1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
5 * Lustre Metadata Server (mds) reintegration routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/obd_support.h>
35 #include <linux/obd_class.h>
36 #include <linux/obd.h>
37 #include <linux/lustre_lib.h>
38 #include <linux/lustre_idl.h>
39 #include <linux/lustre_mds.h>
40 #include <linux/lustre_dlm.h>
41 #include <linux/lustre_fsfilt.h>
43 #include "mds_internal.h"
45 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
48 obd_transno_commit_cb(obd, transno, error);
51 struct mds_logcancel_data {
52 struct lov_mds_md *mlcd_lmm;
56 struct llog_cookie mlcd_cookies[0];
60 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
61 void *cb_data, int error)
63 struct mds_logcancel_data *mlcd = cb_data;
64 struct lov_stripe_md *lsm = NULL;
65 struct llog_ctxt *ctxt;
68 obd_transno_commit_cb(obd, transno, error);
70 CDEBUG(D_HA, "cancelling %d cookies\n",
71 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
73 rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
74 mlcd->mlcd_eadatalen);
76 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
77 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
80 ///* XXX 0 normally, SENDNOW for debug */);
81 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
82 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
83 sizeof(*mlcd->mlcd_cookies),
84 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
86 CERROR("error cancelling %d log cookies: rc %d\n",
87 (int)(mlcd->mlcd_cookielen /
88 sizeof(*mlcd->mlcd_cookies)), rc);
91 OBD_FREE(mlcd, mlcd->mlcd_size);
94 /* Assumes caller has already pushed us into the kernel context. */
95 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
96 struct ptlrpc_request *req, int rc, __u32 op_data)
98 struct mds_export_data *med = &req->rq_export->exp_mds_data;
99 struct mds_client_data *mcd = med->med_mcd;
100 struct obd_device *obd = req->rq_export->exp_obd;
107 /* if the export has already been failed, we have no last_rcvd slot */
108 if (req->rq_export->exp_failed) {
109 CERROR("committing transaction for disconnected client\n");
118 if (handle == NULL) {
119 /* if we're starting our own xaction, use our own inode */
120 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
121 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
122 if (IS_ERR(handle)) {
123 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
124 RETURN(PTR_ERR(handle));
130 transno = req->rq_reqmsg->transno;
132 LASSERT(transno == 0);
133 } else if (transno == 0) {
134 spin_lock(&mds->mds_transno_lock);
135 transno = ++mds->mds_last_transno;
136 spin_unlock(&mds->mds_transno_lock);
138 spin_lock(&mds->mds_transno_lock);
139 if (transno > mds->mds_last_transno)
140 mds->mds_last_transno = transno;
141 spin_unlock(&mds->mds_transno_lock);
143 req->rq_repmsg->transno = req->rq_transno = transno;
144 mcd->mcd_last_transno = cpu_to_le64(transno);
145 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
146 mcd->mcd_last_result = cpu_to_le32(rc);
147 mcd->mcd_last_data = cpu_to_le32(op_data);
149 fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
150 mds_commit_cb, NULL);
151 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
160 DEBUG_REQ(log_pri, req,
161 "wrote trans #"LPU64" client %s at idx %u: err = %d",
162 transno, mcd->mcd_uuid, med->med_idx, err);
164 err = mds_lov_write_objids(obd);
170 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
173 err = fsfilt_commit(obd, inode, handle, 0);
175 CERROR("error committing transaction: %d\n", err);
183 /* this gives the same functionality as the code between
184 * sys_chmod and inode_setattr
185 * chown_common and inode_setattr
186 * utimes and inode_setattr
188 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
190 time_t now = LTIME_S(CURRENT_TIME);
191 struct iattr *attr = &rec->ur_iattr;
192 unsigned int ia_valid = attr->ia_valid;
196 /* only fix up attrs if the client VFS didn't already */
197 if (!(ia_valid & ATTR_RAW))
200 if (!(ia_valid & ATTR_CTIME_SET))
201 LTIME_S(attr->ia_ctime) = now;
202 if (!(ia_valid & ATTR_ATIME_SET))
203 LTIME_S(attr->ia_atime) = now;
204 if (!(ia_valid & ATTR_MTIME_SET))
205 LTIME_S(attr->ia_mtime) = now;
207 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
211 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
212 if (rec->ur_fsuid != inode->i_uid &&
213 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
217 if (ia_valid & ATTR_SIZE) {
218 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
222 if (ia_valid & ATTR_UID) {
225 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
227 if (attr->ia_uid == (uid_t) -1)
228 attr->ia_uid = inode->i_uid;
229 if (attr->ia_gid == (gid_t) -1)
230 attr->ia_gid = inode->i_gid;
231 attr->ia_mode = inode->i_mode;
233 * If the user or group of a non-directory has been
234 * changed by a non-root user, remove the setuid bit.
235 * 19981026 David C Niemi <niemi@tux.org>
237 * Changed this to apply to all users, including root,
238 * to avoid some races. This is the behavior we had in
239 * 2.0. The check for non-root was definitely wrong
240 * for 2.2 anyway, as it should have been using
241 * CAP_FSETID rather than fsuid -- 19990830 SD.
