1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (MDS) filesystem interface code
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of the Lustre file system, http://www.lustre.org
11 * Lustre is a trademark of Cluster File Systems, Inc.
13 * You may have signed or agreed to another license before downloading
14 * this software. If so, you are bound by the terms and conditions
15 * of that agreement, and the following does not apply to you. See the
16 * LICENSE file included with this distribution for more information.
18 * If you did not agree to a different license, then this copy of Lustre
19 * is open source software; you can redistribute it and/or modify it
20 * under the terms of version 2 of the GNU General Public License as
21 * published by the Free Software Foundation.
23 * In either case, Lustre is distributed in the hope that it will be
24 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
25 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
26 * license text for more details.
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/module.h>
35 #include <linux/kmod.h>
36 #include <linux/version.h>
37 #include <linux/sched.h>
38 #include <linux/lustre_quota.h>
39 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
40 #include <linux/mount.h>
42 #include <linux/lustre_mds.h>
43 #include <linux/obd_class.h>
44 #include <linux/obd_support.h>
45 #include <linux/lustre_lib.h>
46 #include <linux/lustre_fsfilt.h>
47 #include <linux/lustre_disk.h>
48 #include <libcfs/list.h>
50 #include "mds_internal.h"
52 #define HEALTH_CHECK "health_check"
54 /* Add client data to the MDS. We use a bitmap to locate a free space
55 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
56 * Otherwise, we have just read the data from the last_rcvd file and
59 * It should not be possible to fail adding an existing client - otherwise
60 * mds_init_server_data() callsite needs to be fixed.
62 int mds_client_add(struct obd_device *obd, struct mds_obd *mds,
63 struct mds_export_data *med, int cl_idx)
65 unsigned long *bitmap = mds->mds_client_bitmap;
66 int new_client = (cl_idx == -1);
69 LASSERT(bitmap != NULL);
70 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
72 /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
73 if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
76 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
77 * there's no need for extra complication here
80 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
82 if (cl_idx >= LR_MAX_CLIENTS ||
83 OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_CLIENT_ADD)) {
84 CERROR("no room for clients - fix LR_MAX_CLIENTS\n");
87 if (test_and_set_bit(cl_idx, bitmap)) {
88 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
93 if (test_and_set_bit(cl_idx, bitmap)) {
94 CERROR("MDS client %d: bit already set in bitmap!!\n",
100 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
101 cl_idx, med->med_mcd->mcd_uuid);
103 med->med_lr_idx = cl_idx;
104 med->med_lr_off = le32_to_cpu(mds->mds_server_data->lsd_client_start) +
105 (cl_idx * le16_to_cpu(mds->mds_server_data->lsd_client_size));
106 LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
109 struct lvfs_run_ctxt saved;
110 loff_t off = med->med_lr_off;
111 struct file *file = mds->mds_rcvd_filp;
114 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
115 rc = fsfilt_write_record(obd, file, med->med_mcd,
116 sizeof(*med->med_mcd), &off, 1);
117 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
121 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
122 med->med_lr_idx, med->med_lr_off,
123 (unsigned int)sizeof(*med->med_mcd));
128 int mds_client_free(struct obd_export *exp)
130 struct mds_export_data *med = &exp->exp_mds_data;
131 struct mds_obd *mds = &exp->exp_obd->u.mds;
132 struct obd_device *obd = exp->exp_obd;
133 struct mds_client_data zero_mcd;
134 struct lvfs_run_ctxt saved;
142 /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
143 if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
146 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
147 med->med_lr_idx, med->med_lr_off, med->med_mcd->mcd_uuid);
149 LASSERT(mds->mds_client_bitmap != NULL);
