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 = -ERESTART);
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 = -ERESTART);
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;
647 if (rc == 0 && nstripes) {
648 /* FIXME: error handling here */
649 mds_try_to_split_dir(obd, dchild,
652 } else if (!DENTRY_VALID(dchild)) {
653 /* inode will be created on another MDS */
654 struct obdo *oa = NULL;
655 struct mds_body *body;
657 /* first, create that inode */
661 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
662 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
663 OBD_MD_FLUID | OBD_MD_FLGID);
664 oa->o_mode = dir->i_mode;
665 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
667 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
668 CWARN("%s: replay dir creation %*s -> %u/%u\n",
669 obd->obd_name, rec->ur_namelen - 1,
670 rec->ur_name, (unsigned) rec->ur_fid2->id,
671 (unsigned) rec->ur_fid2->generation);
672 oa->o_id = rec->ur_fid2->id;
673 oa->o_generation = rec->ur_fid2->generation;
674 oa->o_flags |= OBD_FL_RECREATE_OBJS;
677 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
680 /* now, add new dir entry for it */
681 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
683 GOTO(cleanup, rc = PTR_ERR(handle));
684 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
686 oa->o_id, oa->o_generation,
691 body = lustre_msg_buf(req->rq_repmsg,
692 offset, sizeof (*body));
693 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
694 body->fid1.id = oa->o_id;
696 body->fid1.generation = oa->o_generation;
699 /* requested name exists in the directory */
706 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
708 GOTO(cleanup, rc = PTR_ERR(handle));
709 if (rec->ur_tgt == NULL) /* no target supplied */
710 rc = -EINVAL; /* -EPROTO? */
712 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
720 int rdev = rec->ur_rdev;
721 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
723 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
724 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
729 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
730 dchild->d_fsdata = NULL;
731 GOTO(cleanup, rc = -EINVAL);
734 /* In case we stored the desired inum in here, we want to clean up. */
735 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
736 dchild->d_fsdata = NULL;
739 CDEBUG(D_INODE, "error during create: %d\n", rc);
741 } else if (dchild->d_inode) {
743 struct inode *inode = dchild->d_inode;
744 struct mds_body *body;
747 LTIME_S(iattr.ia_atime) = rec->ur_time;
748 LTIME_S(iattr.ia_ctime) = rec->ur_time;
749 LTIME_S(iattr.ia_mtime) = rec->ur_time;
750 iattr.ia_uid = rec->ur_fsuid;
751 if (dir->i_mode & S_ISGID)
752 iattr.ia_gid = dir->i_gid;
754 iattr.ia_gid = rec->ur_fsgid;
755 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
756 ATTR_MTIME | ATTR_CTIME;
758 if (rec->ur_fid2->id) {
759 LASSERT(rec->ur_fid2->id == inode->i_ino);
760 inode->i_generation = rec->ur_fid2->generation;
761 /* Dirtied and committed by the upcoming setattr. */
762 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
763 inode->i_ino, inode->i_generation);
765 struct lustre_handle child_ino_lockh;
767 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
768 inode->i_ino, inode->i_generation);
770 /* The inode we were allocated may have just been freed
771 * by an unlink operation. We take this lock to
772 * synchronize against the matching reply-ack-lock taken
773 * in unlink, to avoid replay problems if this reply
774 * makes it out to the client but the unlink's does not.
775 * See bug 2029 for more detail.*/
776 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
777 if (rc != ELDLM_OK) {
778 CERROR("error locking for unlink/create sync: "
781 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
785 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
787 CERROR("error on child setattr: rc = %d\n", rc);
789 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
790 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
792 CERROR("error on parent setattr: rc = %d\n", rc);
794 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
795 mds_pack_inode2fid(obd, &body->fid1, inode);
796 mds_pack_inode2body(obd, body, inode);
801 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
804 /* Destroy the file we just created. This should not need
805 * extra journal credits, as we have already modified all of
806 * the blocks needed in order to create the file in the first
811 err = vfs_rmdir(dir, dchild);
813 CERROR("rmdir in error path: %d\n", err);
816 err = vfs_unlink(dir, dchild);
818 CERROR("unlink in error path: %d\n", err);
824 switch (cleanup_phase) {
825 case 2: /* child dentry */
827 case 1: /* locked parent dentry */
829 if (lockh[1].cookie != 0)
830 ldlm_lock_decref(lockh + 1, LCK_CW);
833 ldlm_lock_decref(lockh, LCK_PW);
835 ptlrpc_save_lock (req, lockh, LCK_PW);
841 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
845 OBD_FREE(mea, mea_size);
850 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
851 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
855 for (i = 0; i < RES_NAME_SIZE; i++) {
856 /* return 1 here, because enqueue_ordered will skip resources
857 * of all zeroes if they're sorted to the end of the list. */
858 if (res1->name[i] == 0 && res2->name[i] != 0)
860 if (res2->name[i] == 0 && res1->name[i] != 0)
863 if (res1->name[i] > res2->name[i])
865 if (res1->name[i] < res2->name[i])
872 if (memcmp(p1, p2, sizeof(*p1)) < 0)
878 /* This function doesn't use ldlm_match_or_enqueue because we're always called
879 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
880 * because they take the place of local semaphores.