243 if ((inode->i_mode & S_ISUID) == S_ISUID &&
244 !S_ISDIR(inode->i_mode)) {
245 attr->ia_mode &= ~S_ISUID;
246 attr->ia_valid |= ATTR_MODE;
249 * Likewise, if the user or group of a non-directory
250 * has been changed by a non-root user, remove the
251 * setgid bit UNLESS there is no group execute bit
252 * (this would be a file marked for mandatory
253 * locking). 19981026 David C Niemi <niemi@tux.org>
255 * Removed the fsuid check (see the comment above) --
258 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
259 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
260 attr->ia_mode &= ~S_ISGID;
261 attr->ia_valid |= ATTR_MODE;
263 } else if (ia_valid & ATTR_MODE) {
264 int mode = attr->ia_mode;
266 if (attr->ia_mode == (mode_t) -1)
267 attr->ia_mode = inode->i_mode;
269 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
274 void mds_steal_ack_locks(struct ptlrpc_request *req)
276 struct obd_export *exp = req->rq_export;
277 struct list_head *tmp;
278 struct ptlrpc_reply_state *oldrep;
279 struct ptlrpc_service *svc;
281 char str[PTL_NALFMT_SIZE];
284 /* CAVEAT EMPTOR: spinlock order */
285 spin_lock_irqsave (&exp->exp_lock, flags);
286 list_for_each (tmp, &exp->exp_outstanding_replies) {
287 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
289 if (oldrep->rs_xid != req->rq_xid)
292 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
293 CERROR ("Resent req xid "LPX64" has mismatched opc: "
294 "new %d old %d\n", req->rq_xid,
295 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
297 svc = oldrep->rs_srv_ni->sni_service;
298 spin_lock (&svc->srv_lock);
300 list_del_init (&oldrep->rs_exp_list);
302 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
304 oldrep->rs_nlocks, oldrep,
305 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
306 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
308 for (i = 0; i < oldrep->rs_nlocks; i++)
309 ptlrpc_save_lock(req,
310 &oldrep->rs_locks[i],
311 oldrep->rs_modes[i]);
312 oldrep->rs_nlocks = 0;
314 DEBUG_REQ(D_HA, req, "stole locks for");
315 ptlrpc_schedule_difficult_reply (oldrep);
317 spin_unlock (&svc->srv_lock);
318 spin_unlock_irqrestore (&exp->exp_lock, flags);
321 spin_unlock_irqrestore (&exp->exp_lock, flags);
324 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
326 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
327 mcd->mcd_last_transno, mcd->mcd_last_result);
328 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
329 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
331 mds_steal_ack_locks(req);
334 static void reconstruct_reint_setattr(struct mds_update_record *rec,
335 int offset, struct ptlrpc_request *req)
337 struct mds_export_data *med = &req->rq_export->exp_mds_data;
338 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
340 struct mds_body *body;
342 mds_req_from_mcd(req, med->med_mcd);
344 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
346 LASSERT(PTR_ERR(de) == req->rq_status);
350 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
351 mds_pack_inode2fid(&body->fid1, de->d_inode);
352 mds_pack_inode2body(body, de->d_inode);
354 /* Don't return OST-specific attributes if we didn't just set them */
355 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
356 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
357 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
358 body->valid |= OBD_MD_FLMTIME;
359 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
360 body->valid |= OBD_MD_FLATIME;
365 /* In the raw-setattr case, we lock the child inode.
366 * In the write-back case or if being called from open, the client holds a lock
369 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
370 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
371 struct ptlrpc_request *req,
372 struct lustre_handle *lh)
374 struct mds_obd *mds = mds_req2mds(req);
375 struct obd_device *obd = req->rq_export->exp_obd;
376 struct mds_body *body;
378 struct inode *inode = NULL;
379 struct lustre_handle lockh;
381 struct mds_logcancel_data *mlcd = NULL;
382 int rc = 0, cleanup_phase = 0, err, locked = 0;
385 LASSERT(offset == 0);
387 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
388 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
390 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
392 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
393 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
395 GOTO(cleanup, rc = PTR_ERR(de));
397 __u64 lockpart = MDS_INODELOCK_UPDATE;
398 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
399 lockpart |= MDS_INODELOCK_LOOKUP;
400 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
401 &lockh, NULL, 0, lockpart);
403 GOTO(cleanup, rc = PTR_ERR(de));
411 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
412 rec->ur_eadata != NULL)
415 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
417 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
419 GOTO(cleanup, rc = PTR_ERR(handle));
421 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
422 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
423 LTIME_S(rec->ur_iattr.ia_mtime),
424 LTIME_S(rec->ur_iattr.ia_ctime));
425 rc = mds_fix_attr(inode, rec);
429 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
430 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
431 (long)&rec->ur_iattr.ia_attr_flags);
433 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
435 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
436 rec->ur_eadata != NULL) {
437 struct lov_stripe_md *lsm = NULL;
439 rc = ll_permission(inode, MAY_WRITE, NULL);
443 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
444 mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
448 obd_free_memmd(mds->mds_osc_exp, &lsm);
450 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
456 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
457 mds_pack_inode2fid(&body->fid1, inode);
458 mds_pack_inode2body(body, inode);
460 /* Don't return OST-specific attributes if we didn't just set them */
461 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
462 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
463 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
464 body->valid |= OBD_MD_FLMTIME;
465 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
466 body->valid |= OBD_MD_FLATIME;
468 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
469 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
472 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
474 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
475 mlcd->mlcd_cookielen = rec->ur_cookielen;
476 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
477 mlcd->mlcd_cookielen;
478 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
479 mlcd->mlcd_cookielen);
480 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
481 mlcd->mlcd_eadatalen);
483 CERROR("unable to allocate log cancel data\n");
