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 #include <linux/config.h>
39 #include <linux/module.h>
41 #include <linux/dcache.h>
42 #include <linux/init.h>
43 #include <linux/version.h>
44 #include <linux/sched.h>
45 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
46 # include <linux/mount.h>
47 # 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 <lustre_ver.h>
61 #include <linux/slab.h>
62 #include <lustre_param.h>
64 #include "filter_internal.h"
66 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
67 #define FILTER_MIN_GROUPS 3
68 static struct lvfs_callback_ops filter_lvfs_ops;
69 kmem_cache_t *ll_fmd_cachep;
71 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
72 void *cb_data, int error)
74 obd_transno_commit_cb(obd, transno, error);
77 /* Assumes caller has already pushed us into the kernel context. */
78 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
81 struct filter_obd *filter = &exp->exp_obd->u.filter;
82 struct filter_export_data *fed = &exp->exp_filter_data;
83 struct filter_client_data *fcd = fed->fed_fcd;
86 int err, log_pri = D_HA;
88 /* Propagate error code. */
92 if (!exp->exp_obd->obd_replayable || oti == NULL)
95 /* we don't allocate new transnos for replayed requests */
96 if (oti->oti_transno == 0) {
97 spin_lock(&filter->fo_translock);
98 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
99 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
100 spin_unlock(&filter->fo_translock);
101 oti->oti_transno = last_rcvd;
103 spin_lock(&filter->fo_translock);
104 last_rcvd = oti->oti_transno;
105 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
106 filter->fo_fsd->lsd_last_transno =
107 cpu_to_le64(last_rcvd);
108 spin_unlock(&filter->fo_translock);
110 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
112 /* could get xid from oti, if it's ever needed */
113 fcd->fcd_last_xid = 0;
115 off = fed->fed_lr_off;
117 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
118 fed->fed_lr_idx, fed->fed_lr_off);
121 fsfilt_add_journal_cb(exp->exp_obd, last_rcvd, oti->oti_handle,
122 filter_commit_cb, NULL);
123 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
124 fcd, sizeof(*fcd), &off, 0);
132 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
133 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
138 void f_dput(struct dentry *dentry)
140 /* Can't go inside filter_ddelete because it can block */
141 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
142 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
143 LASSERT(atomic_read(&dentry->d_count) > 0);
148 /* Add client data to the FILTER. We use a bitmap to locate a free space
149 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
150 * Otherwise, we have just read the data from the last_rcvd file and
151 * we know its offset. */
152 static int filter_client_add(struct obd_device *obd, struct filter_obd *filter,
153 struct filter_export_data *fed, int cl_idx)
155 unsigned long *bitmap = filter->fo_last_rcvd_slots;
156 int new_client = (cl_idx == -1);
159 LASSERT(bitmap != NULL);
160 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
162 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
163 if (!strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid))
166 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
167 * there's no need for extra complication here
170 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
172 if (cl_idx >= LR_MAX_CLIENTS) {
173 CERROR("no client slots - fix LR_MAX_CLIENTS\n");
176 if (test_and_set_bit(cl_idx, bitmap)) {
177 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
182 if (test_and_set_bit(cl_idx, bitmap)) {
183 CERROR("FILTER client %d: bit already set in bitmap!\n",
189 fed->fed_lr_idx = cl_idx;
190 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
191 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
192 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
194 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
195 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
198 struct lvfs_run_ctxt saved;
199 loff_t off = fed->fed_lr_off;
203 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
204 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
206 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
207 /* Transaction needed to fix bug 1403 */
208 handle = fsfilt_start(obd,
209 filter->fo_rcvd_filp->f_dentry->d_inode,
210 FSFILT_OP_SETATTR, NULL);
211 if (IS_ERR(handle)) {
212 err = PTR_ERR(handle);
213 CERROR("unable to start transaction: rc %d\n", err);
215 err = fsfilt_write_record(obd, filter->fo_rcvd_filp,
217 sizeof(*fed->fed_fcd),
220 filter->fo_rcvd_filp->f_dentry->d_inode,
223 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
226 CERROR("error writing %s client idx %u: rc %d\n",
227 LAST_RCVD, fed->fed_lr_idx, err);
234 static int filter_client_free(struct obd_export *exp)
236 struct filter_export_data *fed = &exp->exp_filter_data;
237 struct filter_obd *filter = &exp->exp_obd->u.filter;
238 struct obd_device *obd = exp->exp_obd;
239 struct filter_client_data zero_fcd;
240 struct lvfs_run_ctxt saved;
245 if (fed->fed_fcd == NULL)
248 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
249 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
252 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
253 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
255 LASSERT(filter->fo_last_rcvd_slots != NULL);
257 off = fed->fed_lr_off;
259 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
260 * we leak a client slot that will be cleaned on the next recovery. */
262 CERROR("%s: client idx %d has med_off %lld\n",
263 obd->obd_name, fed->fed_lr_idx, off);
264 GOTO(free, rc = -EINVAL);
267 /* Clear the bit _after_ zeroing out the client so we don't
268 race with filter_client_add and zero out new clients.*/
269 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
270 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
275 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
276 memset(&zero_fcd, 0, sizeof zero_fcd);
277 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
278 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
279 sizeof(zero_fcd), &off, 0);
282 /* update server's transno */
283 filter_update_server_data(obd, filter->fo_rcvd_filp,
285 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
287 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
288 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
289 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
293 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
294 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
301 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
307 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
308 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
309 struct filter_mod_data *fmd)
311 LASSERT_SPIN_LOCKED(&fed->fed_lock);
312 if (--fmd->fmd_refcount == 0) {
313 /* XXX when we have persistent reservations and the handle
314 * is stored herein we need to drop it here. */
315 fed->fed_mod_count--;
316 list_del(&fmd->fmd_list);
317 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
321 /* drop fmd reference, free it if last ref */
322 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
324 struct filter_export_data *fed;
329 fed = &exp->exp_filter_data;
330 spin_lock(&fed->fed_lock);
331 filter_fmd_put_nolock(fed, fmd); /* caller reference */
332 spin_unlock(&fed->fed_lock);
335 /* expire entries from the end of the list if there are too many
336 * or they are too old */
337 static void filter_fmd_expire_nolock(struct filter_obd *filter,
338 struct filter_export_data *fed,
339 struct filter_mod_data *keep)
341 struct filter_mod_data *fmd, *tmp;
343 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
347 if (time_before(jiffies, fmd->fmd_expire) &&
348 fed->fed_mod_count < filter->fo_fmd_max_num)
351 list_del_init(&fmd->fmd_list);
352 filter_fmd_put_nolock(fed, fmd); /* list reference */
356 void filter_fmd_expire(struct obd_export *exp)
358 spin_lock(&exp->exp_filter_data.fed_lock);
359 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
360 &exp->exp_filter_data, NULL);
361 spin_unlock(&exp->exp_filter_data.fed_lock);
364 /* find specified objid, group in export fmd list.
365 * caller must hold fed_lock and take fmd reference itself */
366 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
367 struct filter_export_data *fed,
368 obd_id objid, obd_gr group)
370 struct filter_mod_data *found = NULL, *fmd;
372 LASSERT_SPIN_LOCKED(&fed->fed_lock);
374 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
375 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
377 list_del(&fmd->fmd_list);
378 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
379 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
384 filter_fmd_expire_nolock(filter, fed, found);
389 /* Find fmd based on objid and group, or return NULL if not found. */
390 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
391 obd_id objid, obd_gr group)
393 struct filter_mod_data *fmd;
395 spin_lock(&exp->exp_filter_data.fed_lock);
396 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
397 &exp->exp_filter_data, objid, group);
399 fmd->fmd_refcount++; /* caller reference */
400 spin_unlock(&exp->exp_filter_data.fed_lock);
405 /* Find fmd based on objid and group, or create a new one if none is found.
406 * It is possible for this function to return NULL under memory pressure,
407 * or if objid = 0 is passed (which will only cause old entries to expire).
408 * Currently this is not fatal because any fmd state is transient and
409 * may also be freed when it gets sufficiently old. */
410 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
411 obd_id objid, obd_gr group)
413 struct filter_export_data *fed = &exp->exp_filter_data;
414 struct filter_mod_data *found = NULL, *fmd_new = NULL;
416 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, SLAB_NOFS, sizeof(*fmd_new));
418 spin_lock(&fed->fed_lock);
419 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
422 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
423 fmd_new->fmd_id = objid;
424 fmd_new->fmd_gr = group;
425 fmd_new->fmd_refcount++; /* list reference */
427 fed->fed_mod_count++;
429 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
433 found->fmd_refcount++; /* caller reference */
434 found->fmd_expire = jiffies +
435 exp->exp_obd->u.filter.fo_fmd_max_age;
438 spin_unlock(&fed->fed_lock);
444 /* drop fmd list reference so it will disappear when last reference is put.
445 * This isn't so critical because it would in fact only affect the one client
446 * that is doing the unlink and at worst we have an stale entry referencing
447 * an object that should never be used again. */
448 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
450 struct filter_mod_data *found = NULL;
452 spin_lock(&exp->exp_filter_data.fed_lock);
453 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
455 list_del_init(&found->fmd_list);
456 filter_fmd_put_nolock(&exp->exp_filter_data, found);
458 spin_unlock(&exp->exp_filter_data.fed_lock);
461 #define filter_fmd_drop(exp, objid, group)
464 /* remove all entries from fmd list */
465 static void filter_fmd_cleanup(struct obd_export *exp)
467 struct filter_export_data *fed = &exp->exp_filter_data;
468 struct filter_mod_data *fmd = NULL, *tmp;
470 spin_lock(&fed->fed_lock);
471 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
472 list_del_init(&fmd->fmd_list);
473 filter_fmd_put_nolock(fed, fmd);
475 spin_unlock(&fed->fed_lock);
478 static int filter_init_export(struct obd_export *exp)
480 spin_lock_init(&exp->exp_filter_data.fed_lock);
481 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
482 exp->exp_connecting = 1;
487 static int filter_free_server_data(struct filter_obd *filter)
489 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
490 filter->fo_fsd = NULL;
491 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
492 filter->fo_last_rcvd_slots = NULL;
496 /* assumes caller is already in kernel ctxt */
497 int filter_update_server_data(struct obd_device *obd, struct file *filp,
498 struct lr_server_data *fsd, int force_sync)
504 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
505 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
506 le64_to_cpu(fsd->lsd_last_transno));
507 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
508 le64_to_cpu(fsd->lsd_mount_count));
510 fsd->lsd_compat14 = fsd->lsd_last_transno;
511 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off,force_sync);
513 CERROR("error writing lr_server_data: rc = %d\n", rc);
518 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
521 struct filter_obd *filter = &obd->u.filter;
527 if (filter->fo_last_objid_files[group] == NULL) {
528 CERROR("Object group "LPU64" not fully setup; not updating "
529 "last_objid\n", group);
533 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
534 obd->obd_name, group, filter->fo_last_objids[group]);
536 tmp = cpu_to_le64(filter->fo_last_objids[group]);
537 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
538 &tmp, sizeof(tmp), &off, force_sync);
540 CERROR("error writing group "LPU64" last objid: rc = %d\n",
545 /* assumes caller has already in kernel ctxt */
546 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
548 struct filter_obd *filter = &obd->u.filter;
549 struct lr_server_data *fsd;
550 struct filter_client_data *fcd = NULL;
551 struct inode *inode = filp->f_dentry->d_inode;
552 unsigned long last_rcvd_size = inode->i_size;
558 /* ensure padding in the struct is the correct size */
559 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
560 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
561 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
562 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
564 OBD_ALLOC(fsd, sizeof(*fsd));
567 filter->fo_fsd = fsd;
569 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
570 if (filter->fo_last_rcvd_slots == NULL) {
571 OBD_FREE(fsd, sizeof(*fsd));
575 if (last_rcvd_size == 0) {
576 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
578 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
579 fsd->lsd_last_transno = 0;
580 mount_count = fsd->lsd_mount_count = 0;
581 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
582 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
583 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
584 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
585 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
586 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
588 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
590 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
594 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
595 LCONSOLE_ERROR("Trying to start OBD %s using the wrong"
596 " disk %s. Were the /dev/ assignments "
598 obd->obd_uuid.uuid, fsd->lsd_uuid);
599 GOTO(err_fsd, rc = -EINVAL);
601 mount_count = le64_to_cpu(fsd->lsd_mount_count);
602 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
604 /* Assume old last_rcvd format unless I_C_LR is set */
605 if (!(fsd->lsd_feature_incompat &
606 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
607 fsd->lsd_last_transno = fsd->lsd_compat14;
611 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
612 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
613 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
614 ~FILTER_INCOMPAT_SUPP);
615 GOTO(err_fsd, rc = -EINVAL);
617 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
618 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
619 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
620 ~FILTER_ROCOMPAT_SUPP);
621 /* Do something like remount filesystem read-only */
622 GOTO(err_fsd, rc = -EINVAL);
625 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
626 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
627 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
628 obd->obd_name, mount_count + 1);
629 CDEBUG(D_INODE, "%s: server data size: %u\n",
630 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
631 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
632 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
633 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
634 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
635 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
636 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
637 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
638 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
639 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
640 le16_to_cpu(fsd->lsd_client_size));
642 if (!obd->obd_replayable) {
643 CWARN("%s: recovery support OFF\n", obd->obd_name);
647 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
648 off < last_rcvd_size; cl_idx++) {
650 struct obd_export *exp;
651 struct filter_export_data *fed;
654 OBD_ALLOC(fcd, sizeof(*fcd));
656 GOTO(err_client, rc = -ENOMEM);
659 /* Don't assume off is incremented properly by
660 * fsfilt_read_record(), in case sizeof(*fcd)
661 * isn't the same as fsd->lsd_client_size. */
662 off = le32_to_cpu(fsd->lsd_client_start) +
663 cl_idx * le16_to_cpu(fsd->lsd_client_size);
664 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
666 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
667 LAST_RCVD, cl_idx, off, rc);
668 break; /* read error shouldn't cause startup to fail */
671 if (fcd->fcd_uuid[0] == '\0') {
672 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
677 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
679 /* These exports are cleaned up by filter_disconnect(), so they
680 * need to be set up like real exports as filter_connect() does.
