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>
56 #include <portals/list.h>
58 #include "filter_internal.h"
60 static struct lvfs_callback_ops filter_lvfs_ops;
62 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
63 struct lov_stripe_md *ea, struct obd_trans_info *);
65 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
66 void *cb_data, int error)
68 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 if (filter->fo_last_objid_files[group] == NULL) {
323 CERROR("Object group "LPU64" not fully setup; not updating "
324 "last_objid\n", group);
328 CDEBUG(D_INODE, "server last_objid for group "LPU64": "LPU64"\n",
329 group, filter->fo_last_objids[group]);
331 tmp = cpu_to_le64(filter->fo_last_objids[group]);
332 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
333 &tmp, sizeof(tmp), &off, force_sync);
335 CERROR("error writing group "LPU64" last objid: rc = %d\n",
340 /* assumes caller has already in kernel ctxt */
341 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
343 struct filter_obd *filter = &obd->u.filter;
344 struct filter_server_data *fsd;
345 struct filter_client_data *fcd = NULL;
346 struct inode *inode = filp->f_dentry->d_inode;
347 unsigned long last_rcvd_size = inode->i_size;
353 /* ensure padding in the struct is the correct size */
354 LASSERT (offsetof(struct filter_server_data, fsd_padding) +
355 sizeof(fsd->fsd_padding) == FILTER_LR_SERVER_SIZE);
356 LASSERT (offsetof(struct filter_client_data, fcd_padding) +
357 sizeof(fcd->fcd_padding) == FILTER_LR_CLIENT_SIZE);
359 OBD_ALLOC(fsd, sizeof(*fsd));
362 filter->fo_fsd = fsd;
364 OBD_ALLOC(filter->fo_last_rcvd_slots,
365 FILTER_LR_MAX_CLIENT_WORDS * sizeof(unsigned long));
366 if (filter->fo_last_rcvd_slots == NULL) {
367 OBD_FREE(fsd, sizeof(*fsd));
371 if (last_rcvd_size == 0) {
372 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
374 memcpy(fsd->fsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->fsd_uuid));
375 fsd->fsd_last_transno = 0;
376 mount_count = fsd->fsd_mount_count = 0;
377 fsd->fsd_server_size = cpu_to_le32(FILTER_LR_SERVER_SIZE);
378 fsd->fsd_client_start = cpu_to_le32(FILTER_LR_CLIENT_START);
379 fsd->fsd_client_size = cpu_to_le16(FILTER_LR_CLIENT_SIZE);
380 fsd->fsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
381 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
383 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
385 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
389 if (strcmp(fsd->fsd_uuid, obd->obd_uuid.uuid) != 0) {
390 CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
391 obd->obd_uuid.uuid, fsd->fsd_uuid);
392 GOTO(err_fsd, rc = -EINVAL);
394 mount_count = le64_to_cpu(fsd->fsd_mount_count);
395 filter->fo_subdir_count = le16_to_cpu(fsd->fsd_subdir_count);
398 if (fsd->fsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
399 CERROR("unsupported feature %x\n",
400 le32_to_cpu(fsd->fsd_feature_incompat) &
401 ~FILTER_INCOMPAT_SUPP);
402 GOTO(err_fsd, rc = -EINVAL);
404 if (fsd->fsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
405 CERROR("read-only feature %x\n",
406 le32_to_cpu(fsd->fsd_feature_rocompat) &
407 ~FILTER_ROCOMPAT_SUPP);
408 /* Do something like remount filesystem read-only */
409 GOTO(err_fsd, rc = -EINVAL);
412 CDEBUG(D_INODE, "%s: server last_rcvd : "LPU64"\n",
413 obd->obd_name, le64_to_cpu(fsd->fsd_last_transno));
414 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
415 obd->obd_name, mount_count + 1);
416 CDEBUG(D_INODE, "%s: server data size: %u\n",
417 obd->obd_name, le32_to_cpu(fsd->fsd_server_size));
418 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
419 obd->obd_name, le32_to_cpu(fsd->fsd_client_start));
420 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
421 obd->obd_name, le32_to_cpu(fsd->fsd_client_size));
422 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
423 obd->obd_name, le16_to_cpu(fsd->fsd_subdir_count));
424 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
425 last_rcvd_size <= FILTER_LR_CLIENT_START ? 0 :
426 (last_rcvd_size-FILTER_LR_CLIENT_START) /FILTER_LR_CLIENT_SIZE);
428 if (!obd->obd_replayable) {
429 CWARN("%s: recovery support OFF\n", obd->obd_name);
433 for (cl_idx = 0, off = le32_to_cpu(fsd->fsd_client_start);
434 off < last_rcvd_size; cl_idx++) {
436 struct obd_export *exp;
437 struct filter_export_data *fed;
440 OBD_ALLOC(fcd, sizeof(*fcd));
442 GOTO(err_client, rc = -ENOMEM);
445 /* Don't assume off is incremented properly by
446 * fsfilt_read_record(), in case sizeof(*fcd)
447 * isn't the same as fsd->fsd_client_size. */
448 off = le32_to_cpu(fsd->fsd_client_start) +
449 cl_idx * le16_to_cpu(fsd->fsd_client_size);
450 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
452 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
453 LAST_RCVD, cl_idx, off, rc);
454 break; /* read error shouldn't cause startup to fail */
457 if (fcd->fcd_uuid[0] == '\0') {
458 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
463 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
465 /* These exports are cleaned up by filter_disconnect(), so they
466 * need to be set up like real exports as filter_connect() does.
468 exp = class_new_export(obd);
469 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
470 " srv lr: "LPU64"\n", fcd->fcd_uuid, cl_idx,
471 last_rcvd, le64_to_cpu(fsd->fsd_last_transno));
473 GOTO(err_client, rc = -ENOMEM);
475 memcpy(&exp->exp_client_uuid.uuid, fcd->fcd_uuid,
476 sizeof exp->exp_client_uuid.uuid);
477 fed = &exp->exp_filter_data;
479 filter_client_add(obd, filter, fed, cl_idx);
480 /* create helper if export init gets more complex */
481 spin_lock_init(&fed->fed_lock);
484 obd->obd_recoverable_clients++;
485 class_export_put(exp);
487 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
490 if (last_rcvd > le64_to_cpu(fsd->fsd_last_transno))
491 fsd->fsd_last_transno = cpu_to_le64(last_rcvd);
495 obd->obd_last_committed = le64_to_cpu(fsd->fsd_last_transno);
497 if (obd->obd_recoverable_clients) {
498 CWARN("RECOVERY: %d recoverable clients, last_rcvd "
499 LPU64"\n", obd->obd_recoverable_clients,
500 le64_to_cpu(fsd->fsd_last_transno));
501 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
502 obd->obd_recovering = 1;
506 OBD_FREE(fcd, sizeof(*fcd));
509 filter->fo_mount_count = mount_count + 1;
510 fsd->fsd_mount_count = cpu_to_le64(filter->fo_mount_count);
512 /* save it, so mount count and last_transno is current */
513 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
518 class_disconnect_exports(obd, 0);
520 filter_free_server_data(filter);
524 static int filter_cleanup_groups(struct obd_device *obd)
526 struct filter_obd *filter = &obd->u.filter;
527 struct dentry *dentry;
531 for (i = 0; i < filter->fo_group_count; i++) {
532 if (filter->fo_subdirs != NULL) {
533 for (k = 0; k < filter->fo_subdir_count; k++) {
534 dentry = filter->fo_subdirs[i].dentry[k];
538 filter->fo_subdirs[i].dentry[k] = NULL;
541 if (filter->fo_last_objid_files[i] != NULL) {
542 filp_close(filter->fo_last_objid_files[i], 0);
543 filter->fo_last_objid_files[i] = NULL;
545 if (filter->fo_groups[i] != NULL) {
546 dput(filter->fo_groups[i]);
547 filter->fo_groups[i] = NULL;
550 if (filter->fo_subdirs != NULL)
551 OBD_FREE(filter->fo_subdirs,
552 filter->fo_group_count * sizeof(*filter->fo_subdirs));
553 if (filter->fo_groups != NULL)
554 OBD_FREE(filter->fo_groups,
555 filter->fo_group_count * sizeof(*filter->fo_groups));
556 if (filter->fo_last_objids != NULL)
557 OBD_FREE(filter->fo_last_objids,
558 filter->fo_group_count * sizeof(__u64));
