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 *dparent)
907 down(&dparent->d_inode->i_sem);
911 /* We never dget the object parent, so DON'T dput it either */
912 static void filter_parent_unlock(struct dentry *dparent)
914 up(&dparent->d_inode->i_sem);
917 /* We never dget the object parent, so DON'T dput it either */
918 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
920 struct filter_obd *filter = &obd->u.filter;
921 LASSERT(group < FILTER_GROUPS); /* FIXME: object groups */
923 if (group > 0 || filter->fo_subdir_count == 0)
924 return filter->fo_dentry_O_groups[group];
926 return filter->fo_dentry_O_sub[objid & (filter->fo_subdir_count - 1)];
929 /* We never dget the object parent, so DON'T dput it either */
930 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
933 unsigned long now = jiffies;
934 struct dentry *dparent = filter_parent(obd, group, objid);
940 rc = filter_lock_dentry(obd, dparent);
941 if (time_after(jiffies, now + 15 * HZ))
942 CERROR("slow parent lock %lus\n", (jiffies - now) / HZ);
943 return rc ? ERR_PTR(rc) : dparent;
946 /* How to get files, dentries, inodes from object id's.
948 * If dir_dentry is passed, the caller has already locked the parent
949 * appropriately for this operation (normally a write lock). If
950 * dir_dentry is NULL, we do a read lock while we do the lookup to
951 * avoid races with create/destroy and such changing the directory
952 * internal to the filesystem code. */
953 struct dentry *filter_fid2dentry(struct obd_device *obd,
954 struct dentry *dir_dentry,
955 obd_gr group, obd_id id)
957 struct dentry *dparent = dir_dentry;
958 struct dentry *dchild;
964 CERROR("fatal: invalid object id 0\n");
965 RETURN(ERR_PTR(-ESTALE));
968 len = sprintf(name, LPU64, id);
969 if (dir_dentry == NULL) {
970 dparent = filter_parent_lock(obd, group, id);
974 CDEBUG(D_INODE, "looking up object O/%*s/%s\n",
975 dparent->d_name.len, dparent->d_name.name, name);
976 dchild = /*ll_*/lookup_one_len(name, dparent, len);
977 if (dir_dentry == NULL)
978 filter_parent_unlock(dparent);
979 if (IS_ERR(dchild)) {
980 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
984 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
985 name, dchild, atomic_read(&dchild->d_count));
987 LASSERT(atomic_read(&dchild->d_count) > 0);
992 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid)
994 struct lustre_handle lockh;
995 int flags = LDLM_AST_DISCARD_DATA, rc;
996 struct ldlm_res_id res_id = { .name = { objid } };
997 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1000 /* Tell the clients that the object is gone now and that they should
1001 * throw away any cached pages. */
1002 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
1003 LDLM_EXTENT, &policy, LCK_PW,
1004 &flags, filter_blocking_ast, ldlm_completion_ast,
1005 NULL, NULL, NULL, 0, NULL, &lockh);
1007 /* We only care about the side-effects, just drop the lock. */
1009 ldlm_lock_decref(&lockh, LCK_PW);
1014 /* Caller must hold LCK_PW on parent and push us into kernel context.
1015 * Caller is also required to ensure that dchild->d_inode exists. */
1016 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1017 struct dentry *dparent,
1018 struct dentry *dchild)
1020 struct inode *inode = dchild->d_inode;
1024 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1025 CERROR("destroying objid %*s nlink = %lu, count = %d\n",
1026 dchild->d_name.len, dchild->d_name.name,
1027 (unsigned long)inode->i_nlink,
1028 atomic_read(&inode->i_count));
1031 rc = vfs_unlink(dparent->d_inode, dchild);
1034 CERROR("error unlinking objid %*s: rc %d\n",
1035 dchild->d_name.len, dchild->d_name.name, rc);
1040 static int filter_intent_policy(struct ldlm_namespace *ns,
1041 struct ldlm_lock **lockp, void *req_cookie,
1042 ldlm_mode_t mode, int flags, void *data)
1044 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1045 struct ptlrpc_request *req = req_cookie;
1046 struct ldlm_lock *lock = *lockp, *l = NULL;
1047 struct ldlm_resource *res = lock->l_resource;
1048 ldlm_processing_policy policy;
1049 struct ost_lvb *res_lvb, *reply_lvb;
1050 struct list_head *tmp;
1052 int tmpflags = 0, rc, repsize[2] = {sizeof(struct ldlm_reply),
1053 sizeof(struct ost_lvb) };
1056 policy = ldlm_get_processing_policy(res);
1057 LASSERT(policy != NULL);
1058 LASSERT(req != NULL);
1060 rc = lustre_pack_reply(req, 2, repsize, NULL);
1062 RETURN(req->rq_status = rc);
1064 reply_lvb = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*reply_lvb));
1065 LASSERT(reply_lvb != NULL);
1067 //fixup_handle_for_resent_req(req, lock, &lockh);
1069 /* If we grant any lock at all, it will be a whole-file read lock.
