
2909 iattr.ia_mode &= ~S_ISUID;
2911 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2912 if (!(iattr.ia_valid & ATTR_MODE)) {
2913 iattr.ia_mode = inode->i_mode;
2914 iattr.ia_valid |= ATTR_MODE;
2916 iattr.ia_mode &= ~S_ISGID;
2919 orig_ids[USRQUOTA] = inode->i_uid;
2920 orig_ids[GRPQUOTA] = inode->i_gid;
2921 handle = fsfilt_start_log(exp->exp_obd, inode,
2922 FSFILT_OP_SETATTR, oti, 1);
2924 GOTO(out_unlock, rc = PTR_ERR(handle));
2926 /* update inode EA only once when inode is suid bit marked. As
2927 * on 2.6.x UID and GID may be set separately, we check here
2928 * only one of them to avoid double setting. */
2929 if (inode->i_mode & S_ISUID)
2930 filter_update_fidea(exp, inode, handle, oa);
2932 handle = fsfilt_start(exp->exp_obd, inode,
2933 FSFILT_OP_SETATTR, oti);
2935 GOTO(out_unlock, rc = PTR_ERR(handle));
2937 if (oa->o_valid & OBD_MD_FLFLAGS) {
2938 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2939 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2941 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2943 /* set cancel cookie callback function */
2944 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2945 filter_cancel_cookies_cb,
2949 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2950 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2952 /* The truncate might have used up our transaction credits. Make
2953 * sure we have one left for the last_rcvd update. */
2954 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2956 rc = filter_finish_transno(exp, oti, rc, sync);
2958 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2962 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2964 CERROR("error on commit, err = %d\n", err);
2972 /* Let's flush truncated page on disk immediately, then we can
2973 * avoid need to search for page aliases before directio writes
2974 * and this sort of stuff at expense of somewhat slower
2975 * truncates not on a page boundary. I believe this is the only
2976 * place in filter code that can lead to pages getting to
2977 * pagecache so far. */
2978 filter_clear_truncated_page(inode);
2979 UNLOCK_INODE_MUTEX(inode);
2986 UNLOCK_INODE_MUTEX(inode);
2989 OBD_FREE(fcc, sizeof(*fcc));
2991 /* trigger quota release */
2992 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2993 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2994 int rc2 = lquota_adjust(filter_quota_interface_ref,
2995 exp->exp_obd, cur_ids,
2996 orig_ids, rc, FSFILT_OP_SETATTR);
2997 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2998 "filter adjust qunit. (rc:%d)\n", rc2);
3003 /* this is called from filter_truncate() until we have filter_punch() */
3004 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3005 struct obd_trans_info *oti)
3007 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3008 oinfo->oi_oa->o_gr, 0 } };
3009 struct filter_mod_data *fmd;
3010 struct lvfs_run_ctxt saved;
3011 struct filter_obd *filter;
3012 struct ldlm_resource *res;
3013 struct dentry *dentry;
3017 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3018 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3022 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3025 RETURN(PTR_ERR(dentry));
3027 filter = &exp->exp_obd->u.filter;
3028 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3031 if (oinfo->oi_oa->o_valid &
3032 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3033 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3034 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3035 fmd->fmd_mactime_xid = oti->oti_xid;
3036 filter_fmd_put(exp, fmd);
3039 /* setting objects attributes (including owner/group) */
3040 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3042 GOTO(out_unlock, rc);
3044 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3045 &res_id, LDLM_EXTENT, 0);
3048 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3049 ldlm_resource_putref(res);
3052 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3054 /* Quota release need uid/gid info */
3055 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3056 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3062 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3066 /* XXX identical to osc_unpackmd */
3067 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3068 struct lov_mds_md *lmm, int lmm_bytes)
3074 if (lmm_bytes < sizeof (*lmm)) {
3075 CERROR("lov_mds_md too small: %d, need %d\n",
3076 lmm_bytes, (int)sizeof(*lmm));