151 off = med->med_lr_off;
153 /* Don't clear med_lr_idx here as it is likely also unset. At worst
154 * we leak a client slot that will be cleaned on the next recovery. */
156 CERROR("%s: client idx %d has offset %lld\n",
157 obd->obd_name, med->med_lr_idx, off);
158 GOTO(free, rc = -EINVAL);
161 /* Clear the bit _after_ zeroing out the client so we don't
162 race with mds_client_add and zero out new clients.*/
163 if (!test_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
164 CERROR("MDS client %u: bit already clear in bitmap!!\n",
169 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
170 memset(&zero_mcd, 0, sizeof zero_mcd);
171 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
172 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
173 sizeof(zero_mcd), &off, 1);
174 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
176 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
177 "zeroing out client %s idx %u in %s rc %d\n",
178 med->med_mcd->mcd_uuid, med->med_lr_idx, LAST_RCVD, rc);
181 if (!test_and_clear_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
182 CERROR("MDS client %u: bit already clear in bitmap!!\n",
188 /* Make sure the server's last_transno is up to date. Do this
189 * after the client is freed so we know all the client's
190 * transactions have been committed. */
191 mds_update_server_data(exp->exp_obd, 0);
195 OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
201 static int mds_server_free_data(struct mds_obd *mds)
203 OBD_FREE(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
204 OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
205 mds->mds_server_data = NULL;
210 static int mds_init_server_data(struct obd_device *obd, struct file *file)
212 struct mds_obd *mds = &obd->u.mds;
213 struct lr_server_data *lsd;
214 struct mds_client_data *mcd = NULL;
216 unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
221 /* ensure padding in the struct is the correct size */
222 LASSERT(offsetof(struct lr_server_data, lsd_padding) +
223 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
224 LASSERT(offsetof(struct mds_client_data, mcd_padding) +
225 sizeof(mcd->mcd_padding) == LR_CLIENT_SIZE);
227 OBD_ALLOC_WAIT(lsd, sizeof(*lsd));
231 OBD_ALLOC_WAIT(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
232 if (!mds->mds_client_bitmap) {
233 OBD_FREE(lsd, sizeof(*lsd));
237 mds->mds_server_data = lsd;
239 if (last_rcvd_size == 0) {
240 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
242 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
243 lsd->lsd_last_transno = 0;
244 mount_count = lsd->lsd_mount_count = 0;
245 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
246 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
247 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
248 lsd->lsd_feature_rocompat = cpu_to_le32(OBD_ROCOMPAT_LOVOBJID);
249 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_MDT |
250 OBD_INCOMPAT_COMMON_LR);
252 rc = fsfilt_read_record(obd, file, lsd, sizeof(*lsd), &off);
254 CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
257 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
258 CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
259 obd->obd_uuid.uuid, lsd->lsd_uuid);
260 GOTO(err_msd, rc = -EINVAL);
262 mount_count = le64_to_cpu(lsd->lsd_mount_count);
264 if (lsd->lsd_feature_incompat & ~cpu_to_le32(MDT_INCOMPAT_SUPP)) {
265 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
266 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
268 GOTO(err_msd, rc = -EINVAL);
271 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(MDT_ROCOMPAT_SUPP)) {
272 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
273 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
275 /* Do something like remount filesystem read-only */
276 GOTO(err_msd, rc = -EINVAL);
279 if (!(lsd->lsd_feature_incompat & cpu_to_le32(OBD_INCOMPAT_COMMON_LR))){
280 CDEBUG(D_WARNING, "using old last_rcvd format\n");
281 lsd->lsd_mount_count = lsd->lsd_last_transno;
282 lsd->lsd_last_transno = lsd->lsd_unused;
283 /* If we update the last_rcvd, we can never go back to
284 an old install, so leave this in the old format for now.