559 if (filter->fo_last_objid_files != NULL)
560 OBD_FREE(filter->fo_last_objid_files,
561 filter->fo_group_count * sizeof(struct file *));
565 static int filter_read_group_internal(struct obd_device *obd, int group,
568 struct filter_obd *filter = &obd->u.filter;
569 __u64 *new_objids = NULL;
570 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs;
571 struct dentry **new_groups = NULL;
572 struct file **new_files = NULL;
573 struct dentry *dentry;
575 int old_count = filter->fo_group_count, rc, stage = 0, i;
580 snprintf(name, 24, "%d", group);
584 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
586 if (IS_ERR(dentry)) {
587 CERROR("Cannot lookup expected object group %d: %ld\n",
588 group, PTR_ERR(dentry));
589 RETURN(PTR_ERR(dentry));
592 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
593 if (IS_ERR(dentry)) {
594 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
596 RETURN(PTR_ERR(dentry));
601 snprintf(name, 24, "O/%d/LAST_ID", group);
603 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
605 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
606 GOTO(cleanup, rc = PTR_ERR(filp));
610 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
612 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
616 if (filter->fo_subdir_count) {
617 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
618 if (tmp_subdirs == NULL)
619 GOTO(cleanup, rc = -ENOMEM);
622 for (i = 0; i < filter->fo_subdir_count; i++) {
624 snprintf(dir, sizeof(dir), "d%u", i);
626 tmp_subdirs->dentry[i] =
627 simple_mkdir(dentry, dir, 0700, 1);
628 if (IS_ERR(tmp_subdirs->dentry[i])) {
629 rc = PTR_ERR(tmp_subdirs->dentry[i]);
630 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
634 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
635 tmp_subdirs->dentry[i]);
639 /* 'group' is an index; we need an array of length 'group + 1' */
640 if (group + 1 > old_count) {
642 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
643 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
644 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
645 OBD_ALLOC(new_files, len * sizeof(*new_files));
647 if (new_objids == NULL || new_subdirs == NULL ||
648 new_groups == NULL || new_files == NULL)
649 GOTO(cleanup, rc = -ENOMEM);
651 memcpy(new_objids, filter->fo_last_objids,
652 old_count * sizeof(*new_objids));
653 memcpy(new_subdirs, filter->fo_subdirs,
654 old_count * sizeof(*new_subdirs));
655 memcpy(new_groups, filter->fo_groups,
656 old_count * sizeof(*new_groups));
657 memcpy(new_files, filter->fo_last_objid_files,
658 old_count * sizeof(*new_files));
661 OBD_FREE(filter->fo_last_objids,
662 old_count * sizeof(*new_objids));
663 OBD_FREE(filter->fo_subdirs,
664 old_count * sizeof(*new_subdirs));
665 OBD_FREE(filter->fo_groups,
666 old_count * sizeof(*new_groups));
667 OBD_FREE(filter->fo_last_objid_files,
668 old_count * sizeof(*new_files));
670 filter->fo_last_objids = new_objids;
671 filter->fo_subdirs = new_subdirs;
672 filter->fo_groups = new_groups;
673 filter->fo_last_objid_files = new_files;
674 filter->fo_group_count = len;
677 filter->fo_groups[group] = dentry;
678 filter->fo_last_objid_files[group] = filp;
679 if (filter->fo_subdir_count) {
680 filter->fo_subdirs[group] = *tmp_subdirs;
681 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
684 if (filp->f_dentry->d_inode->i_size == 0) {
685 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
689 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
690 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
691 obd->obd_name, group, last_objid);
696 if (new_objids != NULL)
697 OBD_FREE(new_objids, group * sizeof(*new_objids));
698 if (new_subdirs != NULL)
699 OBD_FREE(new_subdirs, group * sizeof(*new_subdirs));
700 if (new_groups != NULL)
701 OBD_FREE(new_groups, group * sizeof(*new_groups));
702 if (new_files != NULL)
703 OBD_FREE(new_files, group * sizeof(*new_files));
705 if (filter->fo_subdir_count) {
706 for (i = 0; i < filter->fo_subdir_count; i++) {
707 if (tmp_subdirs->dentry[i] != NULL)
708 dput(tmp_subdirs->dentry[i]);
710 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
720 static int filter_read_groups(struct obd_device *obd, int last_group,
723 struct filter_obd *filter = &obd->u.filter;
724 int old_count = filter->fo_group_count, group = old_count, rc = 0;
726 for (group = old_count; group <= last_group; group++) {
728 continue; /* no group zero */
730 rc = filter_read_group_internal(obd, group, create);
737 static int filter_prep_groups(struct obd_device *obd)
739 struct filter_obd *filter = &obd->u.filter;
740 struct dentry *dentry, *O_dentry;
741 int rc = 0, cleanup_phase = 0;
744 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
745 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
746 if (IS_ERR(O_dentry)) {
747 rc = PTR_ERR(O_dentry);
748 CERROR("cannot open/create O: rc = %d\n", rc);
751 filter->fo_dentry_O = O_dentry;
752 cleanup_phase = 1; /* O_dentry */
754 /* Lookup "R" to tell if we're on an old OST FS and need to convert
755 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
756 * some time post 1.0 when all old-style OSTs have converted along
757 * with the init_objid hack. */
758 dentry = ll_lookup_one_len("R", O_dentry, 1);
760 GOTO(cleanup, rc = PTR_ERR(dentry));
761 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
762 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
765 CWARN("converting OST to new object layout\n");
766 if (IS_ERR(O0_dentry)) {
767 rc = PTR_ERR(O0_dentry);
768 CERROR("error looking up O/0: rc %d\n", rc);
772 if (O0_dentry->d_inode) {
773 CERROR("Both O/R and O/0 exist. Fix manually.\n");
774 GOTO(cleanup_O0, rc = -EEXIST);
777 down(&O_dentry->d_inode->i_sem);
778 rc = vfs_rename(O_dentry->d_inode, dentry,
779 O_dentry->d_inode, O0_dentry);
780 up(&O_dentry->d_inode->i_sem);
783 CERROR("error renaming O/R to O/0: rc %d\n", rc);
784 GOTO(cleanup_O0, rc);
786 filter->fo_fsd->fsd_feature_incompat |=
787 cpu_to_le32(FILTER_INCOMPAT_GROUPS);
788 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
790 GOTO(cleanup_O0, rc);
802 cleanup_phase = 2; /* groups */
804 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
805 #define FILTER_MIN_GROUPS 3
806 rc = filter_read_groups(obd, FILTER_MIN_GROUPS, 1);
813 switch (cleanup_phase) {
815 filter_cleanup_groups(obd);
817 f_dput(filter->fo_dentry_O);
818 filter->fo_dentry_O = NULL;
825 /* setup the object store with correct subdirectories */
826 static int filter_prep(struct obd_device *obd)
828 struct obd_run_ctxt saved;
829 struct filter_obd *filter = &obd->u.filter;
835 push_ctxt(&saved, &obd->obd_ctxt, NULL);
836 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
837 if (!file || IS_ERR(file)) {
839 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
844 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
845 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
846 file->f_dentry->d_inode->i_mode);
847 GOTO(err_filp, rc = -ENOENT);
850 /* steal operations */
851 inode = file->f_dentry->d_inode;
852 filter->fo_fop = file->f_op;
853 filter->fo_iop = inode->i_op;
854 filter->fo_aops = inode->i_mapping->a_ops;
856 rc = filter_init_server_data(obd, file);
858 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
861 filter->fo_rcvd_filp = file;
863 rc = filter_prep_groups(obd);
865 GOTO(err_server_data, rc);
868 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
873 //class_disconnect_exports(obd, 0);
874 filter_free_server_data(filter);
876 if (filp_close(file, 0))