3079 /* XXX LOV_MAGIC etc check? */
3081 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3082 CERROR("lov_mds_md: zero lmm_object_id\n");
3087 lsm_size = lov_stripe_md_size(1);
3091 if (*lsmp != NULL && lmm == NULL) {
3092 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3093 OBD_FREE(*lsmp, lsm_size);
3098 if (*lsmp == NULL) {
3099 OBD_ALLOC(*lsmp, lsm_size);
3103 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3104 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3105 OBD_FREE(*lsmp, lsm_size);
3108 loi_init((*lsmp)->lsm_oinfo[0]);
3112 /* XXX zero *lsmp? */
3113 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3114 LASSERT((*lsmp)->lsm_object_id);
3117 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3122 /* caller must hold fo_create_locks[oa->o_gr] */
3123 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3124 struct filter_obd *filter)
3126 struct obdo doa; /* XXX obdo on stack */
3132 LASSERT(oa->o_gr != 0);
3133 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3134 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3136 memset(&doa, 0, sizeof(doa));
3138 doa.o_valid |= OBD_MD_FLGROUP;
3139 doa.o_gr = oa->o_gr;
3140 doa.o_mode = S_IFREG;
3142 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3143 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3144 exp->exp_obd->obd_name, doa.o_gr);
3148 last = filter_last_id(filter, doa.o_gr);
3150 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3151 exp->exp_obd->obd_name, oa->o_id + 1, last);
3153 for (id = last; id > oa->o_id; id--) {
3155 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3156 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3157 CEMERG("error destroying precreate objid "LPU64": %d\n",
3159 filter_set_last_id(filter, id - 1, doa.o_gr);
3160 /* update last_id on disk periodically so that if we restart
3161 * we don't need to re-scan all of the just-deleted objects. */
3162 if ((id & 511) == 0)
3163 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3166 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3167 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3169 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3170 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3175 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3176 obd_gr group, int *num);
3177 /* returns a negative error or a nonnegative number of files to create */
3178 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3179 obd_gr group, struct obd_trans_info *oti)
3181 struct obd_device *obd = exp->exp_obd;
3182 struct filter_obd *filter = &obd->u.filter;
3186 /* delete orphans request */
3187 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3188 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3189 CERROR("%s: dropping old orphan cleanup request\n",
3193 /* This causes inflight precreates to abort and drop lock */
3194 set_bit(group, &filter->fo_destroys_in_progress);
3195 down(&filter->fo_create_locks[group]);
3196 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3197 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3198 exp->exp_obd->obd_name, group);
3199 up(&filter->fo_create_locks[group]);
3202 diff = oa->o_id - filter_last_id(filter, group);
3203 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3204 filter_last_id(filter, group), diff);
3206 if (-diff > OST_MAX_PRECREATE) {
3207 CERROR("%s: ignoring bogus orphan destroy request: "
3208 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3209 oa->o_id, filter_last_id(filter, group));
3210 /* FIXME: should reset precreate_next_id on MDS */
3211 GOTO(out, rc = -EINVAL);
3214 rc = filter_destroy_precreated(exp, oa, filter);
3216 CERROR("%s: unable to write lastobjid, but "
3217 "orphans were deleted\n", obd->obd_name);
3220 /* XXX: Used by MDS for the first time! */
3221 clear_bit(group, &filter->fo_destroys_in_progress);
3224 down(&filter->fo_create_locks[group]);
3225 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3226 CERROR("%s: dropping old precreate request\n",
3230 /* only precreate if group == 0 and o_id is specfied */
3231 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3234 diff = oa->o_id - filter_last_id(filter, group);
3235 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3236 filter_last_id(filter, group), diff);
3238 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3239 oa->o_id, filter_last_id(filter, group), diff);
3243 oa->o_id = filter_last_id(&obd->u.filter, group);