285 lsd->lsd_feature_incompat |= cpu_to_le32(LR_INCOMPAT_COMMON_LR);
289 mds->mds_last_transno = le64_to_cpu(lsd->lsd_last_transno);
291 lsd->lsd_feature_compat = cpu_to_le32(OBD_COMPAT_MDT);
292 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
293 obd->obd_name, mds->mds_last_transno);
294 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
295 obd->obd_name, mount_count + 1);
296 CDEBUG(D_INODE, "%s: server data size: %u\n",
297 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
298 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
299 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
300 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
301 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
302 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
303 obd->obd_name, last_rcvd_size);
304 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
305 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
306 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
307 le16_to_cpu(lsd->lsd_client_size));
309 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
310 !lsd->lsd_client_size) {
311 CERROR("Bad last_rcvd contents!\n");
312 GOTO(err_msd, rc = -EINVAL);
315 /* When we do a clean MDS shutdown, we save the last_transno into
316 * the header. If we find clients with higher last_transno values
317 * then those clients may need recovery done. */
318 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
319 off < last_rcvd_size; cl_idx++) {
321 struct obd_export *exp;
322 struct mds_export_data *med;
325 OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
327 GOTO(err_client, rc = -ENOMEM);
330 /* Don't assume off is incremented properly by
331 * fsfilt_read_record(), in case sizeof(*mcd)
332 * isn't the same as lsd->lsd_client_size. */
333 off = le32_to_cpu(lsd->lsd_client_start) +
334 cl_idx * le16_to_cpu(lsd->lsd_client_size);
335 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
337 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
338 LAST_RCVD, cl_idx, off, rc);
339 break; /* read error shouldn't cause startup to fail */
342 if (mcd->mcd_uuid[0] == '\0') {
343 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
348 last_transno = le64_to_cpu(mcd->mcd_last_transno);
350 /* These exports are cleaned up by mds_disconnect(), so they
351 * need to be set up like real exports as mds_connect() does.
353 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
354 " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
355 last_transno, le64_to_cpu(lsd->lsd_last_transno),
356 le64_to_cpu(mcd->mcd_last_xid));
358 exp = class_new_export(obd);
360 GOTO(err_client, rc = -ENOMEM);
362 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
363 sizeof exp->exp_client_uuid.uuid);
364 med = &exp->exp_mds_data;
366 rc = mds_client_add(obd, mds, med, cl_idx);
367 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
369 /* create helper if export init gets more complex */
370 INIT_LIST_HEAD(&med->med_open_head);
371 spin_lock_init(&med->med_open_lock);
374 exp->exp_replay_needed = 1;
375 obd->obd_recoverable_clients++;
376 obd->obd_max_recoverable_clients++;
377 class_export_put(exp);
379 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
380 cl_idx, last_transno);
382 if (last_transno > mds->mds_last_transno)
383 mds->mds_last_transno = last_transno;
387 OBD_FREE(mcd, sizeof(*mcd));
389 obd->obd_last_committed = mds->mds_last_transno;
391 if (obd->obd_recoverable_clients) {
392 CWARN("RECOVERY: service %s, %d recoverable clients, "
393 "last_transno "LPU64"\n", obd->obd_name,
394 obd->obd_recoverable_clients, mds->mds_last_transno);
395 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
396 obd->obd_recovering = 1;
397 obd->obd_recovery_start = CURRENT_SECONDS;
398 /* Only used for lprocfs_status */
399 obd->obd_recovery_end = obd->obd_recovery_start +
400 OBD_RECOVERY_TIMEOUT / HZ;
403 mds->mds_mount_count = mount_count + 1;
404 lsd->lsd_mount_count = cpu_to_le64(mds->mds_mount_count);
406 /* save it, so mount count and last_transno is current */
407 rc = mds_update_server_data(obd, 1);
409 GOTO(err_client, rc);
414 class_disconnect_exports(obd);
416 mds_server_free_data(mds);
420 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
422 struct mds_obd *mds = &obd->u.mds;
423 struct lvfs_run_ctxt saved;
424 struct dentry *dentry;
429 rc = cleanup_group_info();
433 mds->mds_vfsmnt = mnt;
434 obd->u.obt.obt_sb = mnt->mnt_root->d_inode->i_sb;
436 fsfilt_setup(obd, obd->u.obt.obt_sb);