682 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
684 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
685 " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
686 last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
687 le32_to_cpu(fcd->fcd_group));
689 if (PTR_ERR(exp) == -EALREADY) {
690 /* export already exists, zero out this one */
691 CERROR("Zeroing out duplicate export due to "
693 fcd->fcd_uuid[0] = '\0';
695 GOTO(err_client, rc = PTR_ERR(exp));
698 fed = &exp->exp_filter_data;
700 fed->fed_group = le32_to_cpu(fcd->fcd_group);
701 rc = filter_client_add(obd, filter, fed, cl_idx);
702 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
705 exp->exp_replay_needed = 1;
706 exp->exp_connecting = 0;
707 obd->obd_recoverable_clients++;
708 obd->obd_max_recoverable_clients++;
709 class_export_put(exp);
712 /* Need to check last_rcvd even for duplicated exports. */
713 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
716 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
717 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
721 OBD_FREE(fcd, sizeof(*fcd));
723 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
725 if (obd->obd_recoverable_clients) {
726 CWARN("RECOVERY: service %s, %d recoverable clients, "
727 "last_rcvd "LPU64"\n", obd->obd_name,
728 obd->obd_recoverable_clients,
729 le64_to_cpu(fsd->lsd_last_transno));
730 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
731 obd->obd_recovering = 1;
732 obd->obd_recovery_start = CURRENT_SECONDS;
733 /* Only used for lprocfs_status */
734 obd->obd_recovery_end = obd->obd_recovery_start +
735 OBD_RECOVERY_TIMEOUT;
739 filter->fo_mount_count = mount_count + 1;
740 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
742 /* save it, so mount count and last_transno is current */
743 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
745 GOTO(err_client, rc);
750 class_disconnect_exports(obd);
752 filter_free_server_data(filter);
756 static int filter_cleanup_groups(struct obd_device *obd)
758 struct filter_obd *filter = &obd->u.filter;
760 struct dentry *dentry;
764 if (filter->fo_dentry_O_groups != NULL) {
765 for (i = 0; i < filter->fo_group_count; i++) {
766 dentry = filter->fo_dentry_O_groups[i];
770 OBD_FREE(filter->fo_dentry_O_groups,
771 filter->fo_group_count *
772 sizeof(*filter->fo_dentry_O_groups));
773 filter->fo_dentry_O_groups = NULL;
775 if (filter->fo_last_objid_files != NULL) {
776 for (i = 0; i < filter->fo_group_count; i++) {
777 filp = filter->fo_last_objid_files[i];
781 OBD_FREE(filter->fo_last_objid_files,
782 filter->fo_group_count *
783 sizeof(*filter->fo_last_objid_files));
784 filter->fo_last_objid_files = NULL;
786 if (filter->fo_dentry_O_sub != NULL) {
787 for (i = 0; i < filter->fo_group_count; i++) {
788 for (j = 0; j < filter->fo_subdir_count; j++) {
789 dentry = filter->fo_dentry_O_sub[i].dentry[j];
794 OBD_FREE(filter->fo_dentry_O_sub,
795 filter->fo_group_count *
796 sizeof(*filter->fo_dentry_O_sub));
797 filter->fo_dentry_O_sub = NULL;
799 if (filter->fo_last_objids != NULL) {
800 OBD_FREE(filter->fo_last_objids,
801 filter->fo_group_count *
802 sizeof(*filter->fo_last_objids));
803 filter->fo_last_objids = NULL;
805 if (filter->fo_dentry_O != NULL) {
806 f_dput(filter->fo_dentry_O);
807 filter->fo_dentry_O = NULL;
812 static int filter_update_last_group(struct obd_device *obd, int group)
814 struct filter_obd *filter = &obd->u.filter;
815 struct file *filp = NULL;
816 int last_group = 0, rc;
820 if (group <= filter->fo_committed_group)
823 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
827 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
831 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
833 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
836 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
837 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
838 obd->obd_name, last_group, group);
842 /* must be sync: bXXXX */
843 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
845 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
849 filter->fo_committed_group = group;
856 static int filter_read_group_internal(struct obd_device *obd, int group,
859 struct filter_obd *filter = &obd->u.filter;
860 __u64 *new_objids = NULL;
861 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
862 struct dentry **new_groups = NULL;
863 struct file **new_files = NULL;
864 struct dentry *dentry;
866 int old_count = filter->fo_group_count, rc, stage = 0, i;
872 snprintf(name, 24, "%d", group);
876 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
878 if (IS_ERR(dentry)) {
879 CERROR("Cannot lookup expected object group %d: %ld\n",
880 group, PTR_ERR(dentry));
881 RETURN(PTR_ERR(dentry));
884 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
885 if (IS_ERR(dentry)) {
886 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
888 RETURN(PTR_ERR(dentry));
893 snprintf(name, 24, "O/%d/LAST_ID", group);
895 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
897 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
898 GOTO(cleanup, rc = PTR_ERR(filp));
902 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
904 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
908 if (filter->fo_subdir_count) {
909 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
910 if (tmp_subdirs == NULL)
911 GOTO(cleanup, rc = -ENOMEM);
914 for (i = 0; i < filter->fo_subdir_count; i++) {
916 snprintf(dir, sizeof(dir), "d%u", i);
918 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
919 if (IS_ERR(tmp_subdirs->dentry[i])) {
920 rc = PTR_ERR(tmp_subdirs->dentry[i]);
921 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
926 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
927 tmp_subdirs->dentry[i]);
931 /* 'group' is an index; we need an array of length 'group + 1' */
932 if (group + 1 > old_count) {
933 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
934 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
935 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
936 OBD_ALLOC(new_files, len * sizeof(*new_files));
938 if (new_objids == NULL || new_subdirs == NULL ||
939 new_groups == NULL || new_files == NULL)
940 GOTO(cleanup, rc = -ENOMEM);
943 memcpy(new_objids, filter->fo_last_objids,
944 old_count * sizeof(*new_objids));
945 memcpy(new_subdirs, filter->fo_dentry_O_sub,
946 old_count * sizeof(*new_subdirs));
947 memcpy(new_groups, filter->fo_dentry_O_groups,
948 old_count * sizeof(*new_groups));
949 memcpy(new_files, filter->fo_last_objid_files,
950 old_count * sizeof(*new_files));
952 OBD_FREE(filter->fo_last_objids,
953 old_count * sizeof(*new_objids));
954 OBD_FREE(filter->fo_dentry_O_sub,
955 old_count * sizeof(*new_subdirs));
956 OBD_FREE(filter->fo_dentry_O_groups,
957 old_count * sizeof(*new_groups));
958 OBD_FREE(filter->fo_last_objid_files,
959 old_count * sizeof(*new_files));
961 filter->fo_last_objids = new_objids;
962 filter->fo_dentry_O_sub = new_subdirs;
963 filter->fo_dentry_O_groups = new_groups;
964 filter->fo_last_objid_files = new_files;
965 filter->fo_group_count = len;
968 filter->fo_dentry_O_groups[group] = dentry;
969 filter->fo_last_objid_files[group] = filp;
970 if (filter->fo_subdir_count) {
971 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
972 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
975 filter_update_last_group(obd, group);
977 if (filp->f_dentry->d_inode->i_size == 0) {
978 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
979 rc = filter_update_last_objid(obd, group, 1);
983 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
984 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
985 obd->obd_name, group, last_objid);
990 if (new_objids != NULL)
991 OBD_FREE(new_objids, len * sizeof(*new_objids));
992 if (new_subdirs != NULL)
993 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
994 if (new_groups != NULL)
995 OBD_FREE(new_groups, len * sizeof(*new_groups));
996 if (new_files != NULL)
997 OBD_FREE(new_files, len * sizeof(*new_files));
999 if (filter->fo_subdir_count) {
1000 for (i = 0; i < filter->fo_subdir_count; i++) {
1001 if (tmp_subdirs->dentry[i] != NULL)
1002 dput(tmp_subdirs->dentry[i]);
1004 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1007 filp_close(filp, 0);
1014 static int filter_read_groups(struct obd_device *obd, int last_group,
1017 struct filter_obd *filter = &obd->u.filter;
1018 int old_count, group, rc = 0;
1020 down(&filter->fo_init_lock);
1021 old_count = filter->fo_group_count;
1022 for (group = old_count; group <= last_group; group++) {
1024 continue; /* no group zero */
1026 rc = filter_read_group_internal(obd, group, create);
1030 up(&filter->fo_init_lock);
1034 /* FIXME: object groups */
1035 static int filter_prep_groups(struct obd_device *obd)
1037 struct filter_obd *filter = &obd->u.filter;
1038 struct dentry *dentry, *O_dentry;
1040 int last_group, rc = 0, cleanup_phase = 0;
1044 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1045 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1046 if (IS_ERR(O_dentry)) {
1047 rc = PTR_ERR(O_dentry);
1048 CERROR("cannot open/create O: rc = %d\n", rc);
1051 filter->fo_dentry_O = O_dentry;
1052 cleanup_phase = 1; /* O_dentry */
1054 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1055 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1056 * some time post 1.0 when all old-style OSTs have converted along
1057 * with the init_objid hack. */
1058 dentry = ll_lookup_one_len("R", O_dentry, 1);
1060 GOTO(cleanup, rc = PTR_ERR(dentry));
1061 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1062 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1065 CWARN("converting OST to new object layout\n");
1066 if (IS_ERR(O0_dentry)) {
1067 rc = PTR_ERR(O0_dentry);
1068 CERROR("error looking up O/0: rc %d\n", rc);
1069 GOTO(cleanup_R, rc);
1072 if (O0_dentry->d_inode) {
1073 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1074 GOTO(cleanup_O0, rc = -EEXIST);
1077 LOCK_INODE_MUTEX(O_dentry->d_inode);
1078 rc = vfs_rename(O_dentry->d_inode, dentry,
1079 O_dentry->d_inode, O0_dentry);
1080 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1083 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1084 GOTO(cleanup_O0, rc);
1086 filter->fo_fsd->lsd_feature_incompat |=
1087 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1088 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1090 GOTO(cleanup_O0, rc);
1102 cleanup_phase = 2; /* groups */
1104 /* we have to initialize all groups before first connections from
1105 * clients because they may send create/destroy for any group -bzzz */
1106 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1108 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1109 GOTO(cleanup, rc = PTR_ERR(filp));
1111 cleanup_phase = 3; /* filp */
1113 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1115 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1119 last_group = FILTER_MIN_GROUPS;
1121 LASSERT(last_group >= FILTER_MIN_GROUPS);
1124 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1125 FILTER_MIN_GROUPS, last_group);
1126 filter->fo_committed_group = last_group;
1127 rc = filter_read_groups(obd, last_group, 1);
1131 filp_close(filp, 0);
1135 switch (cleanup_phase) {
1137 filp_close(filp, 0);
1139 filter_cleanup_groups(obd);
1141 f_dput(filter->fo_dentry_O);
1142 filter->fo_dentry_O = NULL;
1150 /* setup the object store with correct subdirectories */
1151 static int filter_prep(struct obd_device *obd)
1153 struct lvfs_run_ctxt saved;
1154 struct filter_obd *filter = &obd->u.filter;