882 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
883 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
884 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
885 struct lustre_handle *p1_lockh, int p1_lock_mode,
886 ldlm_policy_data_t *p1_policy,
887 struct ldlm_res_id *p2_res_id,
888 struct lustre_handle *p2_lockh, int p2_lock_mode,
889 ldlm_policy_data_t *p2_policy)
891 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
892 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
893 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
894 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
898 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
900 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
903 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
904 handles[1] = p1_lockh;
905 handles[0] = p2_lockh;
906 res_id[1] = p1_res_id;
907 res_id[0] = p2_res_id;
908 lock_modes[1] = p1_lock_mode;
909 lock_modes[0] = p2_lock_mode;
910 policies[1] = p1_policy;
911 policies[0] = p2_policy;
914 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
915 res_id[0]->name[0], res_id[1]->name[0]);
917 flags = LDLM_FL_LOCAL_ONLY;
918 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
919 LDLM_IBITS, policies[0], lock_modes[0], &flags,
920 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
921 NULL, 0, NULL, handles[0]);
924 ldlm_lock_dump_handle(D_OTHER, handles[0]);
926 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
927 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
928 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
929 ldlm_lock_addref(handles[1], lock_modes[1]);
930 } else if (res_id[1]->name[0] != 0) {
931 flags = LDLM_FL_LOCAL_ONLY;
932 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
933 *res_id[1], LDLM_IBITS, policies[1],
934 lock_modes[1], &flags, mds_blocking_ast,
935 ldlm_completion_ast, NULL, NULL, NULL, 0,
937 if (rc != ELDLM_OK) {
938 ldlm_lock_decref(handles[0], lock_modes[0]);
941 ldlm_lock_dump_handle(D_OTHER, handles[1]);
947 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
948 struct lustre_handle *p1_lockh, int p1_lock_mode,
949 ldlm_policy_data_t *p1_policy,
950 struct ldlm_res_id *p2_res_id,
951 struct lustre_handle *p2_lockh, int p2_lock_mode,
952 ldlm_policy_data_t *p2_policy,
953 struct ldlm_res_id *c1_res_id,
954 struct lustre_handle *c1_lockh, int c1_lock_mode,
955 ldlm_policy_data_t *c1_policy,
956 struct ldlm_res_id *c2_res_id,
957 struct lustre_handle *c2_lockh, int c2_lock_mode,
958 ldlm_policy_data_t *c2_policy)
960 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
961 c1_res_id, c2_res_id };
962 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
963 c1_lockh, c2_lockh };
964 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
965 c1_lock_mode, c2_lock_mode };
966 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
967 c1_policy, c2_policy};
968 int rc, i, j, sorted, flags;
972 "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
973 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
976 /* simple insertion sort - we have at most 4 elements */
977 for (i = 1; i < 4; i++) {
979 dlm_handles[4] = dlm_handles[i];
980 res_id[4] = res_id[i];
981 lock_modes[4] = lock_modes[i];
982 policies[4] = policies[i];
986 if (res_gt(res_id[j], res_id[4], policies[j],
988 dlm_handles[j + 1] = dlm_handles[j];
989 res_id[j + 1] = res_id[j];
990 lock_modes[j + 1] = lock_modes[j];
991 policies[j + 1] = policies[j];
996 } while (j >= 0 && !sorted);
998 dlm_handles[j + 1] = dlm_handles[4];
999 res_id[j + 1] = res_id[4];
1000 lock_modes[j + 1] = lock_modes[4];
1001 policies[j + 1] = policies[4];
1005 "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1006 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1007 res_id[3]->name[0]);
1009 /* XXX we could send ASTs on all these locks first before blocking? */
1010 for (i = 0; i < 4; i++) {
1012 if (res_id[i]->name[0] == 0)
1015 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1016 (policies[i]->l_inodebits.bits &
1017 policies[i-1]->l_inodebits.bits) ) {
1018 memcpy(dlm_handles[i], dlm_handles[i-1],
1019 sizeof(*(dlm_handles[i])));
1020 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1022 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1023 *res_id[i], LDLM_IBITS,
1025 lock_modes[i], &flags,
1027 ldlm_completion_ast, NULL, NULL,
1028 NULL, 0, NULL, dlm_handles[i]);
1030 GOTO(out_err, rc = -EIO);
1031 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1038 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1043 /* In the unlikely case that the child changed while we were waiting
1044 * on the lock, we need to drop the lock on the old child and either:
1045 * - if the child has a lower resource name, then we have to also
1046 * drop the parent lock and regain the locks in the right order
1047 * - in the rename case, if the child has a lower resource name than one of
1048 * the other parent/child resources (maxres) we also need to reget the locks
1049 * - if the child has a higher resource name (this is the common case)
1050 * we can just get the lock on the new child (still in lock order)
1052 * Returns 0 if the child did not change or if it changed but could be locked.
1053 * Returns 1 if the child changed and we need to re-lock (no locks held).
1054 * Returns -ve error with a valid dchild (no locks held). */
1055 static int mds_verify_child(struct obd_device *obd,
1056 struct ldlm_res_id *parent_res_id,
1057 struct lustre_handle *parent_lockh,
1058 struct dentry *dparent, int parent_mode,
1059 struct ldlm_res_id *child_res_id,
1060 struct lustre_handle *child_lockh,
1061 struct dentry **dchildp, int child_mode,
1062 ldlm_policy_data_t *child_policy,
1063 const char *name, int namelen,
1064 struct ldlm_res_id *maxres)
1066 struct dentry *vchild, *dchild = *dchildp;
1067 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1070 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1072 GOTO(cleanup, rc = PTR_ERR(vchild));
1074 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1075 if (child_res_id->name[0] == vchild->d_inum &&
1076 child_res_id->name[1] == vchild->d_generation) {
1085 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1086 (vchild->d_inode != NULL &&
1087 child_res_id->name[0] == vchild->d_inode->i_ino &&
1088 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1097 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1098 vchild->d_inode, dchild ? dchild->d_inode : 0,
1099 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1100 child_res_id->name[0]);
1101 if (child_res_id->name[0] != 0)
1102 ldlm_lock_decref(child_lockh, child_mode);
1106 cleanup_phase = 1; /* parent lock only */
1107 *dchildp = dchild = vchild;
1109 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1111 if (dchild->d_inode) {
1112 child_res_id->name[0] = dchild->d_inode->i_ino;
1113 child_res_id->name[1] = dchild->d_inode->i_generation;
1115 child_res_id->name[0] = dchild->d_inum;
1116 child_res_id->name[1] = dchild->d_generation;