1070 * Call the extent policy function to see if our request can be
1071 * granted, or is blocked. */
1072 lock->l_policy_data.l_extent.start = 0;
1073 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1074 lock->l_req_mode = LCK_PR;
1076 l_lock(&res->lr_namespace->ns_lock);
1078 res->lr_tmp = &rpc_list;
1079 rc = policy(lock, &tmpflags, 0, &err);
1082 /* FIXME: we should change the policy function slightly, to not make
1083 * this list at all, since we just turn around and free it */
1084 while (!list_empty(&rpc_list)) {
1085 struct ldlm_ast_work *w =
1086 list_entry(rpc_list.next, struct ldlm_ast_work, w_list);
1087 list_del(&w->w_list);
1088 LDLM_LOCK_PUT(w->w_lock);
1089 OBD_FREE(w, sizeof(*w));
1092 if (rc == LDLM_ITER_CONTINUE) {
1093 /* The lock met with no resistance; we're finished. */
1094 l_unlock(&res->lr_namespace->ns_lock);
1095 RETURN(ELDLM_LOCK_REPLACED);
1098 /* Do not grant any lock, but instead send GL callbacks. The extent
1099 * policy nicely created a list of all PW locks for us. We will choose
1100 * the highest of those which are larger than the size in the LVB, if
1101 * any, and perform a glimpse callback. */
1102 down(&res->lr_lvb_sem);
1103 res_lvb = res->lr_lvb_data;
1104 LASSERT(res_lvb != NULL);
1105 reply_lvb->lvb_size = res_lvb->lvb_size;
1106 up(&res->lr_lvb_sem);
1108 list_for_each(tmp, &res->lr_granted) {
1109 struct ldlm_lock *tmplock =
1110 list_entry(tmp, struct ldlm_lock, l_res_link);
1112 if (tmplock->l_granted_mode == LCK_PR)
1115 if (tmplock->l_policy_data.l_extent.end <=
1116 reply_lvb->lvb_size)
1120 l = LDLM_LOCK_GET(tmplock);
1124 if (l->l_policy_data.l_extent.start >
1125 tmplock->l_policy_data.l_extent.start)
1129 l = LDLM_LOCK_GET(tmplock);
1131 l_unlock(&res->lr_namespace->ns_lock);
1133 /* There were no PW locks beyond the size in the LVB; finished. */
1135 RETURN(ELDLM_LOCK_ABORTED);
1137 LASSERT(l->l_glimpse_ast != NULL);
1138 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1140 down(&res->lr_lvb_sem);
1142 if (res_lvb->lvb_size == reply_lvb->lvb_size)
1143 LDLM_ERROR(l, "we lost the glimpse race!");
1145 reply_lvb->lvb_size = res_lvb->lvb_size;
1146 up(&res->lr_lvb_sem);
1150 RETURN(ELDLM_LOCK_ABORTED);
1153 /* mount the file system (secretly) */
1154 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1157 struct lustre_cfg* lcfg = buf;
1158 struct filter_obd *filter = &obd->u.filter;
1159 struct vfsmount *mnt;
1163 dev_clear_rdonly(2);
1165 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1168 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1169 if (IS_ERR(obd->obd_fsops))
1170 RETURN(PTR_ERR(obd->obd_fsops));
1172 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, MS_NOATIME | MS_NODIRATIME,
1173 lcfg->lcfg_inlbuf1, option);
1178 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1179 if (*lcfg->lcfg_inlbuf3 == 'f') {
1180 obd->obd_replayable = 1;
1181 obd_sync_filter = 1;
1182 CWARN("%s: recovery enabled\n", obd->obd_name);
1184 if (*lcfg->lcfg_inlbuf3 != 'n') {
1185 CERROR("unrecognised flag '%c'\n",
1186 *lcfg->lcfg_inlbuf3);
1188 // XXX Robert? Why do we get errors here
1189 // GOTO(err_mntput, rc = -EINVAL);
1193 filter->fo_vfsmnt = mnt;
1194 filter->fo_sb = mnt->mnt_sb;
1195 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1196 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1198 OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
1199 obd->obd_ctxt.pwdmnt = mnt;
1200 obd->obd_ctxt.pwd = mnt->mnt_root;
1201 obd->obd_ctxt.fs = get_ds();
1202 obd->obd_ctxt.cb_ops = filter_lvfs_ops;
1204 rc = filter_prep(obd);
1206 GOTO(err_mntput, rc);
1208 spin_lock_init(&filter->fo_translock);
1209 spin_lock_init(&filter->fo_objidlock);
1210 INIT_LIST_HEAD(&filter->fo_export_list);
1211 sema_init(&filter->fo_alloc_lock, 1);
1212 spin_lock_init(&filter->fo_r_pages.oh_lock);
1213 spin_lock_init(&filter->fo_w_pages.oh_lock);
1214 spin_lock_init(&filter->fo_r_discont_pages.oh_lock);
1215 spin_lock_init(&filter->fo_w_discont_pages.oh_lock);
1216 spin_lock_init(&filter->fo_r_discont_blocks.oh_lock);
1217 spin_lock_init(&filter->fo_w_discont_blocks.oh_lock);
1218 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1220 obd->obd_namespace = ldlm_namespace_new("filter-tgt",
1221 LDLM_NAMESPACE_SERVER);
1222 if (obd->obd_namespace == NULL)
1223 GOTO(err_post, rc = -ENOMEM);
1224 obd->obd_namespace->ns_lvbp = obd;
1225 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1226 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1228 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1229 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1231 rc = llog_cat_initialize(obd, 1);
1233 CERROR("failed to setup llogging subsystems\n");
1247 fsfilt_put_ops(obd->obd_fsops);
1251 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1253 struct lustre_cfg* lcfg = buf;
1254 const char *str = NULL;
1255 char *option = NULL;
1259 if (!strcmp(lcfg->lcfg_inlbuf2, "ext3")) {
1260 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1261 /* bug 1577: implement async-delete for 2.5 */
1262 str = "errors=remount-ro,asyncdel";
1264 str = "errors=remount-ro";
1266 n = strlen(str) + 1;
1267 OBD_ALLOC(option, n);
1270 strcpy(option, str);
1273 rc = filter_common_setup(obd, len, buf, option);
1275 OBD_FREE(option, n);
1279 static int filter_cleanup(struct obd_device *obd, int flags)
1281 struct filter_obd *filter = &obd->u.filter;