1156 struct inode *inode;
1160 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1161 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1162 if (!file || IS_ERR(file)) {
1164 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1168 filter->fo_rcvd_filp = file;
1169 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1170 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1171 file->f_dentry->d_inode->i_mode);
1172 GOTO(err_filp, rc = -ENOENT);
1175 inode = file->f_dentry->d_parent->d_inode;
1176 /* We use i_op->unlink directly in filter_vfs_unlink() */
1177 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1178 CERROR("%s: filesystem does not support create/unlink ops\n",
1180 GOTO(err_filp, rc = -EOPNOTSUPP);
1183 rc = filter_init_server_data(obd, file);
1185 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1188 /* open and create health check io file*/
1189 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1192 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1196 filter->fo_health_check_filp = file;
1197 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1198 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1199 file->f_dentry->d_inode->i_mode);
1200 GOTO(err_health_check, rc = -ENOENT);
1202 rc = lvfs_check_io_health(obd, file);
1204 GOTO(err_health_check, rc);
1206 rc = filter_prep_groups(obd);
1208 GOTO(err_server_data, rc);
1210 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1215 //class_disconnect_exports(obd, 0);
1216 filter_free_server_data(filter);
1218 if (filp_close(filter->fo_health_check_filp, 0))
1219 CERROR("can't close %s after error\n", HEALTH_CHECK);
1220 filter->fo_health_check_filp = NULL;
1222 if (filp_close(filter->fo_rcvd_filp, 0))
1223 CERROR("can't close %s after error\n", LAST_RCVD);
1224 filter->fo_rcvd_filp = NULL;
1228 /* cleanup the filter: write last used object id to status file */
1229 static void filter_post(struct obd_device *obd)
1231 struct lvfs_run_ctxt saved;
1232 struct filter_obd *filter = &obd->u.filter;
1235 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1236 * best to start a transaction with h_sync, because we removed this
1239 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1240 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1243 CERROR("error writing server data: rc = %d\n", rc);
1245 for (i = 1; i < filter->fo_group_count; i++) {
1246 rc = filter_update_last_objid(obd, i,
1247 (i == filter->fo_group_count - 1));
1249 CERROR("error writing group %d lastobjid: rc = %d\n",
1253 rc = filp_close(filter->fo_rcvd_filp, 0);
1254 filter->fo_rcvd_filp = NULL;
1256 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1258 rc = filp_close(filter->fo_health_check_filp, 0);
1259 filter->fo_health_check_filp = NULL;
1261 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1263 filter_cleanup_groups(obd);
1264 filter_free_server_data(filter);
1265 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1267 filter_free_capa_keys(filter);
1270 static void filter_set_last_id(struct filter_obd *filter,
1271 obd_id id, obd_gr group)
1273 LASSERT(filter->fo_fsd != NULL);
1274 LASSERT(group <= filter->fo_group_count);
1276 spin_lock(&filter->fo_objidlock);
1277 filter->fo_last_objids[group] = id;
1278 spin_unlock(&filter->fo_objidlock);
1281 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1284 LASSERT(filter->fo_fsd != NULL);
1285 LASSERT(group <= filter->fo_group_count);
1287 /* FIXME: object groups */
1288 spin_lock(&filter->fo_objidlock);
1289 id = filter->fo_last_objids[group];
1290 spin_unlock(&filter->fo_objidlock);
1295 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1297 LOCK_INODE_MUTEX(dparent->d_inode);
1301 /* We never dget the object parent, so DON'T dput it either */
1302 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1304 struct filter_obd *filter = &obd->u.filter;
1305 struct filter_subdirs *subdirs;
1306 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1308 if (group > 0 || filter->fo_subdir_count == 0)
1309 return filter->fo_dentry_O_groups[group];
1311 subdirs = &filter->fo_dentry_O_sub[group];
1312 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1315 /* We never dget the object parent, so DON'T dput it either */
1316 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1319 unsigned long now = jiffies;
1320 struct dentry *dparent = filter_parent(obd, group, objid);
1323 if (IS_ERR(dparent))
1326 rc = filter_lock_dentry(obd, dparent);
1327 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1328 return rc ? ERR_PTR(rc) : dparent;
1331 /* We never dget the object parent, so DON'T dput it either */
1332 static void filter_parent_unlock(struct dentry *dparent)
1334 UNLOCK_INODE_MUTEX(dparent->d_inode);
1337 /* How to get files, dentries, inodes from object id's.
1339 * If dir_dentry is passed, the caller has already locked the parent
1340 * appropriately for this operation (normally a write lock). If
1341 * dir_dentry is NULL, we do a read lock while we do the lookup to
1342 * avoid races with create/destroy and such changing the directory
1343 * internal to the filesystem code. */
1344 struct dentry *filter_fid2dentry(struct obd_device *obd,
1345 struct dentry *dir_dentry,
1346 obd_gr group, obd_id id)
1348 struct dentry *dparent = dir_dentry;
1349 struct dentry *dchild;
1354 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT))
1355 RETURN(ERR_PTR(-ENOENT));
1358 CERROR("fatal: invalid object id 0\n");
1359 RETURN(ERR_PTR(-ESTALE));
1362 len = sprintf(name, LPU64, id);
1363 if (dir_dentry == NULL) {
1364 dparent = filter_parent_lock(obd, group, id);
1365 if (IS_ERR(dparent)) {
1366 CERROR("%s: error getting object "LPU64":"LPU64
1367 " parent: rc %ld\n", obd->obd_name,
1368 id, group, PTR_ERR(dparent));
1372 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1373 dparent->d_name.len, dparent->d_name.name, name);
1374 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1375 if (dir_dentry == NULL)
1376 filter_parent_unlock(dparent);
1377 if (IS_ERR(dchild)) {
1378 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1383 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1384 CERROR("%s: got bad object "LPU64" inode %lu\n",
1385 obd->obd_name, id, dchild->d_inode->i_ino);
1387 RETURN(ERR_PTR(-ENOENT));
1390 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1391 name, dchild, atomic_read(&dchild->d_count));
1393 LASSERT(atomic_read(&dchild->d_count) > 0);
1398 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1401 struct lustre_handle lockh;
1402 int flags = LDLM_AST_DISCARD_DATA, rc;
1403 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1404 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1407 /* Tell the clients that the object is gone now and that they should
1408 * throw away any cached pages. */
1409 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id, LDLM_EXTENT,
1410 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1411 ldlm_completion_ast, NULL, NULL, 0, NULL,
1414 /* We only care about the side-effects, just drop the lock. */
1416 ldlm_lock_decref(&lockh, LCK_PW);
1421 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1422 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1423 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1424 * reverse. Caller must take i_sem before starting the transaction and we
1425 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1426 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1431 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1432 * itself we already hold dir->i_mutex for child create/unlink ops */
1433 LASSERT(dentry->d_inode != NULL);
1434 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1435 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1439 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1440 GOTO(out, rc = -ENOENT);
1442 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1447 GOTO(out, rc = -EPERM);
1449 /* check_sticky() */
1450 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1451 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1452 GOTO(out, rc = -EPERM);
1454 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1455 * that was done because of journal_start (which is already done
1456 * here) or some other ordering issue. */
1459 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1460 rc = security_inode_unlink(dir, dentry);
1465 rc = dir->i_op->unlink(dir, dentry);
1467 /* need to drop i_mutex before we lose inode reference */
1468 UNLOCK_INODE_MUTEX(dentry->d_inode);
1475 /* Caller must hold LCK_PW on parent and push us into kernel context.
1476 * Caller must hold child i_mutex, we drop it always.
1477 * Caller is also required to ensure that dchild->d_inode exists. */
1478 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1479 obd_gr group, struct dentry *dparent,
1480 struct dentry *dchild)
1482 struct inode *inode = dchild->d_inode;
1485 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1486 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1487 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1488 (unsigned long)inode->i_nlink,
1489 atomic_read(&inode->i_count));
1492 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1494 CERROR("error unlinking objid %.*s: rc %d\n",
1495 dchild->d_name.len, dchild->d_name.name, rc);
1499 static int filter_intent_policy(struct ldlm_namespace *ns,
1500 struct ldlm_lock **lockp, void *req_cookie,
1501 ldlm_mode_t mode, int flags, void *data)
1503 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1504 struct ptlrpc_request *req = req_cookie;
1505 struct ldlm_lock *lock = *lockp, *l = NULL;
1506 struct ldlm_resource *res = lock->l_resource;
1507 ldlm_processing_policy policy;
1508 struct ost_lvb *res_lvb, *reply_lvb;
1509 struct ldlm_reply *rep;
1510 struct list_head *tmp;
1512 int rc, tmpflags = 0, only_liblustre = 0;
1513 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1514 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1515 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1518 policy = ldlm_get_processing_policy(res);
1519 LASSERT(policy != NULL);
1520 LASSERT(req != NULL);
1522 rc = lustre_pack_reply(req, 3, repsize, NULL);
1524 RETURN(req->rq_status = rc);
1526 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1527 LASSERT(rep != NULL);
1529 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1530 sizeof(*reply_lvb));
1531 LASSERT(reply_lvb != NULL);
1533 //fixup_handle_for_resent_req(req, lock, &lockh);
1535 /* If we grant any lock at all, it will be a whole-file read lock.
1536 * Call the extent policy function to see if our request can be
1537 * granted, or is blocked. */
1538 lock->l_policy_data.l_extent.start = 0;
1539 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1540 lock->l_req_mode = LCK_PR;
1542 LASSERT(ns == res->lr_namespace);
1544 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1545 check_res_locked(res);
1547 /* FIXME: we should change the policy function slightly, to not make
1548 * this list at all, since we just turn around and free it */
1549 while (!list_empty(&rpc_list)) {
1550 struct ldlm_lock *wlock =
1551 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1552 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1553 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1554 lock->l_flags &= ~LDLM_FL_CP_REQD;
1555 list_del_init(&wlock->l_cp_ast);
1556 LDLM_LOCK_PUT(wlock);
1559 /* The lock met with no resistance; we're finished. */
1560 if (rc == LDLM_ITER_CONTINUE) {
1563 * do not grant locks to the liblustre clients: they cannot
1564 * handle ASTs robustly.