3244 rc = filter_precreate(obd, oa, group, &diff);
3245 oa->o_id = filter_last_id(&obd->u.filter, group);
3247 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3250 /* else diff == 0 */
3253 up(&filter->fo_create_locks[group]);
3257 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3260 struct filter_obd *filter = &obd->u.filter;
3261 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3265 /* at least try to account for cached pages. its still racey and
3266 * might be under-reporting if clients haven't announced their
3267 * caches with brw recently */
3268 spin_lock(&obd->obd_osfs_lock);
3269 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3270 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3271 spin_unlock(&obd->obd_osfs_lock);
3273 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3274 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3275 filter->fo_tot_dirty, filter->fo_tot_granted,
3276 filter->fo_tot_pending,
3277 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3279 filter_grant_sanity_check(obd, __FUNCTION__);
3281 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3282 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3283 osfs->os_bsize - 1) >> blockbits));
3285 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3286 * stop creating files on MDS if OST is not good shape to create
3288 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3293 static int filter_use_existing_obj(struct obd_device *obd,
3294 struct dentry *dchild, void **handle,
3297 struct inode *inode = dchild->d_inode;
3301 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3304 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3305 if (IS_ERR(*handle))
3306 return PTR_ERR(*handle);
3308 iattr.ia_valid = ATTR_MODE;
3309 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3310 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3318 /* We rely on the fact that only one thread will be creating files in a given
3319 * group at a time, which is why we don't need an atomic filter_get_new_id.
3320 * Even if we had that atomic function, the following race would exist:
3322 * thread 1: gets id x from filter_next_id
3323 * thread 2: gets id (x + 1) from filter_next_id
3324 * thread 2: creates object (x + 1)
3325 * thread 1: tries to create object x, gets -ENOSPC
3327 * Caller must hold fo_create_locks[group]
3329 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3330 obd_gr group, int *num)
3332 struct dentry *dchild = NULL, *dparent = NULL;
3333 struct filter_obd *filter;
3334 struct obd_statfs *osfs;
3335 int err = 0, rc = 0, recreate_obj = 0, i;
3336 unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3338 void *handle = NULL;
3341 filter = &obd->u.filter;
3343 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3345 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3347 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3348 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3351 OBD_ALLOC(osfs, sizeof(*osfs));
3354 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3355 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3356 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3357 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3358 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3362 OBD_FREE(osfs, sizeof(*osfs));
3367 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3368 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3370 for (i = 0; i < *num && err == 0; i++) {
3371 int cleanup_phase = 0;
3373 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3374 CWARN("%s: create aborted by destroy\n",
3383 last_id = filter_last_id(filter, group);
3384 if (next_id > last_id) {
3385 CERROR("Error: Trying to recreate obj greater"
3386 "than last id "LPD64" > "LPD64"\n",
3388 GOTO(cleanup, rc = -EINVAL);
3391 next_id = filter_last_id(filter, group) + 1;
3393 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3395 dparent = filter_parent_lock(obd, group, next_id);
3396 if (IS_ERR(dparent))
3397 GOTO(cleanup, rc = PTR_ERR(dparent));
3398 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3400 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3402 GOTO(cleanup, rc = PTR_ERR(dchild));
3403 cleanup_phase = 2; /* f_dput(dchild) */
3405 if (dchild->d_inode != NULL) {
3406 /* This would only happen if lastobjid was bad on disk*/
3407 /* Could also happen if recreating missing obj but it