1119 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1120 res_gt(maxres, child_res_id, NULL, NULL)) {
1121 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1122 child_res_id->name[0], parent_res_id->name[0],
1124 GOTO(cleanup, rc = 1);
1127 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1128 *child_res_id, LDLM_IBITS, child_policy,
1129 child_mode, &flags, mds_blocking_ast,
1130 ldlm_completion_ast, NULL, NULL, NULL, 0,
1133 GOTO(cleanup, rc = -EIO);
1136 memset(child_res_id, 0, sizeof(*child_res_id));
1142 switch(cleanup_phase) {
1144 if (child_res_id->name[0] != 0)
1145 ldlm_lock_decref(child_lockh, child_mode);
1147 ldlm_lock_decref(parent_lockh, parent_mode);
1153 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1155 struct lustre_handle *parent_lockh,
1156 struct dentry **dparentp, int parent_mode,
1157 __u64 parent_lockpart,
1158 char *name, int namelen,
1159 struct lustre_handle *child_lockh,
1160 struct dentry **dchildp, int child_mode,
1161 __u64 child_lockpart)
1163 struct ldlm_res_id child_res_id = { .name = {0} };
1164 struct ldlm_res_id parent_res_id = { .name = {0} };
1165 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1166 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1167 struct inode *inode;
1168 int rc = 0, cleanup_phase = 0;
1171 /* Step 1: Lookup parent */
1172 *dparentp = mds_fid2dentry(mds, fid, NULL);
1173 if (IS_ERR(*dparentp))
1174 RETURN(rc = PTR_ERR(*dparentp));
1175 LASSERT((*dparentp)->d_inode);
1177 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1178 (*dparentp)->d_inode->i_ino, name);
1180 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1181 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1183 parent_lockh[1].cookie = 0;
1184 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1185 /* lock just dir { ino, generation } to flush client cache */
1186 if (parent_mode == LCK_PW) {
1187 struct ldlm_res_id res_id = { .name = {0} };
1188 ldlm_policy_data_t policy;
1190 res_id.name[0] = (*dparentp)->d_inode->i_ino;
1191 res_id.name[1] = (*dparentp)->d_inode->i_generation;
1192 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1193 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1195 &policy, LCK_CW, &flags,
1197 ldlm_completion_ast, NULL, NULL,
1198 NULL, 0, NULL, parent_lockh+1);
1203 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1204 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1205 (*dparentp)->d_inode->i_ino,
1206 (*dparentp)->d_inode->i_generation,
1207 parent_res_id.name[2]);
1211 cleanup_phase = 1; /* parent dentry */
1213 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1214 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1215 if (IS_ERR(*dchildp)) {
1216 rc = PTR_ERR(*dchildp);
1217 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1221 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1222 /* inode lives on another MDS: return * mds/ino/gen
1223 * and LOOKUP lock. drop possible UPDATE lock! */
1224 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1225 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1226 child_res_id.name[0] = (*dchildp)->d_inum;
1227 child_res_id.name[1] = (*dchildp)->d_generation;
1231 inode = (*dchildp)->d_inode;
1233 inode = igrab(inode);
1237 child_res_id.name[0] = inode->i_ino;
1238 child_res_id.name[1] = inode->i_generation;
1243 cleanup_phase = 2; /* child dentry */
1245 /* Step 3: Lock parent and child in resource order. If child doesn't
1246 * exist, we still have to lock the parent and re-lookup. */
1247 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1248 &parent_policy, &child_res_id, child_lockh,
1249 child_mode, &child_policy);
1253 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1254 cleanup_phase = 4; /* child lock */
1256 cleanup_phase = 3; /* parent lock */
1258 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1259 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1260 parent_mode, &child_res_id, child_lockh,
1261 dchildp, child_mode, &child_policy,
1262 name, namelen, &parent_res_id);
1272 switch (cleanup_phase) {
1274 ldlm_lock_decref(child_lockh, child_mode);
1276 ldlm_lock_decref(parent_lockh, parent_mode);
1281 if (parent_lockh[1].cookie)
1282 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1291 void mds_reconstruct_generic(struct ptlrpc_request *req)
1293 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1295 mds_req_from_mcd(req, med->med_mcd);
1298 int mds_create_local_dentry(struct mds_update_record *rec,
1299 struct obd_device *obd)
1301 struct mds_obd *mds = &obd->u.mds;
1302 struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1303 int fidlen = 0, rc, cleanup_phase = 0;
1304 struct dentry *new_child = NULL;
1305 char *fidname = rec->ur_name;
1306 struct dentry *child = NULL;
1307 struct lustre_handle lockh;
1311 down(&fids_dir->i_sem);
1312 fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1313 CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1314 fidname, (unsigned) rec->ur_fid1->id,
1315 (unsigned) rec->ur_fid1->generation);
1317 new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1318 up(&fids_dir->i_sem);
1319 if (IS_ERR(new_child)) {
1320 CERROR("can't lookup %s: %d\n", fidname,
1321 (int) PTR_ERR(new_child));
1322 GOTO(cleanup, rc = PTR_ERR(new_child));
1326 if (new_child->d_inode != NULL) {
1327 /* nice. we've already have local dentry! */
1328 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1329 (unsigned) new_child->d_inode->i_ino,
1330 (unsigned) new_child->d_inode->i_generation);
1331 rec->ur_fid1->id = fids_dir->i_ino;
1332 rec->ur_fid1->generation = fids_dir->i_generation;
1333 rec->ur_namelen = fidlen + 1;
1334 GOTO(cleanup, rc = 0);
1337 /* new, local dentry will be added soon. we need no aliases here */
1340 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
1341 &lockh, NULL, 0, MDS_INODELOCK_UPDATE);
1342 if (IS_ERR(child)) {
1343 CERROR("can't get victim\n");
1344 GOTO(cleanup, rc = PTR_ERR(child));
1348 handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1350 GOTO(cleanup, rc = PTR_ERR(handle));
1352 rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1353 rec->ur_fid1->id, rec->ur_fid1->generation,
1356 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1357 (unsigned long) child->d_inode->i_ino,
1358 (unsigned long) child->d_inode->i_generation, rc);
1360 if (S_ISDIR(child->d_inode->i_mode)) {
1361 fids_dir->i_nlink++;
1362 mark_inode_dirty(fids_dir);
1364 mark_inode_dirty(child->d_inode);
1366 fsfilt_commit(obd, fids_dir, handle, 0);
1368 rec->ur_fid1->id = fids_dir->i_ino;
1369 rec->ur_fid1->generation = fids_dir->i_generation;
1370 rec->ur_namelen = fidlen + 1;
1373 switch(cleanup_phase) {
1375 ldlm_lock_decref(&lockh, LCK_EX);
1385 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1386 struct ptlrpc_request *slave)
1388 void *cookie, *cookie2;
1389 struct mds_body *body2;
1390 struct mds_body *body;
1394 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1395 LASSERT(body != NULL);
1397 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1398 LASSERT(body2 != NULL);