1566 if (lock->l_export->exp_libclient) {
1567 ldlm_resource_unlink_lock(lock);
1568 RETURN(ELDLM_LOCK_ABORTED);
1570 RETURN(ELDLM_LOCK_REPLACED);
1573 /* Do not grant any lock, but instead send GL callbacks. The extent
1574 * policy nicely created a list of all PW locks for us. We will choose
1575 * the highest of those which are larger than the size in the LVB, if
1576 * any, and perform a glimpse callback. */
1577 res_lvb = res->lr_lvb_data;
1578 LASSERT(res_lvb != NULL);
1579 *reply_lvb = *res_lvb;
1581 list_for_each(tmp, &res->lr_granted) {
1582 struct ldlm_lock *tmplock =
1583 list_entry(tmp, struct ldlm_lock, l_res_link);
1585 if (tmplock->l_granted_mode == LCK_PR)
1588 * ->ns_lock guarantees that no new locks are granted, and,
1589 * therefore, that res->lr_lvb_data cannot increase beyond the
1590 * end of already granted lock. As a result, it is safe to
1591 * check against "stale" reply_lvb->lvb_size value without
1594 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1597 /* Don't send glimpse ASTs to liblustre clients. They aren't
1598 * listening for them, and they do entirely synchronous I/O
1600 if (tmplock->l_export == NULL ||
1601 tmplock->l_export->exp_libclient == 1) {
1607 l = LDLM_LOCK_GET(tmplock);
1611 if (l->l_policy_data.l_extent.start >
1612 tmplock->l_policy_data.l_extent.start)
1616 l = LDLM_LOCK_GET(tmplock);
1620 /* There were no PW locks beyond the size in the LVB; finished. */
1622 if (only_liblustre) {
1623 /* If we discovered a liblustre client with a PW lock,
1624 * however, the LVB may be out of date! The LVB is
1625 * updated only on glimpse (which we don't do for
1626 * liblustre clients) and cancel (which the client
1627 * obviously has not yet done). So if it has written
1628 * data but kept the lock, the LVB is stale and needs
1629 * to be updated from disk.
1631 * Of course, this will all disappear when we switch to
1632 * taking liblustre locks on the OST. */
1633 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1634 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1636 RETURN(ELDLM_LOCK_ABORTED);
1640 * This check is for lock taken in filter_prepare_destroy() that does
1641 * not have l_glimpse_ast set. So the logic is: if there is a lock
1642 * with no l_glimpse_ast set, this object is being destroyed already.
1644 * Hence, if you are grabbing DLM locks on the server, always set
1645 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1647 if (l->l_glimpse_ast == NULL) {
1648 /* We are racing with unlink(); just return -ENOENT */
1649 rep->lock_policy_res1 = -ENOENT;
1653 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1654 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1655 /* Update the LVB from disk if the AST failed (this is a legal race) */
1657 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1658 * sending ast is not handled. This can result in lost client writes.
1660 if (rc != 0 && ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1661 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1664 *reply_lvb = *res_lvb;
1670 RETURN(ELDLM_LOCK_ABORTED);
1674 * per-obd_device iobuf pool.
1676 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1677 * allocations in write-path has to be minimized (see bug 5137).
1679 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1680 * OST threads (see ost_thread_{init,done}()).
1682 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1683 * at the OST layer there are only (potentially) multiple obd_device of type
1684 * unknown at the time of OST thread creation.
1686 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1687 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1688 * it's very own iobuf.
1692 * filter_kiobuf_pool_init()
1694 * filter_kiobuf_pool_done()
1696 * filter_iobuf_get()
1698 * operate on this array. They are "generic" in a sense that they don't depend
1699 * on actual type of iobuf's (the latter depending on Linux kernel version).
1703 * destroy pool created by filter_iobuf_pool_init
1705 static void filter_iobuf_pool_done(struct filter_obd *filter)
1707 struct filter_iobuf **pool;
1712 pool = filter->fo_iobuf_pool;
1714 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1715 if (pool[i] != NULL)
1716 filter_free_iobuf(pool[i]);
1718 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1719 filter->fo_iobuf_pool = NULL;
1725 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1727 static int filter_iobuf_pool_init(struct filter_obd *filter)
1733 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OST_MAX_THREADS * sizeof(*pool),
1735 if (filter->fo_iobuf_pool == NULL)
1738 filter->fo_iobuf_count = OST_MAX_THREADS;
1743 /* Return iobuf allocated for @thread_id. We don't know in advance how
1744 * many threads there will be so we allocate a large empty array and only
1745 * fill in those slots that are actually in use.
1746 * If we haven't allocated a pool entry for this thread before, do so now. */
1747 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1749 int thread_id = oti ? oti->oti_thread_id : -1;
1750 struct filter_iobuf *pool = NULL;
1751 struct filter_iobuf **pool_place = NULL;
1753 if (thread_id >= 0) {
1754 LASSERT(thread_id < filter->fo_iobuf_count);
1755 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1758 if (unlikely(pool == NULL)) {
1759 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1760 PTLRPC_MAX_BRW_PAGES);
1761 if (pool_place != NULL)
1768 /* mount the file system (secretly). lustre_cfg parameters are:
1771 * 3 = flags: failover=f, failout=n
1774 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1777 struct filter_obd *filter = &obd->u.filter;
1778 struct vfsmount *mnt;
1779 struct lustre_mount_info *lmi;
1780 struct obd_uuid uuid;
1787 if (lcfg->lcfg_bufcount < 3 ||
1788 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1789 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1792 lmi = server_get_mount(obd->obd_name);
1794 /* We already mounted in lustre_fill_super.
1795 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1796 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1798 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1800 /* old path - used by lctl */
1801 CERROR("Using old MDS mount method\n");
1802 mnt = do_kern_mount(lustre_cfg_string(lcfg, 2),
1803 MS_NOATIME|MS_NODIRATIME,
1804 lustre_cfg_string(lcfg, 1), option);
1807 LCONSOLE_ERROR("Can't mount disk %s (%d)\n",
1808 lustre_cfg_string(lcfg, 1), rc);
1812 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1814 if (IS_ERR(obd->obd_fsops))
1815 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1817 rc = filter_iobuf_pool_init(filter);
1821 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1823 /* failover is the default */
1824 obd->obd_replayable = 1;
1826 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1827 str = lustre_cfg_string(lcfg, 3);
1828 if (strchr(str, 'n')) {
1829 CWARN("%s: recovery disabled\n", obd->obd_name);
1830 obd->obd_replayable = 0;
1834 filter->fo_vfsmnt = mnt;
1835 obd->u.obt.obt_sb = mnt->mnt_sb;
1836 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1837 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1839 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1840 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1841 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1842 obd->obd_lvfs_ctxt.fs = get_ds();
1843 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1845 sema_init(&filter->fo_init_lock, 1);
1846 filter->fo_committed_group = 0;
1848 rc = filter_prep(obd);
1852 filter->fo_destroy_in_progress = 0;
1853 sema_init(&filter->fo_create_lock, 1);
1854 spin_lock_init(&filter->fo_translock);
1855 spin_lock_init(&filter->fo_objidlock);
1856 INIT_LIST_HEAD(&filter->fo_export_list);
1857 sema_init(&filter->fo_alloc_lock, 1);
1858 spin_lock_init(&filter->fo_r_pages.oh_lock);
1859 spin_lock_init(&filter->fo_w_pages.oh_lock);
1860 spin_lock_init(&filter->fo_read_rpc_hist.oh_lock);
1861 spin_lock_init(&filter->fo_write_rpc_hist.oh_lock);
1862 spin_lock_init(&filter->fo_r_io_time.oh_lock);
1863 spin_lock_init(&filter->fo_w_io_time.oh_lock);
1864 spin_lock_init(&filter->fo_r_discont_pages.oh_lock);
1865 spin_lock_init(&filter->fo_w_discont_pages.oh_lock);
1866 spin_lock_init(&filter->fo_r_discont_blocks.oh_lock);
1867 spin_lock_init(&filter->fo_w_discont_blocks.oh_lock);
1868 spin_lock_init(&filter->fo_r_disk_iosize.oh_lock);
1869 spin_lock_init(&filter->fo_w_disk_iosize.oh_lock);
1870 spin_lock_init(&filter->fo_r_dio_frags.oh_lock);
1871 spin_lock_init(&filter->fo_w_dio_frags.oh_lock);
1872 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1873 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1874 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1876 INIT_LIST_HEAD(&filter->fo_llog_list);
1877 spin_lock_init(&filter->fo_llog_list_lock);
1879 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1880 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1881 if (obd->obd_namespace == NULL)
1882 GOTO(err_post, rc = -ENOMEM);
1883 obd->obd_namespace->ns_lvbp = obd;
1884 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1885 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1887 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1888 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1890 rc = llog_cat_initialize(obd, NULL, 1, NULL);
1892 CERROR("failed to setup llogging subsystems\n");
1896 rc = lquota_setup(quota_interface, obd, lcfg);
1900 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1901 if (uuid_ptr != NULL) {
1902 class_uuid_unparse(uuid_ptr, &uuid);
1908 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1910 if (obd->obd_recovering) {
1911 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1912 "recovery until %d %s reconnect, or if no clients"
1913 " reconnect for %d:%.02d; during that time new "
1914 "clients will not be allowed to connect. "
1915 "Recovery progress can be monitored by watching "
1916 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1917 obd->obd_name, lustre_cfg_string(lcfg, 1),
1918 label ?: "", label ? "/" : "", str,
1919 obd->obd_recoverable_clients,
1920 (obd->obd_recoverable_clients == 1)
1921 ? "client" : "clients",
1922 (int)(OBD_RECOVERY_TIMEOUT) / 60,
1923 (int)(OBD_RECOVERY_TIMEOUT) % 60,
1926 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1927 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
1928 label ?: "", label ? "/" : "", str,
1929 obd->obd_replayable ? "enabled" : "disabled");
1932 filter->fo_fl_oss_capa = 0;
1933 INIT_LIST_HEAD(&filter->fo_capa_keys);
1940 fsfilt_put_ops(obd->obd_fsops);
1941 filter_iobuf_pool_done(filter);
1944 server_put_mount(obd->obd_name, mnt);
1951 obd->u.obt.obt_sb = 0;
1955 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1957 struct lprocfs_static_vars lvars;
1961 CLASSERT(offsetof(struct obd_device, u.obt) ==
1962 offsetof(struct obd_device, u.filter.fo_obt));
1964 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
1967 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
1968 page = get_zeroed_page(GFP_KERNEL);
1972 memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
1973 LUSTRE_CFG_BUFLEN(lcfg, 4));
1974 rc = filter_common_setup(obd, lcfg, (void *)page);
1977 lprocfs_init_vars(filter, &lvars);
1978 if (rc == 0 && lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1979 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
1980 /* Init obdfilter private stats here */
1981 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1982 LPROCFS_CNTR_AVGMINMAX,
1983 "read_bytes", "bytes");
1984 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1985 LPROCFS_CNTR_AVGMINMAX,
1986 "write_bytes", "bytes");
1988 lproc_filter_attach_seqstat(obd);
1994 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
1995 static struct llog_operations filter_size_orig_logops = {
1996 lop_setup: llog_obd_origin_setup,
1997 lop_cleanup: llog_obd_origin_cleanup,
1998 lop_add: llog_obd_origin_add
2001 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2002 struct obd_device *tgt, int count,
2003 struct llog_catid *catid,
2004 struct obd_uuid *uuid)
2006 struct llog_ctxt *ctxt;
2010 filter_mds_ost_repl_logops = llog_client_ops;
2011 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2012 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2013 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2015 rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2016 &filter_mds_ost_repl_logops);
2020 /* FIXME - assign unlink_cb for filter's recovery */
2022 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2024 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2026 LASSERT(ctxt != NULL);
2027 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2029 rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2030 &filter_size_orig_logops);
2034 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2039 if (CTXTP(ctxt, cleanup))
2040 rc = CTXTP(ctxt, cleanup)(ctxt);
2043 class_export_put(ctxt->loc_exp);
2044 OBD_FREE(ctxt, sizeof(*ctxt));