489 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
490 mds_cancel_cookies_cb, mlcd);
491 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
492 switch (cleanup_phase) {
494 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
495 rec->ur_eadata != NULL)
500 ldlm_lock_decref(&lockh, LCK_PW);
502 ptlrpc_save_lock (req, &lockh, LCK_PW);
517 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
518 struct ptlrpc_request *req)
520 struct mds_export_data *med = &req->rq_export->exp_mds_data;
521 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
522 struct dentry *parent, *child;
523 struct mds_body *body;
525 mds_req_from_mcd(req, med->med_mcd);
530 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
531 LASSERT(!IS_ERR(parent));
532 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
533 LASSERT(!IS_ERR(child));
534 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
535 mds_pack_inode2fid(&body->fid1, child->d_inode);
536 mds_pack_inode2body(body, child->d_inode);
541 static int mds_reint_create(struct mds_update_record *rec, int offset,
542 struct ptlrpc_request *req,
543 struct lustre_handle *lh)
545 struct dentry *dparent = NULL;
546 struct mds_obd *mds = mds_req2mds(req);
547 struct obd_device *obd = req->rq_export->exp_obd;
548 struct dentry *dchild = NULL;
549 struct inode *dir = NULL;
551 struct lustre_handle lockh;
552 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
554 struct dentry_params dp;
557 LASSERT(offset == 0);
558 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
560 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
561 rec->ur_fid1->id, rec->ur_fid1->generation,
562 rec->ur_name, rec->ur_mode);
564 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
566 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
567 GOTO(cleanup, rc = -ESTALE);
569 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, &lockh,
570 rec->ur_name, rec->ur_namelen - 1,
571 MDS_INODELOCK_UPDATE);
572 if (IS_ERR(dparent)) {
573 rc = PTR_ERR(dparent);
574 CERROR("parent lookup error %d\n", rc);
577 cleanup_phase = 1; /* locked parent dentry */
578 dir = dparent->d_inode;
581 ldlm_lock_dump_handle(D_OTHER, &lockh);
583 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
584 if (IS_ERR(dchild)) {
585 rc = PTR_ERR(dchild);
586 CERROR("child lookup error %d\n", rc);
590 cleanup_phase = 2; /* child dentry */
592 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
594 if (dir->i_mode & S_ISGID) {
595 if (S_ISDIR(rec->ur_mode))
596 rec->ur_mode |= S_ISGID;
599 dchild->d_fsdata = (void *)&dp;
600 dp.p_inum = (unsigned long)rec->ur_fid2->id;
605 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
607 GOTO(cleanup, rc = PTR_ERR(handle));
608 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
613 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
615 GOTO(cleanup, rc = PTR_ERR(handle));
616 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
621 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
623 GOTO(cleanup, rc = PTR_ERR(handle));
624 if (rec->ur_tgt == NULL) /* no target supplied */
625 rc = -EINVAL; /* -EPROTO? */
627 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
635 int rdev = rec->ur_rdev;
636 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
638 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
639 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
644 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
645 dchild->d_fsdata = NULL;
646 GOTO(cleanup, rc = -EINVAL);
649 /* In case we stored the desired inum in here, we want to clean up. */
650 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
651 dchild->d_fsdata = NULL;
654 CDEBUG(D_INODE, "error during create: %d\n", rc);
658 struct inode *inode = dchild->d_inode;
659 struct mds_body *body;
662 LTIME_S(iattr.ia_atime) = rec->ur_time;
663 LTIME_S(iattr.ia_ctime) = rec->ur_time;
664 LTIME_S(iattr.ia_mtime) = rec->ur_time;
665 iattr.ia_uid = rec->ur_fsuid;
666 if (dir->i_mode & S_ISGID)
667 iattr.ia_gid = dir->i_gid;
669 iattr.ia_gid = rec->ur_fsgid;
670 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
671 ATTR_MTIME | ATTR_CTIME;
673 if (rec->ur_fid2->id) {
674 LASSERT(rec->ur_fid2->id == inode->i_ino);
675 inode->i_generation = rec->ur_fid2->generation;
676 /* Dirtied and committed by the upcoming setattr. */
677 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
678 inode->i_ino, inode->i_generation);
680 struct lustre_handle child_ino_lockh;
682 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
683 inode->i_ino, inode->i_generation);
685 /* The inode we were allocated may have just been freed
686 * by an unlink operation. We take this lock to
687 * synchronize against the matching reply-ack-lock taken
688 * in unlink, to avoid replay problems if this reply
689 * makes it out to the client but the unlink's does not.
690 * See bug 2029 for more detail.*/
691 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
692 if (rc != ELDLM_OK) {
693 CERROR("error locking for unlink/create sync: "
696 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
700 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
702 CERROR("error on child setattr: rc = %d\n", rc);
704 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
705 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
707 CERROR("error on parent setattr: rc = %d\n", rc);
709 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
710 mds_pack_inode2fid(&body->fid1, inode);
711 mds_pack_inode2body(body, inode);
716 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
719 /* Destroy the file we just created. This should not need
720 * extra journal credits, as we have already modified all of
721 * the blocks needed in order to create the file in the first
726 err = vfs_rmdir(dir, dchild);
728 CERROR("rmdir in error path: %d\n", err);
731 err = vfs_unlink(dir, dchild);
733 CERROR("unlink in error path: %d\n", err);
739 switch (cleanup_phase) {
740 case 2: /* child dentry */
742 case 1: /* locked parent dentry */
744 ldlm_lock_decref(&lockh, LCK_PW);
746 ptlrpc_save_lock (req, &lockh, LCK_PW);
752 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
759 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
760 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
764 for (i = 0; i < RES_NAME_SIZE; i++) {
765 /* return 1 here, because enqueue_ordered will skip resources
766 * of all zeroes if they're sorted to the end of the list. */
767 if (res1->name[i] == 0 && res2->name[i] != 0)
769 if (res2->name[i] == 0 && res1->name[i] != 0)
772 if (res1->name[i] > res2->name[i])
774 if (res1->name[i] < res2->name[i])
781 if (memcmp(p1, p2, sizeof(*p1)) < 0)
787 /* This function doesn't use ldlm_match_or_enqueue because we're always called
788 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
789 * because they take the place of local semaphores.
