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->obd_llogs,
82 mlcd->mlcd_cookies[0].lgc_subsys + 1);
83 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
84 sizeof(*mlcd->mlcd_cookies),
85 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
87 CERROR("error cancelling %d log cookies: rc %d\n",
88 (int)(mlcd->mlcd_cookielen /
89 sizeof(*mlcd->mlcd_cookies)), rc);
92 OBD_FREE(mlcd, mlcd->mlcd_size);
95 /* Assumes caller has already pushed us into the kernel context. */
96 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
97 struct ptlrpc_request *req, int rc, __u32 op_data)
99 struct mds_export_data *med = &req->rq_export->exp_mds_data;
100 struct mds_client_data *mcd = med->med_mcd;
101 struct obd_device *obd = req->rq_export->exp_obd;
108 /* if the export has already been failed, we have no last_rcvd slot */
109 if (req->rq_export->exp_failed) {
110 CERROR("committing transaction for disconnected client\n");
119 if (handle == NULL) {
120 /* if we're starting our own xaction, use our own inode */
121 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
122 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
123 if (IS_ERR(handle)) {
124 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
125 RETURN(PTR_ERR(handle));
131 transno = req->rq_reqmsg->transno;
133 LASSERT(transno == 0);
134 } else if (transno == 0) {
135 spin_lock(&mds->mds_transno_lock);
136 transno = ++mds->mds_last_transno;
137 spin_unlock(&mds->mds_transno_lock);
139 spin_lock(&mds->mds_transno_lock);
140 if (transno > mds->mds_last_transno)
141 mds->mds_last_transno = transno;
142 spin_unlock(&mds->mds_transno_lock);
144 req->rq_repmsg->transno = req->rq_transno = transno;
145 mcd->mcd_last_transno = cpu_to_le64(transno);
146 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
147 mcd->mcd_last_result = cpu_to_le32(rc);
148 mcd->mcd_last_data = cpu_to_le32(op_data);
150 fsfilt_add_journal_cb(req->rq_export->exp_obd, transno, handle,
151 mds_commit_cb, NULL);
152 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
161 DEBUG_REQ(log_pri, req,
162 "wrote trans #"LPU64" client %s at idx %u: err = %d",
163 transno, mcd->mcd_uuid, med->med_idx, err);
165 err = mds_lov_write_objids(obd);
171 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
174 err = fsfilt_commit(obd, inode, handle, 0);
176 CERROR("error committing transaction: %d\n", err);
184 /* this gives the same functionality as the code between
185 * sys_chmod and inode_setattr
186 * chown_common and inode_setattr
187 * utimes and inode_setattr
189 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
191 time_t now = LTIME_S(CURRENT_TIME);
192 struct iattr *attr = &rec->ur_iattr;
193 unsigned int ia_valid = attr->ia_valid;
197 /* only fix up attrs if the client VFS didn't already */
198 if (!(ia_valid & ATTR_RAW))
201 if (!(ia_valid & ATTR_CTIME_SET))
202 LTIME_S(attr->ia_ctime) = now;
203 if (!(ia_valid & ATTR_ATIME_SET))
204 LTIME_S(attr->ia_atime) = now;
205 if (!(ia_valid & ATTR_MTIME_SET))
206 LTIME_S(attr->ia_mtime) = now;
208 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
212 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
213 if (rec->ur_fsuid != inode->i_uid &&
214 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
218 if (ia_valid & ATTR_SIZE) {
219 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
223 if (ia_valid & ATTR_UID) {
226 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
228 if (attr->ia_uid == (uid_t) -1)
229 attr->ia_uid = inode->i_uid;
230 if (attr->ia_gid == (gid_t) -1)
231 attr->ia_gid = inode->i_gid;
232 attr->ia_mode = inode->i_mode;
234 * If the user or group of a non-directory has been
235 * changed by a non-root user, remove the setuid bit.
236 * 19981026 David C Niemi <niemi@tux.org>
238 * Changed this to apply to all users, including root,
239 * to avoid some races. This is the behavior we had in
240 * 2.0. The check for non-root was definitely wrong
241 * for 2.2 anyway, as it should have been using
242 * CAP_FSETID rather than fsuid -- 19990830 SD.
244 if ((inode->i_mode & S_ISUID) == S_ISUID &&
245 !S_ISDIR(inode->i_mode)) {
246 attr->ia_mode &= ~S_ISUID;
247 attr->ia_valid |= ATTR_MODE;
250 * Likewise, if the user or group of a non-directory
251 * has been changed by a non-root user, remove the
252 * setgid bit UNLESS there is no group execute bit
253 * (this would be a file marked for mandatory
254 * locking). 19981026 David C Niemi <niemi@tux.org>
256 * Removed the fsuid check (see the comment above) --
259 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
260 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
261 attr->ia_mode &= ~S_ISGID;
262 attr->ia_valid |= ATTR_MODE;
264 } else if (ia_valid & ATTR_MODE) {
265 int mode = attr->ia_mode;
267 if (attr->ia_mode == (mode_t) -1)
268 attr->ia_mode = inode->i_mode;
270 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
275 void mds_steal_ack_locks(struct ptlrpc_request *req)
277 struct obd_export *exp = req->rq_export;
278 struct list_head *tmp;
279 struct ptlrpc_reply_state *oldrep;
280 struct ptlrpc_service *svc;
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 exp->exp_connection->c_peer.peer_nid);
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(req2obd(req), &body->fid1, de->d_inode);
352 mds_pack_inode2body(req2obd(req), 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[2];
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) && rec->ur_eadata != NULL)
414 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
416 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
418 GOTO(cleanup, rc = PTR_ERR(handle));
420 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
421 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
422 LTIME_S(rec->ur_iattr.ia_mtime),
423 LTIME_S(rec->ur_iattr.ia_ctime));
424 rc = mds_fix_attr(inode, rec);
428 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
429 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
430 (long)&rec->ur_iattr.ia_attr_flags);
432 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
434 if (rc == 0 && S_ISREG(inode->i_mode) && rec->ur_eadata != NULL) {
435 rc = fsfilt_set_md(obd, inode, handle,
436 rec->ur_eadata, rec->ur_eadatalen);
439 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
440 mds_pack_inode2fid(obd, &body->fid1, inode);
441 mds_pack_inode2body(obd, body, inode);
443 /* Don't return OST-specific attributes if we didn't just set them */
444 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
445 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
446 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
447 body->valid |= OBD_MD_FLMTIME;
448 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
449 body->valid |= OBD_MD_FLATIME;
451 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
452 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
455 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
457 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
458 mlcd->mlcd_cookielen = rec->ur_cookielen;
459 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
460 mlcd->mlcd_cookielen;
461 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
462 mlcd->mlcd_cookielen);
463 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
464 mlcd->mlcd_eadatalen);
466 CERROR("unable to allocate log cancel data\n");
472 fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
473 mds_cancel_cookies_cb, mlcd);
474 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
475 switch (cleanup_phase) {
477 if (S_ISREG(inode->i_mode) && rec->ur_eadata != NULL)
482 if (lockh[1].cookie != 0)
483 ldlm_lock_decref(lockh + 1, LCK_CW);
486 ldlm_lock_decref(lockh, LCK_PW);
488 ptlrpc_save_lock (req, lockh, LCK_PW);
503 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
504 struct ptlrpc_request *req)
506 struct mds_export_data *med = &req->rq_export->exp_mds_data;
507 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
508 struct dentry *parent, *child;
509 struct mds_body *body;
511 mds_req_from_mcd(req, med->med_mcd);
516 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
517 LASSERT(!IS_ERR(parent));
518 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
519 LASSERT(!IS_ERR(child));
520 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
521 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
