1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
6 * Lustre Metadata Server (mds) reintegration routines
8 * Copyright (C) 2002 Cluster File Systems, Inc.
9 * author: Peter Braam <braam@clusterfs.com>
11 * This code is issued under the GNU General Public License.
12 * See the file COPYING in this distribution
16 // XXX - add transaction sequence numbers
20 #include <linux/version.h>
21 #include <linux/module.h>
23 #include <linux/stat.h>
24 #include <linux/locks.h>
25 #include <linux/quotaops.h>
26 #include <asm/unistd.h>
27 #include <asm/uaccess.h>
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/obd_support.h>
32 #include <linux/obd_class.h>
33 #include <linux/obd.h>
34 #include <linux/lustre_lib.h>
35 #include <linux/lustre_idl.h>
36 #include <linux/lustre_mds.h>
37 #include <linux/obd_class.h>
39 struct mds_client_info *mds_uuid_to_mci(struct mds_obd *mds, __u8 *uuid)
46 list_for_each(p, &mds->mds_client_info) {
47 struct mds_client_info *mci;
49 mci = list_entry(p, struct mds_client_info, mci_list);
50 CDEBUG(D_INFO, "checking client UUID '%s'\n",
51 mci->mci_mcd->mcd_uuid);
52 if (!strncmp(mci->mci_mcd->mcd_uuid, uuid,
53 sizeof(mci->mci_mcd->mcd_uuid)))
56 CDEBUG(D_INFO, "no mds client info found for UUID '%s'\n", uuid);
60 int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
61 struct ptlrpc_request *req)
63 /* get from req->rq_connection-> or req->rq_client */
64 struct obd_run_ctxt saved;
65 struct mds_client_info *mci;
69 mci = mds_uuid_to_mci(mds, req->rq_connection->c_remote_uuid);
71 CERROR("unable to locate MDS client data for UUID '%s'\n",
72 ptlrpc_req_to_uuid(req));
73 /* This will be a real error once everything is working */
78 off = MDS_LR_CLIENT + mci->mci_off * MDS_LR_SIZE;
80 ++mds->mds_last_rcvd; /* lock this, or make it an LDLM function? */
81 req->rq_repmsg->transno = HTON__u64(mds->mds_last_rcvd);
82 mci->mci_mcd->mcd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
83 mci->mci_mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
84 mci->mci_mcd->mcd_last_xid = cpu_to_le32(req->rq_reqmsg->xid);
86 mds_fs_set_last_rcvd(mds, handle);
87 push_ctxt(&saved, &mds->mds_ctxt);
88 rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mci->mci_mcd,
89 sizeof(*mci->mci_mcd), &off);
91 CDEBUG(D_INODE, "wrote trans #%Ld for client '%s' at #%d: rc = %d\n",
92 mds->mds_last_rcvd, mci->mci_mcd->mcd_uuid, mci->mci_off, rc);
93 // store new value and last committed value in req struct
95 if (rc == sizeof(*mci->mci_mcd))
98 CERROR("error writing to last_rcvd file: rc = %d\n", rc);
106 static int mds_reint_setattr(struct mds_update_record *rec,
107 struct ptlrpc_request *req)
109 struct mds_obd *mds = &req->rq_obd->u.mds;
115 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
116 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_SETATTR)) {
117 GOTO(out_setattr, rc = -ESTALE);
120 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
122 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE,
123 de->d_inode->i_sb->s_dev);
125 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
127 GOTO(out_setattr_de, rc = PTR_ERR(handle));
128 rc = mds_fs_setattr(mds, de, handle, &rec->ur_iattr);
131 rc = mds_update_last_rcvd(mds, handle, req);
133 err = mds_fs_commit(mds, de->d_inode, handle);
135 CERROR("error on commit: err = %d\n", err);
147 static int mds_reint_recreate(struct mds_update_record *rec,
148 struct ptlrpc_request *req)
150 struct dentry *de = NULL;
151 struct mds_obd *mds = &req->rq_obd->u.mds;
152 struct dentry *dchild = NULL;
157 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
158 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
160 GOTO(out_create_de, rc = -ESTALE);
163 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
166 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
167 if (IS_ERR(dchild)) {
168 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
171 GOTO(out_create_dchild, rc = -ESTALE);
174 if (dchild->d_inode) {
175 struct mds_body *body;
177 body = lustre_msg_buf(req->rq_repmsg, 0);
178 body->ino = dchild->d_inode->i_ino;
179 body->generation = dchild->d_inode->i_generation;