791 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
792 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
793 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
794 struct lustre_handle *p1_lockh, int p1_lock_mode,
795 ldlm_policy_data_t *p1_policy,
796 struct ldlm_res_id *p2_res_id,
797 struct lustre_handle *p2_lockh, int p2_lock_mode,
798 ldlm_policy_data_t *p2_policy)
800 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
801 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
802 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
803 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
807 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
809 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
812 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
813 handles[1] = p1_lockh;
814 handles[0] = p2_lockh;
815 res_id[1] = p1_res_id;
816 res_id[0] = p2_res_id;
817 lock_modes[1] = p1_lock_mode;
818 lock_modes[0] = p2_lock_mode;
819 policies[1] = p1_policy;
820 policies[0] = p2_policy;
823 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
824 res_id[0]->name[0], res_id[1]->name[0]);
826 flags = LDLM_FL_LOCAL_ONLY;
827 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
828 LDLM_IBITS, policies[0], lock_modes[0], &flags,
829 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
830 NULL, 0, NULL, handles[0]);
833 ldlm_lock_dump_handle(D_OTHER, handles[0]);
835 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
836 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
837 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
838 ldlm_lock_addref(handles[1], lock_modes[1]);
839 } else if (res_id[1]->name[0] != 0) {
840 flags = LDLM_FL_LOCAL_ONLY;
841 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
842 *res_id[1], LDLM_IBITS, policies[1],
843 lock_modes[1], &flags, mds_blocking_ast,
844 ldlm_completion_ast, NULL, NULL, NULL, 0,
846 if (rc != ELDLM_OK) {
847 ldlm_lock_decref(handles[0], lock_modes[0]);
850 ldlm_lock_dump_handle(D_OTHER, handles[1]);
856 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
857 struct lustre_handle *p1_lockh, int p1_lock_mode,
858 ldlm_policy_data_t *p1_policy,
859 struct ldlm_res_id *p2_res_id,
860 struct lustre_handle *p2_lockh, int p2_lock_mode,
861 ldlm_policy_data_t *p2_policy,
862 struct ldlm_res_id *c1_res_id,
863 struct lustre_handle *c1_lockh, int c1_lock_mode,
864 ldlm_policy_data_t *c1_policy,
865 struct ldlm_res_id *c2_res_id,
866 struct lustre_handle *c2_lockh, int c2_lock_mode,
867 ldlm_policy_data_t *c2_policy)
869 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
870 c1_res_id, c2_res_id };
871 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
872 c1_lockh, c2_lockh };
873 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
874 c1_lock_mode, c2_lock_mode };
875 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
876 c1_policy, c2_policy};
877 int rc, i, j, sorted, flags;
881 "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
882 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
885 /* simple insertion sort - we have at most 4 elements */
886 for (i = 1; i < 4; i++) {
888 dlm_handles[4] = dlm_handles[i];
889 res_id[4] = res_id[i];
890 lock_modes[4] = lock_modes[i];
891 policies[4] = policies[i];
895 if (res_gt(res_id[j], res_id[4], policies[j],
897 dlm_handles[j + 1] = dlm_handles[j];
898 res_id[j + 1] = res_id[j];
899 lock_modes[j + 1] = lock_modes[j];
900 policies[j + 1] = policies[j];
905 } while (j >= 0 && !sorted);
907 dlm_handles[j + 1] = dlm_handles[4];
908 res_id[j + 1] = res_id[4];
909 lock_modes[j + 1] = lock_modes[4];
910 policies[j + 1] = policies[4];
914 "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
915 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
918 /* XXX we could send ASTs on all these locks first before blocking? */
919 for (i = 0; i < 4; i++) {
921 if (res_id[i]->name[0] == 0)
924 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
925 (policies[i]->l_inodebits.bits &
926 policies[i-1]->l_inodebits.bits) ) {
927 memcpy(dlm_handles[i], dlm_handles[i-1],
928 sizeof(*(dlm_handles[i])));
929 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
931 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
932 *res_id[i], LDLM_IBITS,
934 lock_modes[i], &flags,
936 ldlm_completion_ast, NULL, NULL,
937 NULL, 0, NULL, dlm_handles[i]);
939 GOTO(out_err, rc = -EIO);
940 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
947 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
952 /* In the unlikely case that the child changed while we were waiting
953 * on the lock, we need to drop the lock on the old child and either:
954 * - if the child has a lower resource name, then we have to also
955 * drop the parent lock and regain the locks in the right order
956 * - in the rename case, if the child has a lower resource name than one of
957 * the other parent/child resources (maxres) we also need to reget the locks
958 * - if the child has a higher resource name (this is the common case)
959 * we can just get the lock on the new child (still in lock order)
961 * Returns 0 if the child did not change or if it changed but could be locked.
962 * Returns 1 if the child changed and we need to re-lock (no locks held).