1284 if (flags & OBD_OPT_FAILOVER)
1285 CERROR("%s: shutting down for failover; client state will"
1286 " be preserved.\n", obd->obd_name);
1288 if (!list_empty(&obd->obd_exports)) {
1289 CERROR("%s: still has clients!\n", obd->obd_name);
1290 class_disconnect_exports(obd, flags);
1291 if (!list_empty(&obd->obd_exports)) {
1292 CERROR("still has exports after forced cleanup?\n");
1297 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1299 if (filter->fo_sb == NULL)
1304 shrink_dcache_parent(filter->fo_sb->s_root);
1307 if (atomic_read(&filter->fo_vfsmnt->mnt_count) > 1)
1308 CERROR("%s: mount point %p busy, mnt_count: %d\n",
1309 obd->obd_name, filter->fo_vfsmnt,
1310 atomic_read(&filter->fo_vfsmnt->mnt_count));
1313 mntput(filter->fo_vfsmnt);
1314 //destroy_buffers(filter->fo_sb->s_dev);
1315 filter->fo_sb = NULL;
1316 fsfilt_put_ops(obd->obd_fsops);
1319 dev_clear_rdonly(2);
1324 static int filter_attach(struct obd_device *obd, obd_count len, void *data)
1326 struct lprocfs_static_vars lvars;
1329 lprocfs_init_vars(filter, &lvars);
1330 rc = lprocfs_obd_attach(obd, lvars.obd_vars);
1334 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
1338 /* Init obdfilter private stats here */
1339 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1340 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
1341 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1342 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
1344 return lproc_filter_attach_seqstat(obd);
1347 static int filter_detach(struct obd_device *dev)
1349 lprocfs_free_obd_stats(dev);
1350 return lprocfs_obd_detach(dev);
1353 /* nearly identical to mds_connect */
1354 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
1355 struct obd_uuid *cluuid)
1357 struct obd_export *exp;
1358 struct filter_export_data *fed;
1359 struct filter_client_data *fcd = NULL;
1360 struct filter_obd *filter = &obd->u.filter;
1364 if (conn == NULL || obd == NULL || cluuid == NULL)
1367 rc = class_connect(conn, obd, cluuid);
1370 exp = class_conn2export(conn);
1371 LASSERT(exp != NULL);
1373 fed = &exp->exp_filter_data;
1375 spin_lock_init(&fed->fed_lock);
1377 if (!obd->obd_replayable)
1378 GOTO(cleanup, rc = 0);
1380 OBD_ALLOC(fcd, sizeof(*fcd));
1382 CERROR("filter: out of memory for client data\n");
1383 GOTO(cleanup, rc = -ENOMEM);
1386 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
1389 rc = filter_client_add(obd, filter, fed, -1);
1394 OBD_FREE(fcd, sizeof(*fcd));
1395 class_disconnect(exp, 0);
1397 class_export_put(exp);
1402 static int filter_precleanup(struct obd_device *obd, int flags)
1407 rc = obd_llog_finish(obd, 0);
1409 CERROR("failed to cleanup llogging subsystem\n");
1414 static int filter_destroy_export(struct obd_export *exp)
1418 if (exp->exp_filter_data.fed_pending)
1419 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
1420 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1421 exp, exp->exp_filter_data.fed_pending);
1423 target_destroy_export(exp);
1425 if (exp->exp_obd->obd_replayable)
1426 filter_client_free(exp, exp->exp_flags);
1430 /* also incredibly similar to mds_disconnect */
1431 static int filter_disconnect(struct obd_export *exp, int flags)
1433 struct filter_obd *filter = &exp->exp_obd->u.filter;
1434 struct filter_export_data *fed = &exp->exp_filter_data;
1435 unsigned long irqflags;
1436 struct llog_ctxt *ctxt;
1442 /* This would imply RPCs still in flight or preprw/commitrw imbalance */
1443 if (fed->fed_pending)
1444 CWARN("%s: cli %s has %lu pending at disconnect time\n",
1445 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1448 /* Forget what this client had cached. This is also done on the
1449 * client when it invalidates its import. Do this before unlinking
1450 * from the export list so filter_grant_sanity_check totals are OK. */
1451 spin_lock(&exp->exp_obd->obd_osfs_lock);
1452 LASSERTF(exp->exp_obd->u.filter.fo_tot_dirty >= fed->fed_dirty,
1453 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %lu\n",
1454 exp->exp_obd->obd_name, exp->exp_obd->u.filter.fo_tot_dirty,
1455 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
1456 exp->exp_obd->u.filter.fo_tot_dirty -= fed->fed_dirty;
1457 LASSERTF(exp->exp_obd->u.filter.fo_tot_granted >= fed->fed_grant,
1458 "%s: tot_granted "LPU64" cli %s/%p fed_grant %lu\n",
1459 exp->exp_obd->obd_name, exp->exp_obd->u.filter.fo_tot_granted,
1460 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
1461 exp->exp_obd->u.filter.fo_tot_granted -= fed->fed_grant;
1462 LASSERTF(exp->exp_obd->u.filter.fo_tot_pending >= fed->fed_pending,
1463 "%s: tot_pending "LPU64" cli %s/%p fed_pending %lu\n",
1464 exp->exp_obd->obd_name, exp->exp_obd->u.filter.fo_tot_pending,
1465 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
1468 spin_unlock(&exp->exp_obd->obd_osfs_lock);
1470 ldlm_cancel_locks_for_export(exp);
1472 spin_lock_irqsave(&exp->exp_lock, irqflags);
1473 exp->exp_flags = flags;
1474 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
1476 fsfilt_sync(exp->exp_obd, filter->fo_sb);
1478 /* flush any remaining cancel messages out to the target */
1479 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_REPL_CTXT);
1480 llog_sync(ctxt, exp);
1482 rc = class_disconnect(exp, flags);
1486 struct dentry *__filter_oa2dentry(struct obd_device *obd,
1487 struct obdo *oa, const char *what)
1489 struct dentry *dchild = NULL;
1492 if (oa->o_valid & OBD_MD_FLGROUP)
