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_sem 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_sem 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 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
45 # include <linux/mount.h>
46 # include <linux/buffer_head.h>
49 #include <linux/obd_class.h>
50 #include <linux/obd_lov.h>
51 #include <linux/lustre_dlm.h>
52 #include <linux/lustre_fsfilt.h>
53 #include <linux/lprocfs_status.h>
54 #include <linux/lustre_log.h>
55 #include <linux/lustre_commit_confd.h>
57 #include "filter_internal.h"
59 static struct lvfs_callback_ops filter_lvfs_ops;
61 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
62 struct lov_stripe_md *ea, struct obd_trans_info *);
64 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
65 void *cb_data, int error)
67 obd_transno_commit_cb(obd, transno, error);
71 /* Assumes caller has already pushed us into the kernel context. */
72 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
75 struct filter_obd *filter = &exp->exp_obd->u.filter;
76 struct filter_export_data *fed = &exp->exp_filter_data;
77 struct filter_client_data *fcd = fed->fed_fcd;
80 int err, log_pri = D_HA;
82 /* Propagate error code. */
86 if (!exp->exp_obd->obd_replayable || oti == NULL)
89 /* we don't allocate new transnos for replayed requests */
90 if (oti->oti_transno == 0) {
91 spin_lock(&filter->fo_translock);
92 last_rcvd = le64_to_cpu(filter->fo_fsd->fsd_last_transno) + 1;
93 filter->fo_fsd->fsd_last_transno = cpu_to_le64(last_rcvd);
94 spin_unlock(&filter->fo_translock);
95 oti->oti_transno = last_rcvd;
97 spin_lock(&filter->fo_translock);
98 last_rcvd = oti->oti_transno;
99 if (last_rcvd > le64_to_cpu(filter->fo_fsd->fsd_last_transno))
100 filter->fo_fsd->fsd_last_transno =
101 cpu_to_le64(last_rcvd);
102 spin_unlock(&filter->fo_translock);
104 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
106 /* could get xid from oti, if it's ever needed */
107 fcd->fcd_last_xid = 0;
109 off = fed->fed_lr_off;
110 fsfilt_add_journal_cb(exp->exp_obd, last_rcvd, oti->oti_handle,
111 filter_commit_cb, NULL);
112 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp, fcd,
113 sizeof(*fcd), &off, 0);
120 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
121 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
126 void f_dput(struct dentry *dentry)
128 /* Can't go inside filter_ddelete because it can block */
129 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
130 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
131 LASSERT(atomic_read(&dentry->d_count) > 0);
136 /* Add client data to the FILTER. We use a bitmap to locate a free space
137 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
138 * Otherwise, we have just read the data from the last_rcvd file and
139 * we know its offset. */
140 static int filter_client_add(struct obd_device *obd, struct filter_obd *filter,
141 struct filter_export_data *fed, int cl_idx)
143 unsigned long *bitmap = filter->fo_last_rcvd_slots;
144 int new_client = (cl_idx == -1);
147 LASSERT(bitmap != NULL);
149 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
150 if (!strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid))
153 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
154 * there's no need for extra complication here
157 cl_idx = find_first_zero_bit(bitmap, FILTER_LR_MAX_CLIENTS);
159 if (cl_idx >= FILTER_LR_MAX_CLIENTS) {
160 CERROR("no client slots - fix FILTER_LR_MAX_CLIENTS\n");
163 if (test_and_set_bit(cl_idx, bitmap)) {
164 CERROR("FILTER client %d: found bit is set in bitmap\n",
166 cl_idx = find_next_zero_bit(bitmap,
167 FILTER_LR_MAX_CLIENTS,
172 if (test_and_set_bit(cl_idx, bitmap)) {
173 CERROR("FILTER client %d: bit already set in bitmap!\n",
179 fed->fed_lr_idx = cl_idx;
180 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->fsd_client_start) +
181 cl_idx * le16_to_cpu(filter->fo_fsd->fsd_client_size);
183 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
184 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
187 struct obd_run_ctxt saved;
188 loff_t off = fed->fed_lr_off;
192 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
193 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
195 push_ctxt(&saved, &obd->obd_ctxt, NULL);
196 /* Transaction needed to fix bug 1403 */
197 handle = fsfilt_start(obd,
198 filter->fo_rcvd_filp->f_dentry->d_inode,
199 FSFILT_OP_SETATTR, NULL);
200 if (IS_ERR(handle)) {
201 err = PTR_ERR(handle);
202 CERROR("unable to start transaction: rc %d\n", err);
204 err = fsfilt_write_record(obd, filter->fo_rcvd_filp,
206 sizeof(*fed->fed_fcd),
209 filter->fo_rcvd_filp->f_dentry->d_inode,
212 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
215 CERROR("error writing %s client idx %u: rc %d\n",
216 LAST_RCVD, fed->fed_lr_idx, err);
223 static int filter_client_free(struct obd_export *exp, int flags)
225 struct filter_export_data *fed = &exp->exp_filter_data;
226 struct filter_obd *filter = &exp->exp_obd->u.filter;
227 struct obd_device *obd = exp->exp_obd;
228 struct filter_client_data zero_fcd;
229 struct obd_run_ctxt saved;
234 if (fed->fed_fcd == NULL)
237 if (flags & OBD_OPT_FAILOVER)
240 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
241 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
244 LASSERT(filter->fo_last_rcvd_slots != NULL);
246 off = fed->fed_lr_off;
248 CDEBUG(D_INFO, "freeing client at idx %u (%lld) with UUID '%s'\n",
249 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
251 /* Clear the bit _after_ zeroing out the client so we don't
252 race with filter_client_add and zero out new clients.*/
253 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
254 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
259 memset(&zero_fcd, 0, sizeof zero_fcd);
260 push_ctxt(&saved, &obd->obd_ctxt, NULL);
261 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
262 sizeof(zero_fcd), &off, 1);
263 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
265 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
266 "zeroing disconnecting client %s at idx %u (%llu) in %s rc %d\n",
267 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
270 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
271 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
277 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
282 static int filter_free_server_data(struct filter_obd *filter)
284 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
285 filter->fo_fsd = NULL;
286 OBD_FREE(filter->fo_last_rcvd_slots,
287 FILTER_LR_MAX_CLIENT_WORDS * sizeof(unsigned long));
288 filter->fo_last_rcvd_slots = NULL;
292 /* assumes caller is already in kernel ctxt */