877 CERROR("can't close %s after error\n", LAST_RCVD);
878 filter->fo_rcvd_filp = NULL;
882 /* cleanup the filter: write last used object id to status file */
883 static void filter_post(struct obd_device *obd)
885 struct obd_run_ctxt saved;
886 struct filter_obd *filter = &obd->u.filter;
889 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
890 * best to start a transaction with h_sync, because we removed this
893 push_ctxt(&saved, &obd->obd_ctxt, NULL);
894 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
897 CERROR("error writing server data: rc = %d\n", rc);
899 for (i = 1; i < filter->fo_group_count; i++) {
900 rc = filter_update_last_objid(obd, i,
901 (i == filter->fo_group_count - 1));
903 CERROR("error writing group %d lastobjid: rc = %d\n",
907 filp_close(filter->fo_rcvd_filp, 0);
908 filter->fo_rcvd_filp = NULL;
910 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
912 filter_cleanup_groups(obd);
913 f_dput(filter->fo_dentry_O);
914 filter_free_server_data(filter);
915 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
918 static void filter_set_last_id(struct filter_obd *filter, int group, obd_id id)
920 LASSERT(filter->fo_fsd != NULL);
922 LASSERT(group < filter->fo_group_count);
924 spin_lock(&filter->fo_objidlock);
925 filter->fo_last_objids[group] = id;
926 spin_unlock(&filter->fo_objidlock);
929 __u64 filter_last_id(struct filter_obd *filter, int group)
932 LASSERT(filter->fo_fsd != NULL);
934 LASSERT(group < filter->fo_group_count);
936 spin_lock(&filter->fo_objidlock);
937 id = filter->fo_last_objids[group];
938 spin_unlock(&filter->fo_objidlock);
943 /* direct cut-n-paste of mds_blocking_ast() */
944 static int filter_blocking_ast(struct ldlm_lock *lock,
945 struct ldlm_lock_desc *desc,
946 void *data, int flag)
951 if (flag == LDLM_CB_CANCELING) {
952 /* Don't need to do anything here. */
956 /* XXX layering violation! -phil */
957 l_lock(&lock->l_resource->lr_namespace->ns_lock);
958 /* Get this: if filter_blocking_ast is racing with ldlm_intent_policy,
959 * such that filter_blocking_ast is called just before l_i_p takes the
960 * ns_lock, then by the time we get the lock, we might not be the
961 * correct blocking function anymore. So check, and return early, if
963 if (lock->l_blocking_ast != filter_blocking_ast) {
964 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
968 lock->l_flags |= LDLM_FL_CBPENDING;
969 do_ast = (!lock->l_readers && !lock->l_writers);
970 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
973 struct lustre_handle lockh;
976 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
977 ldlm_lock2handle(lock, &lockh);
978 rc = ldlm_cli_cancel(&lockh);
980 CERROR("ldlm_cli_cancel: %d\n", rc);
982 LDLM_DEBUG(lock, "Lock still has references, will be "
988 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
990 down(&dparent->d_inode->i_sem);
994 /* We never dget the object parent, so DON'T dput it either */
995 static void filter_parent_unlock(struct dentry *dparent)
997 up(&dparent->d_inode->i_sem);
1000 /* We never dget the object parent, so DON'T dput it either */
1001 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1003 struct filter_obd *filter = &obd->u.filter;
1004 LASSERT(group < filter->fo_group_count);
1007 if (/*group > 0 || */filter->fo_subdir_count == 0)
1008 return filter->fo_groups[group];
1010 return filter->fo_subdirs[group].dentry[objid & (filter->fo_subdir_count - 1)];
1013 /* We never dget the object parent, so DON'T dput it either */
1014 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1017 unsigned long now = jiffies;
1018 struct dentry *dparent = filter_parent(obd, group, objid);
1021 if (IS_ERR(dparent))
1025 LASSERT(dparent->d_inode);
1027 rc = filter_lock_dentry(obd, dparent);
1028 if (time_after(jiffies, now + 15 * HZ))
1029 CERROR("slow parent lock %lus\n", (jiffies - now) / HZ);
1030 return rc ? ERR_PTR(rc) : dparent;
1033 /* How to get files, dentries, inodes from object id's.
1035 * If dir_dentry is passed, the caller has already locked the parent
1036 * appropriately for this operation (normally a write lock). If
1037 * dir_dentry is NULL, we do a read lock while we do the lookup to
1038 * avoid races with create/destroy and such changing the directory
1039 * internal to the filesystem code. */
1040 struct dentry *filter_fid2dentry(struct obd_device *obd,
1041 struct dentry *dir_dentry,
1042 obd_gr group, obd_id id)
1044 struct dentry *dparent = dir_dentry;
1045 struct dentry *dchild;
1051 CERROR("fatal: invalid object id 0\n");
1052 RETURN(ERR_PTR(-ESTALE));
1055 len = sprintf(name, LPU64, id);
1056 if (dir_dentry == NULL) {
1057 dparent = filter_parent_lock(obd, group, id);
1058 if (IS_ERR(dparent))
1061 CDEBUG(D_INODE, "looking up object O/%*s/%s\n",
1062 dparent->d_name.len, dparent->d_name.name, name);
1063 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1064 if (dir_dentry == NULL)
1065 filter_parent_unlock(dparent);
1066 if (IS_ERR(dchild)) {
1067 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
1071 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1072 name, dchild, atomic_read(&dchild->d_count));
1074 LASSERT(atomic_read(&dchild->d_count) > 0);
1079 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1082 struct lustre_handle lockh;
1083 int flags = LDLM_AST_DISCARD_DATA, rc;
1084 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0 } };
1085 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1088 /* Tell the clients that the object is gone now and that they should
1089 * throw away any cached pages. */
1090 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
1091 LDLM_EXTENT, &policy, LCK_PW,
1092 &flags, filter_blocking_ast, ldlm_completion_ast,
1093 NULL, NULL, NULL, 0, NULL, &lockh);
1095 /* We only care about the side-effects, just drop the lock. */
1097 ldlm_lock_decref(&lockh, LCK_PW);
1102 /* Caller must hold LCK_PW on parent and push us into kernel context.
1103 * Caller is also required to ensure that dchild->d_inode exists. */
1104 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1105 struct dentry *dparent,
1106 struct dentry *dchild)
1108 struct inode *inode = dchild->d_inode;
1112 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1113 CERROR("destroying objid %*s nlink = %lu, count = %d\n",
1114 dchild->d_name.len, dchild->d_name.name,
1115 (unsigned long)inode->i_nlink,
1116 atomic_read(&inode->i_count));
1119 rc = vfs_unlink(dparent->d_inode, dchild);
1122 CERROR("error unlinking objid %*s: rc %d\n",
1123 dchild->d_name.len, dchild->d_name.name, rc);
1128 static int filter_intent_policy(struct ldlm_namespace *ns,
1129 struct ldlm_lock **lockp, void *req_cookie,
1130 ldlm_mode_t mode, int flags, void *data)
1132 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1133 struct ptlrpc_request *req = req_cookie;
1134 struct ldlm_lock *lock = *lockp, *l = NULL;
1135 struct ldlm_resource *res = lock->l_resource;
1136 ldlm_processing_policy policy;
1137 struct ost_lvb *res_lvb, *reply_lvb;
1138 struct list_head *tmp;
1140 int tmpflags = 0, rc, repsize[2] = {sizeof(struct ldlm_reply),
1141 sizeof(struct ost_lvb) };
1144 policy = ldlm_get_processing_policy(res);
1145 LASSERT(policy != NULL);
1146 LASSERT(req != NULL);
1148 rc = lustre_pack_reply(req, 2, repsize, NULL);
1150 RETURN(req->rq_status = rc);
1152 reply_lvb = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*reply_lvb));
1153 LASSERT(reply_lvb != NULL);
1155 //fixup_handle_for_resent_req(req, lock, &lockh);
1157 /* If we grant any lock at all, it will be a whole-file read lock.