1495 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
1497 if (IS_ERR(dchild)) {
1498 CERROR("%s error looking up object: "LPU64"\n", what, oa->o_id);
1502 if (dchild->d_inode == NULL) {
1503 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
1505 RETURN(ERR_PTR(-ENOENT));
1511 static int filter_getattr(struct obd_export *exp, struct obdo *oa,
1512 struct lov_stripe_md *md)
1514 struct dentry *dentry = NULL;
1515 struct obd_device *obd;
1519 obd = class_exp2obd(exp);
1521 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1522 exp->exp_handle.h_cookie);
1526 dentry = filter_oa2dentry(obd, oa);
1528 RETURN(PTR_ERR(dentry));
1530 /* Limit the valid bits in the return data to what we actually use */
1531 oa->o_valid = OBD_MD_FLID;
1532 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1538 /* this is called from filter_truncate() until we have filter_punch() */
1539 static int filter_setattr(struct obd_export *exp, struct obdo *oa,
1540 struct lov_stripe_md *md, struct obd_trans_info *oti)
1542 struct obd_run_ctxt saved;
1543 struct filter_obd *filter;
1544 struct dentry *dentry;
1546 struct ldlm_res_id res_id = { .name = { oa->o_id } };
1547 struct ldlm_resource *res;
1552 LASSERT(oti != NULL);
1554 dentry = filter_oa2dentry(exp->exp_obd, oa);
1556 RETURN(PTR_ERR(dentry));
1558 filter = &exp->exp_obd->u.filter;
1560 iattr_from_obdo(&iattr, oa, oa->o_valid);
1562 push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1565 if (iattr.ia_valid & ATTR_SIZE)
1566 down(&dentry->d_inode->i_sem);
1567 handle = fsfilt_start(exp->exp_obd, dentry->d_inode, FSFILT_OP_SETATTR,
1570 GOTO(out_unlock, rc = PTR_ERR(handle));
1572 /* XXX this could be a rwsem instead, if filter_preprw played along */
1573 if (iattr.ia_valid & ATTR_ATTR_FLAG)
1574 rc = fsfilt_iocontrol(exp->exp_obd, dentry->d_inode, NULL,
1576 (long)&iattr.ia_attr_flags);
1578 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
1579 rc = filter_finish_transno(exp, oti, rc);
1580 rc2 = fsfilt_commit(exp->exp_obd, dentry->d_inode, handle, 0);
1582 CERROR("error on commit, err = %d\n", rc2);
1587 if (iattr.ia_valid & ATTR_SIZE) {
1588 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
1589 res_id, LDLM_EXTENT, 0);
1591 CERROR("!!! resource_get failed for object "LPU64" -- "
1592 "filter_setattr with no lock?\n", oa->o_id);
1594 if (res->lr_namespace->ns_lvbo &&
1595 res->lr_namespace->ns_lvbo->lvbo_update) {
1596 rc = res->lr_namespace->ns_lvbo->lvbo_update
1599 ldlm_resource_putref(res);
1603 oa->o_valid = OBD_MD_FLID;
1604 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1607 if (iattr.ia_valid & ATTR_SIZE)
1608 up(&dentry->d_inode->i_sem);
1610 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1616 /* XXX identical to osc_unpackmd */
1617 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
1618 struct lov_mds_md *lmm, int lmm_bytes)
1624 if (lmm_bytes < sizeof (*lmm)) {
1625 CERROR("lov_mds_md too small: %d, need %d\n",
1626 lmm_bytes, (int)sizeof(*lmm));
1629 /* XXX LOV_MAGIC etc check? */
1631 if (lmm->lmm_object_id == cpu_to_le64(0)) {
1632 CERROR("lov_mds_md: zero lmm_object_id\n");
1637 lsm_size = lov_stripe_md_size(1);
1641 if (*lsmp != NULL && lmm == NULL) {
1642 OBD_FREE(*lsmp, lsm_size);
1647 if (*lsmp == NULL) {
1648 OBD_ALLOC(*lsmp, lsm_size);
1652 loi_init((*lsmp)->lsm_oinfo);
1656 /* XXX zero *lsmp? */
1657 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
1658 LASSERT((*lsmp)->lsm_object_id);
1661 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
1666 static void filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
1667 struct filter_obd *filter)
1669 struct obdo doa; /* XXX obdo on stack */
1674 memset(&doa, 0, sizeof(doa));
1675 if (oa->o_valid & OBD_MD_FLGROUP)
1676 doa.o_gr = oa->o_gr;
1679 doa.o_mode = S_IFREG;
1680 last = filter_last_id(filter, &doa); /* FIXME: object groups */
1681 CWARN("deleting orphan objects from "LPU64" to "LPU64"\n",
1682 oa->o_id + 1, last);
1683 for (id = oa->o_id + 1; id <= last; id++) {
1685 filter_destroy(exp, &doa, NULL, NULL);
1687 spin_lock(&filter->fo_objidlock);
1688 filter->fo_last_objids[0] = oa->o_id; /* FIXME: object groups */
1689 spin_unlock(&filter->fo_objidlock);
1693 /* returns a negative error or a nonnegative number of files to create */
1694 static int filter_should_precreate(struct obd_export *exp, struct obdo *oa,
1697 struct obd_device *obd = exp->exp_obd;
1698 struct filter_obd *filter = &obd->u.filter;
1702 diff = oa->o_id - filter_last_id(filter, oa);
1703 CDEBUG(D_INFO, "filter_last_id() = "LPU64" -> diff = %d\n",
1704 filter_last_id(filter, oa), diff);
1706 /* delete orphans request */
1707 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1708 (oa->o_flags & OBD_FL_DELORPHAN)) {
1711 if (-diff > 10000) { /* XXX make this smarter */
1712 CERROR("ignoring bogus orphan destroy request: obdid "
1713 LPU64" last_id "LPU64"\n",
1714 oa->o_id, filter_last_id(filter, oa));
1717 filter_destroy_precreated(exp, oa, filter);
1718 rc = filter_update_last_objid(obd, group, 0);
1720 CERROR("unable to write lastobjid, but orphans"
1724 /* only precreate if group == 0 and o_id is specfied */
1725 if (!(oa->o_valid & OBD_FL_DELORPHAN) &&
1726 (group != 0 || oa->o_id == 0))
1734 /* We rely on the fact that only one thread will be creating files in a given
1735 * group at a time, which is why we don't need an atomic filter_get_new_id.
