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
19 #define DEBUG_SUBSYSTEM S_MDS
21 #include <linux/obd_support.h>
22 #include <linux/obd_class.h>
23 #include <linux/obd.h>
24 #include <linux/lustre_lib.h>
25 #include <linux/lustre_idl.h>
26 #include <linux/lustre_mds.h>
27 #include <linux/obd_class.h>
29 struct mds_client_info *mds_uuid_to_mci(struct mds_obd *mds, __u8 *uuid)
36 list_for_each(p, &mds->mds_client_info) {
37 struct mds_client_info *mci;
39 mci = list_entry(p, struct mds_client_info, mci_list);
40 CDEBUG(D_INFO, "checking client UUID '%s'\n",
41 mci->mci_mcd->mcd_uuid);
42 if (!strncmp(mci->mci_mcd->mcd_uuid, uuid,
43 sizeof(mci->mci_mcd->mcd_uuid)))
46 CDEBUG(D_INFO, "no mds client info found for UUID '%s'\n", uuid);
50 int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
51 struct ptlrpc_request *req)
53 /* get from req->rq_connection-> or req->rq_client */
54 struct obd_run_ctxt saved;
55 struct mds_client_info *mci;
59 mci = mds_uuid_to_mci(mds, req->rq_connection->c_remote_uuid);
61 CERROR("unable to locate MDS client data for UUID '%s'\n",
62 ptlrpc_req_to_uuid(req));
63 /* This will be a real error once everything is working */
68 off = MDS_LR_CLIENT + mci->mci_off * MDS_LR_SIZE;
70 ++mds->mds_last_rcvd; /* lock this, or make it an LDLM function? */
71 req->rq_repmsg->transno = HTON__u64(mds->mds_last_rcvd);
72 mci->mci_mcd->mcd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
73 mci->mci_mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
74 mci->mci_mcd->mcd_last_xid = cpu_to_le32(req->rq_reqmsg->xid);
76 mds_fs_set_last_rcvd(mds, handle);
77 push_ctxt(&saved, &mds->mds_ctxt);
78 rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mci->mci_mcd,
79 sizeof(*mci->mci_mcd), &off);
81 CDEBUG(D_INODE, "wrote trans #%Ld for client '%s' at #%d: rc = %d\n",
82 mds->mds_last_rcvd, mci->mci_mcd->mcd_uuid, mci->mci_off, rc);
83 // store new value and last committed value in req struct
85 if (rc == sizeof(*mci->mci_mcd))
88 CERROR("error writing to last_rcvd file: rc = %d\n", rc);
96 static int mds_reint_setattr(struct mds_update_record *rec,
97 struct ptlrpc_request *req)
99 struct mds_obd *mds = &req->rq_obd->u.mds;
105 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
106 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_SETATTR)) {
107 GOTO(out_setattr, rc = -ESTALE);
110 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
112 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE,
113 de->d_inode->i_sb->s_dev);
115 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
117 GOTO(out_setattr_de, rc = PTR_ERR(handle));
118 rc = mds_fs_setattr(mds, de, handle, &rec->ur_iattr);
121 rc = mds_update_last_rcvd(mds, handle, req);
123 err = mds_fs_commit(mds, de->d_inode, handle);
125 CERROR("error on commit: err = %d\n", err);
137 static int mds_reint_recreate(struct mds_update_record *rec,
138 struct ptlrpc_request *req)
140 struct dentry *de = NULL;
141 struct mds_obd *mds = &req->rq_obd->u.mds;
142 struct dentry *dchild = NULL;
147 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
148 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
150 GOTO(out_create_de, rc = -ESTALE);
153 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
156 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
157 if (IS_ERR(dchild)) {
158 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
161 GOTO(out_create_dchild, rc = -ESTALE);
164 if (dchild->d_inode) {
165 struct mds_body *body;
167 body = lustre_msg_buf(req->rq_repmsg, 0);
168 body->ino = dchild->d_inode->i_ino;
169 body->generation = dchild->d_inode->i_generation;
171 CERROR("child doesn't exist (dir %ld, name %s)\n",
172 dir->i_ino, rec->ur_name);
186 static int mds_reint_create(struct mds_update_record *rec,
187 struct ptlrpc_request *req)
189 struct dentry *de = NULL;
190 struct mds_obd *mds = &req->rq_obd->u.mds;
191 struct dentry *dchild = NULL;
194 int rc = 0, type = rec->ur_mode & S_IFMT;