1400 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1404 memcpy(body2, body, sizeof(*body));
1405 body2->valid &= ~OBD_MD_FLCOOKIE;
1407 if (!(body->valid & OBD_MD_FLEASIZE))
1410 if (body->eadatasize == 0) {
1411 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1415 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1417 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1418 LASSERT(ea != NULL);
1420 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1421 LASSERT(ea2 != NULL);
1423 memcpy(ea2, ea, body->eadatasize);
1425 if (body->valid & OBD_MD_FLCOOKIE) {
1426 LASSERT(master->rq_repmsg->buflens[2] >=
1427 slave->rq_repmsg->buflens[2]);
1428 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1429 slave->rq_repmsg->buflens[2]);
1430 LASSERT(cookie != NULL);
1432 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1433 master->rq_repmsg->buflens[2]);
1434 LASSERT(cookie2 != NULL);
1435 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1436 body2->valid |= OBD_MD_FLCOOKIE;
1441 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1442 struct ptlrpc_request *req,
1443 struct lustre_handle *parent_lockh,
1444 struct dentry *dparent,
1445 struct lustre_handle *child_lockh,
1446 struct dentry *dchild)
1448 struct mds_obd *mds = mds_req2mds(req);
1449 struct mdc_op_data op_data;
1450 int rc = 0, cleanup_phase = 0;
1451 struct ptlrpc_request *request = NULL;
1454 LASSERT(offset == 0 || offset == 2);
1456 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1457 rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1458 (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1460 /* time to drop i_nlink on remote MDS */
1461 op_data.fid1.mds = dchild->d_mdsnum;
1462 op_data.fid1.id = dchild->d_inum;
1463 op_data.fid1.generation = dchild->d_generation;
1464 op_data.create_mode = rec->ur_mode;
1465 op_data.namelen = 0;
1466 op_data.name = NULL;
1467 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1470 mds_copy_unlink_reply(req, request);
1471 ptlrpc_req_finished(request);
1474 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1475 req->rq_status = rc;
1478 if (parent_lockh[1].cookie != 0)
1479 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1481 ldlm_lock_decref(child_lockh, LCK_EX);
1483 ldlm_lock_decref(parent_lockh, LCK_PW);
1485 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1492 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1493 struct ptlrpc_request *req,
1494 struct lustre_handle *lh)
1496 struct dentry *dparent, *dchild;
1497 struct mds_obd *mds = mds_req2mds(req);
1498 struct obd_device *obd = req->rq_export->exp_obd;
1499 struct mds_body *body = NULL;
1500 struct inode *child_inode;
1501 struct lustre_handle parent_lockh[2], child_lockh, child_reuse_lockh;
1502 char fidname[LL_FID_NAMELEN];
1503 void *handle = NULL;
1504 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1505 int unlink_by_fid = 0;
1508 LASSERT(offset == 0 || offset == 2);
1510 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1511 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1513 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1515 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1516 GOTO(cleanup, rc = -ENOENT);
1518 if (rec->ur_namelen == 1) {
1519 /* this is request to drop i_nlink on local inode */
1521 rec->ur_name = fidname;
1522 rc = mds_create_local_dentry(rec, obd);
1525 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1526 parent_lockh, &dparent, LCK_PW,
1527 MDS_INODELOCK_UPDATE,
1528 rec->ur_name, rec->ur_namelen,
1529 &child_lockh, &dchild, LCK_EX,
1530 MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE);
1534 if (dchild->d_flags & DCACHE_CROSS_REF) {
1535 /* we should have parent lock only here */
1536 LASSERT(unlink_by_fid == 0);
1537 LASSERT(dchild->d_mdsnum != mds->mds_num);
1538 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1539 dparent, &child_lockh, dchild);
1543 cleanup_phase = 1; /* dchild, dparent, locks */
1546 child_inode = dchild->d_inode;
1547 if (child_inode == NULL) {
1548 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1549 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1550 GOTO(cleanup, rc = -ENOENT);
1553 cleanup_phase = 2; /* dchild has a lock */
1555 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1556 * reuse (see bug 2029). */
1557 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1561 cleanup_phase = 3; /* child inum lock */
1563 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1565 /* ldlm_reply in buf[0] if called via intent */
1569 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1570 LASSERT(body != NULL);
1572 /* If this is the last reference to this inode, get the OBD EA
1573 * data first so the client can destroy OST objects.
1574 * we only do the object removal if no open files remain.
1575 * Nobody can get at this name anymore because of the locks so
1576 * we make decisions here as to whether to remove the inode */
1577 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1578 mds_open_orphan_count(child_inode) == 0) {
1579 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1580 mds_pack_inode2body(obd, body, child_inode);
1581 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1583 if (!(body->valid & OBD_MD_FLEASIZE)) {
1584 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1585 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1591 /* We have to do these checks ourselves, in case we are making an
1592 * orphan. The client tells us whether rmdir() or unlink() was called,
1593 * so we need to return appropriate errors (bug 72).
1595 * We don't have to check permissions, because vfs_rename (called from
1596 * mds_open_unlink_rename) also calls may_delete. */
1597 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1598 if (!S_ISDIR(child_inode->i_mode))
1599 GOTO(cleanup, rc = -ENOTDIR);
1601 if (S_ISDIR(child_inode->i_mode))
1602 GOTO(cleanup, rc = -EISDIR);
1605 if (child_inode->i_nlink == (S_ISDIR(child_inode->i_mode) ? 2 : 1) &&
1606 mds_open_orphan_count(child_inode) > 0) {
1607 rc = mds_open_unlink_rename(rec, obd, dparent, dchild, &handle);
1608 cleanup_phase = 4; /* transaction */
1612 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1613 switch (child_inode->i_mode & S_IFMT) {
1615 /* Drop any lingering child directories before we start our
1616 * transaction, to avoid doing multiple inode dirty/delete
1617 * in our compound transaction (bug 1321). */
1618 shrink_dcache_parent(dchild);
1619 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1622 GOTO(cleanup, rc = PTR_ERR(handle));
1623 cleanup_phase = 4; /* transaction */
1624 rc = vfs_rmdir(dparent->d_inode, dchild);
1627 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1628 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1630 handle = fsfilt_start_log(obd, dparent->d_inode,
1631 FSFILT_OP_UNLINK, NULL,
1632 le32_to_cpu(lmm->lmm_stripe_count));
1634 GOTO(cleanup, rc = PTR_ERR(handle));
1636 cleanup_phase = 4; /* transaction */
1637 rc = vfs_unlink(dparent->d_inode, dchild);
1639 if (!rc && log_unlink)
1640 if (mds_log_op_unlink(obd, child_inode,