438 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
439 obd->obd_lvfs_ctxt.pwdmnt = mnt;
440 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
441 obd->obd_lvfs_ctxt.fs = get_ds();
442 obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
444 /* setup the directory tree */
445 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
446 dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
447 if (IS_ERR(dentry)) {
448 rc = PTR_ERR(dentry);
449 CERROR("cannot create ROOT directory: rc = %d\n", rc);
453 mds->mds_rootfid.id = dentry->d_inode->i_ino;
454 mds->mds_rootfid.generation = dentry->d_inode->i_generation;
455 mds->mds_rootfid.f_type = S_IFDIR;
459 dentry = lookup_one_len("__iopen__", current->fs->pwd,
460 strlen("__iopen__"));
461 if (IS_ERR(dentry)) {
462 rc = PTR_ERR(dentry);
463 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
467 mds->mds_fid_de = dentry;
468 if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
470 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
474 dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
475 if (IS_ERR(dentry)) {
476 rc = PTR_ERR(dentry);
477 CERROR("cannot create PENDING directory: rc = %d\n", rc);
480 mds->mds_pending_dir = dentry;
483 dentry = simple_mkdir(current->fs->pwd, MDT_LOGS_DIR, 0777, 1);
484 if (IS_ERR(dentry)) {
485 rc = PTR_ERR(dentry);
486 CERROR("cannot create %s directory: rc = %d\n",
488 GOTO(err_pending, rc);
490 mds->mds_logs_dir = dentry;
493 dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
494 if (IS_ERR(dentry)) {
495 rc = PTR_ERR(dentry);
496 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
499 mds->mds_objects_dir = dentry;
501 /* open and test the last rcvd file */
502 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
505 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
506 GOTO(err_objects, rc = PTR_ERR(file));
508 mds->mds_rcvd_filp = file;
509 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
510 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
511 file->f_dentry->d_inode->i_mode);
512 GOTO(err_last_rcvd, rc = -ENOENT);
515 rc = mds_init_server_data(obd, file);
517 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
518 GOTO(err_last_rcvd, rc);
521 /* open and test the lov objd file */
522 file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
525 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
526 GOTO(err_client, rc = PTR_ERR(file));
528 mds->mds_lov_objid_filp = file;
529 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
530 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
531 file->f_dentry->d_inode->i_mode);
532 GOTO(err_lov_objid, rc = -ENOENT);
535 /* open and test the check io file junk */
536 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
539 CERROR("cannot open/create %s file: rc = %d\n", HEALTH_CHECK, rc);
540 GOTO(err_lov_objid, rc = PTR_ERR(file));
542 mds->mds_health_check_filp = file;
543 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
544 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
545 file->f_dentry->d_inode->i_mode);
546 GOTO(err_health_check, rc = -ENOENT);
548 rc = lvfs_check_io_health(obd, file);
550 GOTO(err_health_check, rc);
552 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
557 if (mds->mds_health_check_filp &&
558 filp_close(mds->mds_health_check_filp, 0))
559 CERROR("can't close %s after error\n", HEALTH_CHECK);
561 if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
562 CERROR("can't close %s after error\n", LOV_OBJID);
564 class_disconnect_exports(obd);
566 if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
567 CERROR("can't close %s after error\n", LAST_RCVD);
569 dput(mds->mds_objects_dir);
571 dput(mds->mds_logs_dir);
573 dput(mds->mds_pending_dir);
575 dput(mds->mds_fid_de);
580 int mds_fs_cleanup(struct obd_device *obd)
582 struct mds_obd *mds = &obd->u.mds;
583 struct lvfs_run_ctxt saved;
587 LCONSOLE_WARN("%s: shutting down for failover; client state "
588 "will be preserved.\n", obd->obd_name);
590 class_disconnect_exports(obd); /* cleans up client info too */
591 mds_server_free_data(mds);
593 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
594 if (mds->mds_rcvd_filp) {
595 rc = filp_close(mds->mds_rcvd_filp, 0);
596 mds->mds_rcvd_filp = NULL;
598 CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
600 if (mds->mds_lov_objid_filp) {
601 rc = filp_close(mds->mds_lov_objid_filp, 0);
602 mds->mds_lov_objid_filp = NULL;
604 CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