1736 * Even if we had that atomic function, the following race would exist:
1738 * thread 1: gets id x from filter_next_id
1739 * thread 2: gets id (x + 1) from filter_next_id
1740 * thread 2: creates object (x + 1)
1741 * thread 1: tries to create object x, gets -ENOSPC
1743 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
1744 obd_gr group, int *num)
1746 struct dentry *dchild = NULL;
1747 struct filter_obd *filter;
1748 struct dentry *dparent;
1749 int err = 0, rc = 0, i;
1751 int recreate_obj = 0;
1752 void *handle = NULL;
1755 filter = &obd->u.filter;
1757 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1758 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
1762 for (i = 0; i < *num && err == 0; i++) {
1763 int cleanup_phase = 0;
1768 last_id = filter_last_id(filter, NULL);
1769 if (next_id > last_id) {
1770 CERROR("Error: Trying to recreate obj greater"
1771 "than last id "LPD64" > "LPD64"\n",
1776 next_id = filter_last_id(filter, NULL) + 1;
1778 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
1780 dparent = filter_parent_lock(obd, group, next_id);
1781 if (IS_ERR(dparent))
1782 GOTO(cleanup, rc = PTR_ERR(dparent));
1785 dchild = filter_fid2dentry(obd, dparent, group, next_id);
1787 GOTO(cleanup, rc = PTR_ERR(dchild));
1790 if (dchild->d_inode != NULL) {
1791 /* This would only happen if lastobjid was bad on disk*/
1792 /* Could also happen if recreating missing obj but
1796 CERROR("Serious error: recreating obj %*s but "
1797 "obj already exists \n",
1798 dchild->d_name.len, dchild->d_name.name);
1800 CERROR("Serious error: objid %*s already "
1801 "exists; is this filesystem corrupt?\n",
1802 dchild->d_name.len, dchild->d_name.name);
1804 GOTO(cleanup, rc = -EEXIST);
1807 handle = fsfilt_start(obd, dparent->d_inode,
1808 FSFILT_OP_CREATE_LOG, NULL);
1810 GOTO(cleanup, rc = PTR_ERR(handle));
1813 rc = ll_vfs_create(dparent->d_inode, dchild, S_IFREG, NULL);
1815 CERROR("create failed rc = %d\n", rc);
1819 if (!recreate_obj) {
1820 filter_set_last_id(filter, NULL, next_id);
1821 err = filter_update_last_objid(obd, group, 0);
1823 CERROR("unable to write lastobjid "
1824 "but file created\n");
1828 switch(cleanup_phase) {
1830 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
1832 CERROR("error on commit, err = %d\n", err);
1839 filter_parent_unlock(dparent);
1849 CDEBUG(D_INFO, "filter_precreate() created %d objects\n", i);
1853 static int filter_create(struct obd_export *exp, struct obdo *oa,
1854 struct lov_stripe_md **ea, struct obd_trans_info *oti)
1856 struct obd_device *obd = NULL;
1857 struct obd_run_ctxt saved;
1858 struct lov_stripe_md *lsm = NULL;
1863 if (oa->o_valid & OBD_MD_FLGROUP)
1866 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
1871 rc = obd_alloc_memmd(exp, &lsm);
1878 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1880 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1881 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
1882 if (oa->o_id > filter_last_id(&obd->u.filter, oa)) {
1883 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
1884 oa->o_id, filter_last_id(&obd->u.filter, oa));
1888 rc = filter_precreate(obd, oa, group, &diff);
1891 diff = filter_should_precreate(exp, oa, group);
1893 oa->o_id = filter_last_id(&obd->u.filter, oa);
1894 rc = filter_precreate(obd, oa, group, &diff);
1896 oa->o_valid = OBD_MD_FLID;
1900 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1901 if (rc && ea != NULL && *ea != lsm) {
1902 obd_free_memmd(exp, &lsm);
1903 } else if (rc == 0 && ea != NULL) {
1904 /* XXX LOV STACKING: the lsm that is passed to us from
1905 * LOV does not have valid lsm_oinfo data structs, so
1906 * don't go touching that. This needs to be fixed in a
1908 lsm->lsm_object_id = oa->o_id;
1915 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
1916 struct lov_stripe_md *ea, struct obd_trans_info *oti)
1918 struct obd_device *obd;
1919 struct filter_obd *filter;
1920 struct dentry *dchild = NULL, *dparent = NULL;
1921 struct obd_run_ctxt saved;
1922 void *handle = NULL;
1923 struct llog_cookie *fcc = NULL;
1924 int rc, rc2, cleanup_phase = 0, have_prepared = 0;
1928 if (oa->o_valid & OBD_MD_FLGROUP)
1932 filter = &obd->u.filter;
1934 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1937 dparent = filter_parent_lock(obd, group, oa->o_id);
1938 if (IS_ERR(dparent))
1939 GOTO(cleanup, rc = PTR_ERR(dparent));
1942 dchild = filter_fid2dentry(obd, dparent, group, oa->o_id);
1944 GOTO(cleanup, rc = -ENOENT);
1947 if (dchild->d_inode == NULL) {
1948 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
1950 GOTO(cleanup, rc = -ENOENT);
1953 if (!have_prepared) {
1954 /* If we're really going to destroy the object, get ready
1955 * by getting the clients to discard their cached data.