963 * Returns -ve error with a valid dchild (no locks held). */
964 static int mds_verify_child(struct obd_device *obd,
965 struct ldlm_res_id *parent_res_id,
966 struct lustre_handle *parent_lockh,
967 struct dentry *dparent, int parent_mode,
968 struct ldlm_res_id *child_res_id,
969 struct lustre_handle *child_lockh,
970 struct dentry **dchildp, int child_mode,
971 ldlm_policy_data_t *child_policy,
972 const char *name, int namelen,
973 struct ldlm_res_id *maxres)
975 struct dentry *vchild, *dchild = *dchildp;
976 int rc = 0, cleanup_phase = 2; /* parent, child locks */
979 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
981 GOTO(cleanup, rc = PTR_ERR(vchild));
983 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
984 (vchild->d_inode != NULL &&
985 child_res_id->name[0] == vchild->d_inode->i_ino &&
986 child_res_id->name[1] == vchild->d_inode->i_generation))) {
994 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
995 vchild->d_inode, dchild ? dchild->d_inode : 0,
996 vchild->d_inode ? vchild->d_inode->i_ino : 0,
997 child_res_id->name[0]);
998 if (child_res_id->name[0] != 0)
999 ldlm_lock_decref(child_lockh, child_mode);
1003 cleanup_phase = 1; /* parent lock only */
1004 *dchildp = dchild = vchild;
1006 if (dchild->d_inode) {
1008 child_res_id->name[0] = dchild->d_inode->i_ino;
1009 child_res_id->name[1] = dchild->d_inode->i_generation;
1011 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1012 res_gt(maxres, child_res_id, NULL, NULL)) {
1013 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1014 child_res_id->name[0], parent_res_id->name[0],
1016 GOTO(cleanup, rc = 1);
1019 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1020 *child_res_id, LDLM_IBITS, child_policy,
1021 child_mode, &flags, mds_blocking_ast,
1022 ldlm_completion_ast, NULL, NULL, NULL, 0,
1025 GOTO(cleanup, rc = -EIO);
1028 memset(child_res_id, 0, sizeof(*child_res_id));
1034 switch(cleanup_phase) {
1036 if (child_res_id->name[0] != 0)
1037 ldlm_lock_decref(child_lockh, child_mode);
1039 ldlm_lock_decref(parent_lockh, parent_mode);
1045 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1047 struct lustre_handle *parent_lockh,
1048 struct dentry **dparentp, int parent_mode,
1049 __u64 parent_lockpart,
1050 char *name, int namelen,
1051 struct lustre_handle *child_lockh,
1052 struct dentry **dchildp, int child_mode,
1053 __u64 child_lockpart)
1055 struct ldlm_res_id child_res_id = { .name = {0} };
1056 struct ldlm_res_id parent_res_id = { .name = {0} };
1057 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1058 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1059 struct inode *inode;
1060 int rc = 0, cleanup_phase = 0;
1063 /* Step 1: Lookup parent */
1064 *dparentp = mds_fid2dentry(mds, fid, NULL);
1065 if (IS_ERR(*dparentp)) {
1066 rc = PTR_ERR(*dparentp);
1071 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1072 (*dparentp)->d_inode->i_ino, name);
1074 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1075 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1077 cleanup_phase = 1; /* parent dentry */
1079 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1080 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1081 if (IS_ERR(*dchildp)) {
1082 rc = PTR_ERR(*dchildp);
1083 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1087 inode = (*dchildp)->d_inode;
1089 inode = igrab(inode);
1093 child_res_id.name[0] = inode->i_ino;
1094 child_res_id.name[1] = inode->i_generation;
1099 cleanup_phase = 2; /* child dentry */
1101 /* Step 3: Lock parent and child in resource order. If child doesn't
1102 * exist, we still have to lock the parent and re-lookup. */
1103 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1105 &child_res_id, child_lockh, child_mode,
1110 if (!(*dchildp)->d_inode)
1111 cleanup_phase = 3; /* parent lock */
1113 cleanup_phase = 4; /* child lock */
1115 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1116 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1117 parent_mode, &child_res_id, child_lockh,
1118 dchildp, child_mode, &child_policy,
1119 name, namelen, &parent_res_id);
1129 switch (cleanup_phase) {
1131 ldlm_lock_decref(child_lockh, child_mode);
1133 ldlm_lock_decref(parent_lockh, parent_mode);
1144 void mds_reconstruct_generic(struct ptlrpc_request *req)
1146 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1148 mds_req_from_mcd(req, med->med_mcd);
1151 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1152 struct ptlrpc_request *req,
1153 struct lustre_handle *lh)
1155 struct dentry *dparent, *dchild;
1156 struct mds_obd *mds = mds_req2mds(req);
1157 struct obd_device *obd = req->rq_export->exp_obd;
1158 struct mds_body *body = NULL;
1159 struct inode *child_inode;
1160 struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1161 void *handle = NULL;
1162 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1165 LASSERT(offset == 0 || offset == 2);
1167 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1168 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1170 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1172 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1173 GOTO(cleanup, rc = -ENOENT);
1175 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1176 &parent_lockh, &dparent, LCK_PW,
1177 MDS_INODELOCK_UPDATE,
1178 rec->ur_name, rec->ur_namelen,
1179 &child_lockh, &dchild, LCK_EX,
1180 MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1184 cleanup_phase = 1; /* dchild, dparent, locks */
1187 child_inode = dchild->d_inode;
1188 if (child_inode == NULL) {
1189 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1190 dparent->d_inode->i_ino, rec->ur_name);
1191 GOTO(cleanup, rc = -ENOENT);
1194 cleanup_phase = 2; /* dchild has a lock */
1196 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1197 * reuse (see bug 2029). */
1198 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1202 cleanup_phase = 3; /* child inum lock */
1204 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1206 /* ldlm_reply in buf[0] if called via intent */
1210 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1211 LASSERT(body != NULL);
1213 /* If this is the last reference to this inode, get the OBD EA
1214 * data first so the client can destroy OST objects.
1215 * we only do the object removal if no open files remain.
1216 * Nobody can get at this name anymore because of the locks so
1217 * we make decisions here as to whether to remove the inode */
1218 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1219 mds_open_orphan_count(child_inode) == 0) {
1220 mds_pack_inode2fid(&body->fid1, child_inode);
1221 mds_pack_inode2body(body, child_inode);
1222 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1224 if (!(body->valid & OBD_MD_FLEASIZE)) {
1225 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1226 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1232 /* We have to do these checks ourselves, in case we are making an
1233 * orphan. The client tells us whether rmdir() or unlink() was called,
1234 * so we need to return appropriate errors (bug 72).