198 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
199 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
201 GOTO(out_create_de, rc = -ESTALE);
204 CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
205 dir->i_ino, rec->ur_name, rec->ur_mode);
208 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
209 if (IS_ERR(dchild)) {
210 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
213 GOTO(out_create_dchild, rc = -ESTALE);
216 if (dchild->d_inode) {
217 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
218 dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
220 GOTO(out_create_dchild, rc = -EEXIST);
223 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
227 handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
229 GOTO(out_create_dchild, PTR_ERR(handle));
230 rc = vfs_create(dir, dchild, rec->ur_mode);
235 handle = mds_fs_start(mds, dir, MDS_FSOP_MKDIR);
237 GOTO(out_create_dchild, PTR_ERR(handle));
238 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
243 handle = mds_fs_start(mds, dir, MDS_FSOP_SYMLINK);
245 GOTO(out_create_dchild, PTR_ERR(handle));
246 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
254 int rdev = rec->ur_rdev;
255 handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
257 GOTO(out_create_dchild, PTR_ERR(handle));
258 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
263 CERROR("bad file type %d for create of %s\n",type,rec->ur_name);
264 GOTO(out_create_dchild, rc = -EINVAL);
268 CERROR("error during create: %d\n", rc);
269 if (rc != -ENOSPC) LBUG();
270 GOTO(out_create_commit, rc);
273 struct inode *inode = dchild->d_inode;
274 struct mds_body *body;
276 CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
277 if (type == S_IFREG) {
279 obdo = lustre_msg_buf(req->rq_reqmsg, 2);
280 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
282 CERROR("error %d setting objid for %ld\n",
286 iattr.ia_atime = rec->ur_time;
287 iattr.ia_ctime = rec->ur_time;
288 iattr.ia_mtime = rec->ur_time;
289 iattr.ia_uid = rec->ur_uid;
290 iattr.ia_gid = rec->ur_gid;
291 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
292 ATTR_MTIME | ATTR_CTIME;
294 rc = mds_fs_setattr(mds, dchild, handle, &iattr);
296 CERROR("error on setattr: rc = %d\n", rc);
297 /* XXX should we abort here in case of error? */
300 rc = mds_update_last_rcvd(mds, handle, req);
302 CERROR("error on update_last_rcvd: rc = %d\n", rc);
303 /* XXX should we abort here in case of error? */
306 body = lustre_msg_buf(req->rq_repmsg, 0);
307 body->ino = inode->i_ino;
308 body->generation = inode->i_generation;
312 err = mds_fs_commit(mds, dir, handle);
314 CERROR("error on commit: err = %d\n", err);
327 static int mds_reint_unlink(struct mds_update_record *rec,
328 struct ptlrpc_request *req)
330 struct dentry *de = NULL;
331 struct dentry *dchild = NULL;
332 struct mds_obd *mds = &req->rq_obd->u.mds;
333 struct inode *dir, *inode;
339 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
340 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) {
342 GOTO(out_unlink, rc = -ESTALE);
345 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
348 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
349 if (IS_ERR(dchild)) {
350 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
352 GOTO(out_unlink_de, rc = -ESTALE);
355 inode = dchild->d_inode;
357 CERROR("child doesn't exist (dir %ld, name %s\n",
358 dir->i_ino, rec->ur_name);
360 GOTO(out_unlink_dchild, rc = -ESTALE);
363 if (inode->i_ino != rec->ur_fid2->id) {
364 CERROR("inode and FID ID do not match (%ld != %Ld)\n",
365 inode->i_ino, rec->ur_fid2->id);
367 GOTO(out_unlink_dchild, rc = -ESTALE);
369 if (inode->i_generation != rec->ur_fid2->generation) {
370 CERROR("inode and FID GENERATION do not match (%d != %d)\n",
371 inode->i_generation, rec->ur_fid2->generation);
373 GOTO(out_unlink_dchild, rc = -ESTALE);
376 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
378 switch (dchild->d_inode->i_mode & S_IFMT) {
380 handle = mds_fs_start(mds, dir, MDS_FSOP_RMDIR);