522 mds_pack_inode2body(req2obd(req), body, child->d_inode);
527 static int mds_reint_create(struct mds_update_record *rec, int offset,
528 struct ptlrpc_request *req,
529 struct lustre_handle *lh)
531 struct dentry *dparent = NULL;
532 struct mds_obd *mds = mds_req2mds(req);
533 struct obd_device *obd = req->rq_export->exp_obd;
534 struct dentry *dchild = NULL;
535 struct inode *dir = NULL;
537 struct lustre_handle lockh[2];
538 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
540 struct dentry_params dp;
541 struct mea *mea = NULL;
545 LASSERT(offset == 0);
546 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
548 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
549 rec->ur_fid1->id, rec->ur_fid1->generation,
550 rec->ur_name, rec->ur_mode);
552 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
554 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
555 GOTO(cleanup, rc = -ESTALE);
557 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
558 rec->ur_name, rec->ur_namelen - 1,
559 MDS_INODELOCK_UPDATE);
560 if (IS_ERR(dparent)) {
561 rc = PTR_ERR(dparent);
562 CERROR("parent lookup error %d\n", rc);
565 cleanup_phase = 1; /* locked parent dentry */
566 dir = dparent->d_inode;
569 ldlm_lock_dump_handle(D_OTHER, lockh);
571 /* try to retrieve MEA data for this dir */
572 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
574 /* dir is already splitted, check is requested filename
575 * should live at this MDS or at another one */
577 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
578 if (mea->mea_master != i) {
579 CERROR("inapropriate MDS(%d) for %s. should be %d\n",
580 mea->mea_master, rec->ur_name, i);
581 GOTO(cleanup, rc = -ESTALE);
585 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
586 if (IS_ERR(dchild)) {
587 rc = PTR_ERR(dchild);
588 CERROR("child lookup error %d\n", rc);
592 cleanup_phase = 2; /* child dentry */
594 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
596 if (type == S_IFREG || type == S_IFDIR) {
597 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
599 /* dir got splitted */
600 GOTO(cleanup, rc = -ESTALE);
602 /* error happened during spitting */
608 if (dir->i_mode & S_ISGID) {
609 if (S_ISDIR(rec->ur_mode))
610 rec->ur_mode |= S_ISGID;
613 dchild->d_fsdata = (void *)&dp;
614 dp.p_inum = (unsigned long)rec->ur_fid2->id;
619 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
621 GOTO(cleanup, rc = PTR_ERR(handle));
622 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
630 /* as Peter asked, mkdir() should distribute new directories
631 * over the whole cluster in order to distribute namespace
632 * processing load. first, we calculate which MDS to use to
633 * put new directory's inode in */
634 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1);
635 if (i == mds->mds_num) {
636 /* inode will be created locally */
638 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
640 GOTO(cleanup, rc = PTR_ERR(handle));
642 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
645 nstripes = *(u16 *)rec->ur_eadata;
648 /* this is for current testing yet. after the testing
649 * directory will split if size reaches some limite -bzzz */
652 if (rc == 0 && nstripes) {
654 /* FIXME: error handling here */
655 mds_try_to_split_dir(obd, dchild, NULL, nstripes);
657 } else if (!DENTRY_VALID(dchild)) {
658 /* inode will be created on another MDS */
659 struct obdo *oa = NULL;
660 struct mds_body *body;
662 /* first, create that inode */
666 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
667 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
668 OBD_MD_FLUID | OBD_MD_FLGID);
669 oa->o_mode = dir->i_mode;
670 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
672 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
675 /* now, add new dir entry for it */
676 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
678 GOTO(cleanup, rc = PTR_ERR(handle));
679 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
681 oa->o_id, oa->o_generation,
686 body = lustre_msg_buf(req->rq_repmsg,
687 offset, sizeof (*body));
688 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
689 body->fid1.id = oa->o_id;
691 body->fid1.generation = oa->o_generation;
694 /* requested name exists in the directory */
701 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
703 GOTO(cleanup, rc = PTR_ERR(handle));
704 if (rec->ur_tgt == NULL) /* no target supplied */
705 rc = -EINVAL; /* -EPROTO? */
707 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
715 int rdev = rec->ur_rdev;
716 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
718 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
719 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
724 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
725 dchild->d_fsdata = NULL;
726 GOTO(cleanup, rc = -EINVAL);
729 /* In case we stored the desired inum in here, we want to clean up. */
730 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
731 dchild->d_fsdata = NULL;
734 CDEBUG(D_INODE, "error during create: %d\n", rc);
736 } else if (dchild->d_inode) {
738 struct inode *inode = dchild->d_inode;
739 struct mds_body *body;
742 LTIME_S(iattr.ia_atime) = rec->ur_time;
743 LTIME_S(iattr.ia_ctime) = rec->ur_time;
744 LTIME_S(iattr.ia_mtime) = rec->ur_time;
745 iattr.ia_uid = rec->ur_fsuid;
746 if (dir->i_mode & S_ISGID)
747 iattr.ia_gid = dir->i_gid;
749 iattr.ia_gid = rec->ur_fsgid;
750 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
751 ATTR_MTIME | ATTR_CTIME;
753 if (rec->ur_fid2->id) {
754 LASSERT(rec->ur_fid2->id == inode->i_ino);
755 inode->i_generation = rec->ur_fid2->generation;
756 /* Dirtied and committed by the upcoming setattr. */
757 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
758 inode->i_ino, inode->i_generation);
760 struct lustre_handle child_ino_lockh;
762 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
763 inode->i_ino, inode->i_generation);
765 /* The inode we were allocated may have just been freed
766 * by an unlink operation. We take this lock to
767 * synchronize against the matching reply-ack-lock taken
768 * in unlink, to avoid replay problems if this reply
769 * makes it out to the client but the unlink's does not.
770 * See bug 2029 for more detail.*/
771 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
772 if (rc != ELDLM_OK) {
773 CERROR("error locking for unlink/create sync: "
776 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
780 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
782 CERROR("error on child setattr: rc = %d\n", rc);
784 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
785 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
787 CERROR("error on parent setattr: rc = %d\n", rc);
789 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
790 mds_pack_inode2fid(obd, &body->fid1, inode);
791 mds_pack_inode2body(obd, body, inode);
796 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
799 /* Destroy the file we just created. This should not need
800 * extra journal credits, as we have already modified all of
801 * the blocks needed in order to create the file in the first
806 err = vfs_rmdir(dir, dchild);
808 CERROR("rmdir in error path: %d\n", err);
811 err = vfs_unlink(dir, dchild);
813 CERROR("unlink in error path: %d\n", err);
819 switch (cleanup_phase) {
820 case 2: /* child dentry */
822 case 1: /* locked parent dentry */
824 if (lockh[1].cookie != 0)
825 ldlm_lock_decref(lockh + 1, LCK_CW);
828 ldlm_lock_decref(lockh, LCK_PW);
830 ptlrpc_save_lock (req, lockh, LCK_PW);
836 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
840 OBD_FREE(mea, mea_size);
845 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
846 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
850 for (i = 0; i < RES_NAME_SIZE; i++) {
851 /* return 1 here, because enqueue_ordered will skip resources
852 * of all zeroes if they're sorted to the end of the list. */
853 if (res1->name[i] == 0 && res2->name[i] != 0)
855 if (res2->name[i] == 0 && res1->name[i] != 0)
858 if (res1->name[i] > res2->name[i])
860 if (res1->name[i] < res2->name[i])
867 if (memcmp(p1, p2, sizeof(*p1)) < 0)
873 /* This function doesn't use ldlm_match_or_enqueue because we're always called
874 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
875 * because they take the place of local semaphores.