1236 * We don't have to check permissions, because vfs_rename (called from
1237 * mds_open_unlink_rename) also calls may_delete. */
1238 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1239 if (!S_ISDIR(child_inode->i_mode))
1240 GOTO(cleanup, rc = -ENOTDIR);
1242 if (S_ISDIR(child_inode->i_mode))
1243 GOTO(cleanup, rc = -EISDIR);
1246 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1247 switch (child_inode->i_mode & S_IFMT) {
1249 /* Drop any lingering child directories before we start our
1250 * transaction, to avoid doing multiple inode dirty/delete
1251 * in our compound transaction (bug 1321). */
1252 shrink_dcache_parent(dchild);
1253 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1256 GOTO(cleanup, rc = PTR_ERR(handle));
1257 cleanup_phase = 4; /* transaction */
1258 rc = vfs_rmdir(dparent->d_inode, dchild);
1261 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1263 handle = fsfilt_start_log(obd, dparent->d_inode,
1264 FSFILT_OP_UNLINK, NULL,
1265 le32_to_cpu(lmm->lmm_stripe_count));
1267 GOTO(cleanup, rc = PTR_ERR(handle));
1269 cleanup_phase = 4; /* transaction */
1270 rc = vfs_unlink(dparent->d_inode, dchild);
1272 if (!rc && log_unlink)
1273 if (mds_log_op_unlink(obd, child_inode,
1274 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1275 req->rq_repmsg->buflens[offset + 1],
1276 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1277 req->rq_repmsg->buflens[offset + 2]) > 0)
1278 body->valid |= OBD_MD_FLCOOKIE;
1286 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1289 GOTO(cleanup, rc = PTR_ERR(handle));
1290 cleanup_phase = 4; /* transaction */
1291 rc = vfs_unlink(dparent->d_inode, dchild);
1294 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1297 GOTO(cleanup, rc = -EINVAL);
1305 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1306 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1307 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1309 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1311 CERROR("error on parent setattr: rc = %d\n", err);
1314 switch(cleanup_phase) {
1316 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1317 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1318 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1319 mds_open_orphan_count(dchild->d_inode) > 0) {
1320 /* filesystem is really going to destroy an inode
1321 * we have to delay this till inode is opened -bzzz */
1322 mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1324 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1327 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1328 "unlinked", 0, NULL);
1329 case 3: /* child ino-reuse lock */
1330 if (rc && body != NULL) {
1331 // Don't unlink the OST objects if the MDS unlink failed
1335 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1337 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1338 case 2: /* child lock */
1339 ldlm_lock_decref(&child_lockh, LCK_EX);
1340 case 1: /* child and parent dentry, parent lock */
1342 ldlm_lock_decref(&parent_lockh, LCK_PW);
1344 ptlrpc_save_lock(req, &parent_lockh, LCK_PW);
1351 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1354 req->rq_status = rc;
1358 static int mds_reint_link(struct mds_update_record *rec, int offset,
1359 struct ptlrpc_request *req,
1360 struct lustre_handle *lh)
1362 struct obd_device *obd = req->rq_export->exp_obd;
1363 struct dentry *de_src = NULL;
1364 struct dentry *de_tgt_dir = NULL;
1365 struct dentry *dchild = NULL;
1366 struct mds_obd *mds = mds_req2mds(req);
1367 struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1368 struct ldlm_res_id src_res_id = { .name = {0} };
1369 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1370 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1371 ldlm_policy_data_t tgt_dir_policy =
1372 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1374 int rc = 0, cleanup_phase = 0;
1377 LASSERT(offset == 0);
1379 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1380 rec->ur_fid1->id, rec->ur_fid1->generation,
1381 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1383 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1385 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1386 GOTO(cleanup, rc = -ENOENT);
1388 /* Step 1: Lookup the source inode and target directory by FID */
1389 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1391 GOTO(cleanup, rc = PTR_ERR(de_src));
1393 cleanup_phase = 1; /* source dentry */
1395 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1396 if (IS_ERR(de_tgt_dir)) {
1397 rc = PTR_ERR(de_tgt_dir);
1402 cleanup_phase = 2; /* target directory dentry */
1404 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1405 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1406 de_src->d_inode->i_ino);
1408 /* Step 2: Take the two locks */
1409 src_res_id.name[0] = de_src->d_inode->i_ino;
1410 src_res_id.name[1] = de_src->d_inode->i_generation;
1411 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1412 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1414 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1416 &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
1421 cleanup_phase = 3; /* locks */
1423 /* Step 3: Lookup the child */
1424 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1425 if (IS_ERR(dchild)) {
1426 rc = PTR_ERR(dchild);
1427 if (rc != -EPERM && rc != -EACCES)
1428 CERROR("child lookup error %d\n", rc);
1432 cleanup_phase = 4; /* child dentry */
1434 if (dchild->d_inode) {
1435 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1436 de_tgt_dir->d_inode->i_ino, rec->ur_name);
1441 /* Step 4: Do it. */
1442 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1444 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1445 if (IS_ERR(handle)) {
1446 rc = PTR_ERR(handle);
1450 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1451 if (rc && rc != -EPERM && rc != -EACCES)
1452 CERROR("vfs_link error %d\n", rc);
1454 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1455 handle, req, rc, 0);
1458 switch (cleanup_phase) {
1459 case 4: /* child dentry */
1463 ldlm_lock_decref(&src_lockh, LCK_EX);
1464 ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1466 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1467 ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1469 case 2: /* target dentry */
1471 case 1: /* source dentry */
1476 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1479 req->rq_status = rc;
1484 * add a hard link in the PENDING directory, only used by rename()
1486 static int mds_add_link_orphan(struct mds_update_record *rec,
1487 struct obd_device *obd,
1488 struct dentry *dentry)
1490 struct mds_obd *mds = &obd->u.mds;
1491 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1492 struct dentry *pending_child;
1493 char fidname[LL_FID_NAMELEN];
1497 LASSERT(dentry->d_inode);
1498 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
1500 down(&pending_dir->i_sem);
1501 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
1502 dentry->d_inode->i_generation);
1504 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
1505 mds_open_orphan_count(dentry->d_inode),
1506 rec->ur_name, fidname);
1508 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1509 if (IS_ERR(pending_child))
1510 GOTO(out_lock, rc = PTR_ERR(pending_child));
1512 if (pending_child->d_inode != NULL) {
1513 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1514 LASSERT(pending_child->d_inode == dentry->d_inode);
1515 GOTO(out_dput, rc = 0);
1519 rc = vfs_link(dentry, pending_dir, pending_child);
1522 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
1525 mds_inode_set_orphan(dentry->d_inode);
1527 l_dput(pending_child);
1529 up(&pending_dir->i_sem);
1533 /* The idea here is that we need to get four locks in the end:
1534 * one on each parent directory, one on each child. We need to take
1535 * these locks in some kind of order (to avoid deadlocks), and the order
1536 * I selected is "increasing resource number" order. We need to look up
1537 * the children, however, before we know what the resource number(s) are.