382 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
383 rc = vfs_rmdir(dir, dchild);
386 handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
388 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
389 rc = vfs_unlink(dir, dchild);
394 rc = mds_update_last_rcvd(mds, handle, req);
395 err = mds_fs_commit(mds, dir, handle);
397 CERROR("error on commit: err = %d\n", err);
413 static int mds_reint_link(struct mds_update_record *rec,
414 struct ptlrpc_request *req)
416 struct dentry *de_src = NULL;
417 struct dentry *de_tgt_dir = NULL;
418 struct dentry *dchild = NULL;
419 struct mds_obd *mds = &req->rq_obd->u.mds;
425 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
426 if (IS_ERR(de_src) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) {
427 GOTO(out_link, rc = -ESTALE);
430 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
431 if (IS_ERR(de_tgt_dir)) {
432 GOTO(out_link_de_src, rc = -ESTALE);
435 down(&de_tgt_dir->d_inode->i_sem);
436 dchild = lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen - 1);
437 if (IS_ERR(dchild)) {
438 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
439 GOTO(out_link_de_tgt_dir, rc = -ESTALE);
442 if (dchild->d_inode) {
443 CERROR("child exists (dir %ld, name %s\n",
444 de_tgt_dir->d_inode->i_ino, rec->ur_name);
445 GOTO(out_link_dchild, rc = -EEXIST);
448 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
449 de_src->d_inode->i_sb->s_dev);
451 handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
453 GOTO(out_link_dchild, rc = PTR_ERR(handle));
455 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
458 rc = mds_update_last_rcvd(mds, handle, req);
460 err = mds_fs_commit(mds, de_tgt_dir->d_inode, handle);
462 CERROR("error on commit: err = %d\n", err);
471 up(&de_tgt_dir->d_inode->i_sem);
480 static int mds_reint_rename(struct mds_update_record *rec,
481 struct ptlrpc_request *req)
483 struct dentry *de_srcdir = NULL;
484 struct dentry *de_tgtdir = NULL;
485 struct dentry *de_old = NULL;
486 struct dentry *de_new = NULL;
487 struct mds_obd *mds = &req->rq_obd->u.mds;
493 de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
494 if (IS_ERR(de_srcdir)) {
495 GOTO(out_rename, rc = -ESTALE);
498 de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
499 if (IS_ERR(de_tgtdir)) {
500 GOTO(out_rename_srcdir, rc = -ESTALE);
503 de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
504 if (IS_ERR(de_old)) {
505 CERROR("old child lookup error %ld\n", PTR_ERR(de_old));
506 GOTO(out_rename_tgtdir, rc = -ESTALE);
509 de_new = lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
510 if (IS_ERR(de_new)) {
511 CERROR("new child lookup error %ld\n", PTR_ERR(de_new));
512 GOTO(out_rename_deold, rc = -ESTALE);
515 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
516 de_srcdir->d_inode->i_sb->s_dev);
518 handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
520 GOTO(out_rename_denew, rc = PTR_ERR(handle));
521 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new,
525 rc = mds_update_last_rcvd(mds, handle, req);
527 err = mds_fs_commit(mds, de_tgtdir->d_inode, handle);
529 CERROR("error on commit: err = %d\n", err);
548 typedef int (*mds_reinter)(struct mds_update_record *, struct ptlrpc_request*);
550 static mds_reinter reinters[REINT_MAX+1] = {
551 [REINT_SETATTR] mds_reint_setattr,
552 [REINT_CREATE] mds_reint_create,
553 [REINT_UNLINK] mds_reint_unlink,
554 [REINT_LINK] mds_reint_link,
555 [REINT_RENAME] mds_reint_rename,
556 [REINT_RECREATE] mds_reint_recreate,
559 int mds_reint_rec(struct mds_update_record *rec, struct ptlrpc_request *req)
561 int rc, size = sizeof(struct mds_body);
563 if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
564 CERROR("opcode %d not valid\n", rec->ur_opcode);
565 rc = req->rq_status = -EINVAL;
569 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
571 CERROR("mds: out of memory\n");
572 rc = req->rq_status = -ENOMEM;
576 rc = reinters[rec->ur_opcode](rec, req);