877 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
878 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
879 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
880 struct lustre_handle *p1_lockh, int p1_lock_mode,
881 ldlm_policy_data_t *p1_policy,
882 struct ldlm_res_id *p2_res_id,
883 struct lustre_handle *p2_lockh, int p2_lock_mode,
884 ldlm_policy_data_t *p2_policy)
886 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
887 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
888 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
889 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
893 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
895 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
898 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
899 handles[1] = p1_lockh;
900 handles[0] = p2_lockh;
901 res_id[1] = p1_res_id;
902 res_id[0] = p2_res_id;
903 lock_modes[1] = p1_lock_mode;
904 lock_modes[0] = p2_lock_mode;
905 policies[1] = p1_policy;
906 policies[0] = p2_policy;
909 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
910 res_id[0]->name[0], res_id[1]->name[0]);
912 flags = LDLM_FL_LOCAL_ONLY;
913 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
914 LDLM_IBITS, policies[0], lock_modes[0], &flags,
915 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
916 NULL, 0, NULL, handles[0]);
919 ldlm_lock_dump_handle(D_OTHER, handles[0]);
921 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
922 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
923 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
924 ldlm_lock_addref(handles[1], lock_modes[1]);
925 } else if (res_id[1]->name[0] != 0) {
926 flags = LDLM_FL_LOCAL_ONLY;
927 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
928 *res_id[1], LDLM_IBITS, policies[1],
929 lock_modes[1], &flags, mds_blocking_ast,
930 ldlm_completion_ast, NULL, NULL, NULL, 0,
932 if (rc != ELDLM_OK) {
933 ldlm_lock_decref(handles[0], lock_modes[0]);
936 ldlm_lock_dump_handle(D_OTHER, handles[1]);
942 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
943 struct lustre_handle *p1_lockh, int p1_lock_mode,
944 ldlm_policy_data_t *p1_policy,
945 struct ldlm_res_id *p2_res_id,
946 struct lustre_handle *p2_lockh, int p2_lock_mode,
947 ldlm_policy_data_t *p2_policy,
948 struct ldlm_res_id *c1_res_id,
949 struct lustre_handle *c1_lockh, int c1_lock_mode,
950 ldlm_policy_data_t *c1_policy,
951 struct ldlm_res_id *c2_res_id,
952 struct lustre_handle *c2_lockh, int c2_lock_mode,
953 ldlm_policy_data_t *c2_policy)
955 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
956 c1_res_id, c2_res_id };
957 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
958 c1_lockh, c2_lockh };
959 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
960 c1_lock_mode, c2_lock_mode };
961 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
962 c1_policy, c2_policy};
963 int rc, i, j, sorted, flags;
967 "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
968 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
971 /* simple insertion sort - we have at most 4 elements */
972 for (i = 1; i < 4; i++) {
974 dlm_handles[4] = dlm_handles[i];
975 res_id[4] = res_id[i];
976 lock_modes[4] = lock_modes[i];
977 policies[4] = policies[i];
981 if (res_gt(res_id[j], res_id[4], policies[j],
983 dlm_handles[j + 1] = dlm_handles[j];
984 res_id[j + 1] = res_id[j];
985 lock_modes[j + 1] = lock_modes[j];
986 policies[j + 1] = policies[j];
991 } while (j >= 0 && !sorted);
993 dlm_handles[j + 1] = dlm_handles[4];
994 res_id[j + 1] = res_id[4];
995 lock_modes[j + 1] = lock_modes[4];
996 policies[j + 1] = policies[4];
1000 "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1001 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1002 res_id[3]->name[0]);
1004 /* XXX we could send ASTs on all these locks first before blocking? */
1005 for (i = 0; i < 4; i++) {
1007 if (res_id[i]->name[0] == 0)
1010 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1011 (policies[i]->l_inodebits.bits &
1012 policies[i-1]->l_inodebits.bits) ) {
1013 memcpy(dlm_handles[i], dlm_handles[i-1],
1014 sizeof(*(dlm_handles[i])));
1015 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1017 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1018 *res_id[i], LDLM_IBITS,
1020 lock_modes[i], &flags,
1022 ldlm_completion_ast, NULL, NULL,
1023 NULL, 0, NULL, dlm_handles[i]);
1025 GOTO(out_err, rc = -EIO);
1026 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1033 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1038 /* In the unlikely case that the child changed while we were waiting
1039 * on the lock, we need to drop the lock on the old child and either:
1040 * - if the child has a lower resource name, then we have to also
1041 * drop the parent lock and regain the locks in the right order
1042 * - in the rename case, if the child has a lower resource name than one of
1043 * the other parent/child resources (maxres) we also need to reget the locks
1044 * - if the child has a higher resource name (this is the common case)
1045 * we can just get the lock on the new child (still in lock order)
1047 * Returns 0 if the child did not change or if it changed but could be locked.
1048 * Returns 1 if the child changed and we need to re-lock (no locks held).
1049 * Returns -ve error with a valid dchild (no locks held). */
1050 static int mds_verify_child(struct obd_device *obd,
1051 struct ldlm_res_id *parent_res_id,
1052 struct lustre_handle *parent_lockh,
1053 struct dentry *dparent, int parent_mode,
1054 struct ldlm_res_id *child_res_id,
1055 struct lustre_handle *child_lockh,
1056 struct dentry **dchildp, int child_mode,
1057 ldlm_policy_data_t *child_policy,
1058 const char *name, int namelen,
1059 struct ldlm_res_id *maxres)
1061 struct dentry *vchild, *dchild = *dchildp;
1062 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1065 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1067 GOTO(cleanup, rc = PTR_ERR(vchild));
1069 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1070 if (child_res_id->name[0] == vchild->d_inum &&
1071 child_res_id->name[1] == vchild->d_generation) {
1080 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1081 (vchild->d_inode != NULL &&
1082 child_res_id->name[0] == vchild->d_inode->i_ino &&
1083 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1092 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1093 vchild->d_inode, dchild ? dchild->d_inode : 0,
1094 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1095 child_res_id->name[0]);
1096 if (child_res_id->name[0] != 0)
1097 ldlm_lock_decref(child_lockh, child_mode);
1101 cleanup_phase = 1; /* parent lock only */
1102 *dchildp = dchild = vchild;
1104 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1106 if (dchild->d_inode) {
1107 child_res_id->name[0] = dchild->d_inode->i_ino;
1108 child_res_id->name[1] = dchild->d_inode->i_generation;
1110 child_res_id->name[0] = dchild->d_inum;
1111 child_res_id->name[1] = dchild->d_generation;
1114 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1115 res_gt(maxres, child_res_id, NULL, NULL)) {
1116 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1117 child_res_id->name[0], parent_res_id->name[0],
1119 GOTO(cleanup, rc = 1);
1122 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1123 *child_res_id, LDLM_IBITS, child_policy,
1124 child_mode, &flags, mds_blocking_ast,
1125 ldlm_completion_ast, NULL, NULL, NULL, 0,
1128 GOTO(cleanup, rc = -EIO);
1131 memset(child_res_id, 0, sizeof(*child_res_id));
1137 switch(cleanup_phase) {
1139 if (child_res_id->name[0] != 0)
1140 ldlm_lock_decref(child_lockh, child_mode);
1142 ldlm_lock_decref(parent_lockh, parent_mode);
1148 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1150 struct lustre_handle *parent_lockh,
1151 struct dentry **dparentp, int parent_mode,
1152 __u64 parent_lockpart,
1153 char *name, int namelen,
1154 struct lustre_handle *child_lockh,
1155 struct dentry **dchildp, int child_mode,
1156 __u64 child_lockpart)
1158 struct ldlm_res_id child_res_id = { .name = {0} };
1159 struct ldlm_res_id parent_res_id = { .name = {0} };
1160 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1161 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1162 struct inode *inode;
1163 int rc = 0, cleanup_phase = 0;
1166 /* Step 1: Lookup parent */
1167 *dparentp = mds_fid2dentry(mds, fid, NULL);
1168 if (IS_ERR(*dparentp))
1169 RETURN(rc = PTR_ERR(*dparentp));
1170 LASSERT((*dparentp)->d_inode);
1172 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1173 (*dparentp)->d_inode->i_ino, name);
1175 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1176 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1178 parent_lockh[1].cookie = 0;
1179 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1180 /* lock just dir { ino, generation } to flush client cache */
1181 if (parent_mode == LCK_PW) {
1182 struct ldlm_res_id res_id = { .name = {0} };
1183 ldlm_policy_data_t policy;
1185 res_id.name[0] = (*dparentp)->d_inode->i_ino;
1186 res_id.name[1] = (*dparentp)->d_inode->i_generation;
1187 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1188 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1190 &policy, LCK_CW, &flags,
1192 ldlm_completion_ast, NULL, NULL,
1193 NULL, 0, NULL, parent_lockh+1);