3408 * already exists. */
3410 CERROR("%s: recreating existing object %.*s?\n",
3411 obd->obd_name, dchild->d_name.len,
3412 dchild->d_name.name);
3414 /* Use these existing objects if they are
3416 if (dchild->d_inode->i_size == 0) {
3417 rc = filter_use_existing_obj(obd,dchild,
3418 &handle, &cleanup_phase);
3425 CERROR("%s: Serious error: objid %.*s already "
3426 "exists; is this filesystem corrupt?\n",
3427 obd->obd_name, dchild->d_name.len,
3428 dchild->d_name.name);
3431 GOTO(cleanup, rc = -EEXIST);
3434 handle = fsfilt_start_log(obd, dparent->d_inode,
3435 FSFILT_OP_CREATE, NULL, 1);
3437 GOTO(cleanup, rc = PTR_ERR(handle));
3440 /* We mark object SUID+SGID to flag it for accepting UID+GID
3441 * from client on first write. Currently the permission bits
3442 * on the OST are never used, so this is OK. */
3443 rc = ll_vfs_create(dparent->d_inode, dchild,
3444 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3446 CERROR("create failed rc = %d\n", rc);
3451 if (!recreate_obj) {
3452 filter_set_last_id(filter, next_id, group);
3453 err = filter_update_last_objid(obd, group, 0);
3455 CERROR("unable to write lastobjid "
3456 "but file created\n");
3460 switch(cleanup_phase) {
3462 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3464 CERROR("error on commit, err = %d\n", err);
3471 filter_parent_unlock(dparent);
3478 if (time_after(jiffies, enough_time)) {
3480 "%s: precreate slow - want %d got %d \n",
3481 obd->obd_name, *num, i);
3488 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3489 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3494 static int filter_create(struct obd_export *exp, struct obdo *oa,
3495 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3497 struct filter_export_data *fed;
3498 struct obd_device *obd = NULL;
3499 struct filter_obd *filter;
3500 struct lvfs_run_ctxt saved;
3501 struct lov_stripe_md *lsm = NULL;
3502 int rc = 0, diff, group = oa->o_gr;
3505 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3506 CERROR("!!! nid %s sent invalid object group %d\n",
3507 obd_export_nid2str(exp), group);
3512 fed = &exp->exp_filter_data;
3513 filter = &obd->u.filter;
3515 if (fed->fed_group != group) {
3516 CERROR("!!! this export (nid %s) used object group %d "
3517 "earlier; now it's trying to use group %d! This could "
3518 "be a bug in the MDS. Tell CFS.\n",
3519 obd_export_nid2str(exp), fed->fed_group, group);
3523 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3524 oa->o_gr, oa->o_id);
3528 rc = obd_alloc_memmd(exp, &lsm);
3535 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3537 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3538 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3539 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3540 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3541 oa->o_id, filter_last_id(filter,
3546 down(&filter->fo_create_locks[oa->o_gr]);
3547 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3548 up(&filter->fo_create_locks[oa->o_gr]);
3551 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3554 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3555 if (rc && ea != NULL && *ea != lsm) {
3556 obd_free_memmd(exp, &lsm);
3557 } else if (rc == 0 && ea != NULL) {
3558 /* XXX LOV STACKING: the lsm that is passed to us from
3559 * LOV does not have valid lsm_oinfo data structs, so
3560 * don't go touching that. This needs to be fixed in a
3562 lsm->lsm_object_id = oa->o_id;
3569 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3570 struct lov_stripe_md *md, struct obd_trans_info *oti,
3571 struct obd_export *md_exp)
3573 unsigned int qcids[MAXQUOTAS] = {0, 0};
3574 struct obd_device *obd;
3575 struct filter_obd *filter;
3576 struct dentry *dchild = NULL, *dparent = NULL;
3577 struct lvfs_run_ctxt saved;
3578 void *handle = NULL;
3579 struct llog_cookie *fcc = NULL;
3580 int rc, rc2, cleanup_phase = 0, sync = 0;
3584 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3587 filter = &obd->u.filter;
3589 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3592 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3594 GOTO(cleanup, rc = PTR_ERR(dchild));
3597 if (dchild->d_inode == NULL) {
3598 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3600 /* If object already gone, cancel cookie right now */
3601 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3602 fcc = obdo_logcookie(oa);