1641 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1642 req->rq_repmsg->buflens[offset + 1],
1643 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1644 req->rq_repmsg->buflens[offset + 2]) > 0)
1645 body->valid |= OBD_MD_FLCOOKIE;
1653 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1656 GOTO(cleanup, rc = PTR_ERR(handle));
1657 cleanup_phase = 4; /* transaction */
1658 rc = vfs_unlink(dparent->d_inode, dchild);
1661 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1664 GOTO(cleanup, rc = -EINVAL);
1672 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1673 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1674 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1676 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1678 CERROR("error on parent setattr: rc = %d\n", err);
1681 switch(cleanup_phase) {
1683 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1686 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1687 "unlinked", 0, NULL);
1688 case 3: /* child ino-reuse lock */
1689 if (rc && body != NULL) {
1690 // Don't unlink the OST objects if the MDS unlink failed
1694 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1696 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1697 case 2: /* child lock */
1698 ldlm_lock_decref(&child_lockh, LCK_EX);
1699 case 1: /* child and parent dentry, parent lock */
1701 if (parent_lockh[1].cookie != 0)
1702 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1705 ldlm_lock_decref(parent_lockh, LCK_PW);
1707 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1714 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1717 req->rq_status = rc;
1722 * to service requests from remote MDS to increment i_nlink
1724 static int mds_reint_link_acquire(struct mds_update_record *rec,
1725 int offset, struct ptlrpc_request *req,
1726 struct lustre_handle *lh)
1728 struct obd_device *obd = req->rq_export->exp_obd;
1729 struct ldlm_res_id src_res_id = { .name = {0} };
1730 struct lustre_handle *handle = NULL, src_lockh;
1731 struct mds_obd *mds = mds_req2mds(req);
1732 int rc = 0, cleanup_phase = 0;
1733 struct dentry *de_src = NULL;
1734 ldlm_policy_data_t policy;
1738 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1739 obd->obd_name, (unsigned) rec->ur_fid1->mds,
1740 (unsigned) rec->ur_fid1->id,
1741 (unsigned) rec->ur_fid1->generation);
1743 /* Step 1: Lookup the source inode and target directory by FID */
1744 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1746 GOTO(cleanup, rc = PTR_ERR(de_src));
1747 cleanup_phase = 1; /* source dentry */
1749 src_res_id.name[0] = de_src->d_inode->i_ino;
1750 src_res_id.name[1] = de_src->d_inode->i_generation;
1751 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1753 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1754 src_res_id, LDLM_IBITS, &policy,
1755 LCK_EX, &flags, mds_blocking_ast,
1756 ldlm_completion_ast, NULL, NULL,
1757 NULL, 0, NULL, &src_lockh);
1759 GOTO(cleanup, rc = -ENOLCK);
1760 cleanup_phase = 2; /* lock */
1762 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1764 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1765 if (IS_ERR(handle)) {
1766 rc = PTR_ERR(handle);
1769 de_src->d_inode->i_nlink++;
1770 mark_inode_dirty(de_src->d_inode);
1773 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1774 handle, req, rc, 0);
1776 switch (cleanup_phase) {
1779 ldlm_lock_decref(&src_lockh, LCK_EX);
1781 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1787 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1790 req->rq_status = rc;
1795 * request to link to foreign inode:
1796 * - acquire i_nlinks on this inode
1799 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1800 int offset, struct ptlrpc_request *req,
1801 struct lustre_handle *lh)
1803 struct lustre_handle *handle = NULL, tgt_dir_lockh[2];
1804 struct obd_device *obd = req->rq_export->exp_obd;
1805 struct dentry *de_tgt_dir = NULL;
1806 struct mds_obd *mds = mds_req2mds(req);
1807 int rc = 0, cleanup_phase = 0;
1808 struct mdc_op_data op_data;
1809 struct ptlrpc_request *request = NULL;
1812 #define fmt "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1813 DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1814 (unsigned) rec->ur_fid2->mds,
1815 (unsigned) rec->ur_fid2->id,
1816 (unsigned) rec->ur_fid2->generation,
1817 rec->ur_namelen - 1, rec->ur_name,
1818 (unsigned) rec->ur_fid1->mds,
1819 (unsigned) rec->ur_fid1->id,
1820 (unsigned)rec->ur_fid1->generation);
1822 de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
1823 tgt_dir_lockh, rec->ur_name,
1824 rec->ur_namelen - 1,
1825 MDS_INODELOCK_UPDATE);
1826 if (IS_ERR(de_tgt_dir))
1827 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1830 op_data.fid1 = *(rec->ur_fid1);
1831 op_data.namelen = 0;
1832 op_data.name = NULL;
1833 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
1837 ptlrpc_req_finished(request);
1839 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
1841 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1842 if (IS_ERR(handle)) {
1843 rc = PTR_ERR(handle);
1847 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
1848 rec->ur_namelen - 1, rec->ur_fid1->id,
1849 rec->ur_fid1->generation, rec->ur_fid1->mds);
1853 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1854 handle, req, rc, 0);
1857 switch (cleanup_phase) {
1860 /* FIXME: drop i_nlink on remote inode here */
1861 CERROR("MUST drop drop i_nlink here\n");
1866 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
1868 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
1871 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
1873 ptlrpc_save_lock(req, tgt_dir_lockh + 1, LCK_CW);
1879 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1882 req->rq_status = rc;
1886 static int mds_reint_link(struct mds_update_record *rec, int offset,
1887 struct ptlrpc_request *req,
1888 struct lustre_handle *lh)
1890 struct obd_device *obd = req->rq_export->exp_obd;
1891 struct dentry *de_src = NULL;
1892 struct dentry *de_tgt_dir = NULL;
1893 struct dentry *dchild = NULL;
1894 struct mds_obd *mds = mds_req2mds(req);
1895 struct lustre_handle *handle = NULL, tgt_dir_lockh[2], src_lockh;
1896 struct ldlm_res_id src_res_id = { .name = {0} };
1897 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1898 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1899 ldlm_policy_data_t tgt_dir_policy =
1900 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1902 int rc = 0, cleanup_phase = 0;
1905 LASSERT(offset == 0);
1907 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1908 rec->ur_fid1->id, rec->ur_fid1->generation,
1909 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1911 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1913 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1914 GOTO(cleanup, rc = -ENOENT);
1916 if (rec->ur_fid1->mds != mds->mds_num) {
1917 rc = mds_reint_link_to_remote(rec, offset, req, lh);
1921 if (rec->ur_namelen == 1) {
1922 rc = mds_reint_link_acquire(rec, offset, req, lh);
1926 /* Step 1: Lookup the source inode and target directory by FID */
