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/lustre_dlm.h>
28 #include <linux/obd_class.h>
30 struct mds_client_info *mds_uuid_to_mci(struct mds_obd *mds, __u8 *uuid)
37 list_for_each(p, &mds->mds_client_info) {
38 struct mds_client_info *mci;
40 mci = list_entry(p, struct mds_client_info, mci_list);
41 CDEBUG(D_INFO, "checking client UUID '%s'\n",
42 mci->mci_mcd->mcd_uuid);
43 if (!strncmp(mci->mci_mcd->mcd_uuid, uuid,
44 sizeof(mci->mci_mcd->mcd_uuid)))
47 CDEBUG(D_INFO, "no mds client info found for UUID '%s'\n", uuid);
51 int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
52 struct ptlrpc_request *req)
54 /* get from req->rq_connection-> or req->rq_client */
55 struct obd_run_ctxt saved;
56 struct mds_client_info *mci;
60 mci = mds_uuid_to_mci(mds, req->rq_connection->c_remote_uuid);
62 CERROR("unable to locate MDS client data for UUID '%s'\n",
63 ptlrpc_req_to_uuid(req));
64 /* This will be a real error once everything is working */
69 off = MDS_LR_CLIENT + mci->mci_off * MDS_LR_SIZE;
71 ++mds->mds_last_rcvd; /* lock this, or make it an LDLM function? */
72 req->rq_repmsg->transno = HTON__u64(mds->mds_last_rcvd);
73 mci->mci_mcd->mcd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
74 mci->mci_mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
75 mci->mci_mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
77 mds_fs_set_last_rcvd(mds, handle);
78 push_ctxt(&saved, &mds->mds_ctxt);
79 rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mci->mci_mcd,
80 sizeof(*mci->mci_mcd), &off);
82 CDEBUG(D_INODE, "wrote trans #%Ld for client '%s' at #%d: rc = %d\n",
83 mds->mds_last_rcvd, mci->mci_mcd->mcd_uuid, mci->mci_off, rc);
84 // store new value and last committed value in req struct
86 if (rc == sizeof(*mci->mci_mcd))
89 CERROR("error writing to last_rcvd file: rc = %d\n", rc);
97 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
98 struct ptlrpc_request *req)
100 struct mds_obd *mds = &req->rq_obd->u.mds;
106 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
107 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_SETATTR)) {
108 GOTO(out_setattr, rc = -ESTALE);
111 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
113 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE,
114 de->d_inode->i_sb->s_dev);
116 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
118 GOTO(out_setattr_de, rc = PTR_ERR(handle));
119 rc = mds_fs_setattr(mds, de, handle, &rec->ur_iattr);
122 rc = mds_update_last_rcvd(mds, handle, req);
124 err = mds_fs_commit(mds, de->d_inode, handle);
126 CERROR("error on commit: err = %d\n", err);
138 static int mds_reint_recreate(struct mds_update_record *rec, int offset,
139 struct ptlrpc_request *req)
141 struct dentry *de = NULL;
142 struct mds_obd *mds = &req->rq_obd->u.mds;
143 struct dentry *dchild = NULL;
148 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
149 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
151 GOTO(out_create_de, rc = -ESTALE);
154 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
157 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
158 if (IS_ERR(dchild)) {
159 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
162 GOTO(out_create_dchild, rc = -ESTALE);
165 if (dchild->d_inode) {
166 struct mds_body *body;
168 body = lustre_msg_buf(req->rq_repmsg, 0);
169 body->ino = dchild->d_inode->i_ino;
170 body->generation = dchild->d_inode->i_generation;
172 CERROR("child doesn't exist (dir %ld, name %s)\n",
173 dir->i_ino, rec->ur_name);
187 static int mds_reint_create(struct mds_update_record *rec, int offset,
188 struct ptlrpc_request *req)
190 struct dentry *de = NULL;
191 struct mds_obd *mds = &req->rq_obd->u.mds;