3603 llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3605 fcc = NULL; /* we didn't allocate fcc, don't free it */
3607 GOTO(cleanup, rc = -ENOENT);
3610 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3612 /* Our MDC connection is established by the MDS to us */
3613 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3614 OBD_ALLOC(fcc, sizeof(*fcc));
3616 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3618 DQUOT_INIT(dchild->d_inode);
3620 /* we're gonna truncate it first in order to avoid possible deadlock:
3622 * open trasaction open transaction
3623 * down(i_zombie) down(i_zombie)
3624 * restart transaction
3625 * (see BUG 4180) -bzzz
3627 LOCK_INODE_MUTEX(dchild->d_inode);
3628 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3630 if (IS_ERR(handle)) {
3631 UNLOCK_INODE_MUTEX(dchild->d_inode);
3632 GOTO(cleanup, rc = PTR_ERR(handle));
3635 iattr.ia_valid = ATTR_SIZE;
3637 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3638 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3639 UNLOCK_INODE_MUTEX(dchild->d_inode);
3643 GOTO(cleanup, rc = rc2);
3645 /* We don't actually need to lock the parent until we are unlinking
3646 * here, and not while truncating above. That avoids holding the
3647 * parent lock for a long time during truncate, which can block other
3648 * threads from doing anything to objects in that directory. bug 7171 */
3649 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3650 if (IS_ERR(dparent))
3651 GOTO(cleanup, rc = PTR_ERR(dparent));
3652 cleanup_phase = 3; /* filter_parent_unlock */
3654 LOCK_INODE_MUTEX(dchild->d_inode);
3655 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3656 if (IS_ERR(handle)) {
3657 UNLOCK_INODE_MUTEX(dchild->d_inode);
3658 GOTO(cleanup, rc = PTR_ERR(handle));
3660 cleanup_phase = 4; /* fsfilt_commit */
3662 /* Quota release need uid/gid of inode */
3663 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3665 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3667 /* this drops dchild->d_inode->i_mutex unconditionally */
3668 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3672 switch(cleanup_phase) {
3675 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3676 oti->oti_handle : handle,
3677 filter_cancel_cookies_cb,
3679 /* If add_journal_cb failed, then filter_finish_transno
3680 * will commit the handle and we will do a sync
3681 * on commit. then we call callback directly to free
3684 rc = filter_finish_transno(exp, oti, rc, sync);
3686 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3689 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3691 CERROR("error on commit, err = %d\n", rc2);
3698 filter_parent_unlock(dparent);
3702 OBD_FREE(fcc, sizeof(*fcc));
3704 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3707 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3711 /* trigger quota release */
3712 qcids[USRQUOTA] = oa->o_uid;
3713 qcids[GRPQUOTA] = oa->o_gid;
3714 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3718 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3722 /* NB start and end are used for punch, but not truncate */
3723 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3724 struct obd_trans_info *oti,
3725 struct ptlrpc_request_set *rqset)
3730 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3731 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3732 oinfo->oi_policy.l_extent.end);
3736 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3737 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3738 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3740 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3741 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3745 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3746 rc = filter_setattr(exp, oinfo, oti);
3750 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3751 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3754 struct lvfs_run_ctxt saved;
3755 struct filter_obd *filter;
3756 struct dentry *dentry;
3757 struct llog_ctxt *ctxt;
3761 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3762 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3766 filter = &exp->exp_obd->u.filter;
3768 /* an objid of zero is taken to mean "sync whole filesystem" */
3769 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3770 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3771 /* flush any remaining cancel messages out to the target */
3772 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3773 llog_sync(ctxt, exp);