2049 static int filter_group_llog_finish(struct obd_llogs *llogs)
2051 struct llog_ctxt *ctxt;
2052 int rc = 0, rc2 = 0;
2055 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2057 rc = filter_group_llog_cleanup(ctxt);
2059 ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2061 rc2 = filter_group_llog_cleanup(ctxt);
2068 static int filter_llog_finish(struct obd_device *obd, int count)
2070 struct llog_ctxt *ctxt;
2071 int rc = 0, rc2 = 0;
2074 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2076 rc = llog_cleanup(ctxt);
2078 ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2080 rc2 = llog_cleanup(ctxt);
2087 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2088 struct obd_export *export)
2090 struct filter_group_llog *fglog, *nlog;
2091 struct filter_obd *filter;
2092 struct llog_ctxt *ctxt;
2093 struct list_head *cur;
2096 filter = &obd->u.filter;
2098 spin_lock(&filter->fo_llog_list_lock);
2099 list_for_each(cur, &filter->fo_llog_list) {
2100 fglog = list_entry(cur, struct filter_group_llog, list);
2101 if (fglog->group == group) {
2102 if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2103 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2104 obd->obd_name, group, fglog->exp, export);
2105 spin_unlock(&filter->fo_llog_list_lock);
2109 spin_unlock(&filter->fo_llog_list_lock);
2114 OBD_ALLOC(fglog, sizeof(*fglog));
2117 fglog->group = group;
2119 OBD_ALLOC(fglog->llogs, sizeof(struct obd_llogs));
2120 if (fglog->llogs == NULL) {
2121 OBD_FREE(fglog, sizeof(*fglog));
2125 spin_lock(&filter->fo_llog_list_lock);
2126 list_for_each(cur, &filter->fo_llog_list) {
2127 nlog = list_entry(cur, struct filter_group_llog, list);
2128 LASSERT(nlog->group != group);
2130 list_add(&fglog->list, &filter->fo_llog_list);
2131 spin_unlock(&filter->fo_llog_list_lock);
2133 rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2135 OBD_FREE(fglog->llogs, sizeof(*(fglog->llogs)));
2136 OBD_FREE(fglog, sizeof(*fglog));
2142 fglog->exp = export;
2143 ctxt = llog_get_context_from_llogs(fglog->llogs,
2144 LLOG_MDS_OST_REPL_CTXT);
2145 LASSERT(ctxt != NULL);
2147 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2149 CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2150 obd->obd_name, fglog->llogs, group);
2152 RETURN(fglog->llogs);
2155 static int filter_llog_connect(struct obd_export *exp,
2156 struct llogd_conn_body *body)
2158 struct obd_device *obd = exp->exp_obd;
2159 struct llog_ctxt *ctxt;
2160 struct obd_llogs *llog;
2164 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2165 (unsigned) body->lgdc_logid.lgl_ogr,
2166 (unsigned) body->lgdc_logid.lgl_oid,
2167 (unsigned) body->lgdc_logid.lgl_ogen);
2169 llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2170 LASSERT(llog != NULL);
2171 ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2172 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2173 body->lgdc_ctxt_idx);
2174 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2175 &body->lgdc_gen, NULL);
2177 CERROR("failed to connect rc %d idx %d\n", rc,
2178 body->lgdc_ctxt_idx);
2183 static int filter_llog_preclean (struct obd_device *obd)
2185 struct filter_group_llog *log;
2186 struct filter_obd *filter;
2190 filter = &obd->u.filter;
2191 spin_lock(&filter->fo_llog_list_lock);
2192 while (!list_empty(&filter->fo_llog_list)) {
2193 log = list_entry(filter->fo_llog_list.next,
2194 struct filter_group_llog, list);
2195 list_del(&log->list);
2196 spin_unlock(&filter->fo_llog_list_lock);
2198 rc = filter_group_llog_finish(log->llogs);
2200 CERROR("failed to cleanup llogging subsystem for %u\n",
2202 OBD_FREE(log->llogs, sizeof(*(log->llogs)));
2203 OBD_FREE(log, sizeof(*log));
2204 spin_lock(&filter->fo_llog_list_lock);
2206 spin_unlock(&filter->fo_llog_list_lock);
2208 rc = obd_llog_finish(obd, 0);
2210 CERROR("failed to cleanup llogging subsystem\n");
2215 static int filter_precleanup(struct obd_device *obd,
2216 enum obd_cleanup_stage stage)
2222 case OBD_CLEANUP_EARLY:
2224 case OBD_CLEANUP_EXPORTS:
2225 target_cleanup_recovery(obd);
2227 case OBD_CLEANUP_SELF_EXP:
2228 rc = filter_llog_preclean(obd);
2230 case OBD_CLEANUP_OBD:
2236 static int filter_cleanup(struct obd_device *obd)
2238 struct filter_obd *filter = &obd->u.filter;
2239 lvfs_sbdev_type save_dev;
2240 int must_relock = 0, must_put = 0;
2244 LCONSOLE_WARN("%s: shutting down for failover; client state "
2245 "will be preserved.\n", obd->obd_name);
2247 if (!list_empty(&obd->obd_exports)) {
2248 CERROR("%s: still has clients!\n", obd->obd_name);
2249 class_disconnect_exports(obd);
2250 if (!list_empty(&obd->obd_exports)) {
2251 CERROR("still has exports after forced cleanup?\n");
2256 lquota_cleanup(quota_interface, obd);
2258 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2260 if (obd->u.obt.obt_sb == NULL)
2262 save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2264 lprocfs_free_obd_stats(obd);
2265 lprocfs_obd_cleanup(obd);
2269 shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2271 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2273 must_put = server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2274 /* must_put is for old method (l_p_m returns non-0 on err) */
2276 /* We can only unlock kernel if we are in the context of sys_ioctl,
2277 otherwise we never called lock_kernel */
2278 if (ll_kernel_locked()) {
2284 /* In case we didn't mount with lustre_get_mount -- old method*/
2285 mntput(filter->fo_vfsmnt);
2286 lvfs_clear_rdonly(save_dev);
2288 obd->u.obt.obt_sb = NULL;
2293 fsfilt_put_ops(obd->obd_fsops);
2295 filter_iobuf_pool_done(filter);
2297 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2302 static int filter_connect_internal(struct obd_export *exp,
2303 struct obd_connect_data *data)
2308 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2309 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2310 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2311 data->ocd_connect_flags, data->ocd_version,
2312 data->ocd_grant, data->ocd_index);
2314 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2315 exp->exp_connect_flags = data->ocd_connect_flags;
2316 if (exp->exp_imp_reverse)
2317 exp->exp_imp_reverse->imp_connect_data.ocd_connect_flags
2318 = data->ocd_connect_flags;
2319 data->ocd_version = LUSTRE_VERSION_CODE;
2321 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2322 obd_size left, want;
2324 spin_lock(&exp->exp_obd->obd_osfs_lock);
2325 left = filter_grant_space_left(exp);
2326 want = data->ocd_grant;
2327 data->ocd_grant = filter_grant(exp, 0, want, left);
2328 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2330 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2331 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2332 exp->exp_client_uuid.uuid, exp,
2333 data->ocd_grant, want, left);
2336 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2337 struct filter_obd *filter = &exp->exp_obd->u.filter;
2338 struct lr_server_data *lsd = filter->fo_fsd;
2339 int index = le32_to_cpu(lsd->lsd_ost_index);
2341 if (!(lsd->lsd_feature_compat &
2342 cpu_to_le32(OBD_COMPAT_OST))) {
2343 /* this will only happen on the first connect */
2344 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2345 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2346 filter_update_server_data(exp->exp_obd,
2347 filter->fo_rcvd_filp, lsd, 1);
2348 } else if (index != data->ocd_index) {
2349 LCONSOLE_ERROR("Connection from %s to index "
2350 "%u doesn't match actual OST "
2351 "index %u, bad configuration?\n",
2352 obd_export_nid2str(exp), index,
2358 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2359 data->ocd_brw_size = 65536;
2360 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2361 data->ocd_brw_size = min(data->ocd_brw_size,
2362 (__u32)(PTLRPC_MAX_BRW_PAGES << PAGE_SHIFT));
2365 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2366 * FIXME: We don't strictly need the COMPAT flag for that,
2367 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2368 * FIXME: We needed it for the index, as index 0 is valid. */
2373 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2374 struct obd_uuid *cluuid,
2375 struct obd_connect_data *data)
2380 if (exp == NULL || obd == NULL || cluuid == NULL)
2383 rc = filter_connect_internal(exp, data);
2388 /* nearly identical to mds_connect */
2389 static int filter_connect(const struct lu_env *env,
2390 struct lustre_handle *conn, struct obd_device *obd,
2391 struct obd_uuid *cluuid,
2392 struct obd_connect_data *data)
2394 struct lvfs_run_ctxt saved;
2395 struct obd_export *exp;
2396 struct filter_export_data *fed;
2397 struct filter_client_data *fcd = NULL;
2398 struct filter_obd *filter = &obd->u.filter;
2403 if (conn == NULL || obd == NULL || cluuid == NULL)
2406 rc = class_connect(conn, obd, cluuid);
2409 exp = class_conn2export(conn);
2410 LASSERT(exp != NULL);
2412 fed = &exp->exp_filter_data;
2414 rc = filter_connect_internal(exp, data);
2418 group = data->ocd_group;
2419 if (obd->obd_replayable) {
2420 OBD_ALLOC(fcd, sizeof(*fcd));
2422 CERROR("filter: out of memory for client data\n");
2423 GOTO(cleanup, rc = -ENOMEM);
2426 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2428 fed->fed_fcd->fcd_group = group;
2429 rc = filter_client_add(obd, filter, fed, -1);
2433 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2434 obd->obd_name, exp->exp_handle.h_cookie, group);
2438 if (fed->fed_group != 0 && fed->fed_group != group) {
2439 CERROR("!!! This export (nid %s) used object group %d "
2440 "earlier; now it's trying to use group %d! This could "
2441 "be a bug in the MDS. Tell CFS.\n",
2442 obd_export_nid2str(exp), fed->fed_group, group);
2443 GOTO(cleanup, rc = -EPROTO);
2445 fed->fed_group = group;
2447 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2448 rc = filter_read_groups(obd, group, 1);
2449 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2451 CERROR("can't read group %u\n", group);
2460 OBD_FREE(fcd, sizeof(*fcd));
2461 fed->fed_fcd = NULL;
2463 class_disconnect(exp);
2465 class_export_put(exp);
2471 /* Do extra sanity checks for grant accounting. We do this at connect,
2472 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2473 * always get rid of it or turn it off when we know accounting is good. */
2474 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2476 struct filter_export_data *fed;
2477 struct obd_export *exp;
2478 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2479 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2480 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2482 if (list_empty(&obd->obd_exports))
2485 spin_lock(&obd->obd_osfs_lock);
2486 spin_lock(&obd->obd_dev_lock);
2487 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2489 fed = &exp->exp_filter_data;
2490 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2493 if (maxsize > 0) { /* we may not have done a statfs yet */
2494 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2495 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2496 exp->exp_client_uuid.uuid, exp,
2497 fed->fed_grant, fed->fed_pending, maxsize);
2498 LASSERTF(fed->fed_dirty <= maxsize,
2499 "%s: cli %s/%p %ld > "LPU64"\n", func,
2500 exp->exp_client_uuid.uuid, exp,
2501 fed->fed_dirty, maxsize);
2504 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2505 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2506 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2508 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2509 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2510 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2511 tot_granted += fed->fed_grant + fed->fed_pending;
2512 tot_pending += fed->fed_pending;
2513 tot_dirty += fed->fed_dirty;
2515 fo_tot_granted = obd->u.filter.fo_tot_granted;
2516 fo_tot_pending = obd->u.filter.fo_tot_pending;
2517 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2518 spin_unlock(&obd->obd_dev_lock);
2519 spin_unlock(&obd->obd_osfs_lock);
2521 /* Do these assertions outside the spinlocks so we don't kill system */
2522 if (tot_granted != fo_tot_granted)
2523 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2524 func, tot_granted, fo_tot_granted);
2525 if (tot_pending != fo_tot_pending)
2526 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2527 func, tot_pending, fo_tot_pending);
2528 if (tot_dirty != fo_tot_dirty)
2529 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2530 func, tot_dirty, fo_tot_dirty);
2531 if (tot_pending > tot_granted)
2532 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2533 func, tot_pending, tot_granted);
2534 if (tot_granted > maxsize)
2535 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2536 func, tot_granted, maxsize);
2537 if (tot_dirty > maxsize)
2538 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2539 func, tot_dirty, maxsize);
2542 /* Remove this client from the grant accounting totals. We also remove
2543 * the export from the obd device under the osfs and dev locks to ensure
2544 * that the filter_grant_sanity_check() calculations are always valid.