1158 * Call the extent policy function to see if our request can be
1159 * granted, or is blocked. */
1160 lock->l_policy_data.l_extent.start = 0;
1161 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1162 lock->l_req_mode = LCK_PR;
1164 l_lock(&res->lr_namespace->ns_lock);
1166 res->lr_tmp = &rpc_list;
1167 rc = policy(lock, &tmpflags, 0, &err);
1170 /* FIXME: we should change the policy function slightly, to not make
1171 * this list at all, since we just turn around and free it */
1172 while (!list_empty(&rpc_list)) {
1173 struct ldlm_ast_work *w =
1174 list_entry(rpc_list.next, struct ldlm_ast_work, w_list);
1175 list_del(&w->w_list);
1176 LDLM_LOCK_PUT(w->w_lock);
1177 OBD_FREE(w, sizeof(*w));
1180 if (rc == LDLM_ITER_CONTINUE) {
1181 /* The lock met with no resistance; we're finished. */
1182 l_unlock(&res->lr_namespace->ns_lock);
1183 RETURN(ELDLM_LOCK_REPLACED);
1186 /* Do not grant any lock, but instead send GL callbacks. The extent
1187 * policy nicely created a list of all PW locks for us. We will choose
1188 * the highest of those which are larger than the size in the LVB, if
1189 * any, and perform a glimpse callback. */
1190 down(&res->lr_lvb_sem);
1191 res_lvb = res->lr_lvb_data;
1192 LASSERT(res_lvb != NULL);
1193 reply_lvb->lvb_size = res_lvb->lvb_size;
1194 up(&res->lr_lvb_sem);
1196 list_for_each(tmp, &res->lr_granted) {
1197 struct ldlm_lock *tmplock =
1198 list_entry(tmp, struct ldlm_lock, l_res_link);
1200 if (tmplock->l_granted_mode == LCK_PR)
1203 if (tmplock->l_policy_data.l_extent.end <=
1204 reply_lvb->lvb_size)
1208 l = LDLM_LOCK_GET(tmplock);
1212 if (l->l_policy_data.l_extent.start >
1213 tmplock->l_policy_data.l_extent.start)
1217 l = LDLM_LOCK_GET(tmplock);
1219 l_unlock(&res->lr_namespace->ns_lock);
1221 /* There were no PW locks beyond the size in the LVB; finished. */
1223 RETURN(ELDLM_LOCK_ABORTED);
1225 LASSERT(l->l_glimpse_ast != NULL);
1226 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1227 if (rc != 0 && res->lr_namespace->ns_lvbo &&
1228 res->lr_namespace->ns_lvbo->lvbo_update) {
1229 res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1232 down(&res->lr_lvb_sem);
1233 reply_lvb->lvb_size = res_lvb->lvb_size;
1234 up(&res->lr_lvb_sem);
1238 RETURN(ELDLM_LOCK_ABORTED);
1241 /* mount the file system (secretly) */
1242 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1245 struct lustre_cfg* lcfg = buf;
1246 struct filter_obd *filter = &obd->u.filter;
1247 struct vfsmount *mnt;
1251 dev_clear_rdonly(2);
1253 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1256 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1257 if (IS_ERR(obd->obd_fsops))
1258 RETURN(PTR_ERR(obd->obd_fsops));
1260 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, MS_NOATIME | MS_NODIRATIME,
1261 lcfg->lcfg_inlbuf1, option);
1266 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1267 if (*lcfg->lcfg_inlbuf3 == 'f') {
1268 obd->obd_replayable = 1;
1269 obd_sync_filter = 1;
1270 CWARN("%s: recovery enabled\n", obd->obd_name);
1272 if (*lcfg->lcfg_inlbuf3 != 'n') {
1273 CERROR("unrecognised flag '%c'\n",
1274 *lcfg->lcfg_inlbuf3);
1276 // XXX Robert? Why do we get errors here
1277 // GOTO(err_mntput, rc = -EINVAL);
1281 filter->fo_vfsmnt = mnt;
1282 filter->fo_sb = mnt->mnt_sb;
1283 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1284 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1286 OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
1287 obd->obd_ctxt.pwdmnt = mnt;
1288 obd->obd_ctxt.pwd = mnt->mnt_root;
1289 obd->obd_ctxt.fs = get_ds();
1290 obd->obd_ctxt.cb_ops = filter_lvfs_ops;
1292 rc = filter_prep(obd);
1294 GOTO(err_mntput, rc);
1296 spin_lock_init(&filter->fo_translock);
1297 spin_lock_init(&filter->fo_objidlock);
1298 INIT_LIST_HEAD(&filter->fo_export_list);
1299 sema_init(&filter->fo_alloc_lock, 1);
1300 spin_lock_init(&filter->fo_r_pages.oh_lock);
1301 spin_lock_init(&filter->fo_w_pages.oh_lock);
1302 spin_lock_init(&filter->fo_r_discont_pages.oh_lock);
1303 spin_lock_init(&filter->fo_w_discont_pages.oh_lock);
1304 spin_lock_init(&filter->fo_r_discont_blocks.oh_lock);
1305 spin_lock_init(&filter->fo_w_discont_blocks.oh_lock);
1306 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1308 INIT_LIST_HEAD(&filter->fo_llog_list);
1309 spin_lock_init(&filter->fo_llog_list_lock);
1311 obd->obd_namespace = ldlm_namespace_new("filter-tgt",
1312 LDLM_NAMESPACE_SERVER);
1313 if (obd->obd_namespace == NULL)
1314 GOTO(err_post, rc = -ENOMEM);
1315 obd->obd_namespace->ns_lvbp = obd;
1316 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1317 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1319 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1320 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1322 rc = llog_cat_initialize(obd, &obd->obd_llogs, 1);
1324 CERROR("failed to setup llogging subsystems\n");
1338 fsfilt_put_ops(obd->obd_fsops);
1342 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1344 struct lustre_cfg* lcfg = buf;
1345 const char *str = NULL;
1346 char *option = NULL;
1350 if (!strcmp(lcfg->lcfg_inlbuf2, "ext3")) {
1351 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1352 /* bug 1577: implement async-delete for 2.5 */
1353 str = "errors=remount-ro,asyncdel";
1355 str = "errors=remount-ro";
1357 n = strlen(str) + 1;
1358 OBD_ALLOC(option, n);
1361 strcpy(option, str);
1364 rc = filter_common_setup(obd, len, buf, option);
1366 OBD_FREE(option, n);
1370 static int filter_cleanup(struct obd_device *obd, int flags)
1372 struct filter_obd *filter = &obd->u.filter;
1375 if (flags & OBD_OPT_FAILOVER)
1376 CERROR("%s: shutting down for failover; client state will"
1377 " be preserved.\n", obd->obd_name);
1379 if (!list_empty(&obd->obd_exports)) {
1380 CERROR("%s: still has clients!\n", obd->obd_name);
1381 class_disconnect_exports(obd, flags);
1382 if (!list_empty(&obd->obd_exports)) {
1383 CERROR("still has exports after forced cleanup?\n");
1388 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1390 if (filter->fo_sb == NULL)
1395 shrink_dcache_parent(filter->fo_sb->s_root);
1398 if (atomic_read(&filter->fo_vfsmnt->mnt_count) > 1)
1399 CERROR("%s: mount point %p busy, mnt_count: %d\n",
1400 obd->obd_name, filter->fo_vfsmnt,
1401 atomic_read(&filter->fo_vfsmnt->mnt_count));
1404 mntput(filter->fo_vfsmnt);
1405 //destroy_buffers(filter->fo_sb->s_dev);
1406 filter->fo_sb = NULL;
1407 fsfilt_put_ops(obd->obd_fsops);
1410 dev_clear_rdonly(2);
1415 static int filter_attach(struct obd_device *obd, obd_count len, void *data)
1417 struct lprocfs_static_vars lvars;
1420 lprocfs_init_vars(filter, &lvars);
1421 rc = lprocfs_obd_attach(obd, lvars.obd_vars);
1425 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
1429 /* Init obdfilter private stats here */
1430 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1431 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
1432 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1433 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
1435 return lproc_filter_attach_seqstat(obd);
1438 static int filter_detach(struct obd_device *dev)
1440 lprocfs_free_obd_stats(dev);
1441 return lprocfs_obd_detach(dev);
1444 /* nearly identical to mds_connect */
1445 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
1446 struct obd_uuid *cluuid)
1448 struct obd_export *exp;
1449 struct filter_export_data *fed;
1450 struct filter_client_data *fcd = NULL;
1451 struct filter_obd *filter = &obd->u.filter;
1455 if (conn == NULL || obd == NULL || cluuid == NULL)
1458 rc = class_connect(conn, obd, cluuid);
1461 exp = class_conn2export(conn);
1462 LASSERT(exp != NULL);
1464 fed = &exp->exp_filter_data;
1466 spin_lock_init(&fed->fed_lock);
1468 if (!obd->obd_replayable)
1469 GOTO(cleanup, rc = 0);
1471 OBD_ALLOC(fcd, sizeof(*fcd));
1473 CERROR("filter: out of memory for client data\n");
1474 GOTO(cleanup, rc = -ENOMEM);
1477 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
1480 rc = filter_client_add(obd, filter, fed, -1);
1485 OBD_FREE(fcd, sizeof(*fcd));
1486 class_disconnect(exp, 0);
1488 class_export_put(exp);
1493 static int filter_precleanup(struct obd_device *obd, int flags)
1495 struct filter_group_llog *log;
1496 struct filter_obd *filter;
1500 filter = &obd->u.filter;
1502 spin_lock(&filter->fo_llog_list_lock);
1503 while (!list_empty(&filter->fo_llog_list)) {
1504 log = list_entry(filter->fo_llog_list.next,
1505 struct filter_group_llog, list);
1506 list_del(&log->list);
1507 spin_unlock(&filter->fo_llog_list_lock);
1509 rc = obd_llog_finish(obd, log->llogs, 0);
1511 CERROR("failed to cleanup llogging subsystem for %u\n",
1513 OBD_FREE(log->llogs, sizeof(*(log->llogs)));
1514 OBD_FREE(log, sizeof(*log));
1515 spin_lock(&filter->fo_llog_list_lock);
1517 spin_unlock(&filter->fo_llog_list_lock);
1519 rc = obd_llog_finish(obd, &obd->obd_llogs, 0);
1521 CERROR("failed to cleanup llogging subsystem\n");
1526 /* Do extra sanity checks for grant accounting. We do this at connect,
1527 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
1528 * always get rid of it or turn it off when we know accounting is good. */
1529 static void filter_grant_sanity_check(struct obd_device *obd, char *func)
1531 struct filter_export_data *fed;
1532 struct obd_export *exp;
1533 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
1534 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
1535 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
1537 if (list_empty(&obd->obd_exports))
1540 spin_lock(&obd->obd_osfs_lock);
1541 spin_lock(&obd->obd_dev_lock);
1542 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
1543 fed = &exp->exp_filter_data;
1544 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
1545 "cli %s/%p %lu+%lu > "LPU64"\n",
1546 exp->exp_client_uuid.uuid, exp,
1547 fed->fed_grant, fed->fed_pending, maxsize);
1548 LASSERTF(fed->fed_dirty <= maxsize, "cli %s/%p %lu > "LPU64"\n",
1549 exp->exp_client_uuid.uuid, exp,fed->fed_dirty,maxsize);
1550 CDEBUG(D_CACHE,"%s: cli %s/%p dirty %lu pend %lu grant %lu\n",
1551 obd->obd_name, exp->exp_client_uuid.uuid, exp,
1552 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
1553 tot_granted += fed->fed_grant + fed->fed_pending;
1554 tot_pending += fed->fed_pending;
1555 tot_dirty += fed->fed_dirty;
1557 fo_tot_granted = obd->u.filter.fo_tot_granted;
1558 fo_tot_pending = obd->u.filter.fo_tot_pending;
1559 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
1560 spin_unlock(&obd->obd_dev_lock);
1561 spin_unlock(&obd->obd_osfs_lock);
1563 /* Do these assertions outside the spinlocks so we don't kill system */
1564 if (tot_granted != fo_tot_granted)
1565 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
1566 func, tot_granted, fo_tot_granted);
1567 if (tot_pending != fo_tot_pending)
1568 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
1569 func, tot_pending, fo_tot_pending);
1570 if (tot_dirty != fo_tot_dirty)
1571 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
1572 func, tot_dirty, fo_tot_dirty);
1573 if (tot_pending > tot_granted)
1574 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
1575 func, tot_pending, tot_granted);
1576 if (tot_granted > maxsize)
1577 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
1578 func, tot_granted, maxsize);
1579 if (tot_dirty > maxsize)
1580 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
1581 func, tot_dirty, maxsize);
1584 /* Remove this client from the grant accounting totals. We also remove
1585 * the export from the obd device under the osfs and dev locks to ensure
1586 * that the filter_grant_sanity_check() calculations are always valid.