3777 dentry = filter_oa2dentry(exp->exp_obd, oa);
3779 RETURN(PTR_ERR(dentry));
3781 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3783 LOCK_INODE_MUTEX(dentry->d_inode);
3785 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3787 /* just any file to grab fsync method - "file" arg unused */
3788 struct file *file = filter->fo_rcvd_filp;
3790 if (file->f_op && file->f_op->fsync)
3791 rc = file->f_op->fsync(NULL, dentry, 1);
3793 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3797 UNLOCK_INODE_MUTEX(dentry->d_inode);
3799 oa->o_valid = OBD_MD_FLID;
3800 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3802 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3808 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3809 void *key, __u32 *vallen, void *val)
3811 struct obd_device *obd;
3814 obd = class_exp2obd(exp);
3816 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3820 if (keylen == strlen("blocksize") &&
3821 memcmp(key, "blocksize", keylen) == 0) {
3822 __u32 *blocksize = val;
3823 *vallen = sizeof(*blocksize);
3824 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3828 if (keylen == strlen("blocksize_bits") &&
3829 memcmp(key, "blocksize_bits", keylen) == 0) {
3830 __u32 *blocksize_bits = val;
3831 *vallen = sizeof(*blocksize_bits);
3832 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3836 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3837 obd_id *last_id = val;
3838 /* FIXME: object groups */
3839 *last_id = filter_last_id(&obd->u.filter, 0);
3842 CDEBUG(D_IOCTL, "invalid key\n");
3846 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3847 void *key, __u32 vallen, void *val,
3848 struct ptlrpc_request_set *set)
3850 struct obd_device *obd;
3851 struct obd_llogs *llog;
3852 struct llog_ctxt *ctxt;
3858 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3862 if (KEY_IS(KEY_CAPA_KEY)) {
3863 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3865 CERROR("filter update capability key failed: %d\n", rc);
3869 if (KEY_IS(KEY_REVIMP_UPD)) {
3870 filter_revimp_update(exp);
3874 if (keylen < strlen(KEY_MDS_CONN) ||
3875 memcmp(key, KEY_MDS_CONN, keylen) != 0)
3878 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3879 obd_export_nid2str(exp));
3880 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3882 /* setup llog imports */
3883 LASSERT(val != NULL);
3884 group = (int)(*(__u32 *)val);
3885 LASSERT(group >= FILTER_GROUP_MDS0);
3887 llog = filter_grab_llog_for_group(obd, group, exp);
3888 LASSERT(llog != NULL);
3889 ctxt = llog_get_context_from_llogs(llog, LLOG_MDS_OST_REPL_CTXT);
3890 LASSERTF(ctxt != NULL, "ctxt is not null\n"),
3892 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3894 lquota_setinfo(filter_quota_interface_ref, exp, obd);
3899 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3900 int len, void *karg, void *uarg)
3902 struct obd_device *obd = exp->exp_obd;
3903 struct obd_ioctl_data *data = karg;
3907 case OBD_IOC_ABORT_RECOVERY: {
3908 CERROR("aborting recovery for device %s\n", obd->obd_name);
3909 target_stop_recovery_thread(obd);
3913 case OBD_IOC_SYNC: {
3914 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
3915 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3919 case OBD_IOC_SET_READONLY: {
3921 struct super_block *sb = obd->u.obt.obt_sb;
3922 struct inode *inode = sb->s_root->d_inode;
3923 BDEVNAME_DECLARE_STORAGE(tmp);
3924 CERROR("*** setting device %s read-only ***\n",
3925 ll_bdevname(sb, tmp));
3927 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3928 if (!IS_ERR(handle))
3929 rc = fsfilt_commit(obd, inode, handle, 1);
3931 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3932 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3934 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3938 case OBD_IOC_CATLOGLIST: {
3939 rc = llog_catalog_list(obd, 1, data);
3943 case OBD_IOC_LLOG_CANCEL:
3944 case OBD_IOC_LLOG_REMOVE:
3945 case OBD_IOC_LLOG_INFO:
3946 case OBD_IOC_LLOG_PRINT: {
3947 /* FIXME to be finished */
3948 RETURN(-EOPNOTSUPP);
3950 struct llog_ctxt *ctxt = NULL;
3952 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3953 rc = llog_ioctl(ctxt, cmd, data);
3954 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3967 static int filter_health_check(struct obd_device *obd)
3969 #ifdef USE_HEALTH_CHECK_WRITE
3970 struct filter_obd *filter = &obd->u.filter;
3975 * health_check to return 0 on healthy
3976 * and 1 on unhealthy.