2545 * The client should do something similar when it invalidates its import. */
2546 static void filter_grant_discard(struct obd_export *exp)
2548 struct obd_device *obd = exp->exp_obd;
2549 struct filter_obd *filter = &obd->u.filter;
2550 struct filter_export_data *fed = &exp->exp_filter_data;
2552 spin_lock(&obd->obd_osfs_lock);
2553 spin_lock(&obd->obd_dev_lock);
2554 list_del_init(&exp->exp_obd_chain);
2555 spin_unlock(&obd->obd_dev_lock);
2557 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2558 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2559 obd->obd_name, filter->fo_tot_granted,
2560 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2561 filter->fo_tot_granted -= fed->fed_grant;
2562 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2563 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2564 obd->obd_name, filter->fo_tot_pending,
2565 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2566 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2567 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2568 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2569 obd->obd_name, filter->fo_tot_dirty,
2570 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2571 filter->fo_tot_dirty -= fed->fed_dirty;
2575 spin_unlock(&obd->obd_osfs_lock);
2578 static int filter_destroy_export(struct obd_export *exp)
2582 if (exp->exp_filter_data.fed_pending)
2583 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2584 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2585 exp, exp->exp_filter_data.fed_pending);
2587 target_destroy_export(exp);
2589 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2592 if (exp->exp_obd->obd_replayable)
2593 filter_client_free(exp);
2595 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2597 filter_grant_discard(exp);
2598 filter_fmd_cleanup(exp);
2600 if (!(exp->exp_flags & OBD_OPT_FORCE))
2601 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2606 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2608 struct filter_group_llog *fglog, *nlog;
2609 struct filter_obd *filter;
2610 int worked = 0, group;
2611 struct llog_ctxt *ctxt;
2614 filter = &obd->u.filter;
2616 /* we can't sync log holding spinlock. also, we do not want to get
2617 * into livelock. so we do following: loop over MDS's exports in
2618 * group order and skip already synced llogs -bzzz */
2620 /* look for group with min. number, but > worked */
2623 spin_lock(&filter->fo_llog_list_lock);
2624 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2626 if (nlog->group <= worked) {
2627 /* this group is already synced */
2631 if (group < nlog->group) {
2632 /* we have group with smaller number to sync */
2636 /* store current minimal group */
2638 group = nlog->group;
2640 spin_unlock(&filter->fo_llog_list_lock);
2645 worked = fglog->group;
2646 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2647 ctxt = llog_get_context_from_llogs(fglog->llogs,
2648 LLOG_MDS_OST_REPL_CTXT);
2649 LASSERT(ctxt != NULL);
2650 llog_sync(ctxt, fglog->exp);
2652 } while (fglog != NULL);
2655 /* also incredibly similar to mds_disconnect */
2656 static int filter_disconnect(struct obd_export *exp)
2658 struct obd_device *obd = exp->exp_obd;
2663 class_export_get(exp);
2665 if (!(exp->exp_flags & OBD_OPT_FORCE))
2666 filter_grant_sanity_check(obd, __FUNCTION__);
2667 filter_grant_discard(exp);
2669 /* Disconnect early so that clients can't keep using export */
2670 rc = class_disconnect(exp);
2671 if (exp->exp_obd->obd_namespace != NULL)
2672 ldlm_cancel_locks_for_export(exp);
2674 fsfilt_sync(obd, obd->u.obt.obt_sb);
2676 /* flush any remaining cancel messages out to the target */
2677 filter_sync_llogs(obd, exp);
2678 class_export_put(exp);
2682 static int filter_ping(struct obd_export *exp)
2684 filter_fmd_expire(exp);
2689 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2690 const char *what, int quiet)
2692 struct dentry *dchild = NULL;
2695 if (oa->o_valid & OBD_MD_FLGROUP)
2698 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2700 if (IS_ERR(dchild)) {
2701 CERROR("%s error looking up object: "LPU64"\n",
2706 if (dchild->d_inode == NULL) {
2708 CERROR("%s: %s on non-existent object: "LPU64"\n",
2709 obd->obd_name, what, oa->o_id);
2711 RETURN(ERR_PTR(-ENOENT));
2717 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2719 struct dentry *dentry = NULL;
2720 struct obd_device *obd;
2724 rc = filter_verify_capa(exp, NULL, oinfo_mdsno(oinfo),
2725 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2729 obd = class_exp2obd(exp);
2731 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2735 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2737 RETURN(PTR_ERR(dentry));
2739 /* Limit the valid bits in the return data to what we actually use */
2740 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2741 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2747 /* this should be enabled/disabled in condition to enabled/disabled large
2748 * inodes (fast EAs) in backing store FS. */
2749 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2750 void *handle, struct obdo *oa)
2752 struct obd_device *obd = exp->exp_obd;
2756 if (oa->o_valid & OBD_MD_FLFID) {
2757 struct filter_fid ff;
2759 if (!(oa->o_valid & OBD_MD_FLGROUP))
2761 /* packing fid and converting it to LE for storing into EA.
2762 * Here ->o_stripe_idx should be filled by LOV and rest of
2763 * fields - by client. */
2764 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2765 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2766 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2767 ff.ff_objid = cpu_to_le64(oa->o_id);
2768 ff.ff_group = cpu_to_le64(oa->o_gr);
2770 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2771 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2772 oa->o_generation, oa->o_id, oa->o_gr);
2774 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2776 CERROR("store fid in object failed! rc: %d\n", rc);
2778 CDEBUG(D_HA, "OSS object without fid info!\n");
2784 /* this is called from filter_truncate() until we have filter_punch() */
2785 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2786 struct obdo *oa, struct obd_trans_info *oti)
2788 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2789 struct llog_cookie *fcc = NULL;
2790 struct filter_obd *filter;
2791 int rc, err, locked = 0;
2792 unsigned int ia_valid;
2793 struct inode *inode;
2798 LASSERT(dentry != NULL);
2799 LASSERT(!IS_ERR(dentry));
2801 inode = dentry->d_inode;
2802 LASSERT(inode != NULL);
2804 filter = &exp->exp_obd->u.filter;
2805 iattr_from_obdo(&iattr, oa, oa->o_valid);
2806 ia_valid = iattr.ia_valid;
2808 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2809 OBD_ALLOC(fcc, sizeof(*fcc));
2811 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2814 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2815 LOCK_INODE_MUTEX(inode);
2819 /* If the inode still has SUID+SGID bits set (see filter_precreate())
2820 * then we will accept the UID+GID sent by the client during write for
2821 * initializing the ownership of this inode. We only allow this to
2822 * happen once so clear these bits in setattr. In 2.6 kernels it is
2823 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2824 * the flags that are actually being set. */
2825 if (ia_valid & (ATTR_UID | ATTR_GID)) {
2826 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2827 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2829 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2830 if (!(ia_valid & ATTR_MODE)) {
2831 iattr.ia_mode = inode->i_mode;
2832 iattr.ia_valid |= ATTR_MODE;
2834 iattr.ia_mode &= ~S_ISUID;
2836 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2837 if (!(iattr.ia_valid & ATTR_MODE)) {
2838 iattr.ia_mode = inode->i_mode;
2839 iattr.ia_valid |= ATTR_MODE;
2841 iattr.ia_mode &= ~S_ISGID;
2844 orig_ids[USRQUOTA] = inode->i_uid;
2845 orig_ids[GRPQUOTA] = inode->i_gid;
2846 handle = fsfilt_start_log(exp->exp_obd, inode,
2847 FSFILT_OP_SETATTR, oti, 1);
2849 /* update inode EA only once when inode is suid bit marked. As
2850 * on 2.6.x UID and GID may be set separately, we check here
2851 * only one of them to avoid double setting. */
2852 if (inode->i_mode & S_ISUID)
2853 filter_update_fidea(exp, inode, handle, oa);
2855 handle = fsfilt_start(exp->exp_obd, inode,
2856 FSFILT_OP_SETATTR, oti);
2860 GOTO(out_unlock, rc = PTR_ERR(handle));
2862 if (oa->o_valid & OBD_MD_FLFLAGS) {
2863 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2864 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2866 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2868 /* set cancel cookie callback function */
2869 fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2870 filter_cancel_cookies_cb,
2876 rc = filter_finish_transno(exp, oti, rc);
2878 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2880 CERROR("error on commit, err = %d\n", err);
2886 /* Let's flush truncated page on disk immediately, then we can
2887 * avoid need to search for page aliases before directio writes
2888 * and this sort of stuff at expense of somewhat slower
2889 * truncates not on a page boundary. I believe this is the only
2890 * place in filter code that can lead to pages getting to
2891 * pagecache so far. */
2892 filter_clear_truncated_page(inode);
2893 UNLOCK_INODE_MUTEX(inode);
2900 UNLOCK_INODE_MUTEX(inode);
2903 OBD_FREE(fcc, sizeof(*fcc));
2905 /* trigger quota release */
2906 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2907 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2908 int rc2 = lquota_adjust(quota_interface, exp->exp_obd, cur_ids,
2909 orig_ids, rc, FSFILT_OP_SETATTR);
2911 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2912 "filter adjust qunit. (rc:%d)\n", rc2);
2917 /* this is called from filter_truncate() until we have filter_punch() */
2918 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
2919 struct obd_trans_info *oti)
2921 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
2922 oinfo->oi_oa->o_gr, 0 } };
2923 struct ldlm_valblock_ops *ns_lvbo;
2924 struct filter_mod_data *fmd;
2925 struct lvfs_run_ctxt saved;
2926 struct filter_obd *filter;
2927 struct ldlm_resource *res;
2928 struct dentry *dentry;
2932 rc = filter_verify_capa(exp, NULL, oinfo_mdsno(oinfo),
2933 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
2937 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
2940 RETURN(PTR_ERR(dentry));
2942 filter = &exp->exp_obd->u.filter;
2943 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2946 if (oinfo->oi_oa->o_valid &
2947 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
2948 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
2949 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
2950 fmd->fmd_mactime_xid = oti->oti_xid;
2951 filter_fmd_put(exp, fmd);
2954 /* setting objects attributes (including owner/group) */
2955 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
2957 GOTO(out_unlock, rc);
2959 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
2960 res_id, LDLM_EXTENT, 0);
2963 ns_lvbo = res->lr_namespace->ns_lvbo;
2964 if (ns_lvbo && ns_lvbo->lvbo_update)
2965 rc = ns_lvbo->lvbo_update(res, NULL, 0, 0);
2966 ldlm_resource_putref(res);
2969 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2971 /* Quota release need uid/gid info */
2972 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
2973 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
2979 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2983 /* XXX identical to osc_unpackmd */
2984 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2985 struct lov_mds_md *lmm, int lmm_bytes)
2991 if (lmm_bytes < sizeof (*lmm)) {
2992 CERROR("lov_mds_md too small: %d, need %d\n",
2993 lmm_bytes, (int)sizeof(*lmm));
2996 /* XXX LOV_MAGIC etc check? */
2998 if (lmm->lmm_object_id == cpu_to_le64(0)) {
2999 CERROR("lov_mds_md: zero lmm_object_id\n");
3004 lsm_size = lov_stripe_md_size(1);
3008 if (*lsmp != NULL && lmm == NULL) {
3009 OBD_FREE(*lsmp, lsm_size);