192 struct dentry *dchild = NULL;
195 struct ldlm_lock *lock;
196 struct lustre_handle lockh;
197 int rc = 0, err, flags, lock_mode, type = rec->ur_mode & S_IFMT;
198 __u64 res_id[3] = {0,0,0};
201 /* requests were at offset 2, replies go back at 1 */
205 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
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);
217 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
218 res_id[0] = dir->i_ino;
220 rc = ldlm_local_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
221 NULL, 0, lock_mode, &lockh);
223 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
224 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
225 NULL, mds->mds_local_namespace, NULL,
226 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
227 &flags, (void *)mds_lock_callback, NULL,
229 if (rc != ELDLM_OK) {
230 CERROR("lock enqueue: err: %d\n", rc);
231 GOTO(out_create_de, rc = -EIO);
234 lock = lustre_handle2object(&lockh);
235 LDLM_DEBUG(lock, "matched");
237 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
240 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
241 if (IS_ERR(dchild)) {
242 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
245 GOTO(out_create_dchild, rc = -ESTALE);
248 if (dchild->d_inode) {
249 struct mds_body *body;
251 struct inode *inode = dchild->d_inode;
252 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
253 dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
255 body = lustre_msg_buf(req->rq_repmsg, offset);
256 mds_pack_inode2fid(&body->fid1, inode);
257 mds_pack_inode2body(body, inode);
258 #warning FIXME: This ext3/N-specific code does not belong here
259 /* If i_file_acl is set, this inode has an EA */
260 if (S_ISREG(inode->i_mode) && inode->u.ext3_i.i_file_acl) {
261 obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
262 mds_fs_get_obdo(mds, inode, obdo);
264 /* now a normal case for intent locking */
265 GOTO(out_create_dchild, rc = -EEXIST);
268 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
272 handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
274 GOTO(out_create_dchild, PTR_ERR(handle));
275 rc = vfs_create(dir, dchild, rec->ur_mode);
280 handle = mds_fs_start(mds, dir, MDS_FSOP_MKDIR);
282 GOTO(out_create_dchild, PTR_ERR(handle));
283 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
288 handle = mds_fs_start(mds, dir, MDS_FSOP_SYMLINK);
290 GOTO(out_create_dchild, PTR_ERR(handle));
291 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
299 int rdev = rec->ur_rdev;
300 handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
302 GOTO(out_create_dchild, PTR_ERR(handle));
303 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
308 CERROR("bad file type %d for create of %s\n",type,rec->ur_name);
309 GOTO(out_create_dchild, rc = -EINVAL);
313 CERROR("error during create: %d\n", rc);
314 if (rc != -ENOSPC) LBUG();
315 GOTO(out_create_commit, rc);
318 struct inode *inode = dchild->d_inode;
319 struct mds_body *body;
321 CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
322 if (!offset && type == S_IFREG) {
324 obdo = lustre_msg_buf(req->rq_reqmsg, 2);
325 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
327 CERROR("error %d setting objid for %ld\n",
331 iattr.ia_atime = rec->ur_time;
332 iattr.ia_ctime = rec->ur_time;
333 iattr.ia_mtime = rec->ur_time;
334 iattr.ia_uid = rec->ur_uid;
335 iattr.ia_gid = rec->ur_gid;
336 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
337 ATTR_MTIME | ATTR_CTIME;
339 rc = mds_fs_setattr(mds, dchild, handle, &iattr);
341 CERROR("error on setattr: rc = %d\n", rc);
342 /* XXX should we abort here in case of error? */
345 rc = mds_update_last_rcvd(mds, handle, req);