293 int filter_update_server_data(struct obd_device *obd, struct file *filp,
294 struct filter_server_data *fsd, int force_sync)
300 CDEBUG(D_INODE, "server uuid : %s\n", fsd->fsd_uuid);
301 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
302 le64_to_cpu(fsd->fsd_last_transno));
303 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
304 le64_to_cpu(fsd->fsd_mount_count));
306 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off,force_sync);
308 CERROR("error writing filter_server_data: rc = %d\n", rc);
313 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
316 struct filter_obd *filter = &obd->u.filter;
322 CDEBUG(D_INODE, "server last_objid for group "LPU64": "LPU64"\n",
323 group, filter->fo_last_objids[group]);
325 tmp = cpu_to_le64(filter->fo_last_objids[group]);
326 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
327 &tmp, sizeof(tmp), &off, force_sync);
329 CERROR("error writing group "LPU64" last objid: rc = %d\n",
334 /* assumes caller has already in kernel ctxt */
335 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
337 struct filter_obd *filter = &obd->u.filter;
338 struct filter_server_data *fsd;
339 struct filter_client_data *fcd = NULL;
340 struct inode *inode = filp->f_dentry->d_inode;
341 unsigned long last_rcvd_size = inode->i_size;
347 /* ensure padding in the struct is the correct size */
348 LASSERT (offsetof(struct filter_server_data, fsd_padding) +
349 sizeof(fsd->fsd_padding) == FILTER_LR_SERVER_SIZE);
350 LASSERT (offsetof(struct filter_client_data, fcd_padding) +
351 sizeof(fcd->fcd_padding) == FILTER_LR_CLIENT_SIZE);
353 OBD_ALLOC(fsd, sizeof(*fsd));
356 filter->fo_fsd = fsd;
358 OBD_ALLOC(filter->fo_last_rcvd_slots,
359 FILTER_LR_MAX_CLIENT_WORDS * sizeof(unsigned long));
360 if (filter->fo_last_rcvd_slots == NULL) {
361 OBD_FREE(fsd, sizeof(*fsd));
365 if (last_rcvd_size == 0) {
366 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
368 memcpy(fsd->fsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->fsd_uuid));
369 fsd->fsd_last_transno = 0;
370 mount_count = fsd->fsd_mount_count = 0;
371 fsd->fsd_server_size = cpu_to_le32(FILTER_LR_SERVER_SIZE);
372 fsd->fsd_client_start = cpu_to_le32(FILTER_LR_CLIENT_START);
373 fsd->fsd_client_size = cpu_to_le16(FILTER_LR_CLIENT_SIZE);
374 fsd->fsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
375 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
377 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
379 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
383 if (strcmp(fsd->fsd_uuid, obd->obd_uuid.uuid) != 0) {
384 CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
385 obd->obd_uuid.uuid, fsd->fsd_uuid);
386 GOTO(err_fsd, rc = -EINVAL);
388 mount_count = le64_to_cpu(fsd->fsd_mount_count);
389 filter->fo_subdir_count = le16_to_cpu(fsd->fsd_subdir_count);
392 if (fsd->fsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
393 CERROR("unsupported feature %x\n",
394 le32_to_cpu(fsd->fsd_feature_incompat) &
395 ~FILTER_INCOMPAT_SUPP);
396 GOTO(err_fsd, rc = -EINVAL);
398 if (fsd->fsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
399 CERROR("read-only feature %x\n",
400 le32_to_cpu(fsd->fsd_feature_rocompat) &
401 ~FILTER_ROCOMPAT_SUPP);
402 /* Do something like remount filesystem read-only */
403 GOTO(err_fsd, rc = -EINVAL);
406 CDEBUG(D_INODE, "%s: server last_rcvd : "LPU64"\n",
407 obd->obd_name, le64_to_cpu(fsd->fsd_last_transno));
408 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
409 obd->obd_name, mount_count + 1);
410 CDEBUG(D_INODE, "%s: server data size: %u\n",
411 obd->obd_name, le32_to_cpu(fsd->fsd_server_size));
412 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
413 obd->obd_name, le32_to_cpu(fsd->fsd_client_start));
414 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
415 obd->obd_name, le32_to_cpu(fsd->fsd_client_size));
416 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
417 obd->obd_name, le16_to_cpu(fsd->fsd_subdir_count));
418 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
419 last_rcvd_size <= FILTER_LR_CLIENT_START ? 0 :
420 (last_rcvd_size-FILTER_LR_CLIENT_START) /FILTER_LR_CLIENT_SIZE);
422 if (!obd->obd_replayable) {
423 CWARN("%s: recovery support OFF\n", obd->obd_name);
427 for (cl_idx = 0, off = le32_to_cpu(fsd->fsd_client_start);
428 off < last_rcvd_size; cl_idx++) {
430 struct obd_export *exp;
431 struct filter_export_data *fed;
434 OBD_ALLOC(fcd, sizeof(*fcd));
436 GOTO(err_client, rc = -ENOMEM);
439 /* Don't assume off is incremented properly by
440 * fsfilt_read_record(), in case sizeof(*fcd)
441 * isn't the same as fsd->fsd_client_size. */
442 off = le32_to_cpu(fsd->fsd_client_start) +
443 cl_idx * le16_to_cpu(fsd->fsd_client_size);
444 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
446 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
447 LAST_RCVD, cl_idx, off, rc);
448 break; /* read error shouldn't cause startup to fail */
451 if (fcd->fcd_uuid[0] == '\0') {
452 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
457 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
459 /* These exports are cleaned up by filter_disconnect(), so they
460 * need to be set up like real exports as filter_connect() does.
462 exp = class_new_export(obd);
463 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
464 " srv lr: "LPU64"\n", fcd->fcd_uuid, cl_idx,
465 last_rcvd, le64_to_cpu(fsd->fsd_last_transno));
467 GOTO(err_client, rc = -ENOMEM);
469 memcpy(&exp->exp_client_uuid.uuid, fcd->fcd_uuid,
470 sizeof exp->exp_client_uuid.uuid);
471 fed = &exp->exp_filter_data;
473 filter_client_add(obd, filter, fed, cl_idx);
474 /* create helper if export init gets more complex */
475 spin_lock_init(&fed->fed_lock);
478 obd->obd_recoverable_clients++;
479 class_export_put(exp);
481 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
484 if (last_rcvd > le64_to_cpu(fsd->fsd_last_transno))
485 fsd->fsd_last_transno = cpu_to_le64(last_rcvd);
489 obd->obd_last_committed = le64_to_cpu(fsd->fsd_last_transno);
491 if (obd->obd_recoverable_clients) {
492 CWARN("RECOVERY: %d recoverable clients, last_rcvd "
493 LPU64"\n", obd->obd_recoverable_clients,
494 le64_to_cpu(fsd->fsd_last_transno));
495 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
496 obd->obd_recovering = 1;
500 OBD_FREE(fcd, sizeof(*fcd));
503 filter->fo_mount_count = mount_count + 1;
504 fsd->fsd_mount_count = cpu_to_le64(filter->fo_mount_count);
506 /* save it, so mount count and last_transno is current */
507 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
512 class_disconnect_exports(obd, 0);
514 filter_free_server_data(filter);