3014 if (*lsmp == NULL) {
3015 OBD_ALLOC(*lsmp, lsm_size);
3019 loi_init((*lsmp)->lsm_oinfo);
3023 /* XXX zero *lsmp? */
3024 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3025 LASSERT((*lsmp)->lsm_object_id);
3028 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3033 /* caller must hold fo_create_lock */
3034 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3035 struct filter_obd *filter)
3037 struct obdo doa; /* XXX obdo on stack */
3043 LASSERT(down_trylock(&filter->fo_create_lock) != 0);
3045 memset(&doa, 0, sizeof(doa));
3047 LASSERT(oa->o_gr != 0);
3048 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3050 doa.o_valid |= OBD_MD_FLGROUP;
3051 doa.o_gr = oa->o_gr;
3052 doa.o_mode = S_IFREG;
3054 if (!filter->fo_destroy_in_progress) {
3055 CERROR("%s: destroy_in_progress already cleared\n",
3056 exp->exp_obd->obd_name);
3060 last = filter_last_id(filter, doa.o_gr);
3061 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3062 exp->exp_obd->obd_name, oa->o_id + 1, last);
3063 for (id = last; id > oa->o_id; id--) {
3065 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3066 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3067 CEMERG("error destroying precreate objid "LPU64": %d\n",
3069 filter_set_last_id(filter, id - 1, doa.o_gr);
3070 /* update last_id on disk periodically so that if we restart
3071 * we don't need to re-scan all of the just-deleted objects. */
3072 if ((id & 511) == 0)
3073 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3076 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3077 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3079 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3080 filter->fo_destroy_in_progress = 0;
3085 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3086 obd_gr group, int *num);
3087 /* returns a negative error or a nonnegative number of files to create */
3088 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3089 obd_gr group, struct obd_trans_info *oti)
3091 struct obd_device *obd = exp->exp_obd;
3092 struct filter_obd *filter = &obd->u.filter;
3096 /* delete orphans request */
3097 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3098 /* This causes inflight precreates to abort and drop lock */
3099 filter->fo_destroy_in_progress = 1;
3100 down(&filter->fo_create_lock);
3101 diff = oa->o_id - filter_last_id(filter, group);
3102 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3103 filter_last_id(filter, group), diff);
3105 if (-diff > OST_MAX_PRECREATE) {
3106 CERROR("%s: ignoring bogus orphan destroy request: "
3107 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3108 oa->o_id, filter_last_id(filter, group));
3109 /* FIXME: should reset precreate_next_id on MDS */
3110 GOTO(out, rc = -EINVAL);
3113 rc = filter_destroy_precreated(exp, oa, filter);
3115 CERROR("%s: unable to write lastobjid, but "
3116 "orphans were deleted\n", obd->obd_name);
3119 /*XXX used by MDS for the first time! */
3120 filter->fo_destroy_in_progress = 0;
3123 down(&filter->fo_create_lock);
3124 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3125 CERROR("%s: dropping old precreate request\n",
3129 /* only precreate if group == 0 and o_id is specfied */
3130 if (group != 0 || oa->o_id == 0)
3133 diff = oa->o_id - filter_last_id(filter, group);
3134 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3135 filter_last_id(filter, group), diff);
3137 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3138 oa->o_id, filter_last_id(filter, group), diff);
3142 oa->o_id = filter_last_id(&obd->u.filter, group);
3143 rc = filter_precreate(obd, oa, group, &diff);
3144 oa->o_id = filter_last_id(&obd->u.filter, group);
3146 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3149 /* else diff == 0 */
3152 up(&filter->fo_create_lock);
3156 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3159 struct filter_obd *filter = &obd->u.filter;
3160 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3164 /* at least try to account for cached pages. its still racey and
3165 * might be under-reporting if clients haven't announced their
3166 * caches with brw recently */
3167 spin_lock(&obd->obd_osfs_lock);
3168 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3169 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3170 spin_unlock(&obd->obd_osfs_lock);
3172 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3173 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3174 filter->fo_tot_dirty, filter->fo_tot_granted,
3175 filter->fo_tot_pending,
3176 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3178 filter_grant_sanity_check(obd, __FUNCTION__);
3180 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3181 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3182 osfs->os_bsize - 1) >> blockbits));
3184 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3185 * stop creating files on MDS if OST is not good shape to create
3187 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3192 /* We rely on the fact that only one thread will be creating files in a given
3193 * group at a time, which is why we don't need an atomic filter_get_new_id.
3194 * Even if we had that atomic function, the following race would exist:
3196 * thread 1: gets id x from filter_next_id
3197 * thread 2: gets id (x + 1) from filter_next_id
3198 * thread 2: creates object (x + 1)
3199 * thread 1: tries to create object x, gets -ENOSPC
3201 * Caller must hold fo_create_lock
3203 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3204 obd_gr group, int *num)
3206 struct dentry *dchild = NULL, *dparent = NULL;
3207 struct filter_obd *filter;
3208 struct obd_statfs *osfs;
3209 int err = 0, rc = 0, recreate_obj = 0, i;
3210 unsigned long enough_time = jiffies + (obd_timeout * HZ) / 4;
3212 void *handle = NULL;
3215 filter = &obd->u.filter;
3217 LASSERT(down_trylock(&filter->fo_create_lock) != 0);
3219 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3220 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3223 OBD_ALLOC(osfs, sizeof(*osfs));
3226 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3227 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3228 CDEBUG(D_HA, "OST out of space! avail "LPU64"\n",
3230 filter->fo_obt.obt_sb->s_blocksize_bits);
3234 OBD_FREE(osfs, sizeof(*osfs));
3239 CDEBUG(D_HA, "%s: precreating %d objects in group "LPU64" at "LPU64"\n",
3240 obd->obd_name, *num, group, oa->o_id);
3242 for (i = 0; i < *num && err == 0; i++) {
3243 int cleanup_phase = 0;
3245 if (filter->fo_destroy_in_progress) {
3246 CWARN("%s: precreate aborted by destroy\n",
3255 last_id = filter_last_id(filter, group);
3256 if (next_id > last_id) {
3257 CERROR("Error: Trying to recreate obj greater"
3258 "than last id "LPD64" > "LPD64"\n",
3260 GOTO(cleanup, rc = -EINVAL);
3263 next_id = filter_last_id(filter, group) + 1;
3265 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3267 dparent = filter_parent_lock(obd, group, next_id);
3268 if (IS_ERR(dparent))
3269 GOTO(cleanup, rc = PTR_ERR(dparent));
3272 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3274 GOTO(cleanup, rc = PTR_ERR(dchild));
3277 if (dchild->d_inode != NULL) {
3278 /* This would only happen if lastobjid was bad on disk*/
3279 /* Could also happen if recreating missing obj but
3283 CERROR("%s: recreating existing object %.*s?\n",
3284 obd->obd_name, dchild->d_name.len,
3285 dchild->d_name.name);
3287 CERROR("%s: Serious error: objid %.*s already "
3288 "exists; is this filesystem corrupt?\n",
3289 obd->obd_name, dchild->d_name.len,
3290 dchild->d_name.name);
3293 GOTO(cleanup, rc = -EEXIST);
3296 handle = fsfilt_start_log(obd, dparent->d_inode,
3297 FSFILT_OP_CREATE, NULL, 1);
3299 GOTO(cleanup, rc = PTR_ERR(handle));
3302 /* We mark object SUID+SGID to flag it for accepting UID+GID
3303 * from client on first write. Currently the permission bits
3304 * on the OST are never used, so this is OK. */
3305 rc = ll_vfs_create(dparent->d_inode, dchild,
3306 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3308 CERROR("create failed rc = %d\n", rc);
3312 if (!recreate_obj) {
3313 filter_set_last_id(filter, next_id, group);
3314 err = filter_update_last_objid(obd, group, 0);
3316 CERROR("unable to write lastobjid "
3317 "but file created\n");
3321 switch(cleanup_phase) {
3323 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3325 CERROR("error on commit, err = %d\n", err);
3332 filter_parent_unlock(dparent);
3339 if (time_after(jiffies, enough_time)) {
3340 CDEBUG(D_HA, "%s: precreate slow - want %d got %d \n",
3341 obd->obd_name, *num, i);
3347 CDEBUG(D_HA, "%s: created %d objects for group "LPU64": "LPU64"\n",
3348 obd->obd_name, i, group, filter->fo_last_objids[group]);
3350 CDEBUG(D_HA, "%s: filter_precreate() created %d objects: rc %d\n",
3351 obd->obd_name, i, rc);
3356 static int filter_create(struct obd_export *exp, struct obdo *oa,
3357 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3359 struct filter_export_data *fed;
3360 struct obd_device *obd = NULL;
3361 struct filter_obd *filter;
3362 struct lvfs_run_ctxt saved;
3363 struct lov_stripe_md *lsm = NULL;
3364 int rc = 0, diff, group = oa->o_gr;
3367 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3368 CERROR("!!! nid %s sent invalid object group %d\n",
3369 obd_export_nid2str(exp), group);
3374 fed = &exp->exp_filter_data;
3375 filter = &obd->u.filter;
3377 if (fed->fed_group != group) {
3378 CERROR("!!! this export (nid %s) used object group %d "
3379 "earlier; now it's trying to use group %d! This could "
3380 "be a bug in the MDS. Tell CFS.\n",
3381 obd_export_nid2str(exp), fed->fed_group, group);
3386 if (!(oa->o_valid & OBD_MD_FLGROUP))
3389 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3390 oa->o_gr, oa->o_id);
3394 rc = obd_alloc_memmd(exp, &lsm);
3401 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3403 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3404 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3405 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3406 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3407 oa->o_id, filter_last_id(filter,
3412 down(&filter->fo_create_lock);
3413 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3414 up(&filter->fo_create_lock);
3417 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3420 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3421 if (rc && ea != NULL && *ea != lsm) {
3422 obd_free_memmd(exp, &lsm);
3423 } else if (rc == 0 && ea != NULL) {
3424 /* XXX LOV STACKING: the lsm that is passed to us from
3425 * LOV does not have valid lsm_oinfo data structs, so
3426 * don't go touching that. This needs to be fixed in a
3428 lsm->lsm_object_id = oa->o_id;
3435 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3436 struct lov_stripe_md *md, struct obd_trans_info *oti,
3437 struct obd_export *md_exp, void *capa)
3439 unsigned int qcids[MAXQUOTAS] = {0, 0};
3440 struct obd_device *obd;
3441 struct filter_obd *filter;
3442 struct dentry *dchild = NULL, *dparent = NULL;
3443 struct lvfs_run_ctxt saved;
3444 void *handle = NULL;
3445 struct llog_cookie *fcc = NULL;
3446 int rc, rc2, cleanup_phase = 0;
3450 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3452 #if 0 /* some places don't support capability yet */
3453 rc = filter_verify_capa(exp, NULL, obdo_mdsno(oa),
3454 (struct lustre_capa *)capa,
3455 CAPA_OPC_INDEX_LOOKUP);
3461 if (!(oa->o_valid & OBD_MD_FLGROUP))
3465 filter = &obd->u.filter;
3467 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3470 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3472 GOTO(cleanup, rc = PTR_ERR(dchild));
3475 if (dchild->d_inode == NULL) {
3476 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3478 /* If object already gone, cancel cookie right now */
3479 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3480 fcc = obdo_logcookie(oa);
3481 llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3483 fcc = NULL; /* we didn't allocate fcc, don't free it */
3485 GOTO(cleanup, rc = -ENOENT);