1538 * Thus the following plan:
1540 * 1,2. Look up the parents
1541 * 3,4. Look up the children
1542 * 5. Take locks on the parents and children, in order
1543 * 6. Verify that the children haven't changed since they were looked up
1545 * If there was a race and the children changed since they were first looked
1546 * up, it is possible that mds_verify_child() will be able to just grab the
1547 * lock on the new child resource (if it has a higher resource than any other)
1548 * but we need to compare against not only its parent, but also against the
1549 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1551 * We need the fancy igrab() on the child inodes because we aren't holding a
1552 * lock on the parent after the lookup is done, so dentry->d_inode may change
1553 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1555 static int mds_get_parents_children_locked(struct obd_device *obd,
1556 struct mds_obd *mds,
1557 struct ll_fid *p1_fid,
1558 struct dentry **de_srcdirp,
1559 struct ll_fid *p2_fid,
1560 struct dentry **de_tgtdirp,
1562 const char *old_name, int old_len,
1563 struct dentry **de_oldp,
1564 const char *new_name, int new_len,
1565 struct dentry **de_newp,
1566 struct lustre_handle *dlm_handles,
1569 struct ldlm_res_id p1_res_id = { .name = {0} };
1570 struct ldlm_res_id p2_res_id = { .name = {0} };
1571 struct ldlm_res_id c1_res_id = { .name = {0} };
1572 struct ldlm_res_id c2_res_id = { .name = {0} };
1573 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1574 /* Only dentry should change, but the inode itself would be
1576 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
1577 /* If something is going to be replaced, both dentry and inode locks are
1579 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
1580 MDS_INODELOCK_UPDATE}};
1581 struct ldlm_res_id *maxres_src, *maxres_tgt;
1582 struct inode *inode;
1583 int rc = 0, cleanup_phase = 0;
1586 /* Step 1: Lookup the source directory */
1587 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1588 if (IS_ERR(*de_srcdirp))
1589 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1591 cleanup_phase = 1; /* source directory dentry */
1593 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1594 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1596 /* Step 2: Lookup the target directory */
1597 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1598 *de_tgtdirp = dget(*de_srcdirp);
1600 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
1601 if (IS_ERR(*de_tgtdirp)) {
1602 rc = PTR_ERR(*de_tgtdirp);
1608 cleanup_phase = 2; /* target directory dentry */
1610 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
1611 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
1613 /* Step 3: Lookup the source child entry */
1614 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
1615 if (IS_ERR(*de_oldp)) {
1616 rc = PTR_ERR(*de_oldp);
1617 CERROR("old child lookup error (%*s): %d\n",
1618 old_len - 1, old_name, rc);
1622 cleanup_phase = 3; /* original name dentry */
1624 inode = (*de_oldp)->d_inode;
1626 inode = igrab(inode);
1628 GOTO(cleanup, rc = -ENOENT);
1630 c1_res_id.name[0] = inode->i_ino;
1631 c1_res_id.name[1] = inode->i_generation;
1634 /* Step 4: Lookup the target child entry */
1635 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
1636 if (IS_ERR(*de_newp)) {
1637 rc = PTR_ERR(*de_newp);
1638 CERROR("new child lookup error (%*s): %d\n",
1639 old_len - 1, old_name, rc);
1643 cleanup_phase = 4; /* target dentry */
1645 inode = (*de_newp)->d_inode;
1647 inode = igrab(inode);
1651 c2_res_id.name[0] = inode->i_ino;
1652 c2_res_id.name[1] = inode->i_generation;
1657 /* Step 5: Take locks on the parents and child(ren) */
1658 maxres_src = &p1_res_id;
1659 maxres_tgt = &p2_res_id;
1660 cleanup_phase = 4; /* target dentry */
1662 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
1663 maxres_src = &c1_res_id;
1664 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
1665 maxres_tgt = &c2_res_id;
1667 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
1669 &p2_res_id, &dlm_handles[1], parent_mode,
1671 &c1_res_id, &dlm_handles[2], child_mode,
1673 &c2_res_id, &dlm_handles[3], child_mode,
1678 cleanup_phase = 6; /* parent and child(ren) locks */
1680 /* Step 6a: Re-lookup source child to verify it hasn't changed */
1681 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
1682 parent_mode, &c1_res_id, &dlm_handles[2],
1683 de_oldp, child_mode, &c1_policy, old_name,old_len,
1686 if (c2_res_id.name[0] != 0)
1687 ldlm_lock_decref(&dlm_handles[3], child_mode);
1688 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1695 if ((*de_oldp)->d_inode == NULL)
1696 GOTO(cleanup, rc = -ENOENT);
1698 /* Step 6b: Re-lookup target child to verify it hasn't changed */
1699 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
1700 parent_mode, &c2_res_id, &dlm_handles[3],
1701 de_newp, child_mode, &c2_policy, new_name,
1702 new_len, maxres_src);
1704 ldlm_lock_decref(&dlm_handles[2], child_mode);
1705 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1715 switch (cleanup_phase) {
1716 case 6: /* child lock(s) */
1717 if (c2_res_id.name[0] != 0)