1927 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1929 GOTO(cleanup, rc = PTR_ERR(de_src));
1931 cleanup_phase = 1; /* source dentry */
1933 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1934 if (IS_ERR(de_tgt_dir))
1935 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
1937 cleanup_phase = 2; /* target directory dentry */
1939 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
1940 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1941 de_src->d_inode->i_ino);
1943 /* Step 2: Take the two locks */
1944 src_res_id.name[0] = de_src->d_inode->i_ino;
1945 src_res_id.name[1] = de_src->d_inode->i_generation;
1946 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1947 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1949 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
1951 /* Get a temp lock on just ino, gen to flush client cache */
1952 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1953 tgt_dir_res_id, LDLM_IBITS, &src_policy,
1954 LCK_CW, &flags, mds_blocking_ast,
1955 ldlm_completion_ast, NULL, NULL,
1956 NULL, 0, NULL, tgt_dir_lockh + 1);
1958 GOTO(cleanup, rc = -ENOLCK);
1960 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
1961 rec->ur_namelen - 1);
1962 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1963 de_tgt_dir->d_inode->i_ino,
1964 de_tgt_dir->d_inode->i_generation,
1965 tgt_dir_res_id.name[2]);
1968 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1970 &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
1975 cleanup_phase = 3; /* locks */
1977 /* Step 3: Lookup the child */
1978 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1979 if (IS_ERR(dchild)) {
1980 rc = PTR_ERR(dchild);
1981 if (rc != -EPERM && rc != -EACCES)
1982 CERROR("child lookup error %d\n", rc);
1986 cleanup_phase = 4; /* child dentry */
1988 if (dchild->d_inode) {
1989 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1990 de_tgt_dir->d_inode->i_ino, rec->ur_name);
1995 /* Step 4: Do it. */
1996 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1998 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1999 if (IS_ERR(handle)) {
2000 rc = PTR_ERR(handle);
2004 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2005 if (rc && rc != -EPERM && rc != -EACCES)
2006 CERROR("vfs_link error %d\n", rc);
2008 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2009 handle, req, rc, 0);
2012 switch (cleanup_phase) {
2013 case 4: /* child dentry */
2017 ldlm_lock_decref(&src_lockh, LCK_EX);
2018 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2020 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2021 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2023 case 2: /* target dentry */
2025 if (tgt_dir_lockh[1].cookie)
2026 ldlm_lock_decref(tgt_dir_lockh + 1, LCK_CW);
2030 case 1: /* source dentry */
2035 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2038 req->rq_status = rc;
2043 * add a hard link in the PENDING directory, only used by rename()
2045 static int mds_add_link_orphan(struct mds_update_record *rec,
2046 struct obd_device *obd,
2047 struct dentry *dentry)
2049 struct mds_obd *mds = &obd->u.mds;
2050 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2051 struct dentry *pending_child;
2052 char fidname[LL_FID_NAMELEN];
2056 LASSERT(dentry->d_inode);
2057 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2059 down(&pending_dir->i_sem);
2060 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2061 dentry->d_inode->i_generation);
2063 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2064 mds_open_orphan_count(dentry->d_inode),
2065 rec->ur_name, fidname);
2067 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2068 if (IS_ERR(pending_child))
2069 GOTO(out_lock, rc = PTR_ERR(pending_child));
2071 if (pending_child->d_inode != NULL) {
2072 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2073 LASSERT(pending_child->d_inode == dentry->d_inode);
2074 GOTO(out_dput, rc = 0);
2078 rc = vfs_link(dentry, pending_dir, pending_child);
2081 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2084 mds_inode_set_orphan(dentry->d_inode);
2086 l_dput(pending_child);
2088 up(&pending_dir->i_sem);
2092 /* The idea here is that we need to get four locks in the end:
2093 * one on each parent directory, one on each child. We need to take
2094 * these locks in some kind of order (to avoid deadlocks), and the order
2095 * I selected is "increasing resource number" order. We need to look up
2096 * the children, however, before we know what the resource number(s) are.
2097 * Thus the following plan:
2099 * 1,2. Look up the parents
2100 * 3,4. Look up the children
2101 * 5. Take locks on the parents and children, in order
2102 * 6. Verify that the children haven't changed since they were looked up
2104 * If there was a race and the children changed since they were first looked
2105 * up, it is possible that mds_verify_child() will be able to just grab the
2106 * lock on the new child resource (if it has a higher resource than any other)
2107 * but we need to compare against not only its parent, but also against the
2108 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2110 * We need the fancy igrab() on the child inodes because we aren't holding a
2111 * lock on the parent after the lookup is done, so dentry->d_inode may change
2112 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2114 static int mds_get_parents_children_locked(struct obd_device *obd,
2115 struct mds_obd *mds,
2116 struct ll_fid *p1_fid,
2117 struct dentry **de_srcdirp,
2118 struct ll_fid *p2_fid,
2119 struct dentry **de_tgtdirp,
2121 const char *old_name, int old_len,
2122 struct dentry **de_oldp,
2123 const char *new_name, int new_len,
2124 struct dentry **de_newp,
2125 struct lustre_handle *dlm_handles,
2128 struct ldlm_res_id p1_res_id = { .name = {0} };
2129 struct ldlm_res_id p2_res_id = { .name = {0} };
2130 struct ldlm_res_id c1_res_id = { .name = {0} };
2131 struct ldlm_res_id c2_res_id = { .name = {0} };
2132 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2133 /* Only dentry should change, but the inode itself would be
2135 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2136 /* If something is going to be replaced, both dentry and inode locks are
2138 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2139 MDS_INODELOCK_UPDATE}};
2140 struct ldlm_res_id *maxres_src, *maxres_tgt;
2141 struct inode *inode;
2142 int rc = 0, cleanup_phase = 0;
2145 /* Step 1: Lookup the source directory */
2146 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2147 if (IS_ERR(*de_srcdirp))
2148 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2150 cleanup_phase = 1; /* source directory dentry */
2152 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2153 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2155 /* Step 2: Lookup the target directory */
2156 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2157 *de_tgtdirp = dget(*de_srcdirp);
2159 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2160 if (IS_ERR(*de_tgtdirp))