3488 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3490 /* Our MDC connection is established by the MDS to us */
3491 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3492 OBD_ALLOC(fcc, sizeof(*fcc));
3494 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3497 /* we're gonna truncate it first in order to avoid possible deadlock:
3499 * open trasaction open transaction
3500 * down(i_zombie) down(i_zombie)
3501 * restart transaction
3502 * (see BUG 4180) -bzzz
3504 LOCK_INODE_MUTEX(dchild->d_inode);
3505 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3507 if (IS_ERR(handle)) {
3508 UNLOCK_INODE_MUTEX(dchild->d_inode);
3509 GOTO(cleanup, rc = PTR_ERR(handle));
3512 iattr.ia_valid = ATTR_SIZE;
3514 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3515 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3516 UNLOCK_INODE_MUTEX(dchild->d_inode);
3520 GOTO(cleanup, rc = rc2);
3522 /* We don't actually need to lock the parent until we are unlinking
3523 * here, and not while truncating above. That avoids holding the
3524 * parent lock for a long time during truncate, which can block other
3525 * threads from doing anything to objects in that directory. bug 7171 */
3526 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3527 if (IS_ERR(dparent))
3528 GOTO(cleanup, rc = PTR_ERR(dparent));
3529 cleanup_phase = 3; /* filter_parent_unlock */
3531 LOCK_INODE_MUTEX(dchild->d_inode);
3532 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3533 if (IS_ERR(handle)) {
3534 UNLOCK_INODE_MUTEX(dchild->d_inode);
3535 GOTO(cleanup, rc = PTR_ERR(handle));
3537 cleanup_phase = 4; /* fsfilt_commit */
3539 /* Quota release need uid/gid of inode */
3540 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3542 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3544 /* this drops dchild->d_inode->i_mutex unconditionally */
3545 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3549 switch(cleanup_phase) {
3552 fsfilt_add_journal_cb(obd, 0,
3553 oti ? oti->oti_handle : handle,
3554 filter_cancel_cookies_cb, fcc);
3557 rc = filter_finish_transno(exp, oti, rc);
3558 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3560 CERROR("error on commit, err = %d\n", rc2);
3565 filter_parent_unlock(dparent);
3569 OBD_FREE(fcc, sizeof(*fcc));
3571 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3574 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3578 /* trigger quota release */
3579 qcids[USRQUOTA] = oa->o_uid;
3580 qcids[GRPQUOTA] = oa->o_gid;
3581 rc2 = lquota_adjust(quota_interface, obd, qcids, NULL, rc,
3584 CDEBUG(rc ? D_ERROR : D_QUOTA,
3585 "filter adjust qunit! (rc:%d)\n", rc2);
3589 /* NB start and end are used for punch, but not truncate */
3590 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3591 struct obd_trans_info *oti,
3592 struct ptlrpc_request_set *rqset)
3597 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3598 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3599 oinfo->oi_policy.l_extent.end);
3603 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3604 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3605 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3607 rc = filter_verify_capa(exp, NULL, oinfo_mdsno(oinfo),
3608 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3612 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3613 rc = filter_setattr(exp, oinfo, oti);
3617 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3618 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3621 struct lvfs_run_ctxt saved;
3622 struct filter_obd *filter;
3623 struct dentry *dentry;
3624 struct llog_ctxt *ctxt;
3628 rc = filter_verify_capa(exp, NULL, obdo_mdsno(oa),
3629 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3633 filter = &exp->exp_obd->u.filter;
3635 /* an objid of zero is taken to mean "sync whole filesystem" */
3636 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3637 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3638 /* flush any remaining cancel messages out to the target */
3639 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3640 llog_sync(ctxt, exp);
3644 dentry = filter_oa2dentry(exp->exp_obd, oa);
3646 RETURN(PTR_ERR(dentry));
3648 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3650 LOCK_INODE_MUTEX(dentry->d_inode);
3652 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3654 /* just any file to grab fsync method - "file" arg unused */
3655 struct file *file = filter->fo_rcvd_filp;
3657 if (file->f_op && file->f_op->fsync)
3658 rc = file->f_op->fsync(NULL, dentry, 1);
3660 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3664 UNLOCK_INODE_MUTEX(dentry->d_inode);
3666 oa->o_valid = OBD_MD_FLID;
3667 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3669 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3675 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3676 void *key, __u32 *vallen, void *val)
3678 struct obd_device *obd;
3681 obd = class_exp2obd(exp);
3683 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3687 if (keylen == strlen("blocksize") &&
3688 memcmp(key, "blocksize", keylen) == 0) {
3689 __u32 *blocksize = val;
3690 *vallen = sizeof(*blocksize);
3691 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3695 if (keylen == strlen("blocksize_bits") &&
3696 memcmp(key, "blocksize_bits", keylen) == 0) {
3697 __u32 *blocksize_bits = val;
3698 *vallen = sizeof(*blocksize_bits);
3699 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3703 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3704 obd_id *last_id = val;
3705 /* FIXME: object groups */
3706 *last_id = filter_last_id(&obd->u.filter, 0);
3709 CDEBUG(D_IOCTL, "invalid key\n");
3713 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3714 void *key, __u32 vallen, void *val,
3715 struct ptlrpc_request_set *set)
3717 struct obd_device *obd;
3718 struct obd_llogs *llog;
3719 struct llog_ctxt *ctxt;
3725 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3729 if (KEY_IS(KEY_CAPA_KEY)) {
3730 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3732 CERROR("filter update capability key failed: %d\n", rc);
3736 if (keylen < strlen(KEY_MDS_CONN) ||
3737 memcmp(key, KEY_MDS_CONN, keylen) != 0)
3740 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3741 obd_export_nid2str(exp));
3742 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3744 /* setup llog imports */
3745 LASSERT(val != NULL);
3746 group = (int)(*(__u32 *)val);
3747 LASSERT(group >= FILTER_GROUP_MDS0);
3749 llog = filter_grab_llog_for_group(obd, group, exp);
3750 LASSERT(llog != NULL);
3751 ctxt = llog_get_context_from_llogs(llog, LLOG_MDS_OST_REPL_CTXT);
3752 LASSERTF(ctxt != NULL, "ctxt is not null\n"),
3754 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3756 lquota_setinfo(quota_interface, exp, obd);
3761 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3762 int len, void *karg, void *uarg)
3764 struct obd_device *obd = exp->exp_obd;
3765 struct obd_ioctl_data *data = karg;
3769 case OBD_IOC_ABORT_RECOVERY: {
3770 CERROR("aborting recovery for device %s\n", obd->obd_name);
3771 target_abort_recovery(obd);
3775 case OBD_IOC_SYNC: {
3776 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3777 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3781 case OBD_IOC_SET_READONLY: {
3783 struct super_block *sb = obd->u.obt.obt_sb;
3784 struct inode *inode = sb->s_root->d_inode;
3785 BDEVNAME_DECLARE_STORAGE(tmp);
3786 CERROR("*** setting device %s read-only ***\n",
3787 ll_bdevname(sb, tmp));
3789 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3790 if (!IS_ERR(handle))
3791 rc = fsfilt_commit(obd, inode, handle, 1);
3793 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3794 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3796 lvfs_set_rdonly(lvfs_sbdev(obd->u.obt.obt_sb));
3800 case OBD_IOC_CATLOGLIST: {
3801 rc = llog_catalog_list(obd, 1, data);
3805 case OBD_IOC_LLOG_CANCEL:
3806 case OBD_IOC_LLOG_REMOVE:
3807 case OBD_IOC_LLOG_INFO:
3808 case OBD_IOC_LLOG_PRINT: {
3809 /* FIXME to be finished */
3810 RETURN(-EOPNOTSUPP);
3812 struct llog_ctxt *ctxt = NULL;
3814 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3815 rc = llog_ioctl(ctxt, cmd, data);
3816 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3829 static int filter_health_check(struct obd_device *obd)
3831 struct filter_obd *filter = &obd->u.filter;
3835 * health_check to return 0 on healthy
3836 * and 1 on unhealthy.
3838 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3841 LASSERT(filter->fo_health_check_filp != NULL);
3842 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3847 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3850 return filter_fid2dentry(data, NULL, gr, id);
3853 static int filter_process_config(struct obd_device *obd, obd_count len, void *buf)
3855 struct lustre_cfg *lcfg = buf;
3856 struct lprocfs_static_vars lvars;
3859 lprocfs_init_vars(filter, &lvars);
3861 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
3865 static struct lvfs_callback_ops filter_lvfs_ops = {
3866 l_fid2dentry: filter_lvfs_fid2dentry,
3869 static struct obd_ops filter_obd_ops = {
3870 .o_owner = THIS_MODULE,
3871 .o_get_info = filter_get_info,
3872 .o_set_info_async = filter_set_info_async,
3873 .o_setup = filter_setup,
3874 .o_precleanup = filter_precleanup,
3875 .o_cleanup = filter_cleanup,
3876 .o_connect = filter_connect,
3877 .o_reconnect = filter_reconnect,
3878 .o_disconnect = filter_disconnect,
3879 .o_ping = filter_ping,
3880 .o_init_export = filter_init_export,
3881 .o_destroy_export = filter_destroy_export,
3882 .o_statfs = filter_statfs,
3883 .o_getattr = filter_getattr,
3884 .o_unpackmd = filter_unpackmd,
3885 .o_create = filter_create,
3886 .o_setattr = filter_setattr,
3887 .o_destroy = filter_destroy,
3888 .o_brw = filter_brw,
3889 .o_punch = filter_truncate,
3890 .o_sync = filter_sync,
3891 .o_preprw = filter_preprw,
3892 .o_commitrw = filter_commitrw,
3893 .o_llog_init = filter_llog_init,
3894 .o_llog_connect = filter_llog_connect,
3895 .o_llog_finish = filter_llog_finish,
3896 .o_iocontrol = filter_iocontrol,
3897 .o_health_check = filter_health_check,
3898 .o_process_config = filter_process_config,
3901 quota_interface_t *quota_interface;
3902 extern quota_interface_t filter_quota_interface;
3904 static int __init obdfilter_init(void)
3906 struct lprocfs_static_vars lvars;
3909 printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
3911 lprocfs_init_vars(filter, &lvars);
3913 request_module("lquota");
3914 OBD_ALLOC(obdfilter_created_scratchpad,
3915 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3916 sizeof(*obdfilter_created_scratchpad));
3917 if (obdfilter_created_scratchpad == NULL)
3920 ll_fmd_cachep = kmem_cache_create("ll_fmd_cache",
3921 sizeof(struct filter_mod_data),
3924 GOTO(out, rc = -ENOMEM);
3926 quota_interface = PORTAL_SYMBOL_GET(filter_quota_interface);
3927 init_obd_quota_ops(quota_interface, &filter_obd_ops);
3929 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
3930 LUSTRE_OST_NAME, NULL);
3934 err = kmem_cache_destroy(ll_fmd_cachep);
3935 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
3936 ll_fmd_cachep = NULL;
3938 if (quota_interface)
3939 PORTAL_SYMBOL_PUT(filter_quota_interface);
3941 OBD_FREE(obdfilter_created_scratchpad,
3942 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3943 sizeof(*obdfilter_created_scratchpad));
3949 static void __exit obdfilter_exit(void)
3951 if (quota_interface)
3952 PORTAL_SYMBOL_PUT(filter_quota_interface);
3954 if (ll_fmd_cachep) {
3955 int rc = kmem_cache_destroy(ll_fmd_cachep);
3956 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
3957 ll_fmd_cachep = NULL;
3960 class_unregister_type(LUSTRE_OST_NAME);
3961 OBD_FREE(obdfilter_created_scratchpad,
3962 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3963 sizeof(*obdfilter_created_scratchpad));
3966 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3967 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
3968 MODULE_LICENSE("GPL");
3970 module_init(obdfilter_init);
3971 module_exit(obdfilter_exit);