1198 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1199 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1200 (*dparentp)->d_inode->i_ino,
1201 (*dparentp)->d_inode->i_generation,
1202 parent_res_id.name[2]);
1206 cleanup_phase = 1; /* parent dentry */
1208 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1209 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1210 if (IS_ERR(*dchildp)) {
1211 rc = PTR_ERR(*dchildp);
1212 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1216 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1217 /* inode lives on another MDS: return * mds/ino/gen
1218 * and LOOKUP lock. drop possible UPDATE lock! */
1219 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1220 child_res_id.name[0] = (*dchildp)->d_inum;
1221 child_res_id.name[1] = (*dchildp)->d_generation;
1225 inode = (*dchildp)->d_inode;
1227 inode = igrab(inode);
1231 child_res_id.name[0] = inode->i_ino;
1232 child_res_id.name[1] = inode->i_generation;
1237 cleanup_phase = 2; /* child dentry */
1239 /* Step 3: Lock parent and child in resource order. If child doesn't
1240 * exist, we still have to lock the parent and re-lookup. */
1241 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1242 &parent_policy, &child_res_id, child_lockh,
1243 child_mode, &child_policy);
1247 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1248 cleanup_phase = 4; /* child lock */
1250 cleanup_phase = 3; /* parent lock */
1252 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1253 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1254 parent_mode, &child_res_id, child_lockh,
1255 dchildp, child_mode, &child_policy,
1256 name, namelen, &parent_res_id);
1266 switch (cleanup_phase) {
1268 ldlm_lock_decref(child_lockh, child_mode);
1270 ldlm_lock_decref(parent_lockh, parent_mode);
1275 if (parent_lockh[1].cookie)
1276 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1285 void mds_reconstruct_generic(struct ptlrpc_request *req)
1287 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1289 mds_req_from_mcd(req, med->med_mcd);
1292 int mds_create_local_dentry(struct mds_update_record *rec,
1293 struct obd_device *obd)
1295 struct mds_obd *mds = &obd->u.mds;
1296 struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1297 int fidlen = 0, rc, cleanup_phase = 0;
1298 struct dentry *new_child = NULL;
1299 char *fidname = rec->ur_name;
1300 struct dentry *child = NULL;
1301 struct lustre_handle lockh;
1305 down(&fids_dir->i_sem);
1306 fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1307 CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1308 fidname, (unsigned) rec->ur_fid1->id,
1309 (unsigned) rec->ur_fid1->generation);
1311 new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1312 up(&fids_dir->i_sem);
1313 if (IS_ERR(new_child)) {
1314 CERROR("can't lookup %s: %d\n", fidname,
1315 (int) PTR_ERR(new_child));
1316 GOTO(cleanup, rc = PTR_ERR(new_child));
1320 if (new_child->d_inode != NULL) {
1321 /* nice. we've already have local dentry! */
1322 CERROR("found dentry in FIDS/: %u/%u\n",
1323 (unsigned) new_child->d_inode->i_ino,
1324 (unsigned) new_child->d_inode->i_generation);
1325 rec->ur_fid1->id = fids_dir->i_ino;
1326 rec->ur_fid1->generation = fids_dir->i_generation;
1327 rec->ur_namelen = fidlen + 1;
1328 GOTO(cleanup, rc = 0);
1331 /* new, local dentry will be added soon. we need no aliases here */
1334 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
1335 &lockh, NULL, 0, MDS_INODELOCK_UPDATE);
1336 if (IS_ERR(child)) {
1337 CERROR("can't get victim\n");
1338 GOTO(cleanup, rc = PTR_ERR(child));
1342 handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1344 GOTO(cleanup, rc = PTR_ERR(handle));
1346 rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1347 rec->ur_fid1->id, rec->ur_fid1->generation,
1350 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1351 (unsigned long) child->d_inode->i_ino,
1352 (unsigned long) child->d_inode->i_generation, rc);
1354 if (S_ISDIR(child->d_inode->i_mode)) {
1355 fids_dir->i_nlink++;
1356 mark_inode_dirty(fids_dir);
1358 mark_inode_dirty(child->d_inode);
1360 fsfilt_commit(obd, fids_dir, handle, 0);
1362 rec->ur_fid1->id = fids_dir->i_ino;
1363 rec->ur_fid1->generation = fids_dir->i_generation;
1364 rec->ur_namelen = fidlen + 1;
1367 switch(cleanup_phase) {
1369 ldlm_lock_decref(&lockh, LCK_EX);
1379 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1380 struct ptlrpc_request *slave)
1382 void *cookie, *cookie2;
1383 struct mds_body *body2;
1384 struct mds_body *body;
1388 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1389 LASSERT(body != NULL);
1391 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1392 LASSERT(body2 != NULL);
1394 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1398 memcpy(body2, body, sizeof(*body));
1399 body2->valid &= ~OBD_MD_FLCOOKIE;
1401 if (!(body->valid & OBD_MD_FLEASIZE))
1404 if (body->eadatasize == 0) {
1405 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1409 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1411 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1412 LASSERT(ea != NULL);
1414 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1415 LASSERT(ea2 != NULL);
1417 memcpy(ea2, ea, body->eadatasize);
1419 if (body->valid & OBD_MD_FLCOOKIE) {
1420 LASSERT(master->rq_repmsg->buflens[2] >=
1421 slave->rq_repmsg->buflens[2]);
1422 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1423 slave->rq_repmsg->buflens[2]);
1424 LASSERT(cookie != NULL);
1426 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1427 master->rq_repmsg->buflens[2]);
1428 LASSERT(cookie2 != NULL);
1429 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1430 body2->valid |= OBD_MD_FLCOOKIE;
1435 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1436 struct ptlrpc_request *req,
1437 struct lustre_handle *parent_lockh,
1438 struct dentry *dparent,
1439 struct lustre_handle *child_lockh,
1440 struct dentry *dchild)
1442 struct mds_obd *mds = mds_req2mds(req);
1443 struct mdc_op_data op_data;
1444 int rc = 0, cleanup_phase = 0;
1445 struct ptlrpc_request *request = NULL;
1448 LASSERT(offset == 0 || offset == 2);
1450 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)\n",
1451 rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1452 (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1454 /* time to drop i_nlink on remote MDS */
1455 op_data.fid1.mds = dchild->d_mdsnum;
1456 op_data.fid1.id = dchild->d_inum;
1457 op_data.fid1.generation = dchild->d_generation;
1458 op_data.create_mode = rec->ur_mode;
1459 op_data.namelen = 0;
1460 op_data.name = NULL;
1461 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1464 mds_copy_unlink_reply(req, request);
1465 ptlrpc_req_finished(request);
1468 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1469 req->rq_status = rc;
1472 if (parent_lockh[1].cookie != 0)
1473 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1475 ldlm_lock_decref(child_lockh, LCK_EX);
1477 ldlm_lock_decref(parent_lockh, LCK_PW);
1479 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1486 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1487 struct ptlrpc_request *req,
1488 struct lustre_handle *lh)
1490 struct dentry *dparent, *dchild;
1491 struct mds_obd *mds = mds_req2mds(req);
1492 struct obd_device *obd = req->rq_export->exp_obd;
1493 struct mds_body *body = NULL;
1494 struct inode *child_inode;
1495 struct lustre_handle parent_lockh[2], child_lockh, child_reuse_lockh;
1496 char fidname[LL_FID_NAMELEN];
1497 void *handle = NULL;
1498 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1499 int unlink_by_fid = 0;
1502 LASSERT(offset == 0 || offset == 2);
1504 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1505 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1507 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1509 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1510 GOTO(cleanup, rc = -ENOENT);
1512 if (rec->ur_namelen == 1) {
1513 /* this is request to drop i_nlink on local inode */
1515 rec->ur_name = fidname;
1516 rc = mds_create_local_dentry(rec, obd);
1519 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1520 parent_lockh, &dparent, LCK_PW,
1521 MDS_INODELOCK_UPDATE,
1522 rec->ur_name, rec->ur_namelen,
1523 &child_lockh, &dchild, LCK_EX,
1524 MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1528 if (dchild->d_flags & DCACHE_CROSS_REF) {
1529 /* we should have parent lock only here */
1530 LASSERT(unlink_by_fid == 0);
1531 LASSERT(dchild->d_mdsnum != mds->mds_num);
1532 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1533 dparent, &child_lockh, dchild);
1537 cleanup_phase = 1; /* dchild, dparent, locks */
1540 child_inode = dchild->d_inode;
1541 if (child_inode == NULL) {
1542 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1543 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1544 GOTO(cleanup, rc = -ENOENT);
1547 cleanup_phase = 2; /* dchild has a lock */
1549 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1550 * reuse (see bug 2029). */
1551 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1555 cleanup_phase = 3; /* child inum lock */
1557 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1559 /* ldlm_reply in buf[0] if called via intent */
1563 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1564 LASSERT(body != NULL);
1566 /* If this is the last reference to this inode, get the OBD EA
1567 * data first so the client can destroy OST objects.