2161 GOTO(cleanup, rc = PTR_ERR(*de_tgtdirp));
2164 cleanup_phase = 2; /* target directory dentry */
2166 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2167 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2170 dlm_handles[5].cookie = 0;
2171 dlm_handles[6].cookie = 0;
2172 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2173 /* Get a temp lock on just ino, gen to flush client cache */
2174 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2175 LCK_CW, &p_policy, &p2_res_id,
2176 &(dlm_handles[6]),LCK_CW,&p_policy);
2180 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2181 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2182 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2183 (*de_srcdirp)->d_inode->i_ino,
2184 (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2185 (*de_tgtdirp)->d_inode->i_ino,
2186 (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2190 /* Step 3: Lookup the source child entry */
2191 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2192 if (IS_ERR(*de_oldp)) {
2193 rc = PTR_ERR(*de_oldp);
2194 CERROR("old child lookup error (%*s): %d\n",
2195 old_len - 1, old_name, rc);
2199 cleanup_phase = 3; /* original name dentry */
2201 inode = (*de_oldp)->d_inode;
2202 if (inode != NULL) {
2203 inode = igrab(inode);
2205 GOTO(cleanup, rc = -ENOENT);
2207 c1_res_id.name[0] = inode->i_ino;
2208 c1_res_id.name[1] = inode->i_generation;
2210 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2211 c1_res_id.name[0] = (*de_oldp)->d_inum;
2212 c1_res_id.name[1] = (*de_oldp)->d_generation;
2215 /* Step 4: Lookup the target child entry */
2216 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2217 if (IS_ERR(*de_newp)) {
2218 rc = PTR_ERR(*de_newp);
2219 CERROR("new child lookup error (%*s): %d\n",
2220 old_len - 1, old_name, rc);
2224 cleanup_phase = 4; /* target dentry */
2226 inode = (*de_newp)->d_inode;
2227 if (inode != NULL) {
2228 inode = igrab(inode);
2232 c2_res_id.name[0] = inode->i_ino;
2233 c2_res_id.name[1] = inode->i_generation;
2235 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2236 c2_res_id.name[0] = (*de_newp)->d_inum;
2237 c2_res_id.name[1] = (*de_newp)->d_generation;
2241 /* Step 5: Take locks on the parents and child(ren) */
2242 maxres_src = &p1_res_id;
2243 maxres_tgt = &p2_res_id;
2244 cleanup_phase = 4; /* target dentry */
2246 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2247 maxres_src = &c1_res_id;
2248 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2249 maxres_tgt = &c2_res_id;
2251 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2253 &p2_res_id, &dlm_handles[1], parent_mode,
2255 &c1_res_id, &dlm_handles[2], child_mode,
2257 &c2_res_id, &dlm_handles[3], child_mode,
2262 cleanup_phase = 6; /* parent and child(ren) locks */
2264 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2265 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2266 parent_mode, &c1_res_id, &dlm_handles[2],
2267 de_oldp, child_mode, &c1_policy, old_name,old_len,
2270 if (c2_res_id.name[0] != 0)
2271 ldlm_lock_decref(&dlm_handles[3], child_mode);
2272 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2279 if (!DENTRY_VALID(*de_oldp))
2280 GOTO(cleanup, rc = -ENOENT);
2282 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2283 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2284 parent_mode, &c2_res_id, &dlm_handles[3],
2285 de_newp, child_mode, &c2_policy, new_name,
2286 new_len, maxres_src);
2288 ldlm_lock_decref(&dlm_handles[2], child_mode);
2289 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2299 switch (cleanup_phase) {
2300 case 6: /* child lock(s) */
2301 if (c2_res_id.name[0] != 0)
2302 ldlm_lock_decref(&dlm_handles[3], child_mode);
2303 if (c1_res_id.name[0] != 0)
2304 ldlm_lock_decref(&dlm_handles[2], child_mode);
2305 case 5: /* parent locks */
2306 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2307 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2308 case 4: /* target dentry */
2310 case 3: /* source dentry */
2312 case 2: /* target directory dentry */
2313 l_dput(*de_tgtdirp);
2314 case 1: /* source directry dentry */
2315 l_dput(*de_srcdirp);
2322 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2323 int offset, struct ptlrpc_request *req)
2325 struct obd_device *obd = req->rq_export->exp_obd;
2326 struct dentry *de_srcdir = NULL;
2327 struct dentry *de_new = NULL;
2328 struct mds_obd *mds = mds_req2mds(req);
2329 struct lustre_handle parent_lockh[2];
2330 struct lustre_handle child_lockh;
2331 int cleanup_phase = 0;
2332 void *handle = NULL;
2336 /* another MDS executing rename operation has asked us
2337 * to create target name. such a creation should destroy
2338 * existing target name */
2340 CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2341 obd->obd_name, rec->ur_tgt,
2342 (unsigned long) rec->ur_fid1->mds,
2343 (unsigned long) rec->ur_fid1->id,
2344 (unsigned long) rec->ur_fid1->generation);
2346 /* first, lookup the target */
2347 child_lockh.cookie = 0;
2348 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2349 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2350 rec->ur_tgt, rec->ur_tgtlen,
2351 &child_lockh, &de_new, LCK_EX,
2352 MDS_INODELOCK_LOOKUP);
2359 LASSERT(de_srcdir->d_inode);
2362 if (de_new->d_inode) {
2363 /* name exists and points to local inode
2364 * try to unlink this name and create new one */
2365 CERROR("%s: %s points to local inode %lu/%lu\n",
2366 obd->obd_name, rec->ur_tgt,
2367 (unsigned long) de_new->d_inode->i_ino,
2368 (unsigned long) de_new->d_inode->i_generation);
2369 handle = fsfilt_start(obd, de_srcdir->d_inode,
2370 FSFILT_OP_RENAME, NULL);
2372 GOTO(cleanup, rc = PTR_ERR(handle));
2373 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2376 } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2377 /* name exists adn points to remove inode */
2378 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2379 obd->obd_name, rec->ur_tgt,
2380 (unsigned long) de_new->d_mdsnum,
2381 (unsigned long) de_new->d_inum,
2382 (unsigned long) de_new->d_generation);
2384 /* name doesn't exist. the simplest case */
2385 handle = fsfilt_start(obd, de_srcdir->d_inode,
2386 FSFILT_OP_LINK, NULL);
2388 GOTO(cleanup, rc = PTR_ERR(handle));
2392 rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2393 rec->ur_tgtlen - 1, rec->ur_fid1->id,
2394 rec->ur_fid1->generation, rec->ur_fid1->mds);
2396 CERROR("add_dir_entry() returned error %d\n", rc);
2399 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2400 handle, req, rc, 0);
2401 switch(cleanup_phase) {
2405 if (parent_lockh[1].cookie != 0)
2406 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2408 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2409 if (child_lockh.cookie != 0)
2410 ldlm_lock_decref(&child_lockh, LCK_EX);
2418 req->rq_status = rc;
2423 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2424 struct ptlrpc_request *req)
2426 struct obd_device *obd = req->rq_export->exp_obd;
2427 struct ptlrpc_request *req2 = NULL;
2428 struct dentry *de_srcdir = NULL;
2429 struct dentry *de_old = NULL;