181 CERROR("child doesn't exist (dir %ld, name %s)\n",
182 dir->i_ino, rec->ur_name);
196 static int mds_reint_create(struct mds_update_record *rec,
197 struct ptlrpc_request *req)
199 struct dentry *de = NULL;
200 struct mds_obd *mds = &req->rq_obd->u.mds;
201 struct dentry *dchild = NULL;
204 int rc = 0, type = rec->ur_mode & S_IFMT;
208 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
209 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
211 GOTO(out_create_de, rc = -ESTALE);
214 CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
215 dir->i_ino, rec->ur_name, rec->ur_mode);
218 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
219 if (IS_ERR(dchild)) {
220 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
223 GOTO(out_create_dchild, rc = -ESTALE);
226 if (dchild->d_inode) {
227 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
228 dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
230 GOTO(out_create_dchild, rc = -EEXIST);
233 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
237 handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
239 GOTO(out_create_dchild, PTR_ERR(handle));
240 rc = vfs_create(dir, dchild, rec->ur_mode);
245 handle = mds_fs_start(mds, dir, MDS_FSOP_MKDIR);
247 GOTO(out_create_dchild, PTR_ERR(handle));
248 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
253 handle = mds_fs_start(mds, dir, MDS_FSOP_SYMLINK);
255 GOTO(out_create_dchild, PTR_ERR(handle));
256 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
264 int rdev = rec->ur_id;
265 handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
267 GOTO(out_create_dchild, PTR_ERR(handle));
268 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
273 CERROR("bad file type %d for create of %s\n",type,rec->ur_name);
274 GOTO(out_create_dchild, rc = -EINVAL);
278 CERROR("error during create: %d\n", rc);
279 if (rc != -ENOSPC) LBUG();
280 GOTO(out_create_commit, rc);
283 struct inode *inode = dchild->d_inode;
284 struct mds_body *body;
286 CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
287 if (type == S_IFREG) {
288 rc = mds_fs_set_objid(mds, inode, handle, rec->ur_id);
290 CERROR("error %d setting objid for %ld\n",
294 iattr.ia_atime = rec->ur_time;
295 iattr.ia_ctime = rec->ur_time;
296 iattr.ia_mtime = rec->ur_time;
297 iattr.ia_uid = rec->ur_uid;
298 iattr.ia_gid = rec->ur_gid;
299 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
300 ATTR_MTIME | ATTR_CTIME;
302 rc = mds_fs_setattr(mds, dchild, handle, &iattr);
304 CERROR("error on setattr: rc = %d\n", rc);
305 /* XXX should we abort here in case of error? */
308 rc = mds_update_last_rcvd(mds, handle, req);
310 CERROR("error on update_last_rcvd: rc = %d\n", rc);
311 /* XXX should we abort here in case of error? */
314 body = lustre_msg_buf(req->rq_repmsg, 0);
315 body->ino = inode->i_ino;
316 body->generation = inode->i_generation;
320 err = mds_fs_commit(mds, dir, handle);
322 CERROR("error on commit: err = %d\n", err);
335 static int mds_reint_unlink(struct mds_update_record *rec,
336 struct ptlrpc_request *req)
338 struct dentry *de = NULL;
339 struct dentry *dchild = NULL;
340 struct mds_obd *mds = &req->rq_obd->u.mds;
341 struct inode *dir, *inode;
347 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
348 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) {
350 GOTO(out_unlink, rc = -ESTALE);
353 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
356 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
357 if (IS_ERR(dchild)) {
358 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
360 GOTO(out_unlink_de, rc = -ESTALE);
363 inode = dchild->d_inode;
365 CERROR("child doesn't exist (dir %ld, name %s\n",
366 dir->i_ino, rec->ur_name);
368 GOTO(out_unlink_dchild, rc = -ESTALE);
371 if (inode->i_ino != rec->ur_fid2->id) {
372 CERROR("inode and FID ID do not match (%ld != %Ld)\n",
373 inode->i_ino, rec->ur_fid2->id);
375 GOTO(out_unlink_dchild, rc = -ESTALE);
377 if (inode->i_generation != rec->ur_fid2->generation) {
378 CERROR("inode and FID GENERATION do not match (%d != %d)\n",
379 inode->i_generation, rec->ur_fid2->generation);
381 GOTO(out_unlink_dchild, rc = -ESTALE);