518 static int filter_cleanup_groups(struct obd_device *obd)
520 struct filter_obd *filter = &obd->u.filter;
524 if (filter->fo_dentry_O_groups != NULL &&
525 filter->fo_last_objids != NULL &&
526 filter->fo_last_objid_files != NULL) {
527 for (i = 0; i < FILTER_GROUPS; i++) {
528 struct dentry *dentry = filter->fo_dentry_O_groups[i];
529 struct file *filp = filter->fo_last_objid_files[i];
530 if (dentry != NULL) {
532 filter->fo_dentry_O_groups[i] = NULL;
536 filter->fo_last_objid_files[i] = NULL;
540 if (filter->fo_dentry_O_sub != NULL && filter->fo_subdir_count) {
541 for (i = 0; i < filter->fo_subdir_count; i++) {
542 struct dentry *dentry = filter->fo_dentry_O_sub[i];
543 if (dentry != NULL) {
545 filter->fo_dentry_O_sub[i] = NULL;
548 OBD_FREE(filter->fo_dentry_O_sub,
549 filter->fo_subdir_count *
550 sizeof(*filter->fo_dentry_O_sub));
552 if (filter->fo_dentry_O_groups != NULL)
553 OBD_FREE(filter->fo_dentry_O_groups,
554 FILTER_GROUPS * sizeof(struct dentry *));
555 if (filter->fo_last_objids != NULL)
556 OBD_FREE(filter->fo_last_objids,
557 FILTER_GROUPS * sizeof(__u64));
558 if (filter->fo_last_objid_files != NULL)
559 OBD_FREE(filter->fo_last_objid_files,
560 FILTER_GROUPS * sizeof(struct file *));
564 /* FIXME: object groups */
565 static int filter_prep_groups(struct obd_device *obd)
567 struct filter_obd *filter = &obd->u.filter;
568 struct dentry *dentry, *O_dentry;
570 int i, rc = 0, cleanup_phase = 0;
573 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700);
574 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
575 if (IS_ERR(O_dentry)) {
576 rc = PTR_ERR(O_dentry);
577 CERROR("cannot open/create O: rc = %d\n", rc);
580 filter->fo_dentry_O = O_dentry;
581 cleanup_phase = 1; /* O_dentry */
583 /* Lookup "R" to tell if we're on an old OST FS and need to convert
584 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
585 * some time post 1.0 when all old-style OSTs have converted along
586 * with the init_objid hack. */
587 dentry = ll_lookup_one_len("R", O_dentry, 1);
589 GOTO(cleanup, rc = PTR_ERR(dentry));
590 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
591 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
594 CWARN("converting OST to new object layout\n");
595 if (IS_ERR(O0_dentry)) {
596 rc = PTR_ERR(O0_dentry);
597 CERROR("error looking up O/0: rc %d\n", rc);
601 if (O0_dentry->d_inode) {
602 CERROR("Both O/R and O/0 exist. Fix manually.\n");
603 GOTO(cleanup_O0, rc = -EEXIST);
606 down(&O_dentry->d_inode->i_sem);
607 rc = vfs_rename(O_dentry->d_inode, dentry,
608 O_dentry->d_inode, O0_dentry);
609 up(&O_dentry->d_inode->i_sem);
612 CERROR("error renaming O/R to O/0: rc %d\n", rc);
613 GOTO(cleanup_O0, rc);
615 filter->fo_fsd->fsd_feature_incompat |=
616 cpu_to_le32(FILTER_INCOMPAT_GROUPS);
617 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
619 GOTO(cleanup_O0, rc);
631 OBD_ALLOC(filter->fo_last_objids, FILTER_GROUPS * sizeof(__u64));
632 if (filter->fo_last_objids == NULL)
633 GOTO(cleanup, rc = -ENOMEM);
634 cleanup_phase = 2; /* groups */
636 OBD_ALLOC(filter->fo_dentry_O_groups, FILTER_GROUPS * sizeof(dentry));
637 if (filter->fo_dentry_O_groups == NULL)
638 GOTO(cleanup, rc = -ENOMEM);
639 OBD_ALLOC(filter->fo_last_objid_files, FILTER_GROUPS * sizeof(filp));
640 if (filter->fo_last_objid_files == NULL)
641 GOTO(cleanup, rc = -ENOMEM);
643 for (i = 0; i < FILTER_GROUPS; i++) {
647 sprintf(name, "%d", i);
648 dentry = simple_mkdir(O_dentry, name, 0700);
649 CDEBUG(D_INODE, "got/created O/%s: %p\n", name, dentry);
650 if (IS_ERR(dentry)) {
651 rc = PTR_ERR(dentry);
652 CERROR("cannot lookup/create O/%s: rc = %d\n",
656 filter->fo_dentry_O_groups[i] = dentry;
658 sprintf(name, "O/%d/LAST_ID", i);
659 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
660 if (IS_ERR(dentry)) {
661 rc = PTR_ERR(dentry);
662 CERROR("cannot create %s: rc = %d\n", name, rc);
665 filter->fo_last_objid_files[i] = filp;
667 if (filp->f_dentry->d_inode->i_size == 0) {
668 if (i == 0 && filter->fo_fsd->fsd_unused != 0) {
669 /* OST conversion, remove sometime post 1.0 */
670 filter->fo_last_objids[i] =
671 le64_to_cpu(filter->fo_fsd->fsd_unused);
672 CWARN("saving old objid "LPU64" to LAST_ID\n",
673 filter->fo_last_objids[i]);
674 rc = filter_update_last_objid(obd, 0, 1);
678 filter->fo_last_objids[i] = FILTER_INIT_OBJID;
683 rc = fsfilt_read_record(obd, filp, &filter->fo_last_objids[i],
684 sizeof(__u64), &off);
686 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
690 filter->fo_last_objids[i] =
691 le64_to_cpu(filter->fo_last_objids[i]);
692 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
693 obd->obd_name, i, filter->fo_last_objids[i]);
696 if (filter->fo_subdir_count) {
697 O_dentry = filter->fo_dentry_O_groups[0];
698 OBD_ALLOC(filter->fo_dentry_O_sub,
699 filter->fo_subdir_count * sizeof(dentry));
700 if (filter->fo_dentry_O_sub == NULL)
701 GOTO(cleanup, rc = -ENOMEM);
703 for (i = 0; i < filter->fo_subdir_count; i++) {
705 snprintf(dir, sizeof(dir), "d%u", i);
707 dentry = simple_mkdir(O_dentry, dir, 0700);
708 CDEBUG(D_INODE, "got/created O/0/%s: %p\n", dir,dentry);
709 if (IS_ERR(dentry)) {
710 rc = PTR_ERR(dentry);
711 CERROR("can't lookup/create O/0/%s: rc = %d\n",
715 filter->fo_dentry_O_sub[i] = dentry;
721 switch (cleanup_phase) {
723 filter_cleanup_groups(obd);
725 f_dput(filter->fo_dentry_O);
726 filter->fo_dentry_O = NULL;
733 /* setup the object store with correct subdirectories */
734 static int filter_prep(struct obd_device *obd)
736 struct obd_run_ctxt saved;
737 struct filter_obd *filter = &obd->u.filter;
743 push_ctxt(&saved, &obd->obd_ctxt, NULL);
744 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
745 if (!file || IS_ERR(file)) {
747 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
752 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
753 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
754 file->f_dentry->d_inode->i_mode);
755 GOTO(err_filp, rc = -ENOENT);
758 /* steal operations */
759 inode = file->f_dentry->d_inode;
760 filter->fo_fop = file->f_op;
761 filter->fo_iop = inode->i_op;
762 filter->fo_aops = inode->i_mapping->a_ops;
764 rc = filter_init_server_data(obd, file);
766 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
769 filter->fo_rcvd_filp = file;
771 rc = filter_prep_groups(obd);
773 GOTO(err_server_data, rc);
776 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
781 //class_disconnect_exports(obd, 0);
782 filter_free_server_data(filter);
784 if (filp_close(file, 0))
785 CERROR("can't close %s after error\n", LAST_RCVD);
786 filter->fo_rcvd_filp = NULL;
790 /* cleanup the filter: write last used object id to status file */