3978 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3981 #ifdef USE_HEALTH_CHECK_WRITE
3982 LASSERT(filter->fo_health_check_filp != NULL);
3983 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3988 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3991 return filter_fid2dentry(data, NULL, gr, id);
3994 static int filter_process_config(struct obd_device *obd, obd_count len,
3997 struct lustre_cfg *lcfg = buf;
3998 struct lprocfs_static_vars lvars;
4001 lprocfs_init_vars(filter, &lvars);
4003 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
4007 static struct lvfs_callback_ops filter_lvfs_ops = {
4008 l_fid2dentry: filter_lvfs_fid2dentry,
4011 static struct obd_ops filter_obd_ops = {
4012 .o_owner = THIS_MODULE,
4013 .o_get_info = filter_get_info,
4014 .o_set_info_async = filter_set_info_async,
4015 .o_setup = filter_setup,
4016 .o_precleanup = filter_precleanup,
4017 .o_cleanup = filter_cleanup,
4018 .o_connect = filter_connect,
4019 .o_reconnect = filter_reconnect,
4020 .o_disconnect = filter_disconnect,
4021 .o_ping = filter_ping,
4022 .o_init_export = filter_init_export,
4023 .o_destroy_export = filter_destroy_export,
4024 .o_statfs = filter_statfs,
4025 .o_getattr = filter_getattr,
4026 .o_unpackmd = filter_unpackmd,
4027 .o_create = filter_create,
4028 .o_setattr = filter_setattr,
4029 .o_destroy = filter_destroy,
4030 .o_brw = filter_brw,
4031 .o_punch = filter_truncate,
4032 .o_sync = filter_sync,
4033 .o_preprw = filter_preprw,
4034 .o_commitrw = filter_commitrw,
4035 .o_llog_init = filter_llog_init,
4036 .o_llog_connect = filter_llog_connect,
4037 .o_llog_finish = filter_llog_finish,
4038 .o_iocontrol = filter_iocontrol,
4039 .o_health_check = filter_health_check,
4040 .o_process_config = filter_process_config,
4043 quota_interface_t *filter_quota_interface_ref;
4044 extern quota_interface_t filter_quota_interface;
4046 static int __init obdfilter_init(void)
4048 struct lprocfs_static_vars lvars;
4051 lprocfs_init_vars(filter, &lvars);
4053 request_module("lquota");
4054 OBD_ALLOC(obdfilter_created_scratchpad,
4055 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4056 sizeof(*obdfilter_created_scratchpad));
4057 if (obdfilter_created_scratchpad == NULL)
4060 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4061 sizeof(struct filter_mod_data),
4064 GOTO(out, rc = -ENOMEM);
4066 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4067 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4069 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4070 LUSTRE_OST_NAME, NULL);
4074 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4075 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4076 ll_fmd_cachep = NULL;
4078 if (filter_quota_interface_ref)
4079 PORTAL_SYMBOL_PUT(filter_quota_interface);
4081 OBD_FREE(obdfilter_created_scratchpad,
4082 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4083 sizeof(*obdfilter_created_scratchpad));
4089 static void __exit obdfilter_exit(void)
4091 if (filter_quota_interface_ref)
4092 PORTAL_SYMBOL_PUT(filter_quota_interface);
4094 if (ll_fmd_cachep) {
4095 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4096 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4097 ll_fmd_cachep = NULL;
4100 class_unregister_type(LUSTRE_OST_NAME);
4101 OBD_FREE(obdfilter_created_scratchpad,
4102 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4103 sizeof(*obdfilter_created_scratchpad));
4106 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4107 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4108 MODULE_LICENSE("GPL");
4110 module_init(obdfilter_init);
4111 module_exit(obdfilter_exit);