1587 * The client should do something similar when it invalidates its import. */
1588 static void filter_grant_discard(struct obd_export *exp)
1590 struct obd_device *obd = exp->exp_obd;
1591 struct filter_obd *filter = &obd->u.filter;
1592 struct filter_export_data *fed = &exp->exp_filter_data;
1594 spin_lock(&obd->obd_osfs_lock);
1595 spin_lock(&exp->exp_obd->obd_dev_lock);
1596 list_del_init(&exp->exp_obd_chain);
1597 spin_unlock(&exp->exp_obd->obd_dev_lock);
1599 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %lu pend %lu grant %lu\n",
1600 obd->obd_name, exp->exp_client_uuid.uuid, exp,
1601 fed->fed_dirty, fed->fed_pending, fed->fed_grant);
1603 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
1604 "%s: tot_granted "LPU64" cli %s/%p fed_grant %lu\n",
1605 obd->obd_name, filter->fo_tot_granted,
1606 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
1607 filter->fo_tot_granted -= fed->fed_grant;
1608 LASSERTF(exp->exp_obd->u.filter.fo_tot_pending >= fed->fed_pending,
1609 "%s: tot_pending "LPU64" cli %s/%p fed_pending %lu\n",
1610 obd->obd_name, filter->fo_tot_pending,
1611 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
1612 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
1613 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %lu\n",
1614 obd->obd_name, filter->fo_tot_dirty,
1615 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
1616 filter->fo_tot_dirty -= fed->fed_dirty;
1620 spin_unlock(&obd->obd_osfs_lock);
1623 static int filter_destroy_export(struct obd_export *exp)
1627 if (exp->exp_filter_data.fed_pending)
1628 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
1629 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
1630 exp, exp->exp_filter_data.fed_pending);
1632 target_destroy_export(exp);
1634 if (exp->exp_obd->obd_replayable)
1635 filter_client_free(exp, exp->exp_flags);
1637 filter_grant_discard(exp);
1638 if (!(exp->exp_flags & OBD_OPT_FORCE))
1639 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
1644 static void filter_sync_llogs(struct obd_export *dexp)
1646 struct filter_group_llog *fglog, *nlog;
1647 struct obd_device *obd = dexp->exp_obd;
1648 struct filter_obd *filter;
1649 int worked = 0, group;
1650 struct llog_ctxt *ctxt;
1652 filter = &obd->u.filter;
1654 /* we can't sync log holding spinlock. also, we do not want to get
1655 * into livelock. so we do following: loop over MDS's exports in
1656 * group order and skip already synced llogs -bzzz */
1658 /* look for group with min. number, but > worked */
1661 spin_lock(&filter->fo_llog_list_lock);
1662 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
1664 if (nlog->group <= worked) {
1665 /* this group is already synced */
1669 if (group < nlog->group) {
1670 /* we have group with smaller number to sync */
1674 /* store current minimal group */
1676 group = nlog->group;
1678 spin_lock(&filter->fo_llog_list_lock);
1681 worked = fglog->group;
1682 ctxt = llog_get_context(fglog->llogs,
1683 LLOG_UNLINK_REPL_CTXT);
1684 llog_sync(ctxt, dexp);
1686 } while (fglog != NULL);
1689 /* also incredibly similar to mds_disconnect */
1690 static int filter_disconnect(struct obd_export *exp, int flags)
1692 struct obd_device *obd = exp->exp_obd;
1693 unsigned long irqflags;
1698 class_export_get(exp);
1700 spin_lock_irqsave(&exp->exp_lock, irqflags);
1701 exp->exp_flags = flags;
1702 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
1704 if (!(flags & OBD_OPT_FORCE))
1705 filter_grant_sanity_check(obd, __FUNCTION__);
1706 filter_grant_discard(exp);
1708 /* Disconnect early so that clients can't keep using export */
1709 rc = class_disconnect(exp, flags);
1711 ldlm_cancel_locks_for_export(exp);
1713 fsfilt_sync(obd, obd->u.filter.fo_sb);
1715 /* flush any remaining cancel messages out to the target */
1716 filter_sync_llogs(exp);
1718 class_export_put(exp);
1722 struct dentry *__filter_oa2dentry(struct obd_device *obd,
1723 struct obdo *oa, const char *what)
1725 struct dentry *dchild = NULL;
1728 if (oa->o_valid & OBD_MD_FLGROUP)
1731 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
1733 if (IS_ERR(dchild)) {
1734 CERROR("%s error looking up object: "LPU64"\n", what, oa->o_id);
1738 if (dchild->d_inode == NULL) {
1739 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
1741 RETURN(ERR_PTR(-ENOENT));
1747 static int filter_getattr(struct obd_export *exp, struct obdo *oa,
1748 struct lov_stripe_md *md)
1750 struct dentry *dentry = NULL;
1751 struct obd_device *obd;
1755 obd = class_exp2obd(exp);
1757 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1758 exp->exp_handle.h_cookie);
1762 dentry = filter_oa2dentry(obd, oa);
1764 RETURN(PTR_ERR(dentry));
1766 /* Limit the valid bits in the return data to what we actually use */
1767 oa->o_valid = OBD_MD_FLID;
1768 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1774 /* this is called from filter_truncate() until we have filter_punch() */
1775 static int filter_setattr(struct obd_export *exp, struct obdo *oa,
1776 struct lov_stripe_md *md, struct obd_trans_info *oti)
1778 struct obd_run_ctxt saved;
1779 struct filter_obd *filter;
1780 struct dentry *dentry;
1782 struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0 } };
1783 struct ldlm_resource *res;
1788 LASSERT(oti != NULL);
1790 dentry = filter_oa2dentry(exp->exp_obd, oa);
1792 RETURN(PTR_ERR(dentry));
1794 filter = &exp->exp_obd->u.filter;
1796 iattr_from_obdo(&iattr, oa, oa->o_valid);
1798 push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1801 if (iattr.ia_valid & ATTR_SIZE)
1802 down(&dentry->d_inode->i_sem);
1803 handle = fsfilt_start(exp->exp_obd, dentry->d_inode, FSFILT_OP_SETATTR,
1806 GOTO(out_unlock, rc = PTR_ERR(handle));
1808 /* XXX this could be a rwsem instead, if filter_preprw played along */
1809 if (iattr.ia_valid & ATTR_ATTR_FLAG)
1810 rc = fsfilt_iocontrol(exp->exp_obd, dentry->d_inode, NULL,
1812 (long)&iattr.ia_attr_flags);
1814 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
1815 rc = filter_finish_transno(exp, oti, rc);
1816 rc2 = fsfilt_commit(exp->exp_obd, dentry->d_inode, handle, 0);
1818 CERROR("error on commit, err = %d\n", rc2);
1823 if (iattr.ia_valid & ATTR_SIZE) {
1824 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
1825 res_id, LDLM_EXTENT, 0);
1827 CERROR("!!! resource_get failed for object "LPU64" -- "
1828 "filter_setattr with no lock?\n", oa->o_id);
1830 if (res->lr_namespace->ns_lvbo &&
1831 res->lr_namespace->ns_lvbo->lvbo_update) {
1832 rc = res->lr_namespace->ns_lvbo->lvbo_update
1835 ldlm_resource_putref(res);
1839 oa->o_valid = OBD_MD_FLID;
1840 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1843 if (iattr.ia_valid & ATTR_SIZE)
1844 up(&dentry->d_inode->i_sem);
1846 pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1852 /* XXX identical to osc_unpackmd */
1853 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
1854 struct lov_mds_md *lmm, int lmm_bytes)
1860 if (lmm_bytes < sizeof (*lmm)) {
1861 CERROR("lov_mds_md too small: %d, need %d\n",
1862 lmm_bytes, (int)sizeof(*lmm));
1865 /* XXX LOV_MAGIC etc check? */
1867 if (lmm->lmm_object_id == cpu_to_le64(0)) {
1868 CERROR("lov_mds_md: zero lmm_object_id\n");
1873 lsm_size = lov_stripe_md_size(1);
1877 if (*lsmp != NULL && lmm == NULL) {
1878 OBD_FREE(*lsmp, lsm_size);
1883 if (*lsmp == NULL) {
1884 OBD_ALLOC(*lsmp, lsm_size);
1888 loi_init((*lsmp)->lsm_oinfo);
1892 /* XXX zero *lsmp? */
1893 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
1894 LASSERT((*lsmp)->lsm_object_id);
1897 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