791 static void filter_post(struct obd_device *obd)
793 struct obd_run_ctxt saved;
794 struct filter_obd *filter = &obd->u.filter;
797 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
798 * best to start a transaction with h_sync, because we removed this
801 push_ctxt(&saved, &obd->obd_ctxt, NULL);
802 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
805 CERROR("error writing server data: rc = %d\n", rc);
807 for (i = 0; i < FILTER_GROUPS; i++) {
808 rc = filter_update_last_objid(obd, i, (i == FILTER_GROUPS - 1));
810 CERROR("error writing group %d lastobjid: rc = %d\n",
814 filp_close(filter->fo_rcvd_filp, 0);
815 filter->fo_rcvd_filp = NULL;
817 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
819 filter_cleanup_groups(obd);
820 f_dput(filter->fo_dentry_O);
821 filter_free_server_data(filter);
822 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
825 static void filter_set_last_id(struct filter_obd *filter, struct obdo *oa,
829 LASSERT(filter->fo_fsd != NULL);
832 LASSERT(oa->o_gr <= FILTER_GROUPS);
836 spin_lock(&filter->fo_objidlock);
837 filter->fo_last_objids[group] = id;
838 spin_unlock(&filter->fo_objidlock);
841 __u64 filter_last_id(struct filter_obd *filter, struct obdo *oa)
845 LASSERT(filter->fo_fsd != NULL);
848 LASSERT(oa->o_gr <= FILTER_GROUPS);
852 /* FIXME: object groups */
853 spin_lock(&filter->fo_objidlock);
854 id = filter->fo_last_objids[group];
855 spin_unlock(&filter->fo_objidlock);
860 /* direct cut-n-paste of mds_blocking_ast() */
861 static int filter_blocking_ast(struct ldlm_lock *lock,
862 struct ldlm_lock_desc *desc,
863 void *data, int flag)
868 if (flag == LDLM_CB_CANCELING) {
869 /* Don't need to do anything here. */
873 /* XXX layering violation! -phil */
874 l_lock(&lock->l_resource->lr_namespace->ns_lock);
875 /* Get this: if filter_blocking_ast is racing with ldlm_intent_policy,
876 * such that filter_blocking_ast is called just before l_i_p takes the
877 * ns_lock, then by the time we get the lock, we might not be the
878 * correct blocking function anymore. So check, and return early, if
880 if (lock->l_blocking_ast != filter_blocking_ast) {
881 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
885 lock->l_flags |= LDLM_FL_CBPENDING;
886 do_ast = (!lock->l_readers && !lock->l_writers);
887 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
890 struct lustre_handle lockh;
893 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
894 ldlm_lock2handle(lock, &lockh);
895 rc = ldlm_cli_cancel(&lockh);
897 CERROR("ldlm_cli_cancel: %d\n", rc);
899 LDLM_DEBUG(lock, "Lock still has references, will be "
905 static int filter_lock_dentry(struct obd_device *obd, struct dentry *de,
906 ldlm_mode_t lock_mode,struct lustre_handle *lockh)
908 struct ldlm_res_id res_id = { .name = {0} };
912 res_id.name[0] = de->d_inode->i_ino;
913 res_id.name[1] = de->d_inode->i_generation;
914 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
915 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
916 &flags, ldlm_completion_ast,
917 filter_blocking_ast, NULL, lockh);
919 RETURN(rc == ELDLM_OK ? 0 : -EIO); /* XXX translate ldlm code */
922 /* We never dget the object parent, so DON'T dput it either */
923 static void filter_parent_unlock(struct dentry *dparent,
924 struct lustre_handle *lockh,
925 ldlm_mode_t lock_mode)
927 ldlm_lock_decref(lockh, lock_mode);
930 /* We never dget the object parent, so DON'T dput it either */
931 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
933 struct filter_obd *filter = &obd->u.filter;
934 LASSERT(group < FILTER_GROUPS); /* FIXME: object groups */
936 if (group > 0 || filter->fo_subdir_count == 0)
937 return filter->fo_dentry_O_groups[group];
939 return filter->fo_dentry_O_sub[objid & (filter->fo_subdir_count - 1)];
942 /* We never dget the object parent, so DON'T dput it either */
943 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
944 obd_id objid, ldlm_mode_t lock_mode,
945 struct lustre_handle *lockh)
947 unsigned long now = jiffies;
948 struct dentry *de = filter_parent(obd, group, objid);
954 rc = filter_lock_dentry(obd, de, lock_mode, lockh);
955 if (time_after(jiffies, now + 15 * HZ))
956 CERROR("slow parent lock %lus\n", (jiffies - now) / HZ);
957 return rc ? ERR_PTR(rc) : de;
960 /* How to get files, dentries, inodes from object id's.
962 * If dir_dentry is passed, the caller has already locked the parent
963 * appropriately for this operation (normally a write lock). If
964 * dir_dentry is NULL, we do a read lock while we do the lookup to
965 * avoid races with create/destroy and such changing the directory
966 * internal to the filesystem code. */
967 struct dentry *filter_fid2dentry(struct obd_device *obd,
968 struct dentry *dir_dentry,
969 obd_gr group, obd_id id)
971 struct lustre_handle lockh;
972 struct dentry *dparent = dir_dentry;
973 struct dentry *dchild;
979 CERROR("fatal: invalid object id 0\n");
980 RETURN(ERR_PTR(-ESTALE));
983 len = sprintf(name, LPU64, id);
984 if (dir_dentry == NULL) {
985 dparent = filter_parent_lock(obd, group, id, LCK_PR, &lockh);
989 CDEBUG(D_INODE, "looking up object O/%*s/%s\n",
990 dparent->d_name.len, dparent->d_name.name, name);
991 dchild = ll_lookup_one_len(name, dparent, len);
992 if (dir_dentry == NULL)
993 filter_parent_unlock(dparent, &lockh, LCK_PR);
994 if (IS_ERR(dchild)) {
995 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
999 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1000 name, dchild, atomic_read(&dchild->d_count));
1002 LASSERT(atomic_read(&dchild->d_count) > 0);
1007 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid)
1009 struct lustre_handle lockh;
1010 int flags = LDLM_AST_DISCARD_DATA, rc;
1011 struct ldlm_res_id res_id = { .name = { objid } };
1012 struct ldlm_extent extent = { 0, OBD_OBJECT_EOF };
1015 /* Tell the clients that the object is gone now and that they should
1016 * throw away any cached pages. If we're the OST at stripe 0 in the
1017 * file then this enqueue will communicate the DISCARD to all the
1018 * clients. This assumes that we always destroy all the objects for
1019 * a file at a time, as is currently the case. If we're not the
1020 * OST at stripe 0 then we'll harmlessly get a very lonely lock in
1021 * the local DLM and immediately drop it. */
1022 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
1023 res_id, LDLM_EXTENT, &extent,
1024 sizeof(extent), LCK_PW, &flags,
1025 ldlm_completion_ast, filter_blocking_ast,
1028 /* We only care about the side-effects, just drop the lock. */
1030 ldlm_lock_decref(&lockh, LCK_PW);
1035 /* Caller must hold LCK_PW on parent and push us into kernel context.