347 CERROR("error on update_last_rcvd: rc = %d\n", rc);
348 /* XXX should we abort here in case of error? */
351 body = lustre_msg_buf(req->rq_repmsg, offset);
352 body->ino = inode->i_ino;
353 body->generation = inode->i_generation;
357 err = mds_fs_commit(mds, dir, handle);
359 CERROR("error on commit: err = %d\n", err);
366 lock = lustre_handle2object(&lockh);
367 ldlm_lock_decref(lock, lock_mode);
374 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
375 struct ptlrpc_request *req)
377 struct dentry *de = NULL;
378 struct dentry *dchild = NULL;
379 struct mds_obd *mds = &req->rq_obd->u.mds;
381 struct inode *dir, *inode;
382 int lock_mode, flags;
383 __u64 res_id[3] = {0};
384 struct lustre_handle lockh;
385 struct ldlm_lock *lock;
391 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
392 if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) {
394 GOTO(out_unlink, rc = -ESTALE);
398 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
399 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
400 res_id[0] = dir->i_ino;
402 rc = ldlm_local_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
403 NULL, 0, lock_mode, &lockh);
405 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
406 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
407 NULL, mds->mds_local_namespace, NULL,
408 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
409 &flags, (void *)mds_lock_callback, NULL,
411 if (rc != ELDLM_OK) {
412 CERROR("lock enqueue: err: %d\n", rc);
413 GOTO(out_unlink_de, rc = -EIO);
416 lock = lustre_handle2object(&lockh);
417 LDLM_DEBUG(lock, "matched");
419 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
422 dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
423 if (IS_ERR(dchild)) {
424 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
426 GOTO(out_unlink_de, rc = -ESTALE);
429 inode = dchild->d_inode;
431 CERROR("child doesn't exist (dir %ld, name %s\n",
432 dir->i_ino, rec->ur_name);
433 GOTO(out_unlink_dchild, rc = -ENOENT);
436 #if 0 /* in intent case the client doesn't have the inode */
437 if (inode->i_ino != rec->ur_fid2->id) {
438 CERROR("inode and FID ID do not match (%ld != %Ld)\n",
439 inode->i_ino, rec->ur_fid2->id);
441 GOTO(out_unlink_dchild, rc = -ESTALE);
443 if (inode->i_generation != rec->ur_fid2->generation) {
444 CERROR("inode and FID GENERATION do not match (%d != %d)\n",
445 inode->i_generation, rec->ur_fid2->generation);
447 GOTO(out_unlink_dchild, rc = -ESTALE);
451 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
453 switch (dchild->d_inode->i_mode & S_IFMT) {
455 handle = mds_fs_start(mds, dir, MDS_FSOP_RMDIR);
457 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
458 rc = vfs_rmdir(dir, dchild);
462 obdo = lustre_msg_buf(req->rq_repmsg, 1);
463 rc = mds_fs_get_obdo(mds, inode, obdo);
465 CDEBUG(D_INFO, "No obdo for ino %ld err %d\n",
469 handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
471 GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
472 rc = vfs_unlink(dir, dchild);
477 rc = mds_update_last_rcvd(mds, handle, req);
478 err = mds_fs_commit(mds, dir, handle);
480 CERROR("error on commit: err = %d\n", err);
488 res_id[0] = inode->i_ino;
492 lock = lustre_handle2object(&lockh);
493 ldlm_lock_decref(lock, lock_mode);
495 /* Take an exclusive lock on the resource that we're
496 * about to free, to force everyone to drop their
498 LDLM_DEBUG_NOLOCK("getting EX lock res %Lu", res_id[0]);
499 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
500 NULL, mds->mds_local_namespace, NULL,
502 LDLM_PLAIN, NULL, 0, LCK_EX, &flags,
503 (void *)mds_lock_callback, NULL, 0,
506 CERROR("failed to get child inode lock (child ino %Ld, "
508 res_id[0], de->d_inode->i_ino);