1902 static void filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
1903 struct filter_obd *filter)
1905 struct obdo doa; /* XXX obdo on stack */
1910 memset(&doa, 0, sizeof(doa));
1911 doa.o_mode = S_IFREG;
1912 doa.o_gr = oa->o_gr;
1913 doa.o_valid = oa->o_valid & OBD_MD_FLGROUP;
1914 last = filter_last_id(filter, oa->o_gr);
1915 CWARN("deleting orphan objects from "LPU64" to "LPU64"\n",
1916 oa->o_id + 1, last);
1917 for (id = oa->o_id + 1; id <= last; id++) {
1919 filter_destroy(exp, &doa, NULL, NULL);
1921 spin_lock(&filter->fo_objidlock);
1922 filter->fo_last_objids[doa.o_gr] = oa->o_id;
1923 spin_unlock(&filter->fo_objidlock);
1927 /* returns a negative error or a nonnegative number of files to create */
1928 static int filter_should_precreate(struct obd_export *exp, struct obdo *oa,
1931 struct obd_device *obd = exp->exp_obd;
1932 struct filter_obd *filter = &obd->u.filter;
1936 diff = oa->o_id - filter_last_id(filter, oa->o_gr);
1937 CDEBUG(D_INFO, "filter_last_id() = "LPU64" -> diff = %d\n",
1938 filter_last_id(filter, oa->o_gr), diff);
1940 /* delete orphans request */
1941 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1942 (oa->o_flags & OBD_FL_DELORPHAN)) {
1945 if (-diff > 10000) { /* XXX make this smarter */
1946 CERROR("ignoring bogus orphan destroy request: obdid "
1947 LPU64" last_id "LPU64"\n",
1948 oa->o_id, filter_last_id(filter, oa->o_gr));
1951 filter_destroy_precreated(exp, oa, filter);
1952 rc = filter_update_last_objid(obd, group, 0);
1954 CERROR("unable to write lastobjid, but orphans"
1958 /* only precreate if group == 0 and o_id is specfied */
1959 if (!(oa->o_valid & OBD_FL_DELORPHAN) &&
1960 (group != 0 || oa->o_id == 0))
1968 /* We rely on the fact that only one thread will be creating files in a given
1969 * group at a time, which is why we don't need an atomic filter_get_new_id.
1970 * Even if we had that atomic function, the following race would exist:
1972 * thread 1: gets id x from filter_next_id
1973 * thread 2: gets id (x + 1) from filter_next_id
1974 * thread 2: creates object (x + 1)
1975 * thread 1: tries to create object x, gets -ENOSPC
1977 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
1978 obd_gr group, int *num)
1980 struct dentry *dchild = NULL;
1981 struct filter_obd *filter;
1982 struct dentry *dparent;
1983 int err = 0, rc = 0, i;
1985 int recreate_obj = 0;
1986 void *handle = NULL;
1989 filter = &obd->u.filter;
1991 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1992 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
1996 for (i = 0; i < *num && err == 0; i++) {
1997 int cleanup_phase = 0;
2002 last_id = filter_last_id(filter, group);
2003 if (next_id > last_id) {
2004 CERROR("Error: Trying to recreate obj greater"
2005 "than last id "LPD64" > "LPD64"\n",
2010 next_id = filter_last_id(filter, group) + 1;
2013 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
2015 dparent = filter_parent_lock(obd, group, next_id);
2016 if (IS_ERR(dparent))
2017 GOTO(cleanup, rc = PTR_ERR(dparent));
2020 dchild = filter_fid2dentry(obd, dparent, group, next_id);
2022 GOTO(cleanup, rc = PTR_ERR(dchild));
2025 if (dchild->d_inode != NULL) {
2026 /* This would only happen if lastobjid was bad on disk*/
2027 /* Could also happen if recreating missing obj but
2031 CERROR("Serious error: recreating obj %*s but "
2032 "obj already exists \n",
2033 dchild->d_name.len, dchild->d_name.name);
2035 CERROR("Serious error: objid %*s already "
2036 "exists; is this filesystem corrupt?\n",
2037 dchild->d_name.len, dchild->d_name.name);
2039 GOTO(cleanup, rc = -EEXIST);
2042 handle = fsfilt_start_log(obd, dparent->d_inode,
2043 FSFILT_OP_CREATE, NULL, 1);
2045 GOTO(cleanup, rc = PTR_ERR(handle));
2048 rc = ll_vfs_create(dparent->d_inode, dchild, S_IFREG, NULL);
2050 CERROR("create failed rc = %d\n", rc);
2054 if (!recreate_obj) {
2055 filter_set_last_id(filter, group, next_id);
2056 err = filter_update_last_objid(obd, group, 0);
2058 CERROR("unable to write lastobjid "
2059 "but file created\n");
2063 switch(cleanup_phase) {
2065 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
2067 CERROR("error on commit, err = %d\n", err);
2074 filter_parent_unlock(dparent);
2084 CDEBUG(D_INFO, "filter_precreate() created %d objects\n", i);
2088 static int filter_create(struct obd_export *exp, struct obdo *oa,
2089 struct lov_stripe_md **ea, struct obd_trans_info *oti)
2091 struct obd_device *obd = NULL;
2092 struct filter_obd *filter;
2093 struct obd_run_ctxt saved;
2094 struct lov_stripe_md *lsm = NULL;
2095 struct filter_export_data *fed;
2096 int group = oa->o_gr;
2097 char str[PTL_NALFMT_SIZE];
2101 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
2102 portals_nid2str(exp->exp_connection->c_peer.peer_ni->pni_number,
2103 exp->exp_connection->c_peer.peer_nid, str);
2104 CERROR("!!! nid "LPX64"/%s sent invalid object group %d\n",
2105 exp->exp_connection->c_peer.peer_nid, str, group);
2110 fed = &exp->exp_filter_data;
2111 filter = &obd->u.filter;
2113 if (fed->fed_group != group) {
2114 portals_nid2str(exp->exp_connection->c_peer.peer_ni->pni_number,
2115 exp->exp_connection->c_peer.peer_nid, str);
2116 CERROR("!!! This export (nid "LPX64"/%s) used object group %d "
2117 "earlier; now it's trying to use group %d! This could "
2118 "be a bug in the MDS. Tell CFS.\n",
2119 exp->exp_connection->c_peer.peer_nid, str,
2120 fed->fed_group, group);
2124 CDEBUG(D_INFO, "filter_create(od->o_gr=%d,od->o_id="LPU64")\n",
2129 rc = obd_alloc_memmd(exp, &lsm);
2135 push_ctxt(&saved, &obd->obd_ctxt, NULL);
2137 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
2138 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
2139 if (oa->o_id > filter_last_id(&obd->u.filter, group)) {
2140 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
2141 oa->o_id, filter_last_id(&obd->u.filter, group));
2145 rc = filter_precreate(obd, oa, group, &diff);
2148 diff = filter_should_precreate(exp, oa, group);
2150 oa->o_id = filter_last_id(&obd->u.filter, group);
2151 rc = filter_precreate(obd, oa, group, &diff);
2153 oa->o_valid = OBD_MD_FLID;
2157 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
2158 if (rc && ea != NULL && *ea != lsm) {
2159 obd_free_memmd(exp, &lsm);
2160 } else if (rc == 0 && ea != NULL) {
2161 /* XXX LOV STACKING: the lsm that is passed to us from
2162 * LOV does not have valid lsm_oinfo data structs, so
2163 * don't go touching that. This needs to be fixed in a
2165 lsm->lsm_object_id = oa->o_id;
2166 lsm->lsm_object_gr = oa->o_gr;
2173 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
2174 struct lov_stripe_md *ea, struct obd_trans_info *oti)
2176 struct obd_device *obd;
2177 struct filter_obd *filter;
2178 struct dentry *dchild = NULL, *dparent = NULL;
2179 struct obd_run_ctxt saved;
2180 void *handle = NULL;
2181 struct llog_cookie *fcc = NULL;
2182 int rc, rc2, cleanup_phase = 0, have_prepared = 0;
2185 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
2188 filter = &obd->u.filter;
2190 push_ctxt(&saved, &obd->obd_ctxt, NULL);
2193 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
2194 if (IS_ERR(dparent))
2195 GOTO(cleanup, rc = PTR_ERR(dparent));
2198 dchild = filter_fid2dentry(obd, dparent, oa->o_gr, oa->o_id);
2200 GOTO(cleanup, rc = -ENOENT);
2203 if (dchild->d_inode == NULL) {
2204 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
2206 GOTO(cleanup, rc = -ENOENT);
2209 if (!have_prepared) {
2210 /* If we're really going to destroy the object, get ready
2211 * by getting the clients to discard their cached data.