1036 * Caller is also required to ensure that dchild->d_inode exists. */
1037 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1038 struct dentry *dparent,
1039 struct dentry *dchild)
1041 struct inode *inode = dchild->d_inode;
1045 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1046 CERROR("destroying objid %*s nlink = %lu, count = %d\n",
1047 dchild->d_name.len, dchild->d_name.name,
1048 (unsigned long)inode->i_nlink,
1049 atomic_read(&inode->i_count));
1052 rc = vfs_unlink(dparent->d_inode, dchild);
1055 CERROR("error unlinking objid %*s: rc %d\n",
1056 dchild->d_name.len, dchild->d_name.name, rc);
1061 /* mount the file system (secretly) */
1062 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1065 struct lustre_cfg* lcfg = buf;
1066 struct filter_obd *filter = &obd->u.filter;
1067 struct vfsmount *mnt;
1071 dev_clear_rdonly(2);
1073 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1076 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1077 if (IS_ERR(obd->obd_fsops))
1078 RETURN(PTR_ERR(obd->obd_fsops));
1080 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, MS_NOATIME | MS_NODIRATIME,
1081 lcfg->lcfg_inlbuf1, option);
1086 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1087 if (*lcfg->lcfg_inlbuf3 == 'f') {
1088 obd->obd_replayable = 1;
1089 obd_sync_filter = 1;
1090 CWARN("%s: recovery enabled\n", obd->obd_name);
1092 if (*lcfg->lcfg_inlbuf3 != 'n') {
1093 CERROR("unrecognised flag '%c'\n",
1094 *lcfg->lcfg_inlbuf3);
1096 // XXX Robert? Why do we get errors here
1097 // GOTO(err_mntput, rc = -EINVAL);
1101 filter->fo_vfsmnt = mnt;
1102 filter->fo_sb = mnt->mnt_sb;
1103 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1104 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1106 OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
1107 obd->obd_ctxt.pwdmnt = mnt;
1108 obd->obd_ctxt.pwd = mnt->mnt_root;
1109 obd->obd_ctxt.fs = get_ds();
1110 obd->obd_ctxt.cb_ops = filter_lvfs_ops;
1112 rc = filter_prep(obd);
1114 GOTO(err_mntput, rc);
1116 spin_lock_init(&filter->fo_translock);
1117 spin_lock_init(&filter->fo_objidlock);
1118 INIT_LIST_HEAD(&filter->fo_export_list);
1119 sema_init(&filter->fo_alloc_lock, 1);
1120 spin_lock_init(&filter->fo_r_pages.oh_lock);
1121 spin_lock_init(&filter->fo_w_pages.oh_lock);
1122 spin_lock_init(&filter->fo_r_discont_pages.oh_lock);
1123 spin_lock_init(&filter->fo_w_discont_pages.oh_lock);
1124 spin_lock_init(&filter->fo_r_discont_blocks.oh_lock);
1125 spin_lock_init(&filter->fo_w_discont_blocks.oh_lock);
1126 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1128 obd->obd_namespace = ldlm_namespace_new("filter-tgt",
1129 LDLM_NAMESPACE_SERVER);
1130 if (obd->obd_namespace == NULL)
1131 GOTO(err_post, rc = -ENOMEM);
1133 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1134 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1136 rc = llog_cat_initialize(obd, 1);
1138 CERROR("failed to setup llogging subsystems\n");
1152 fsfilt_put_ops(obd->obd_fsops);
1156 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1158 struct lustre_cfg* lcfg = buf;
1159 const char *str = NULL;
1160 char *option = NULL;
1164 if (!strcmp(lcfg->lcfg_inlbuf2, "ext3")) {
1165 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1166 /* bug 1577: implement async-delete for 2.5 */
1167 str = "errors=remount-ro,asyncdel";
1169 str = "errors=remount-ro";
1171 n = strlen(str) + 1;
1172 OBD_ALLOC(option, n);
1175 strcpy(option, str);
1178 rc = filter_common_setup(obd, len, buf, option);
1180 OBD_FREE(option, n);
1184 static int filter_cleanup(struct obd_device *obd, int flags)
1186 struct filter_obd *filter = &obd->u.filter;
1189 if (flags & OBD_OPT_FAILOVER)
1190 CERROR("%s: shutting down for failover; client state will"
1191 " be preserved.\n", obd->obd_name);
1193 if (!list_empty(&obd->obd_exports)) {
1194 CERROR("%s: still has clients!\n", obd->obd_name);
1195 class_disconnect_exports(obd, flags);
1196 if (!list_empty(&obd->obd_exports)) {
1197 CERROR("still has exports after forced cleanup?\n");
1202 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1204 if (filter->fo_sb == NULL)
1209 shrink_dcache_parent(filter->fo_sb->s_root);
1212 if (atomic_read(&filter->fo_vfsmnt->mnt_count) > 1)
1213 CERROR("%s: mount point %p busy, mnt_count: %d\n",
1214 obd->obd_name, filter->fo_vfsmnt,
1215 atomic_read(&filter->fo_vfsmnt->mnt_count));
1218 mntput(filter->fo_vfsmnt);
1219 //destroy_buffers(filter->fo_sb->s_dev);
1220 filter->fo_sb = NULL;
1221 fsfilt_put_ops(obd->obd_fsops);
1224 dev_clear_rdonly(2);
1229 static int filter_attach(struct obd_device *obd, obd_count len, void *data)
1231 struct lprocfs_static_vars lvars;
1234 lprocfs_init_vars(filter, &lvars);
1235 rc = lprocfs_obd_attach(obd, lvars.obd_vars);
1239 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
1243 /* Init obdfilter private stats here */
1244 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1245 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
1246 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1247 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
1249 return lproc_filter_attach_seqstat(obd);
1252 static int filter_detach(struct obd_device *dev)
1254 lprocfs_free_obd_stats(dev);
1255 return lprocfs_obd_detach(dev);
1258 /* nearly identical to mds_connect */
1259 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
1260 struct obd_uuid *cluuid)
1262 struct obd_export *exp;
1263 struct filter_export_data *fed;
1264 struct filter_client_data *fcd = NULL;
1265 struct filter_obd *filter = &obd->u.filter;
1269 if (conn == NULL || obd == NULL || cluuid == NULL)
1272 rc = class_connect(conn, obd, cluuid);
1275 exp = class_conn2export(conn);
1276 LASSERT(exp != NULL);
1278 fed = &exp->exp_filter_data;
1280 spin_lock_init(&fed->fed_lock);
1282 if (!obd->obd_replayable)
1283 GOTO(cleanup, rc = 0);
1285 OBD_ALLOC(fcd, sizeof(*fcd));
1287 CERROR("filter: out of memory for client data\n");
1288 GOTO(cleanup, rc = -ENOMEM);
1291 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