1957 * We have to drop the parent lock, because
1958 * filter_prepare_destroy will acquire a PW on the object, and
1959 * we don't want to deadlock with an incoming write to the
1960 * object, which has the extent PW and then wants to get the
1961 * parent dentry to do the lookup.
1963 * We dput the child because it's not worth the extra
1964 * complication of condition the above code to skip it on the
1965 * second time through. */
1967 filter_parent_unlock(dparent);
1969 filter_prepare_destroy(obd, oa->o_id);
1974 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK_LOG, oti);
1976 GOTO(cleanup, rc = PTR_ERR(handle));
1979 /* Our MDC connection is established by the MDS to us */
1980 if (oa->o_valid & OBD_MD_FLCOOKIE) {
1981 OBD_ALLOC(fcc, sizeof(*fcc));
1983 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
1986 rc = filter_destroy_internal(obd, oa->o_id, dparent, dchild);
1989 switch(cleanup_phase) {
1993 fsfilt_add_journal_cb(obd, 0, oti->oti_handle,
1994 filter_cancel_cookies_cb,
1997 fsfilt_add_journal_cb(obd, 0, handle,
1998 filter_cancel_cookies_cb,
2001 rc = filter_finish_transno(exp, oti, rc);
2002 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
2004 CERROR("error on commit, err = %d\n", rc2);
2011 filter_parent_unlock(dparent);
2013 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
2016 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2023 /* NB start and end are used for punch, but not truncate */
2024 static int filter_truncate(struct obd_export *exp, struct obdo *oa,
2025 struct lov_stripe_md *lsm,
2026 obd_off start, obd_off end,
2027 struct obd_trans_info *oti)
2032 if (end != OBD_OBJECT_EOF)
2033 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
2036 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = %x, "
2037 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
2039 error = filter_setattr(exp, oa, NULL, oti);
2043 static int filter_sync(struct obd_export *exp, struct obdo *oa,
2044 struct lov_stripe_md *lsm, obd_off start, obd_off end)
2046 struct obd_run_ctxt saved;
2047 struct filter_obd *filter;
2048 struct dentry *dentry;
2049 struct llog_ctxt *ctxt;
2053 filter = &exp->exp_obd->u.filter;
2055 /* an objid of zero is taken to mean "sync whole filesystem" */
2056 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
2057 rc = fsfilt_sync(exp->exp_obd, filter->fo_sb);
2058 /* flush any remaining cancel messages out to the target */
2059 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_REPL_CTXT);
2060 llog_sync(ctxt, exp);
2064 dentry = filter_oa2dentry(exp->exp_obd, oa);
2066 RETURN(PTR_ERR(dentry));
2068 push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
2070 down(&dentry->d_inode->i_sem);
2071 rc = filemap_fdatasync(dentry->d_inode->i_mapping);
2073 /* just any file to grab fsync method - "file" arg unused */
2074 struct file *file = filter->fo_rcvd_filp;
2076 if (file->f_op && file->f_op->fsync)
2077 rc = file->f_op->fsync(NULL, dentry, 1);
2079 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
2083 up(&dentry->d_inode->i_sem);
2085 oa->o_valid = OBD_MD_FLID;
2086 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
2088 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
2094 /* debugging to make sure that nothing bad happens, can be turned off soon.