1568 * we only do the object removal if no open files remain.
1569 * Nobody can get at this name anymore because of the locks so
1570 * we make decisions here as to whether to remove the inode */
1571 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1572 mds_open_orphan_count(child_inode) == 0) {
1573 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1574 mds_pack_inode2body(obd, body, child_inode);
1575 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1577 if (!(body->valid & OBD_MD_FLEASIZE)) {
1578 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1579 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1585 /* We have to do these checks ourselves, in case we are making an
1586 * orphan. The client tells us whether rmdir() or unlink() was called,
1587 * so we need to return appropriate errors (bug 72).
1589 * We don't have to check permissions, because vfs_rename (called from
1590 * mds_open_unlink_rename) also calls may_delete. */
1591 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1592 if (!S_ISDIR(child_inode->i_mode))
1593 GOTO(cleanup, rc = -ENOTDIR);
1595 if (S_ISDIR(child_inode->i_mode))
1596 GOTO(cleanup, rc = -EISDIR);
1599 if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1600 mds_open_orphan_count(child_inode) > 0) {
1601 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1602 cleanup_phase = 4; /* transaction */
1606 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1607 switch (child_inode->i_mode & S_IFMT) {
1609 /* Drop any lingering child directories before we start our
1610 * transaction, to avoid doing multiple inode dirty/delete
1611 * in our compound transaction (bug 1321). */
1612 shrink_dcache_parent(dchild);
1613 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1616 GOTO(cleanup, rc = PTR_ERR(handle));
1617 cleanup_phase = 4; /* transaction */
1618 rc = vfs_rmdir(dparent->d_inode, dchild);
1621 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1622 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1624 handle = fsfilt_start_log(obd, dparent->d_inode,
1625 FSFILT_OP_UNLINK, NULL,
1626 le32_to_cpu(lmm->lmm_stripe_count));
1628 GOTO(cleanup, rc = PTR_ERR(handle));
1630 cleanup_phase = 4; /* transaction */
1631 rc = vfs_unlink(dparent->d_inode, dchild);
1633 if (!rc && log_unlink)
1634 if (mds_log_op_unlink(obd, child_inode,
1635 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1636 req->rq_repmsg->buflens[offset + 1],
1637 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1638 req->rq_repmsg->buflens[offset + 2]) > 0)
1639 body->valid |= OBD_MD_FLCOOKIE;
1647 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1650 GOTO(cleanup, rc = PTR_ERR(handle));
1651 cleanup_phase = 4; /* transaction */
1652 rc = vfs_unlink(dparent->d_inode, dchild);
1655 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1658 GOTO(cleanup, rc = -EINVAL);
1666 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1667 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1668 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1670 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1672 CERROR("error on parent setattr: rc = %d\n", err);
1675 switch(cleanup_phase) {
1677 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1680 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1681 "unlinked", 0, NULL);
1682 case 3: /* child ino-reuse lock */
1683 if (rc && body != NULL) {
1684 // Don't unlink the OST objects if the MDS unlink failed
1688 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1690 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1691 case 2: /* child lock */
1692 ldlm_lock_decref(&child_lockh, LCK_EX);
1693 case 1: /* child and parent dentry, parent lock */
1695 if (parent_lockh[1].cookie != 0)
1696 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1699 ldlm_lock_decref(parent_lockh, LCK_PW);
1701 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1708 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1711 req->rq_status = rc;
1716 * to service requests from remote MDS to increment i_nlink
1718 static int mds_reint_link_acquire(struct mds_update_record *rec,
1719 int offset, struct ptlrpc_request *req,
1720 struct lustre_handle *lh)
1722 struct obd_device *obd = req->rq_export->exp_obd;
1723 struct ldlm_res_id src_res_id = { .name = {0} };
1724 struct lustre_handle *handle = NULL, src_lockh;
1725 struct mds_obd *mds = mds_req2mds(req);
1726 int rc = 0, cleanup_phase = 0;
1727 struct dentry *de_src = NULL;
1728 ldlm_policy_data_t policy;
1732 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1733 obd->obd_name, (unsigned) rec->ur_fid1->mds,
1734 (unsigned) rec->ur_fid1->id,
1735 (unsigned) rec->ur_fid1->generation);
1737 /* Step 1: Lookup the source inode and target directory by FID */
1738 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1740 GOTO(cleanup, rc = PTR_ERR(de_src));
1741 cleanup_phase = 1; /* source dentry */
1743 src_res_id.name[0] = de_src->d_inode->i_ino;
1744 src_res_id.name[1] = de_src->d_inode->i_generation;
1745 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1747 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1748 src_res_id, LDLM_IBITS, &policy,
1749 LCK_EX, &flags, mds_blocking_ast,
1750 ldlm_completion_ast, NULL, NULL,
1751 NULL, 0, NULL, &src_lockh);
1753 GOTO(cleanup, rc = -ENOLCK);
1754 cleanup_phase = 2; /* lock */
1756 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1758 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1759 if (IS_ERR(handle)) {
1760 rc = PTR_ERR(handle);
1763 de_src->d_inode->i_nlink++;
1764 mark_inode_dirty(de_src->d_inode);
1767 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1768 handle, req, rc, 0);
1770 switch (cleanup_phase) {
1773 ldlm_lock_decref(&src_lockh, LCK_EX);
1775 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1781 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1784 req->rq_status = rc;
1789 * request to link to foreign inode:
1790 * - acquire i_nlinks on this inode
1793 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1794 int offset, struct ptlrpc_request *req,
1795 struct lustre_handle *lh)
1797 struct lustre_handle *handle = NULL, tgt_dir_lockh[2];
1798 struct obd_device *obd = req->rq_export->exp_obd;
1799 struct dentry *de_tgt_dir = NULL;
1800 struct mds_obd *mds = mds_req2mds(req);
1801 int rc = 0, cleanup_phase = 0;
1802 struct mdc_op_data op_data;
1803 struct ptlrpc_request *request = NULL;
1806 #define fmt "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1807 DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1808 (unsigned) rec->ur_fid2->mds,
1809 (unsigned) rec->ur_fid2->id,
1810 (unsigned) rec->ur_fid2->generation,
1811 rec->ur_namelen - 1, rec->ur_name,
1812 (unsigned) rec->ur_fid1->mds,
1813 (unsigned) rec->ur_fid1->id,
1814 (unsigned)rec->ur_fid1->generation);
1816 de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1817 tgt_dir_lockh, rec->ur_name,
1818 rec->ur_namelen - 1,
1819 MDS_INODELOCK_UPDATE);