2430 struct mds_obd *mds = mds_req2mds(req);
2431 struct lustre_handle parent_lockh[2];
2432 struct lustre_handle child_lockh;
2433 struct mdc_op_data opdata;
2434 int cleanup_phase = 0;
2435 void *handle = NULL;
2439 CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2440 obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2441 memset(&opdata, 0, sizeof(opdata));
2443 child_lockh.cookie = 0;
2444 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2445 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2446 rec->ur_name, rec->ur_namelen,
2447 &child_lockh, &de_old, LCK_EX,
2448 MDS_INODELOCK_LOOKUP);
2451 LASSERT(de_srcdir->d_inode);
2454 /* we already know the target should be created on another MDS
2455 * so, we have to request that MDS to do it */
2457 /* prepare source fid */
2458 if (de_old->d_flags & DCACHE_CROSS_REF) {
2459 LASSERT(de_old->d_inode == NULL);
2460 CDEBUG(D_OTHER, "request to move remote name\n");
2461 opdata.fid1.mds = de_old->d_mdsnum;
2462 opdata.fid1.id = de_old->d_inum;
2463 opdata.fid1.generation = de_old->d_generation;
2464 } else if (de_old->d_inode == NULL) {
2465 /* oh, source doesn't exist */
2466 GOTO(cleanup, rc = -ENOENT);
2468 LASSERT(de_old->d_inode != NULL);
2469 CDEBUG(D_OTHER, "request to move local name\n");
2470 opdata.fid1.mds = mds->mds_num;
2471 opdata.fid1.id = de_old->d_inode->i_ino;
2472 opdata.fid1.generation = de_old->d_inode->i_generation;
2475 opdata.fid2 = *(rec->ur_fid2);
2476 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2477 rec->ur_tgtlen - 1, &req2);
2482 handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2484 GOTO(cleanup, rc = PTR_ERR(handle));
2485 rc = fsfilt_del_dir_entry(obd, de_old);
2490 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2491 handle, req, rc, 0);
2493 ptlrpc_req_finished(req2);
2496 if (parent_lockh[1].cookie != 0)
2497 ldlm_lock_decref(&parent_lockh[1], LCK_CW);
2499 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2500 if (child_lockh.cookie != 0)
2501 ldlm_lock_decref(&child_lockh, LCK_EX);
2506 req->rq_status = rc;
2511 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2512 struct ptlrpc_request *req,
2513 struct lustre_handle *lockh)
2515 struct obd_device *obd = req->rq_export->exp_obd;
2516 struct dentry *de_srcdir = NULL;
2517 struct dentry *de_tgtdir = NULL;
2518 struct dentry *de_old = NULL;
2519 struct dentry *de_new = NULL;
2520 struct mds_obd *mds = mds_req2mds(req);
2521 struct lustre_handle dlm_handles[7];
2522 struct mds_body *body = NULL;
2523 int rc = 0, lock_count = 3;
2524 int cleanup_phase = 0;
2525 void *handle = NULL;
2528 LASSERT(offset == 0);
2530 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2531 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2532 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2534 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2536 if (rec->ur_namelen == 1) {
2537 rc = mds_reint_rename_create_name(rec, offset, req);
2541 if (rec->ur_fid2->mds != mds->mds_num) {
2542 rc = mds_reint_rename_to_remote(rec, offset, req);
2546 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2547 rec->ur_fid2, &de_tgtdir, LCK_PW,
2548 rec->ur_name, rec->ur_namelen,
2549 &de_old, rec->ur_tgt,
2550 rec->ur_tgtlen, &de_new,
2551 dlm_handles, LCK_EX);
2555 cleanup_phase = 1; /* parent(s), children, locks */
2557 if (de_new->d_inode)
2560 /* sanity check for src inode */
2561 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2562 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2563 GOTO(cleanup, rc = -EINVAL);
2565 /* sanity check for dest inode */
2566 if (de_new->d_inode &&
2567 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2568 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2569 GOTO(cleanup, rc = -EINVAL);
2571 if (de_old->d_inode == de_new->d_inode) {
2572 GOTO(cleanup, rc = 0);
2575 /* if we are about to remove the target at first, pass the EA of
2576 * that inode to client to perform and cleanup on OST */
2577 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2578 LASSERT(body != NULL);
2580 if (de_new->d_inode &&
2581 S_ISREG(de_new->d_inode->i_mode) &&
2582 de_new->d_inode->i_nlink == 1 &&
2583 mds_open_orphan_count(de_new->d_inode) == 0) {
2584 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2585 mds_pack_inode2body(obd, body, de_new->d_inode);
2586 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2587 if (!(body->valid & OBD_MD_FLEASIZE)) {
2588 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2589 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2591 /* XXX need log unlink? */
2595 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2596 de_srcdir->d_inode->i_sb);
2598 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2600 GOTO(cleanup, rc = PTR_ERR(handle));
2602 /* FIXME need adjust the journal block count? */
2603 /* if the target should be moved to PENDING, we at first increase the
2604 * link and later vfs_rename() will decrease the link count again */
2605 if (de_new->d_inode &&
2606 S_ISREG(de_new->d_inode->i_mode) &&
2607 de_new->d_inode->i_nlink == 1 &&
2608 mds_open_orphan_count(de_new->d_inode) > 0) {
2609 rc = mds_add_link_orphan(rec, obd, de_new);
2615 de_old->d_fsdata = req;
2616 de_new->d_fsdata = req;
2617 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2622 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2623 handle, req, rc, 0);
2624 switch (cleanup_phase) {
2627 if (dlm_handles[5].cookie != 0)
2628 ldlm_lock_decref(&(dlm_handles[5]), LCK_CW);
2629 if (dlm_handles[6].cookie != 0)
2630 ldlm_lock_decref(&(dlm_handles[6]), LCK_CW);
2633 if (lock_count == 4)
2634 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2635 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2636 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2637 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2639 if (lock_count == 4)
2640 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2641 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2642 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2643 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2652 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2655 req->rq_status = rc;
2659 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2660 struct ptlrpc_request *, struct lustre_handle *);
2662 static mds_reinter reinters[REINT_MAX + 1] = {
2663 [REINT_SETATTR] mds_reint_setattr,
2664 [REINT_CREATE] mds_reint_create,
2665 [REINT_LINK] mds_reint_link,
2666 [REINT_UNLINK] mds_reint_unlink,
2667 [REINT_RENAME] mds_reint_rename,
2668 [REINT_OPEN] mds_open
2671 int mds_reint_rec(struct mds_update_record *rec, int offset,
2672 struct ptlrpc_request *req, struct lustre_handle *lockh)
2674 struct obd_device *obd = req->rq_export->exp_obd;
2675 struct obd_run_ctxt saved;
2678 /* checked by unpacker */
2679 LASSERT(rec->ur_opcode <= REINT_MAX &&
2680 reinters[rec->ur_opcode] != NULL);
2682 push_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);
2683 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2684 pop_ctxt(&saved, &obd->obd_ctxt, &rec->ur_uc);