2095 * caller must hold osfs lock */
2096 static void filter_grant_total_exports(struct obd_device *obd,
2097 obd_size *tot_dirty,
2098 obd_size *tot_pending,
2099 obd_size *tot_granted,
2102 struct filter_export_data *fed;
2103 struct obd_export *exp_pos;
2105 spin_lock(&obd->obd_dev_lock);
2106 list_for_each_entry(exp_pos, &obd->obd_exports, exp_obd_chain) {
2107 fed = &exp_pos->exp_filter_data;
2108 LASSERTF(fed->fed_dirty <= maxsize, "cli %s/%p %lu > "LPU64"\n",
2109 exp_pos->exp_client_uuid.uuid, exp_pos,
2110 fed->fed_dirty, maxsize);
2111 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2112 "cli %s/%p %lu+%lu > "LPU64"\n",
2113 exp_pos->exp_client_uuid.uuid, exp_pos,
2114 fed->fed_grant, fed->fed_pending, maxsize);
2115 *tot_dirty += fed->fed_dirty;
2116 *tot_pending += fed->fed_pending;
2117 *tot_granted += fed->fed_grant + fed->fed_pending;
2119 spin_unlock(&obd->obd_dev_lock);
2122 static void filter_grant_sanity_check(obd_size tot_dirty, obd_size tot_pending,
2123 obd_size tot_granted,
2124 obd_size fo_tot_dirty,
2125 obd_size fo_tot_pending,
2126 obd_size fo_tot_granted, obd_size maxsize)
2128 LASSERTF(tot_dirty == fo_tot_dirty, LPU64" != "LPU64"\n",
2129 tot_dirty, fo_tot_dirty);
2130 LASSERTF(tot_pending == fo_tot_pending, LPU64" != "LPU64"\n",
2131 tot_pending, fo_tot_pending);
2132 LASSERTF(tot_granted == fo_tot_granted, LPU64" != "LPU64"\n",
2133 tot_granted, fo_tot_granted);
2134 LASSERTF(tot_dirty <= maxsize, LPU64" > "LPU64"\n", tot_dirty, maxsize);
2135 LASSERTF(tot_pending <= tot_granted, LPU64" > "LPU64"\n", tot_pending,
2137 LASSERTF(tot_granted <= maxsize, LPU64" > "LPU64"\n",
2138 tot_granted, maxsize);
2141 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2142 unsigned long max_age)
2144 struct filter_obd *filter = &obd->u.filter;
2145 obd_size tot_cached = 0, tot_pending = 0, tot_granted = 0;
2146 obd_size fo_tot_cached, fo_tot_pending, fo_tot_granted;
2147 int blockbits = filter->fo_sb->s_blocksize_bits;
2151 /* at least try to account for cached pages. its still racey and
2152 * might be under-reporting if clients haven't announced their
2153 * caches with brw recently */
2154 spin_lock(&obd->obd_osfs_lock);
2155 rc = fsfilt_statfs(obd, filter->fo_sb, max_age);
2156 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
2157 filter_grant_total_exports(obd, &tot_cached, &tot_pending, &tot_granted,
2158 osfs->os_blocks << blockbits);
2159 fo_tot_cached = filter->fo_tot_dirty;
2160 fo_tot_pending = filter->fo_tot_pending;
2161 fo_tot_granted = filter->fo_tot_granted;
2162 spin_unlock(&obd->obd_osfs_lock);
2164 /* Do check outside spinlock, to avoid wedging system on failure */
2165 filter_grant_sanity_check(tot_cached, tot_pending, tot_granted,
2166 fo_tot_cached, fo_tot_pending,
2167 fo_tot_granted, osfs->os_blocks << blockbits);
2169 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
2170 "pending "LPU64" free "LPU64" avail "LPU64"\n",
2171 tot_cached >> blockbits, tot_granted >> blockbits,
2172 tot_pending >> blockbits, osfs->os_bfree, osfs->os_bavail);
2174 osfs->os_bavail -= min(osfs->os_bavail,
2175 (tot_cached +tot_pending +osfs->os_bsize -1) >>
2181 static int filter_get_info(struct obd_export *exp, __u32 keylen,
2182 void *key, __u32 *vallen, void *val)
2184 struct obd_device *obd;
2187 obd = class_exp2obd(exp);
2189 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2190 exp->exp_handle.h_cookie);
2194 if (keylen == strlen("blocksize") &&
2195 memcmp(key, "blocksize", keylen) == 0) {
2196 __u32 *blocksize = val;
2197 *vallen = sizeof(*blocksize);
2198 *blocksize = obd->u.filter.fo_sb->s_blocksize;
2202 if (keylen == strlen("blocksize_bits") &&
2203 memcmp(key, "blocksize_bits", keylen) == 0) {
2204 __u32 *blocksize_bits = val;
2205 *vallen = sizeof(*blocksize_bits);
2206 *blocksize_bits = obd->u.filter.fo_sb->s_blocksize_bits;
2210 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
2211 obd_id *last_id = val;
2212 /* FIXME: object groups */
2213 *last_id = filter_last_id(&obd->u.filter, 0);
2216 CDEBUG(D_IOCTL, "invalid key\n");
2220 static int filter_set_info(struct obd_export *exp, __u32 keylen,
2221 void *key, __u32 vallen, void *val)
2223 struct obd_device *obd;
2224 struct lustre_handle conn;
2225 struct llog_ctxt *ctxt;
2229 conn.cookie = exp->exp_handle.h_cookie;
2233 CDEBUG(D_IOCTL, "invalid exp %p cookie "LPX64"\n",
2238 if (keylen < strlen("mds_conn") ||
2239 memcmp(key, "mds_conn", keylen) != 0)
2242 CWARN("Received MDS connection ("LPX64")\n", conn.cookie);
2243 memcpy(&obd->u.filter.fo_mdc_conn, &conn, sizeof(conn));
2244 ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
2245 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
2249 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
2250 int len, void *karg, void *uarg)
2252 struct obd_device *obd = exp->exp_obd;
2253 struct obd_ioctl_data *data = karg;
2257 case OBD_IOC_ABORT_RECOVERY:
2258 CERROR("aborting recovery for device %s\n", obd->obd_name);
2259 target_abort_recovery(obd);
2262 case OBD_IOC_SET_READONLY: {
2264 struct super_block *sb = obd->u.filter.fo_sb;
2265 struct inode *inode = sb->s_root->d_inode;
2266 BDEVNAME_DECLARE_STORAGE(tmp);
2267 CERROR("setting device %s read-only\n",
2268 ll_bdevname(sb, tmp));
2270 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
2272 (void)fsfilt_commit(obd, inode, handle, 1);
2274 dev_set_rdonly(ll_sbdev(obd->u.filter.fo_sb), 2);
2278 case OBD_IOC_CATLOGLIST: {
2279 rc = llog_catlog_list(obd, 1, data);
2283 case OBD_IOC_LLOG_CANCEL:
2284 case OBD_IOC_LLOG_REMOVE:
2285 case OBD_IOC_LLOG_INFO:
2286 case OBD_IOC_LLOG_PRINT: {
2287 /* FIXME to be finished */
2288 RETURN(-EOPNOTSUPP);
2290 struct llog_ctxt *ctxt = NULL;
2292 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2293 rc = llog_ioctl(ctxt, cmd, data);
2294 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2307 static struct llog_operations filter_unlink_repl_logops;
2308 static struct llog_operations filter_size_orig_logops = {
2309 lop_setup: llog_obd_origin_setup,
2310 lop_cleanup: llog_obd_origin_cleanup,
2311 lop_add: llog_obd_origin_add
2314 static int filter_llog_init(struct obd_device *obd, struct obd_device *tgt,
2315 int count, struct llog_logid *logid)
2317 struct llog_ctxt *ctxt;
2321 filter_unlink_repl_logops = llog_client_ops;
2322 filter_unlink_repl_logops.lop_cancel = llog_obd_repl_cancel;
2323 filter_unlink_repl_logops.lop_connect = llog_repl_connect;
2324 filter_unlink_repl_logops.lop_sync = llog_obd_repl_sync;
2326 rc = llog_setup(obd, LLOG_UNLINK_REPL_CTXT, tgt, 0, NULL,
2327 &filter_unlink_repl_logops);
2330 /* FIXME - assign unlink_cb for filter's recovery */
2331 ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
2332 ctxt->llog_proc_cb = filter_recov_log_unlink_cb;
2334 rc = llog_setup(obd, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2335 &filter_size_orig_logops);
2339 static int filter_llog_finish(struct obd_device *obd, int count)
2344 rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_REPL_CTXT));
2348 rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_ORIG_CTXT));
2352 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2355 return filter_fid2dentry(data, NULL, gr, id);
2358 static struct lvfs_callback_ops filter_lvfs_ops = {
2359 l_fid2dentry: filter_lvfs_fid2dentry,
2362 static struct obd_ops filter_obd_ops = {
2363 o_owner: THIS_MODULE,
2364 o_attach: filter_attach,
2365 o_detach: filter_detach,
2366 o_get_info: filter_get_info,
2367 o_set_info: filter_set_info,
2368 o_setup: filter_setup,
2369 o_precleanup: filter_precleanup,
2370 o_cleanup: filter_cleanup,
2371 o_connect: filter_connect,
2372 o_disconnect: filter_disconnect,
2373 o_statfs: filter_statfs,
2374 o_getattr: filter_getattr,
2375 o_unpackmd: filter_unpackmd,
2376 o_create: filter_create,
2377 o_setattr: filter_setattr,
2378 o_destroy: filter_destroy,
2380 o_punch: filter_truncate,
2381 o_sync: filter_sync,
2382 o_preprw: filter_preprw,
2383 o_commitrw: filter_commitrw,
2384 o_destroy_export: filter_destroy_export,
2385 o_llog_init: filter_llog_init,
2386 o_llog_finish: filter_llog_finish,
2387 o_iocontrol: filter_iocontrol,
2390 static struct obd_ops filter_sanobd_ops = {
2391 o_owner: THIS_MODULE,
2392 o_attach: filter_attach,
2393 o_detach: filter_detach,
2394 o_get_info: filter_get_info,
2395 o_set_info: filter_set_info,
2396 o_setup: filter_san_setup,
2397 o_precleanup: filter_precleanup,
2398 o_cleanup: filter_cleanup,
2399 o_connect: filter_connect,
2400 o_disconnect: filter_disconnect,
2401 o_statfs: filter_statfs,
2402 o_getattr: filter_getattr,
2403 o_unpackmd: filter_unpackmd,
2404 o_create: filter_create,
2405 o_setattr: filter_setattr,
2406 o_destroy: filter_destroy,
2408 o_punch: filter_truncate,
2409 o_sync: filter_sync,
2410 o_preprw: filter_preprw,
2411 o_commitrw: filter_commitrw,
2412 o_san_preprw: filter_san_preprw,
2413 o_destroy_export: filter_destroy_export,
2414 o_llog_init: filter_llog_init,
2415 o_llog_finish: filter_llog_finish,
2416 o_iocontrol: filter_iocontrol,
2419 static int __init obdfilter_init(void)
2421 struct lprocfs_static_vars lvars;
2424 printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
2426 lprocfs_init_vars(filter, &lvars);
2428 rc = class_register_type(&filter_obd_ops, lvars.module_vars,
2429 OBD_FILTER_DEVICENAME);
2433 rc = class_register_type(&filter_sanobd_ops, lvars.module_vars,
2434 OBD_FILTER_SAN_DEVICENAME);
2436 class_unregister_type(OBD_FILTER_DEVICENAME);
2440 static void __exit obdfilter_exit(void)
2442 class_unregister_type(OBD_FILTER_SAN_DEVICENAME);
2443 class_unregister_type(OBD_FILTER_DEVICENAME);
2446 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2447 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
2448 MODULE_LICENSE("GPL");
2450 module_init(obdfilter_init);
2451 module_exit(obdfilter_exit);