1820 if (IS_ERR(de_tgt_dir))
1821 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1824 op_data.fid1 = *(rec->ur_fid1);
1825 op_data.namelen = 0;
1826 op_data.name = NULL;
1827 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1831 ptlrpc_req_finished(request);
1833 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1835 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1836 if (IS_ERR(handle)) {
1837 rc = PTR_ERR(handle);
1841 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1842 rec->ur_namelen - 1, rec->ur_fid1->id,
1843 rec->ur_fid1->generation, rec->ur_fid1->mds);
1847 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1848 handle, req, rc, 0);
1851 switch (cleanup_phase) {
1854 /* FIXME: drop i_nlink on remote inode here */
1855 CERROR("MUST drop drop i_nlink here\n");
1860 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1862 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
1865 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1867 ptlrpc_save_lock(req, tgt_dir_lockh + 1, LCK_CW);
1873 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1876 req->rq_status = rc;
1880 static int mds_reint_link(struct mds_update_record *rec, int offset,
1881 struct ptlrpc_request *req,
1882 struct lustre_handle *lh)
1884 struct obd_device *obd = req->rq_export->exp_obd;
1885 struct dentry *de_src = NULL;
1886 struct dentry *de_tgt_dir = NULL;
1887 struct dentry *dchild = NULL;
1888 struct mds_obd *mds = mds_req2mds(req);
1889 struct lustre_handle *handle = NULL, tgt_dir_lockh[2], src_lockh;
1890 struct ldlm_res_id src_res_id = { .name = {0} };
1891 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1892 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1893 ldlm_policy_data_t tgt_dir_policy =
1894 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1896 int rc = 0, cleanup_phase = 0;
1899 LASSERT(offset == 0);
1901 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1902 rec->ur_fid1->id, rec->ur_fid1->generation,
1903 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1905 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1907 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1908 GOTO(cleanup, rc = -ENOENT);
1910 if (rec->ur_fid1->mds != mds->mds_num) {
1911 rc = mds_reint_link_to_remote(rec, offset, req, lh);
1915 if (rec->ur_namelen == 1) {
1916 rc = mds_reint_link_acquire(rec, offset, req, lh);
1920 /* Step 1: Lookup the source inode and target directory by FID */
1921 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1923 GOTO(cleanup, rc = PTR_ERR(de_src));
1925 cleanup_phase = 1; /* source dentry */
1927 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1928 if (IS_ERR(de_tgt_dir))
1929 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1931 cleanup_phase = 2; /* target directory dentry */
1933 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1934 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1935 de_src->d_inode->i_ino);
1937 /* Step 2: Take the two locks */
1938 src_res_id.name[0] = de_src->d_inode->i_ino;
1939 src_res_id.name[1] = de_src->d_inode->i_generation;
1940 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1941 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1943 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
1945 /* Get a temp lock on just ino, gen to flush client cache */
1946 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1947 tgt_dir_res_id, LDLM_IBITS, &src_policy,
1948 LCK_CW, &flags, mds_blocking_ast,
1949 ldlm_completion_ast, NULL, NULL,
1950 NULL, 0, NULL, tgt_dir_lockh + 1);
1952 GOTO(cleanup, rc = -ENOLCK);
1954 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
1955 rec->ur_namelen - 1);
1956 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1957 de_tgt_dir->d_inode->i_ino,
1958 de_tgt_dir->d_inode->i_generation,
1959 tgt_dir_res_id.name[2]);
1962 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1964 &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
1969 cleanup_phase = 3; /* locks */
1971 /* Step 3: Lookup the child */
1972 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1973 if (IS_ERR(dchild)) {
1974 rc = PTR_ERR(dchild);
1975 if (rc != -EPERM && rc != -EACCES)
1976 CERROR("child lookup error %d\n", rc);
1980 cleanup_phase = 4; /* child dentry */
1982 if (dchild->d_inode) {
1983 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1984 de_tgt_dir->d_inode->i_ino, rec->ur_name);
1989 /* Step 4: Do it. */
1990 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1992 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1993 if (IS_ERR(handle)) {
1994 rc = PTR_ERR(handle);
1998 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1999 if (rc && rc != -EPERM && rc != -EACCES)
2000 CERROR("vfs_link error %d\n", rc);
2002 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2003 handle, req, rc, 0);
2006 switch (cleanup_phase) {
2007 case 4: /* child dentry */
2011 ldlm_lock_decref(&src_lockh, LCK_EX);
2012 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2014 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2015 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2017 case 2: /* target dentry */
2019 if (tgt_dir_lockh[1].cookie)
2020 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
2024 case 1: /* source dentry */
2029 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2032 req->rq_status = rc;
2037 * add a hard link in the PENDING directory, only used by rename()
2039 static int mds_add_link_orphan(struct mds_update_record *rec,
2040 struct obd_device *obd,
2041 struct dentry *dentry)
2043 struct mds_obd *mds = &obd->u.mds;
2044 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2045 struct dentry *pending_child;
2046 char fidname[LL_FID_NAMELEN];
2050 LASSERT(dentry->d_inode);
2051 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2053 down(&pending_dir->i_sem);
2054 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2055 dentry->d_inode->i_generation);
2057 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2058 mds_open_orphan_count(dentry->d_inode),
2059 rec->ur_name, fidname);
2061 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2062 if (IS_ERR(pending_child))
2063 GOTO(out_lock, rc = PTR_ERR(pending_child));
2065 if (pending_child->d_inode != NULL) {
2066 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2067 LASSERT(pending_child->d_inode == dentry->d_inode);
2068 GOTO(out_dput, rc = 0);
2072 rc = vfs_link(dentry, pending_dir, pending_child);
2075 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2078 mds_inode_set_orphan(dentry->d_inode);
2080 l_dput(pending_child);
2082 up(&pending_dir->i_sem);
2086 /* The idea here is that we need to get four locks in the end:
2087 * one on each parent directory, one on each child. We need to take
2088 * these locks in some kind of order (to avoid deadlocks), and the order
2089 * I selected is "increasing resource number" order. We need to look up
2090 * the children, however, before we know what the resource number(s) are.