1294 rc = filter_client_add(obd, filter, fed, -1);
1299 OBD_FREE(fcd, sizeof(*fcd));
1300 class_disconnect(exp, 0);
1302 class_export_put(exp);
1307 static int filter_precleanup(struct obd_device *obd, int flags)
1312 rc = obd_llog_finish(obd, 0);
1314 CERROR("failed to cleanup llogging subsystem\n");
1319 static int filter_destroy_export(struct obd_export *exp)
1323 target_destroy_export(exp);
1325 if (exp->exp_obd->obd_replayable)
1326 filter_client_free(exp, exp->exp_flags);
1330 /* also incredibly similar to mds_disconnect */
1331 static int filter_disconnect(struct obd_export *exp, int flags)
1333 unsigned long irqflags;
1334 struct llog_ctxt *ctxt;
1339 ldlm_cancel_locks_for_export(exp);
1341 spin_lock_irqsave(&exp->exp_lock, irqflags);
1342 exp->exp_flags = flags;
1343 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
1345 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.filter.fo_sb);
1346 /* XXX cleanup preallocated inodes */
1348 /* flush any remaining cancel messages out to the target */
1349 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_REPL_CTXT);
1350 llog_sync(ctxt, exp);
1352 rc = class_disconnect(exp, flags);
1356 struct dentry *__filter_oa2dentry(struct obd_device *obd,
1357 struct obdo *oa, const char *what)
1359 struct dentry *dchild = NULL;
1362 if (oa->o_valid & OBD_MD_FLGROUP)
1365 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
1367 if (IS_ERR(dchild)) {
1368 CERROR("%s error looking up object: "LPU64"\n", what, oa->o_id);
1372 if (dchild->d_inode == NULL) {
1373 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
1375 RETURN(ERR_PTR(-ENOENT));
1381 static int filter_getattr(struct obd_export *exp, struct obdo *oa,
1382 struct lov_stripe_md *md)
1384 struct dentry *dentry = NULL;
1385 struct obd_device *obd;
1389 obd = class_exp2obd(exp);
1391 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1392 exp->exp_handle.h_cookie);
1396 dentry = filter_oa2dentry(obd, oa);
1398 RETURN(PTR_ERR(dentry));
1400 /* Limit the valid bits in the return data to what we actually use */
1401 oa->o_valid = OBD_MD_FLID;
1402 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1408 /* this is called from filter_truncate() until we have filter_punch() */
1409 static int filter_setattr(struct obd_export *exp, struct obdo *oa,
1410 struct lov_stripe_md *md, struct obd_trans_info *oti)
1412 struct obd_run_ctxt saved;
1413 struct filter_obd *filter;
1414 struct dentry *dentry;
1420 LASSERT(oti != NULL);
1422 dentry = filter_oa2dentry(exp->exp_obd, oa);
1424 RETURN(PTR_ERR(dentry));
1426 filter = &exp->exp_obd->u.filter;
1428 iattr_from_obdo(&iattr, oa, oa->o_valid);
1430 push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1433 if (iattr.ia_valid & ATTR_SIZE)
1434 down(&dentry->d_inode->i_sem);
1435 handle = fsfilt_start(exp->exp_obd, dentry->d_inode, FSFILT_OP_SETATTR,
1438 GOTO(out_unlock, rc = PTR_ERR(handle));
1440 /* XXX this could be a rwsem instead, if filter_preprw played along */
1441 if (iattr.ia_valid & ATTR_ATTR_FLAG)
1442 rc = fsfilt_iocontrol(exp->exp_obd, dentry->d_inode, NULL,
1444 (long)&iattr.ia_attr_flags);
1446 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
1447 rc = filter_finish_transno(exp, oti, rc);
1448 rc2 = fsfilt_commit(exp->exp_obd, dentry->d_inode, handle, 0);
1450 CERROR("error on commit, err = %d\n", rc2);
1455 oa->o_valid = OBD_MD_FLID;
1456 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1459 if (iattr.ia_valid & ATTR_SIZE)
1460 up(&dentry->d_inode->i_sem);
1462 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1468 /* XXX identical to osc_unpackmd */
1469 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
1470 struct lov_mds_md *lmm, int lmm_bytes)
1476 if (lmm_bytes < sizeof (*lmm)) {
1477 CERROR("lov_mds_md too small: %d, need %d\n",
1478 lmm_bytes, (int)sizeof(*lmm));
1481 /* XXX LOV_MAGIC etc check? */
1483 if (lmm->lmm_object_id == cpu_to_le64(0)) {
1484 CERROR("lov_mds_md: zero lmm_object_id\n");
1489 lsm_size = lov_stripe_md_size(1);
1493 if (*lsmp != NULL && lmm == NULL) {
1494 OBD_FREE(*lsmp, lsm_size);
1499 if (*lsmp == NULL) {
1500 OBD_ALLOC(*lsmp, lsm_size);
1504 loi_init((*lsmp)->lsm_oinfo);
1508 /* XXX zero *lsmp? */
1509 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
1510 LASSERT((*lsmp)->lsm_object_id);
1513 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
1518 static void filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
1519 struct filter_obd *filter)
1521 struct obdo doa; /* XXX obdo on stack */
1526 memset(&doa, 0, sizeof(doa));
1527 if (oa->o_valid & OBD_MD_FLGROUP)
1528 doa.o_gr = oa->o_gr;
1531 doa.o_mode = S_IFREG;
1532 last = filter_last_id(filter, &doa); /* FIXME: object groups */
1533 CWARN("deleting orphan objects from "LPU64" to "LPU64"\n",
1534 oa->o_id + 1, last);
1535 for (id = oa->o_id + 1; id <= last; id++) {
1537 filter_destroy(exp, &doa, NULL, NULL);
1539 spin_lock(&filter->fo_objidlock);
1540 filter->fo_last_objids[0] = oa->o_id; /* FIXME: object groups */
1541 spin_unlock(&filter->fo_objidlock);
1545 /* returns a negative error or a nonnegative number of files to create */
1546 static int filter_should_precreate(struct obd_export *exp, struct obdo *oa,
1549 struct obd_device *obd = exp->exp_obd;
1550 struct filter_obd *filter = &obd->u.filter;
1554 diff = oa->o_id - filter_last_id(filter, oa);
1555 CDEBUG(D_INFO, "filter_last_id() = "LPU64" -> diff = %d\n",
1556 filter_last_id(filter, oa), diff);
1558 /* delete orphans request */
1559 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1560 (oa->o_flags & OBD_FL_DELORPHAN)) {
1563 filter_destroy_precreated(exp, oa, filter);
1564 rc = filter_update_last_objid(obd, group, 0);
1566 CERROR("unable to write lastobjid, but orphans"
1570 /* only precreate if group == 0 and o_id is specfied */
1571 if (!(oa->o_valid & OBD_FL_DELORPHAN) &&
1572 (group != 0 || oa->o_id == 0))
1581 /* We rely on the fact that only one thread will be creating files in a given
1582 * group at a time, which is why we don't need an atomic filter_get_new_id.