384 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
386 switch (dchild->d_inode->i_mode & S_IFMT) {
388 handle = mds_fs_start(mds, dir, MDS_FSOP_RMDIR);
390 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
391 rc = vfs_rmdir(dir, dchild);
394 handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
396 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
397 rc = vfs_unlink(dir, dchild);
402 rc = mds_update_last_rcvd(mds, handle, req);
403 err = mds_fs_commit(mds, dir, handle);
405 CERROR("error on commit: err = %d\n", err);
421 static int mds_reint_link(struct mds_update_record *rec,
422 struct ptlrpc_request *req)
424 struct dentry *de_src = NULL;
425 struct dentry *de_tgt_dir = NULL;
426 struct dentry *dchild = NULL;
427 struct mds_obd *mds = &req->rq_obd->u.mds;
433 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
434 if (IS_ERR(de_src) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) {
435 GOTO(out_link, rc = -ESTALE);
438 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
439 if (IS_ERR(de_tgt_dir)) {
440 GOTO(out_link_de_src, rc = -ESTALE);
443 down(&de_tgt_dir->d_inode->i_sem);
444 dchild = lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen - 1);
445 if (IS_ERR(dchild)) {
446 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
447 GOTO(out_link_de_tgt_dir, rc = -ESTALE);
450 if (dchild->d_inode) {
451 CERROR("child exists (dir %ld, name %s\n",
452 de_tgt_dir->d_inode->i_ino, rec->ur_name);
453 GOTO(out_link_dchild, rc = -EEXIST);
456 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
457 dchild->d_inode->i_sb->s_dev);
459 handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
461 GOTO(out_link_dchild, rc = PTR_ERR(handle));
463 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
466 rc = mds_update_last_rcvd(mds, handle, req);
468 err = mds_fs_commit(mds, de_tgt_dir->d_inode, handle);
470 CERROR("error on commit: err = %d\n", err);
479 up(&de_tgt_dir->d_inode->i_sem);
488 static int mds_reint_rename(struct mds_update_record *rec,
489 struct ptlrpc_request *req)
491 struct dentry *de_srcdir = NULL;
492 struct dentry *de_tgtdir = NULL;
493 struct dentry *de_old = NULL;
494 struct dentry *de_new = NULL;
495 struct mds_obd *mds = &req->rq_obd->u.mds;
501 de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
502 if (IS_ERR(de_srcdir)) {
503 GOTO(out_rename, rc = -ESTALE);
506 de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
507 if (IS_ERR(de_tgtdir)) {
508 GOTO(out_rename_srcdir, rc = -ESTALE);
511 de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
512 if (IS_ERR(de_old)) {
513 CERROR("old child lookup error %ld\n", PTR_ERR(de_old));
514 GOTO(out_rename_tgtdir, rc = -ESTALE);
517 de_new = lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
518 if (IS_ERR(de_new)) {
519 CERROR("new child lookup error %ld\n", PTR_ERR(de_new));
520 GOTO(out_rename_deold, rc = -ESTALE);
523 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
524 de_srcdir->d_inode->i_sb->s_dev);
526 handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
528 GOTO(out_rename_denew, rc = PTR_ERR(handle));
529 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
532 rc = mds_update_last_rcvd(mds, handle, req);
534 err = mds_fs_commit(mds, de_tgtdir->d_inode, handle);
536 CERROR("error on commit: err = %d\n", err);
555 typedef int (*mds_reinter)(struct mds_update_record *, struct ptlrpc_request*);
557 static mds_reinter reinters[REINT_MAX+1] = {
558 [REINT_SETATTR] mds_reint_setattr,
559 [REINT_CREATE] mds_reint_create,
560 [REINT_UNLINK] mds_reint_unlink,
561 [REINT_LINK] mds_reint_link,
562 [REINT_RENAME] mds_reint_rename,
563 [REINT_RECREATE] mds_reint_recreate,
566 int mds_reint_rec(struct mds_update_record *rec, struct ptlrpc_request *req)
568 int rc, size = sizeof(struct mds_body);
570 if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
571 CERROR("opcode %d not valid\n", rec->ur_opcode);
572 rc = req->rq_status = -EINVAL;
576 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
578 CERROR("mds: out of memory\n");
579 rc = req->rq_status = -ENOMEM;
583 rc = reinters[rec->ur_opcode](rec, req);