1718 ldlm_lock_decref(&dlm_handles[3], child_mode);
1719 if (c1_res_id.name[0] != 0)
1720 ldlm_lock_decref(&dlm_handles[2], child_mode);
1721 case 5: /* parent locks */
1722 ldlm_lock_decref(&dlm_handles[1], parent_mode);
1723 ldlm_lock_decref(&dlm_handles[0], parent_mode);
1724 case 4: /* target dentry */
1726 case 3: /* source dentry */
1728 case 2: /* target directory dentry */
1729 l_dput(*de_tgtdirp);
1730 case 1: /* source directry dentry */
1731 l_dput(*de_srcdirp);
1738 static int mds_reint_rename(struct mds_update_record *rec, int offset,
1739 struct ptlrpc_request *req,
1740 struct lustre_handle *lockh)
1742 struct obd_device *obd = req->rq_export->exp_obd;
1743 struct dentry *de_srcdir = NULL;
1744 struct dentry *de_tgtdir = NULL;
1745 struct dentry *de_old = NULL;
1746 struct dentry *de_new = NULL;
1747 struct mds_obd *mds = mds_req2mds(req);
1748 struct lustre_handle dlm_handles[4];
1749 struct mds_body *body = NULL;
1750 int rc = 0, lock_count = 3;
1751 int cleanup_phase = 0;
1752 void *handle = NULL;
1755 LASSERT(offset == 0);
1757 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
1758 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
1759 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
1761 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1763 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
1764 rec->ur_fid2, &de_tgtdir, LCK_PW,
1765 rec->ur_name, rec->ur_namelen,
1766 &de_old, rec->ur_tgt,
1767 rec->ur_tgtlen, &de_new,
1768 dlm_handles, LCK_EX);
1772 cleanup_phase = 1; /* parent(s), children, locks */
1774 if (de_new->d_inode)
1777 /* sanity check for src inode */
1778 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1779 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
1780 GOTO(cleanup, rc = -EINVAL);
1782 /* sanity check for dest inode */
1783 if (de_new->d_inode &&
1784 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
1785 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
1786 GOTO(cleanup, rc = -EINVAL);
1788 if (de_old->d_inode == de_new->d_inode) {
1789 GOTO(cleanup, rc = 0);
1792 /* if we are about to remove the target at first, pass the EA of
1793 * that inode to client to perform and cleanup on OST */
1794 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1795 LASSERT(body != NULL);
1797 if (de_new->d_inode &&
1798 S_ISREG(de_new->d_inode->i_mode) &&
1799 de_new->d_inode->i_nlink == 1 &&
1800 mds_open_orphan_count(de_new->d_inode) == 0) {
1801 mds_pack_inode2fid(&body->fid1, de_new->d_inode);
1802 mds_pack_inode2body(body, de_new->d_inode);
1803 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
1804 if (!(body->valid & OBD_MD_FLEASIZE)) {
1805 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1806 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1808 /* XXX need log unlink? */
1812 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
1813 de_srcdir->d_inode->i_sb);
1815 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
1817 GOTO(cleanup, rc = PTR_ERR(handle));
1819 /* FIXME need adjust the journal block count? */
1820 /* if the target should be moved to PENDING, we at first increase the
1821 * link and later vfs_rename() will decrease the link count again */
1822 if (de_new->d_inode &&
1823 S_ISREG(de_new->d_inode->i_mode) &&
1824 de_new->d_inode->i_nlink == 1 &&
1825 mds_open_orphan_count(de_new->d_inode) > 0) {
1826 rc = mds_add_link_orphan(rec, obd, de_new);
1832 de_old->d_fsdata = req;
1833 de_new->d_fsdata = req;
1834 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
1839 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
1840 handle, req, rc, 0);
1841 switch (cleanup_phase) {
1844 if (lock_count == 4)
1845 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
1846 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
1847 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
1848 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
1850 if (lock_count == 4)
1851 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
1852 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
1853 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
1854 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
1863 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1866 req->rq_status = rc;
1870 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
1871 struct ptlrpc_request *, struct lustre_handle *);
1873 static mds_reinter reinters[REINT_MAX + 1] = {
1874 [REINT_SETATTR] mds_reint_setattr,
1875 [REINT_CREATE] mds_reint_create,
1876 [REINT_LINK] mds_reint_link,
1877 [REINT_UNLINK] mds_reint_unlink,
1878 [REINT_RENAME] mds_reint_rename,
1879 [REINT_OPEN] mds_open
1882 int mds_reint_rec(struct mds_update_record *rec, int offset,
1883 struct ptlrpc_request *req, struct lustre_handle *lockh)
1885 struct obd_device *obd = req->rq_export->exp_obd;
1886 struct obd_run_ctxt saved;
1890 /* checked by unpacker */
1891 LASSERT(rec->ur_opcode <= REINT_MAX &&
1892 reinters[rec->ur_opcode] != NULL);
1894 push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
1895 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
1896 pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);