2091 * Thus the following plan:
2093 * 1,2. Look up the parents
2094 * 3,4. Look up the children
2095 * 5. Take locks on the parents and children, in order
2096 * 6. Verify that the children haven't changed since they were looked up
2098 * If there was a race and the children changed since they were first looked
2099 * up, it is possible that mds_verify_child() will be able to just grab the
2100 * lock on the new child resource (if it has a higher resource than any other)
2101 * but we need to compare against not only its parent, but also against the
2102 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2104 * We need the fancy igrab() on the child inodes because we aren't holding a
2105 * lock on the parent after the lookup is done, so dentry->d_inode may change
2106 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2108 static int mds_get_parents_children_locked(struct obd_device *obd,
2109 struct mds_obd *mds,
2110 struct ll_fid *p1_fid,
2111 struct dentry **de_srcdirp,
2112 struct ll_fid *p2_fid,
2113 struct dentry **de_tgtdirp,
2115 const char *old_name, int old_len,
2116 struct dentry **de_oldp,
2117 const char *new_name, int new_len,
2118 struct dentry **de_newp,
2119 struct lustre_handle *dlm_handles,
2122 struct ldlm_res_id p1_res_id = { .name = {0} };
2123 struct ldlm_res_id p2_res_id = { .name = {0} };
2124 struct ldlm_res_id c1_res_id = { .name = {0} };
2125 struct ldlm_res_id c2_res_id = { .name = {0} };
2126 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2127 /* Only dentry should change, but the inode itself would be
2129 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2130 /* If something is going to be replaced, both dentry and inode locks are
2132 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2133 MDS_INODELOCK_UPDATE}};
2134 struct ldlm_res_id *maxres_src, *maxres_tgt;
2135 struct inode *inode;
2136 int rc = 0, cleanup_phase = 0;
2139 /* Step 1: Lookup the source directory */
2140 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2141 if (IS_ERR(*de_srcdirp))
2142 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2144 cleanup_phase = 1; /* source directory dentry */
2146 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2147 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2149 /* Step 2: Lookup the target directory */
2150 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2151 *de_tgtdirp = dget(*de_srcdirp);
2153 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2154 if (IS_ERR(*de_tgtdirp))
2155 GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
2158 cleanup_phase = 2; /* target directory dentry */
2160 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2161 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2164 dlm_handles[5].cookie = 0;
2165 dlm_handles[6].cookie = 0;
2166 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2167 /* Get a temp lock on just ino, gen to flush client cache */
2168 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2169 LCK_CW, &p_policy, &p2_res_id,
2170 &(dlm_handles[6]),LCK_CW,&p_policy);
2174 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2175 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2176 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2177 (*de_srcdirp)->d_inode->i_ino,
2178 (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2179 (*de_tgtdirp)->d_inode->i_ino,
2180 (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2184 /* Step 3: Lookup the source child entry */
2185 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2186 if (IS_ERR(*de_oldp)) {
2187 rc = PTR_ERR(*de_oldp);
2188 CERROR("old child lookup error (%*s): %d\n",
2189 old_len - 1, old_name, rc);
2193 cleanup_phase = 3; /* original name dentry */
2195 inode = (*de_oldp)->d_inode;
2196 if (inode != NULL) {
2197 inode = igrab(inode);
2199 GOTO(cleanup, rc = -ENOENT);
2201 c1_res_id.name[0] = inode->i_ino;
2202 c1_res_id.name[1] = inode->i_generation;
2204 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2205 c1_res_id.name[0] = (*de_oldp)->d_inum;
2206 c1_res_id.name[1] = (*de_oldp)->d_generation;
2209 /* Step 4: Lookup the target child entry */
2210 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2211 if (IS_ERR(*de_newp)) {
2212 rc = PTR_ERR(*de_newp);
2213 CERROR("new child lookup error (%*s): %d\n",
2214 old_len - 1, old_name, rc);
2218 cleanup_phase = 4; /* target dentry */
2220 inode = (*de_newp)->d_inode;
2221 if (inode != NULL) {
2222 inode = igrab(inode);
2226 c2_res_id.name[0] = inode->i_ino;
2227 c2_res_id.name[1] = inode->i_generation;
2229 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2230 c2_res_id.name[0] = (*de_newp)->d_inum;
2231 c2_res_id.name[1] = (*de_newp)->d_generation;
2235 /* Step 5: Take locks on the parents and child(ren) */
2236 maxres_src = &p1_res_id;
2237 maxres_tgt = &p2_res_id;
2238 cleanup_phase = 4; /* target dentry */
2240 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2241 maxres_src = &c1_res_id;
2242 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2243 maxres_tgt = &c2_res_id;
2245 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2247 &p2_res_id, &dlm_handles[1], parent_mode,
2249 &c1_res_id, &dlm_handles[2], child_mode,
2251 &c2_res_id, &dlm_handles[3], child_mode,
2256 cleanup_phase = 6; /* parent and child(ren) locks */
2258 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2259 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2260 parent_mode, &c1_res_id, &dlm_handles[2],
2261 de_oldp, child_mode, &c1_policy, old_name,old_len,
2264 if (c2_res_id.name[0] != 0)
2265 ldlm_lock_decref(&dlm_handles[3], child_mode);
2266 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2273 if (!DENTRY_VALID(*de_oldp))
2274 GOTO(cleanup, rc = -ENOENT);
2276 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2277 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2278 parent_mode, &c2_res_id, &dlm_handles[3],
2279 de_newp, child_mode, &c2_policy, new_name,
2280 new_len, maxres_src);
2282 ldlm_lock_decref(&dlm_handles[2], child_mode);
2283 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2293 switch (cleanup_phase) {
2294 case 6: /* child lock(s) */
2295 if (c2_res_id.name[0] != 0)
2296 ldlm_lock_decref(&dlm_handles[3], child_mode);
2297 if (c1_res_id.name[0] != 0)
2298 ldlm_lock_decref(&dlm_handles[2], child_mode);
2299 case 5: /* parent locks */
2300 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2301 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2302 case 4: /* target dentry */
2304 case 3: /* source dentry */
2306 case 2: /* target directory dentry */
2307 l_dput(*de_tgtdirp);
2308 case 1: /* source directry dentry */
2309 l_dput(*de_srcdirp);
2316 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2317 int offset, struct ptlrpc_request *req)
2319 struct obd_device *obd = req->rq_export->exp_obd;
2320 struct dentry *de_srcdir = NULL;
2321 struct dentry *de_new = NULL;
2322 struct mds_obd *mds = mds_req2mds(req);
2323 struct lustre_handle parent_lockh[2];
2324 struct lustre_handle child_lockh;
2325 int cleanup_phase = 0;
2326 void *handle = NULL;
2330 /* another MDS executing rename operation has asked us
2331 * to create target name. such a creation should destroy
2332 * existing target name */
2334 CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2335 obd->obd_name, rec->ur_tgt,
2336 (unsigned long) rec->ur_fid1->mds,
2337 (unsigned long) rec->ur_fid1->id,
2338 (unsigned long) rec->ur_fid1->generation);
2340 /* first, lookup the target */
2341 child_lockh.cookie = 0;
2342 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2343 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2344 rec->ur_tgt, rec->ur_tgtlen,
2345 &child_lockh, &de_new, LCK_EX,
2346 MDS_INODELOCK_LOOKUP);
2353 LASSERT(de_srcdir->d_inode);
2356 if (de_new->d_inode) {
2357 /* name exists and points to local inode
2358 * try to unlink this name and create new one */
2359 CERROR("%s: %s points to local inode %lu/%lu\n",
2360 obd->obd_name, rec->ur_tgt,
2361 (unsigned long) de_new->d_inode->i_ino,
2362 (unsigned long) de_new->d_inode->i_generation);
2363 handle = fsfilt_start(obd, de_srcdir->d_inode,
2364 FSFILT_OP_RENAME, NULL);
2366 GOTO(cleanup, rc = PTR_ERR(handle));
2367 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2370 } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2371 /* name exists adn points to remove inode */