1583 * Even if we had that atomic function, the following race would exist:
1585 * thread 1: gets id x from filter_next_id
1586 * thread 2: gets id (x + 1) from filter_next_id
1587 * thread 2: creates object (x + 1)
1588 * thread 1: tries to create object x, gets -ENOSPC
1590 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
1591 obd_gr group, int *num)
1593 struct lustre_handle parent_lockh;
1594 struct dentry *dchild = NULL;
1595 struct filter_obd *filter;
1596 struct dentry *dparent;
1597 int err = 0, rc = 0, i;
1599 void *handle = NULL;
1602 filter = &obd->u.filter;
1604 for (i = 0; i < *num && err == 0; i++) {
1605 int cleanup_phase = 0;
1607 next_id = filter_last_id(filter, oa) + 1;
1608 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
1610 dparent = filter_parent_lock(obd, group, next_id, LCK_PW,
1612 if (IS_ERR(dparent))
1613 GOTO(cleanup, rc = PTR_ERR(dparent));
1616 dchild = filter_fid2dentry(obd, dparent, group, next_id);
1618 GOTO(cleanup, rc = PTR_ERR(dchild));
1621 if (dchild->d_inode != NULL) {
1622 /* This would only happen if lastobjid was bad on disk*/
1623 CERROR("Serious error: objid %*s already exists; is "
1624 "this filesystem corrupt?\n",
1625 dchild->d_name.len, dchild->d_name.name);
1626 GOTO(cleanup, rc = -EEXIST);
1629 handle = fsfilt_start(obd, dparent->d_inode,
1630 FSFILT_OP_CREATE_LOG, NULL);
1632 GOTO(cleanup, rc = PTR_ERR(handle));
1635 rc = ll_vfs_create(dparent->d_inode, dchild, S_IFREG, NULL);
1637 CERROR("create failed rc = %d\n", rc);
1641 filter_set_last_id(filter, oa, next_id);
1642 err = filter_update_last_objid(obd, group, 0);
1644 CERROR("unable to write lastobjid but file created\n");
1647 switch(cleanup_phase) {
1649 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
1651 CERROR("error on commit, err = %d\n", err);
1658 filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1668 CDEBUG(D_INFO, "filter_precreate() created %d objects\n", i);
1672 static int filter_create(struct obd_export *exp, struct obdo *oa,
1673 struct lov_stripe_md **ea, struct obd_trans_info *oti)
1675 struct obd_device *obd = NULL;
1676 struct obd_run_ctxt saved;
1677 struct lov_stripe_md *lsm = NULL;
1682 if (oa->o_valid & OBD_MD_FLGROUP)
1685 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
1690 rc = obd_alloc_memmd(exp, &lsm);
1697 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1699 diff = filter_should_precreate(exp, oa, group);
1701 oa->o_id = filter_last_id(&obd->u.filter, oa);
1702 rc = filter_precreate(obd, oa, group, &diff);
1704 oa->o_valid = OBD_MD_FLID;
1707 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1708 if (rc && ea != NULL && *ea != lsm) {
1709 obd_free_memmd(exp, &lsm);
1710 } else if (rc == 0 && ea != NULL) {
1711 /* XXX LOV STACKING: the lsm that is passed to us from
1712 * LOV does not have valid lsm_oinfo data structs, so
1713 * don't go touching that. This needs to be fixed in a
1715 lsm->lsm_object_id = oa->o_id;
1722 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
1723 struct lov_stripe_md *ea, struct obd_trans_info *oti)
1725 struct obd_device *obd;
1726 struct filter_obd *filter;
1727 struct dentry *dchild = NULL, *dparent = NULL;
1728 struct obd_run_ctxt saved;
1729 void *handle = NULL;
1730 struct lustre_handle parent_lockh;
1731 struct llog_cookie *fcc = NULL;
1732 int rc, rc2, cleanup_phase = 0, have_prepared = 0;
1736 if (oa->o_valid & OBD_MD_FLGROUP)
1740 filter = &obd->u.filter;
1742 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1745 dparent = filter_parent_lock(obd, group, oa->o_id, LCK_PW,
1747 if (IS_ERR(dparent))
1748 GOTO(cleanup, rc = PTR_ERR(dparent));
1751 dchild = filter_fid2dentry(obd, dparent, group, oa->o_id);
1753 GOTO(cleanup, rc = -ENOENT);
1756 if (dchild->d_inode == NULL) {
1757 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
1759 GOTO(cleanup, rc = -ENOENT);
1762 if (!have_prepared) {
1763 /* If we're really going to destroy the object, get ready
1764 * by getting the clients to discard their cached data.
1766 * We have to drop the parent lock, because
1767 * filter_prepare_destroy will acquire a PW on the object, and
1768 * we don't want to deadlock with an incoming write to the
1769 * object, which has the extent PW and then wants to get the
1770 * parent dentry to do the lookup.
1772 * We dput the child because it's not worth the extra
1773 * complication of condition the above code to skip it on the
1774 * second time through. */
1776 filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1778 filter_prepare_destroy(obd, oa->o_id);
1783 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK_LOG, oti);
1785 GOTO(cleanup, rc = PTR_ERR(handle));
1788 /* Our MDC connection is established by the MDS to us */
1789 if (oa->o_valid & OBD_MD_FLCOOKIE) {
1790 OBD_ALLOC(fcc, sizeof(*fcc));
1792 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
1795 rc = filter_destroy_internal(obd, oa->o_id, dparent, dchild);
1798 switch(cleanup_phase) {
1802 fsfilt_add_journal_cb(obd, 0, oti->oti_handle,
1803 filter_cancel_cookies_cb, fcc);
1805 fsfilt_add_journal_cb(obd, 0, handle,
1806 filter_cancel_cookies_cb, fcc);
1809 rc = filter_finish_transno(exp, oti, rc);
1810 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
1812 CERROR("error on commit, err = %d\n", rc2);
1819 if (rc || oti == NULL) {
1820 filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1822 memcpy(&oti->oti_ack_locks[0].lock, &parent_lockh,
1823 sizeof(parent_lockh));
1824 oti->oti_ack_locks[0].mode = LCK_PW;
1827 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1830 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1837 /* NB start and end are used for punch, but not truncate */
1838 static int filter_truncate(struct obd_export *exp, struct obdo *oa,
1839 struct lov_stripe_md *lsm,
1840 obd_off start, obd_off end,
1841 struct obd_trans_info *oti)
1846 if (end != OBD_OBJECT_EOF)
1847 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
1850 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = %x, "
1851 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
1853 error = filter_setattr(exp, oa, NULL, oti);
1857 static int filter_sync(struct obd_export *exp, struct obdo *oa,
1858 struct lov_stripe_md *lsm, obd_off start, obd_off end)
1860 struct obd_run_ctxt saved;
1861 struct filter_obd *filter;
1862 struct dentry *dentry;
1863 struct llog_ctxt *ctxt;
1867 filter = &exp->exp_obd->u.filter;
1869 /* an objid of zero is taken to mean "sync whole filesystem" */
1870 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
1871 rc = fsfilt_sync(exp->exp_obd, filter->fo_sb);
1872 /* flush any remaining cancel messages out to the target */
1873 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_REPL_CTXT);
1874 llog_sync(ctxt, exp);
1878 dentry = filter_oa2dentry(exp->exp_obd, oa);
1880 RETURN(PTR_ERR(dentry));
1882 push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1884 down(&dentry->d_inode->i_sem);
1885 rc = filemap_fdatasync(dentry->d_inode->i_mapping);
1887 /* just any file to grab fsync method - "file" arg unused */
1888 struct file *file = filter->fo_rcvd_filp;
1890 if (file->f_op && file->f_op->fsync)