2213 * We have to drop the parent lock, because
2214 * filter_prepare_destroy will acquire a PW on the object, and
2215 * we don't want to deadlock with an incoming write to the
2216 * object, which has the extent PW and then wants to get the
2217 * parent dentry to do the lookup.
2219 * We dput the child because it's not worth the extra
2220 * complication of condition the above code to skip it on the
2221 * second time through. */
2223 filter_parent_unlock(dparent);
2225 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
2230 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
2232 GOTO(cleanup, rc = PTR_ERR(handle));
2235 /* Our MDC connection is established by the MDS to us */
2236 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2237 OBD_ALLOC(fcc, sizeof(*fcc));
2239 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2242 rc = filter_destroy_internal(obd, oa->o_id, dparent, dchild);
2245 switch(cleanup_phase) {
2249 fsfilt_add_journal_cb(obd, 0, oti->oti_handle,
2250 filter_cancel_cookies_cb,
2253 fsfilt_add_journal_cb(obd, 0, handle,
2254 filter_cancel_cookies_cb,
2257 rc = filter_finish_transno(exp, oti, rc);
2258 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
2260 CERROR("error on commit, err = %d\n", rc2);
2267 filter_parent_unlock(dparent);
2269 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
2272 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2279 /* NB start and end are used for punch, but not truncate */
2280 static int filter_truncate(struct obd_export *exp, struct obdo *oa,
2281 struct lov_stripe_md *lsm,
2282 obd_off start, obd_off end,
2283 struct obd_trans_info *oti)
2288 if (end != OBD_OBJECT_EOF)
2289 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
2292 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = %x, "
2293 "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
2295 error = filter_setattr(exp, oa, NULL, oti);
2299 static int filter_sync(struct obd_export *exp, struct obdo *oa,
2300 struct lov_stripe_md *lsm, obd_off start, obd_off end)
2302 struct obd_device *obd = exp->exp_obd;
2303 struct obd_run_ctxt saved;
2304 struct filter_obd *filter;
2305 struct dentry *dentry;
2306 struct llog_ctxt *ctxt;
2310 filter = &obd->u.filter;
2312 /* an objid of zero is taken to mean "sync whole filesystem" */
2313 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
2314 rc = fsfilt_sync(obd, filter->fo_sb);
2315 /* flush any remaining cancel messages out to the target */
2316 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_REPL_CTXT);
2317 llog_sync(ctxt, exp);
2321 dentry = filter_oa2dentry(obd, oa);
2323 RETURN(PTR_ERR(dentry));
2325 push_ctxt(&saved, &obd->obd_ctxt, NULL);
2327 down(&dentry->d_inode->i_sem);
2328 rc = filemap_fdatasync(dentry->d_inode->i_mapping);
2330 /* just any file to grab fsync method - "file" arg unused */
2331 struct file *file = filter->fo_rcvd_filp;
2333 if (file->f_op && file->f_op->fsync)
2334 rc = file->f_op->fsync(NULL, dentry, 1);
2336 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
2340 up(&dentry->d_inode->i_sem);
2342 oa->o_valid = OBD_MD_FLID;
2343 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
2345 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
2351 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2352 unsigned long max_age)
2354 struct filter_obd *filter = &obd->u.filter;
2355 int blockbits = filter->fo_sb->s_blocksize_bits;
2359 /* at least try to account for cached pages. its still racey and
2360 * might be under-reporting if clients haven't announced their
2361 * caches with brw recently */
2362 spin_lock(&obd->obd_osfs_lock);
2363 rc = fsfilt_statfs(obd, filter->fo_sb, max_age);
2364 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
2365 spin_unlock(&obd->obd_osfs_lock);
2367 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
2368 " pending "LPU64" free "LPU64" avail "LPU64"\n",
2369 filter->fo_tot_dirty, filter->fo_tot_granted,
2370 filter->fo_tot_pending,
2371 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
2373 filter_grant_sanity_check(obd, __FUNCTION__);
2375 osfs->os_bavail -= min(osfs->os_bavail,
2376 (filter->fo_tot_dirty + filter->fo_tot_pending +
2377 osfs->os_bsize -1) >> blockbits);
2382 static int filter_get_info(struct obd_export *exp, __u32 keylen,
2383 void *key, __u32 *vallen, void *val)
2385 struct filter_export_data *fed = &exp->exp_filter_data;
2386 struct obd_device *obd;
2389 obd = class_exp2obd(exp);
2391 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2392 exp->exp_handle.h_cookie);
2396 if (keylen == strlen("blocksize") &&
2397 memcmp(key, "blocksize", keylen) == 0) {
2398 __u32 *blocksize = val;
2399 *vallen = sizeof(*blocksize);
2400 *blocksize = obd->u.filter.fo_sb->s_blocksize;
2404 if (keylen == strlen("blocksize_bits") &&
2405 memcmp(key, "blocksize_bits", keylen) == 0) {
2406 __u32 *blocksize_bits = val;
2407 *vallen = sizeof(*blocksize_bits);
2408 *blocksize_bits = obd->u.filter.fo_sb->s_blocksize_bits;
2412 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
2413 obd_id *last_id = val;
2414 *last_id = filter_last_id(&obd->u.filter, fed->fed_group);
2417 CDEBUG(D_IOCTL, "invalid key\n");
2421 struct obd_llogs * filter_grab_llog_for_group(struct obd_device *obd, int group)
2423 struct filter_group_llog *fglog, *nlog;
2424 struct filter_obd *filter;
2425 struct list_head *cur;
2428 filter = &obd->u.filter;
2430 spin_lock(&filter->fo_llog_list_lock);
2431 list_for_each(cur, &filter->fo_llog_list) {
2432 fglog = list_entry(cur, struct filter_group_llog, list);
2433 if (fglog->group == group) {
2434 spin_unlock(&filter->fo_llog_list_lock);
2435 RETURN(fglog->llogs);
2438 spin_unlock(&filter->fo_llog_list_lock);
2440 OBD_ALLOC(fglog, sizeof(*fglog));
2443 fglog->group = group;
2445 OBD_ALLOC(fglog->llogs, sizeof(struct obd_llogs));
2446 if (fglog->llogs == NULL) {
2447 OBD_FREE(fglog, sizeof(*fglog));
2451 spin_lock(&filter->fo_llog_list_lock);
2452 list_for_each(cur, &filter->fo_llog_list) {
2453 nlog = list_entry(cur, struct filter_group_llog, list);
2454 LASSERT(nlog->group != group);
2456 list_add(&fglog->list, &filter->fo_llog_list);
2457 spin_unlock(&filter->fo_llog_list_lock);
2459 rc = llog_cat_initialize(obd, fglog->llogs, 1);
2461 OBD_FREE(fglog->llogs, sizeof(*(fglog->llogs)));
2462 OBD_FREE(fglog, sizeof(*fglog));
2466 CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n", obd->obd_name,
2467 fglog->llogs, group);
2469 RETURN(fglog->llogs);
2472 static int filter_set_info(struct obd_export *exp, __u32 keylen,
2473 void *key, __u32 vallen, void *val)
2475 struct obd_run_ctxt saved;
2476 struct filter_export_data *fed = &exp->exp_filter_data;
2477 struct obd_device *obd;
2478 struct lustre_handle conn;
2479 struct obd_llogs *llog;
2480 struct llog_ctxt *ctxt;
2485 conn.cookie = exp->exp_handle.h_cookie;
2489 CDEBUG(D_IOCTL, "invalid exp %p cookie "LPX64"\n",
2494 if (keylen < strlen("mds_conn") ||