606 if (mds->mds_health_check_filp) {
607 rc = filp_close(mds->mds_health_check_filp, 0);
608 mds->mds_health_check_filp = NULL;
610 CERROR("%s file won't close, rc=%d\n", HEALTH_CHECK, rc);
612 if (mds->mds_objects_dir != NULL) {
613 l_dput(mds->mds_objects_dir);
614 mds->mds_objects_dir = NULL;
616 if (mds->mds_logs_dir) {
617 l_dput(mds->mds_logs_dir);
618 mds->mds_logs_dir = NULL;
620 if (mds->mds_pending_dir) {
621 l_dput(mds->mds_pending_dir);
622 mds->mds_pending_dir = NULL;
625 lquota_fs_cleanup(quota_interface, obd);
627 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
628 shrink_dcache_parent(mds->mds_fid_de);
629 dput(mds->mds_fid_de);
630 LL_DQUOT_OFF(obd->u.obt.obt_sb);
635 /* Creates an object with the same name as its fid. Because this is not at all
636 * performance sensitive, it is accomplished by creating a file, checking the
637 * fid, and renaming it. */
638 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
639 struct lov_stripe_md **ea, struct obd_trans_info *oti)
641 struct mds_obd *mds = &exp->exp_obd->u.mds;
642 struct inode *parent_inode = mds->mds_objects_dir->d_inode;
643 unsigned int tmpname = ll_rand();
645 struct dentry *new_child;
646 struct lvfs_run_ctxt saved;
647 char fidname[LL_FID_NAMELEN];
649 struct lvfs_ucred ucred = { 0 };
650 int rc = 0, err, namelen;
653 /* the owner of object file should always be root */
654 ucred.luc_cap = current->cap_effective | CAP_SYS_RESOURCE;
656 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
658 sprintf(fidname, "OBJECTS/%u.%u", tmpname, current->pid);
659 filp = filp_open(fidname, O_CREAT | O_EXCL, 0666);
663 CERROR("impossible object name collision %u\n",
667 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
671 LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
673 oa->o_id = filp->f_dentry->d_inode->i_ino;
674 oa->o_generation = filp->f_dentry->d_inode->i_generation;
675 namelen = lu_fid2str(fidname, oa->o_id, oa->o_generation);
677 down(&parent_inode->i_sem);
678 new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
680 if (IS_ERR(new_child)) {
681 CERROR("getting neg dentry for obj rename: %d\n", rc);
682 GOTO(out_close, rc = PTR_ERR(new_child));
684 if (new_child->d_inode != NULL) {
685 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
686 oa->o_id, oa->o_generation);
690 handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
691 FSFILT_OP_RENAME, NULL);
693 GOTO(out_dput, rc = PTR_ERR(handle));
696 rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
697 mds->mds_objects_dir->d_inode, new_child);
700 CERROR("error renaming new object "LPU64":%u: rc %d\n",
701 oa->o_id, oa->o_generation, rc);
703 err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
706 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
712 up(&parent_inode->i_sem);
713 err = filp_close(filp, 0);
715 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
720 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
724 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
725 struct lov_stripe_md *ea, struct obd_trans_info *oti,
726 struct obd_export *md_exp)
728 struct mds_obd *mds = &exp->exp_obd->u.mds;
729 struct inode *parent_inode = mds->mds_objects_dir->d_inode;
730 struct obd_device *obd = exp->exp_obd;
731 struct lvfs_run_ctxt saved;
732 struct lvfs_ucred ucred = { 0 };
733 char fidname[lu_fid_NAMELEN];
736 int err, namelen, rc = 0;
739 ucred.luc_cap = current->cap_effective | CAP_SYS_RESOURCE;
740 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
742 namelen = lu_fid2str(fidname, oa->o_id, oa->o_generation);
744 down(&parent_inode->i_sem);
745 de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
749 CERROR("error looking up object "LPU64" %s: rc %d\n",
750 oa->o_id, fidname, rc);
753 if (de->d_inode == NULL) {
754 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
755 oa->o_id, fidname, rc);
756 GOTO(out_dput, rc = -ENOENT);
759 /* Stripe count is 1 here since this is some MDS specific stuff
760 that is unlinked, not spanned across multiple OSTs */
761 handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
762 FSFILT_OP_UNLINK, oti, 1);
765 GOTO(out_dput, rc = PTR_ERR(handle));
767 rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
769 CERROR("error destroying object "LPU64":%u: rc %d\n",
770 oa->o_id, oa->o_generation, rc);
772 err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
778 up(&parent_inode->i_sem);
780 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);