2372 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2373 obd->obd_name, rec->ur_tgt,
2374 (unsigned long) de_new->d_mdsnum,
2375 (unsigned long) de_new->d_inum,
2376 (unsigned long) de_new->d_generation);
2378 /* name doesn't exist. the simplest case */
2379 handle = fsfilt_start(obd, de_srcdir->d_inode,
2380 FSFILT_OP_LINK, NULL);
2382 GOTO(cleanup, rc = PTR_ERR(handle));
2386 rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2387 rec->ur_tgtlen - 1, rec->ur_fid1->id,
2388 rec->ur_fid1->generation, rec->ur_fid1->mds);
2390 CERROR("add_dir_entry() returned error %d\n", rc);
2393 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2394 handle, req, rc, 0);
2395 switch(cleanup_phase) {
2399 if (parent_lockh[1].cookie != 0)
2400 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2402 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2403 if (child_lockh.cookie != 0)
2404 ldlm_lock_decref(&child_lockh, LCK_EX);
2412 req->rq_status = rc;
2417 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2418 struct ptlrpc_request *req)
2420 struct obd_device *obd = req->rq_export->exp_obd;
2421 struct ptlrpc_request *req2 = NULL;
2422 struct dentry *de_srcdir = NULL;
2423 struct dentry *de_old = NULL;
2424 struct mds_obd *mds = mds_req2mds(req);
2425 struct lustre_handle parent_lockh[2];
2426 struct lustre_handle child_lockh;
2427 struct mdc_op_data opdata;
2428 int cleanup_phase = 0;
2429 void *handle = NULL;
2433 CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2434 obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2435 memset(&opdata, 0, sizeof(opdata));
2437 child_lockh.cookie = 0;
2438 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2439 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2440 rec->ur_name, rec->ur_namelen,
2441 &child_lockh, &de_old, LCK_EX,
2442 MDS_INODELOCK_LOOKUP);
2445 LASSERT(de_srcdir->d_inode);
2448 /* we already know the target should be created on another MDS
2449 * so, we have to request that MDS to do it */
2451 /* prepare source fid */
2452 if (de_old->d_flags & DCACHE_CROSS_REF) {
2453 LASSERT(de_old->d_inode == NULL);
2454 CDEBUG(D_OTHER, "request to move remote name\n");
2455 opdata.fid1.mds = de_old->d_mdsnum;
2456 opdata.fid1.id = de_old->d_inum;
2457 opdata.fid1.generation = de_old->d_generation;
2458 } else if (de_old->d_inode == NULL) {
2459 /* oh, source doesn't exist */
2460 GOTO(cleanup, rc = -ENOENT);
2462 LASSERT(de_old->d_inode != NULL);
2463 CDEBUG(D_OTHER, "request to move local name\n");
2464 opdata.fid1.mds = mds->mds_num;
2465 opdata.fid1.id = de_old->d_inode->i_ino;
2466 opdata.fid1.generation = de_old->d_inode->i_generation;
2469 opdata.fid2 = *(rec->ur_fid2);
2470 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2471 rec->ur_tgtlen - 1, &req2);
2476 handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2478 GOTO(cleanup, rc = PTR_ERR(handle));
2479 rc = fsfilt_del_dir_entry(obd, de_old);
2484 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2485 handle, req, rc, 0);
2487 ptlrpc_req_finished(req2);
2490 if (parent_lockh[1].cookie != 0)
2491 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2493 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2494 if (child_lockh.cookie != 0)
2495 ldlm_lock_decref(&child_lockh, LCK_EX);
2500 req->rq_status = rc;
2505 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2506 struct ptlrpc_request *req,
2507 struct lustre_handle *lockh)
2509 struct obd_device *obd = req->rq_export->exp_obd;
2510 struct dentry *de_srcdir = NULL;
2511 struct dentry *de_tgtdir = NULL;
2512 struct dentry *de_old = NULL;
2513 struct dentry *de_new = NULL;
2514 struct mds_obd *mds = mds_req2mds(req);
2515 struct lustre_handle dlm_handles[7];
2516 struct mds_body *body = NULL;
2517 int rc = 0, lock_count = 3;
2518 int cleanup_phase = 0;
2519 void *handle = NULL;
2522 LASSERT(offset == 0);
2524 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2525 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2526 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2528 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2530 if (rec->ur_namelen == 1) {
2531 rc = mds_reint_rename_create_name(rec, offset, req);
2535 if (rec->ur_fid2->mds != mds->mds_num) {
2536 rc = mds_reint_rename_to_remote(rec, offset, req);
2540 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2541 rec->ur_fid2, &de_tgtdir, LCK_PW,
2542 rec->ur_name, rec->ur_namelen,
2543 &de_old, rec->ur_tgt,
2544 rec->ur_tgtlen, &de_new,
2545 dlm_handles, LCK_EX);
2549 cleanup_phase = 1; /* parent(s), children, locks */
2551 if (de_new->d_inode)
2554 /* sanity check for src inode */
2555 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2556 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2557 GOTO(cleanup, rc = -EINVAL);
2559 /* sanity check for dest inode */
2560 if (de_new->d_inode &&
2561 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2562 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2563 GOTO(cleanup, rc = -EINVAL);
2565 if (de_old->d_inode == de_new->d_inode) {
2566 GOTO(cleanup, rc = 0);
2569 /* if we are about to remove the target at first, pass the EA of
2570 * that inode to client to perform and cleanup on OST */
2571 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2572 LASSERT(body != NULL);
2574 if (de_new->d_inode &&
2575 S_ISREG(de_new->d_inode->i_mode) &&
2576 de_new->d_inode->i_nlink == 1 &&
2577 mds_open_orphan_count(de_new->d_inode) == 0) {
2578 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2579 mds_pack_inode2body(obd, body, de_new->d_inode);
2580 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2581 if (!(body->valid & OBD_MD_FLEASIZE)) {
2582 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2583 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2585 /* XXX need log unlink? */
2589 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2590 de_srcdir->d_inode->i_sb);
2592 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2594 GOTO(cleanup, rc = PTR_ERR(handle));
2596 /* FIXME need adjust the journal block count? */
2597 /* if the target should be moved to PENDING, we at first increase the
2598 * link and later vfs_rename() will decrease the link count again */
2599 if (de_new->d_inode &&
2600 S_ISREG(de_new->d_inode->i_mode) &&
2601 de_new->d_inode->i_nlink == 1 &&
2602 mds_open_orphan_count(de_new->d_inode) > 0) {
2603 rc = mds_add_link_orphan(rec, obd, de_new);
2609 de_old->d_fsdata = req;
2610 de_new->d_fsdata = req;
2611 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2616 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2617 handle, req, rc, 0);
2618 switch (cleanup_phase) {
2621 if (dlm_handles[5].cookie != 0)
2622 ldlm_lock_decref(&(dlm_handles[5]), LCK_CW);
2623 if (dlm_handles[6].cookie != 0)
2624 ldlm_lock_decref(&(dlm_handles[6]), LCK_CW);
2627 if (lock_count == 4)
2628 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2629 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2630 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2631 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2633 if (lock_count == 4)
2634 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2635 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2636 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2637 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2646 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2649 req->rq_status = rc;
2653 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2654 struct ptlrpc_request *, struct lustre_handle *);
2656 static mds_reinter reinters[REINT_MAX + 1] = {
2657 [REINT_SETATTR] mds_reint_setattr,
2658 [REINT_CREATE] mds_reint_create,
2659 [REINT_LINK] mds_reint_link,
2660 [REINT_UNLINK] mds_reint_unlink,
2661 [REINT_RENAME] mds_reint_rename,
2662 [REINT_OPEN] mds_open
2665 int mds_reint_rec(struct mds_update_record *rec, int offset,
2666 struct ptlrpc_request *req, struct lustre_handle *lockh)
2668 struct obd_device *obd = req->rq_export->exp_obd;
2669 struct obd_run_ctxt saved;
2672 /* checked by unpacker */
2673 LASSERT(rec->ur_opcode <= REINT_MAX &&
2674 reinters[rec->ur_opcode] != NULL);
2676 push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2677 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2678 pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);