2495 memcmp(key, "mds_conn", keylen) != 0)
2498 group = *((__u32 *)val);
2499 if (fed->fed_group != 0 && fed->fed_group != group) {
2500 char str[PTL_NALFMT_SIZE];
2501 portals_nid2str(exp->exp_connection->c_peer.peer_ni->pni_number,
2502 exp->exp_connection->c_peer.peer_nid, str);
2503 CERROR("!!! This export (nid "LPX64"/%s) used object group %d "
2504 "earlier; now it's trying to use group %d! This could "
2505 "be a bug in the MDS. Tell CFS.\n",
2506 exp->exp_connection->c_peer.peer_nid, str,
2507 fed->fed_group, group);
2510 fed->fed_group = group;
2511 CWARN("Received MDS connection ("LPX64"); group %d\n", conn.cookie,
2516 push_ctxt(&saved, &obd->obd_ctxt, NULL);
2517 rc = filter_read_groups(obd, group, 1);
2518 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
2520 CERROR("can't read group %u\n", group);
2524 llog = filter_grab_llog_for_group(obd, group);
2525 LASSERT(llog != NULL);
2527 ctxt = llog_get_context(llog, LLOG_UNLINK_REPL_CTXT);
2528 LASSERT(ctxt != NULL);
2529 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
2533 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
2534 int len, void *karg, void *uarg)
2536 struct obd_device *obd = exp->exp_obd;
2537 struct obd_ioctl_data *data = karg;
2541 case OBD_IOC_ABORT_RECOVERY:
2542 CERROR("aborting recovery for device %s\n", obd->obd_name);
2543 target_abort_recovery(obd);
2546 case OBD_IOC_SET_READONLY: {
2548 struct super_block *sb = obd->u.filter.fo_sb;
2549 struct inode *inode = sb->s_root->d_inode;
2550 BDEVNAME_DECLARE_STORAGE(tmp);
2551 CERROR("setting device %s read-only\n",
2552 ll_bdevname(sb, tmp));
2554 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
2556 (void)fsfilt_commit(obd, inode, handle, 1);
2558 dev_set_rdonly(ll_sbdev(obd->u.filter.fo_sb), 2);
2562 case OBD_IOC_CATLOGLIST: {
2563 rc = llog_catlog_list(obd, 1, data);
2567 case OBD_IOC_LLOG_CANCEL:
2568 case OBD_IOC_LLOG_REMOVE:
2569 case OBD_IOC_LLOG_INFO:
2570 case OBD_IOC_LLOG_PRINT: {
2571 /* FIXME to be finished */
2572 RETURN(-EOPNOTSUPP);
2574 struct llog_ctxt *ctxt = NULL;
2576 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2577 rc = llog_ioctl(ctxt, cmd, data);
2578 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2591 static struct llog_operations filter_unlink_repl_logops;
2592 static struct llog_operations filter_size_orig_logops = {
2593 lop_setup: llog_obd_origin_setup,
2594 lop_cleanup: llog_obd_origin_cleanup,
2595 lop_add: llog_obd_origin_add
2598 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2599 struct obd_device *tgt, int count,
2600 struct llog_catid *logid)
2602 struct llog_ctxt *ctxt;
2606 filter_unlink_repl_logops = llog_client_ops;
2607 filter_unlink_repl_logops.lop_cancel = llog_obd_repl_cancel;
2608 filter_unlink_repl_logops.lop_connect = llog_repl_connect;
2609 filter_unlink_repl_logops.lop_sync = llog_obd_repl_sync;
2611 rc = llog_setup(obd, llogs, LLOG_UNLINK_REPL_CTXT, tgt, 0, NULL,
2612 &filter_unlink_repl_logops);
2615 /* FIXME - assign unlink_cb for filter's recovery */
2616 ctxt = llog_get_context(llogs, LLOG_UNLINK_REPL_CTXT);
2617 ctxt->llog_proc_cb = filter_recov_log_unlink_cb;
2619 rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2620 &filter_size_orig_logops);
2624 static int filter_llog_finish(struct obd_device *obd,
2625 struct obd_llogs *llogs, int count)
2630 rc = llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_REPL_CTXT));
2634 rc = llog_cleanup(llog_get_context(llogs, LLOG_SIZE_ORIG_CTXT));
2638 static int filter_llog_connect(struct obd_device *obd,
2639 struct llogd_conn_body *body)
2641 struct llog_ctxt *ctxt;
2642 struct obd_llogs *llog;
2646 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2647 (unsigned) body->lgdc_logid.lgl_ogr,
2648 (unsigned) body->lgdc_logid.lgl_oid,
2649 (unsigned) body->lgdc_logid.lgl_ogen);
2650 llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr);
2651 LASSERT(llog != NULL);
2652 ctxt = llog_get_context(llog, body->lgdc_ctxt_idx);
2653 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2654 &body->lgdc_gen, NULL);
2656 CERROR("failed to connect\n");
2661 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2664 return filter_fid2dentry(data, NULL, gr, id);
2667 static struct lvfs_callback_ops filter_lvfs_ops = {
2668 l_fid2dentry: filter_lvfs_fid2dentry,
2671 static struct obd_ops filter_obd_ops = {
2672 o_owner: THIS_MODULE,
2673 o_attach: filter_attach,
2674 o_detach: filter_detach,
2675 o_get_info: filter_get_info,
2676 o_set_info: filter_set_info,
2677 o_setup: filter_setup,
2678 o_precleanup: filter_precleanup,
2679 o_cleanup: filter_cleanup,
2680 o_connect: filter_connect,
2681 o_disconnect: filter_disconnect,
2682 o_statfs: filter_statfs,
2683 o_getattr: filter_getattr,
2684 o_unpackmd: filter_unpackmd,
2685 o_create: filter_create,
2686 o_setattr: filter_setattr,
2687 o_destroy: filter_destroy,
2689 o_punch: filter_truncate,
2690 o_sync: filter_sync,
2691 o_preprw: filter_preprw,
2692 o_commitrw: filter_commitrw,
2693 o_destroy_export: filter_destroy_export,
2694 o_llog_init: filter_llog_init,
2695 o_llog_finish: filter_llog_finish,
2696 o_llog_connect: filter_llog_connect,
2697 o_iocontrol: filter_iocontrol,
2700 static struct obd_ops filter_sanobd_ops = {
2701 o_owner: THIS_MODULE,
2702 o_attach: filter_attach,
2703 o_detach: filter_detach,
2704 o_get_info: filter_get_info,
2705 o_set_info: filter_set_info,
2706 o_setup: filter_san_setup,
2707 o_precleanup: filter_precleanup,
2708 o_cleanup: filter_cleanup,
2709 o_connect: filter_connect,
2710 o_disconnect: filter_disconnect,
2711 o_statfs: filter_statfs,
2712 o_getattr: filter_getattr,
2713 o_unpackmd: filter_unpackmd,
2714 o_create: filter_create,
2715 o_setattr: filter_setattr,
2716 o_destroy: filter_destroy,
2718 o_punch: filter_truncate,
2719 o_sync: filter_sync,
2720 o_preprw: filter_preprw,
2721 o_commitrw: filter_commitrw,
2722 o_san_preprw: filter_san_preprw,
2723 o_destroy_export: filter_destroy_export,
2724 o_llog_init: filter_llog_init,
2725 o_llog_finish: filter_llog_finish,
2726 o_llog_connect: filter_llog_connect,
2727 o_iocontrol: filter_iocontrol,
2730 static int __init obdfilter_init(void)
2732 struct lprocfs_static_vars lvars;
2735 printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
2737 lprocfs_init_vars(filter, &lvars);
2739 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
2740 OBD_FILTER_DEVICENAME);
2744 rc = class_register_type(&filter_sanobd_ops, NULL, lvars.module_vars,
2745 OBD_FILTER_SAN_DEVICENAME);
2747 class_unregister_type(OBD_FILTER_DEVICENAME);
2751 static void __exit obdfilter_exit(void)
2753 class_unregister_type(OBD_FILTER_SAN_DEVICENAME);
2754 class_unregister_type(OBD_FILTER_DEVICENAME);
2757 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2758 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
2759 MODULE_LICENSE("GPL");
2761 module_init(obdfilter_init);
2762 module_exit(obdfilter_exit);