514 lock = lustre_handle2object(&lockh);
515 ldlm_lock_decref(lock, LCK_EX);
516 rc = ldlm_cli_cancel(lock->l_client, lock);
518 CERROR("failed to cancel child inode lock ino "
519 "%Ld: %d\n", res_id[0], rc);
527 static int mds_reint_link(struct mds_update_record *rec, int offset,
528 struct ptlrpc_request *req)
530 struct dentry *de_src = NULL;
531 struct dentry *de_tgt_dir = NULL;
532 struct dentry *dchild = NULL;
533 struct mds_obd *mds = &req->rq_obd->u.mds;
539 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
540 if (IS_ERR(de_src) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) {
541 GOTO(out_link, rc = -ESTALE);
544 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
545 if (IS_ERR(de_tgt_dir)) {
546 GOTO(out_link_de_src, rc = -ESTALE);
549 down(&de_tgt_dir->d_inode->i_sem);
550 dchild = lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen - 1);
551 if (IS_ERR(dchild)) {
552 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
553 GOTO(out_link_de_tgt_dir, rc = -ESTALE);
556 if (dchild->d_inode) {
557 CERROR("child exists (dir %ld, name %s\n",
558 de_tgt_dir->d_inode->i_ino, rec->ur_name);
559 GOTO(out_link_dchild, rc = -EEXIST);
562 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
563 de_src->d_inode->i_sb->s_dev);
565 handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
567 GOTO(out_link_dchild, rc = PTR_ERR(handle));
569 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
572 rc = mds_update_last_rcvd(mds, handle, req);
574 err = mds_fs_commit(mds, de_tgt_dir->d_inode, handle);
576 CERROR("error on commit: err = %d\n", err);
585 up(&de_tgt_dir->d_inode->i_sem);
594 static int mds_reint_rename(struct mds_update_record *rec, int offset,
595 struct ptlrpc_request *req)
597 struct dentry *de_srcdir = NULL;
598 struct dentry *de_tgtdir = NULL;
599 struct dentry *de_old = NULL;
600 struct dentry *de_new = NULL;
601 struct mds_obd *mds = &req->rq_obd->u.mds;
607 de_srcdir = mds_fid2dentry(mds, rec->ur_fid1, NULL);
608 if (IS_ERR(de_srcdir)) {
609 GOTO(out_rename, rc = -ESTALE);
612 de_tgtdir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
613 if (IS_ERR(de_tgtdir)) {
614 GOTO(out_rename_srcdir, rc = -ESTALE);
617 #warning FIXME: This needs locking attention
619 de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
620 if (IS_ERR(de_old)) {
621 CERROR("old child lookup error %ld\n", PTR_ERR(de_old));
622 GOTO(out_rename_tgtdir, rc = -ESTALE);
625 de_new = lookup_one_len(rec->ur_tgt, de_tgtdir, rec->ur_tgtlen - 1);
626 if (IS_ERR(de_new)) {
627 CERROR("new child lookup error %ld\n", PTR_ERR(de_new));
628 GOTO(out_rename_deold, rc = -ESTALE);
631 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
632 de_srcdir->d_inode->i_sb->s_dev);
634 handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
636 GOTO(out_rename_denew, rc = PTR_ERR(handle));
637 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new,
641 rc = mds_update_last_rcvd(mds, handle, req);
643 err = mds_fs_commit(mds, de_tgtdir->d_inode, handle);
645 CERROR("error on commit: err = %d\n", err);
664 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
665 struct ptlrpc_request *);
667 static mds_reinter reinters[REINT_MAX+1] = {
668 [REINT_SETATTR] mds_reint_setattr,
669 [REINT_CREATE] mds_reint_create,
670 [REINT_UNLINK] mds_reint_unlink,
671 [REINT_LINK] mds_reint_link,
672 [REINT_RENAME] mds_reint_rename,
673 [REINT_RECREATE] mds_reint_recreate,
676 int mds_reint_rec(struct mds_update_record *rec, int offset,
677 struct ptlrpc_request *req)
681 if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
682 CERROR("opcode %d not valid\n", rec->ur_opcode);
683 rc = req->rq_status = -EINVAL;
687 rc = reinters[rec->ur_opcode](rec, offset, req);