1891 rc = file->f_op->fsync(NULL, dentry, 1);
1893 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
1897 up(&dentry->d_inode->i_sem);
1899 oa->o_valid = OBD_MD_FLID;
1900 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1902 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1908 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1909 unsigned long max_age)
1912 RETURN(fsfilt_statfs(obd, obd->u.filter.fo_sb, osfs));
1915 static int filter_get_info(struct obd_export *exp, __u32 keylen,
1916 void *key, __u32 *vallen, void *val)
1918 struct obd_device *obd;
1921 obd = class_exp2obd(exp);
1923 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1924 exp->exp_handle.h_cookie);
1928 if (keylen == strlen("blocksize") &&
1929 memcmp(key, "blocksize", keylen) == 0) {
1930 __u32 *blocksize = val;
1931 *vallen = sizeof(*blocksize);
1932 *blocksize = obd->u.filter.fo_sb->s_blocksize;
1936 if (keylen == strlen("blocksize_bits") &&
1937 memcmp(key, "blocksize_bits", keylen) == 0) {
1938 __u32 *blocksize_bits = val;
1939 *vallen = sizeof(*blocksize_bits);
1940 *blocksize_bits = obd->u.filter.fo_sb->s_blocksize_bits;
1944 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
1945 obd_id *last_id = val;
1946 /* FIXME: object groups */
1947 *last_id = filter_last_id(&obd->u.filter, 0);
1950 CDEBUG(D_IOCTL, "invalid key\n");
1954 static int filter_set_info(struct obd_export *exp, __u32 keylen,
1955 void *key, __u32 vallen, void *val)
1957 struct obd_device *obd;
1958 struct lustre_handle conn;
1959 struct llog_ctxt *ctxt;
1963 conn.cookie = exp->exp_handle.h_cookie;
1967 CDEBUG(D_IOCTL, "invalid exp %p cookie "LPX64"\n",
1972 if (keylen < strlen("mds_conn") ||
1973 memcmp(key, "mds_conn", keylen) != 0)
1976 CWARN("Received MDS connection ("LPX64")\n", conn.cookie);
1977 memcpy(&obd->u.filter.fo_mdc_conn, &conn, sizeof(conn));
1978 ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
1979 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
1983 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
1984 int len, void *karg, void *uarg)
1986 struct obd_device *obd = exp->exp_obd;
1987 struct obd_ioctl_data *data = karg;
1991 case OBD_IOC_ABORT_RECOVERY:
1992 CERROR("aborting recovery for device %s\n", obd->obd_name);
1993 target_abort_recovery(obd);
1996 case OBD_IOC_SET_READONLY: {
1998 struct super_block *sb = obd->u.filter.fo_sb;
1999 struct inode *inode = sb->s_root->d_inode;
2000 BDEVNAME_DECLARE_STORAGE(tmp);
2001 CERROR("setting device %s read-only\n",
2002 ll_bdevname(sb, tmp));
2004 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
2006 (void)fsfilt_commit(obd, inode, handle, 1);
2008 dev_set_rdonly(ll_sbdev(obd->u.filter.fo_sb), 2);
2012 case OBD_IOC_CATLOGLIST: {
2013 rc = llog_catlog_list(obd, 1, data);
2017 case OBD_IOC_LLOG_CANCEL:
2018 case OBD_IOC_LLOG_REMOVE:
2019 case OBD_IOC_LLOG_INFO:
2020 case OBD_IOC_LLOG_PRINT: {
2021 /* FIXME to be finished */
2022 RETURN(-EOPNOTSUPP);
2024 struct llog_ctxt *ctxt = NULL;
2026 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2027 rc = llog_ioctl(ctxt, cmd, data);
2028 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2041 static struct llog_operations filter_unlink_repl_logops;
2042 static struct llog_operations filter_size_orig_logops = {
2043 lop_setup: llog_obd_origin_setup,
2044 lop_cleanup: llog_obd_origin_cleanup,
2045 lop_add: llog_obd_origin_add
2048 static int filter_llog_init(struct obd_device *obd, struct obd_device *tgt,
2049 int count, struct llog_logid *logid)
2051 struct llog_ctxt *ctxt;
2055 filter_unlink_repl_logops = llog_client_ops;
2056 filter_unlink_repl_logops.lop_cancel = llog_obd_repl_cancel;
2057 filter_unlink_repl_logops.lop_connect = llog_repl_connect;
2058 filter_unlink_repl_logops.lop_sync = llog_obd_repl_sync;
2060 rc = llog_setup(obd, LLOG_UNLINK_REPL_CTXT, tgt, 0, NULL,
2061 &filter_unlink_repl_logops);
2064 /* FIXME - assign unlink_cb for filter's recovery */
2065 ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
2066 ctxt->llog_proc_cb = filter_recov_log_unlink_cb;
2068 rc = llog_setup(obd, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2069 &filter_size_orig_logops);
2073 static int filter_llog_finish(struct obd_device *obd, int count)
2078 rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_REPL_CTXT));
2082 rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_ORIG_CTXT));
2086 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2089 return filter_fid2dentry(data, NULL, gr, id);
2092 static struct lvfs_callback_ops filter_lvfs_ops = {
2093 l_fid2dentry: filter_lvfs_fid2dentry,
2096 static struct obd_ops filter_obd_ops = {
2097 o_owner: THIS_MODULE,
2098 o_attach: filter_attach,
2099 o_detach: filter_detach,
2100 o_get_info: filter_get_info,
2101 o_set_info: filter_set_info,
2102 o_setup: filter_setup,
2103 o_precleanup: filter_precleanup,
2104 o_cleanup: filter_cleanup,
2105 o_connect: filter_connect,
2106 o_disconnect: filter_disconnect,
2107 o_statfs: filter_statfs,
2108 o_getattr: filter_getattr,
2109 o_unpackmd: filter_unpackmd,
2110 o_create: filter_create,
2111 o_setattr: filter_setattr,
2112 o_destroy: filter_destroy,
2114 o_punch: filter_truncate,
2115 o_sync: filter_sync,
2116 o_preprw: filter_preprw,
2117 o_commitrw: filter_commitrw,
2118 o_destroy_export: filter_destroy_export,
2119 o_llog_init: filter_llog_init,
2120 o_llog_finish: filter_llog_finish,
2121 o_iocontrol: filter_iocontrol,
2124 static struct obd_ops filter_sanobd_ops = {
2125 o_owner: THIS_MODULE,
2126 o_attach: filter_attach,
2127 o_detach: filter_detach,
2128 o_get_info: filter_get_info,
2129 o_set_info: filter_set_info,
2130 o_setup: filter_san_setup,
2131 o_precleanup: filter_precleanup,
2132 o_cleanup: filter_cleanup,
2133 o_connect: filter_connect,
2134 o_disconnect: filter_disconnect,
2135 o_statfs: filter_statfs,
2136 o_getattr: filter_getattr,
2137 o_unpackmd: filter_unpackmd,
2138 o_create: filter_create,
2139 o_setattr: filter_setattr,
2140 o_destroy: filter_destroy,
2142 o_punch: filter_truncate,
2143 o_sync: filter_sync,
2144 o_preprw: filter_preprw,
2145 o_commitrw: filter_commitrw,
2146 o_san_preprw: filter_san_preprw,
2147 o_destroy_export: filter_destroy_export,
2148 o_llog_init: filter_llog_init,
2149 o_llog_finish: filter_llog_finish,
2150 o_iocontrol: filter_iocontrol,
2153 static int __init obdfilter_init(void)
2155 struct lprocfs_static_vars lvars;
2158 printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
2160 lprocfs_init_vars(filter, &lvars);
2162 rc = class_register_type(&filter_obd_ops, lvars.module_vars,
2163 OBD_FILTER_DEVICENAME);
2167 rc = class_register_type(&filter_sanobd_ops, lvars.module_vars,
2168 OBD_FILTER_SAN_DEVICENAME);
2170 class_unregister_type(OBD_FILTER_DEVICENAME);
2174 static void __exit obdfilter_exit(void)
2176 class_unregister_type(OBD_FILTER_SAN_DEVICENAME);
2177 class_unregister_type(OBD_FILTER_DEVICENAME);
2180 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2181 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
2182 MODULE_LICENSE("GPL");
2184 module_init(obdfilter_init);
2185 module_exit(obdfilter_exit);