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) 2002-2005 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 the Lustre file system, http://www.lustre.org
13 * Lustre is a trademark of Cluster File Systems, Inc.
15 * You may have signed or agreed to another license before downloading
16 * this software. If so, you are bound by the terms and conditions
17 * of that agreement, and the following does not apply to you. See the
18 * LICENSE file included with this distribution for more information.
20 * If you did not agree to a different license, then this copy of Lustre
21 * is open source software; you can redistribute it and/or modify it
22 * under the terms of version 2 of the GNU General Public License as
23 * published by the Free Software Foundation.
25 * In either case, Lustre is distributed in the hope that it will be
26 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
27 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
28 * license text for more details.
32 # define EXPORT_SYMTAB
34 #define DEBUG_SUBSYSTEM S_MDS
37 #include <obd_support.h>
38 #include <obd_class.h>
40 #include <lustre_lib.h>
41 #include <lustre/lustre_idl.h>
42 #include <lustre_mds.h>
43 #include <lustre_dlm.h>
44 #include <lustre_fsfilt.h>
45 #include <lustre_ucache.h>
47 #include "mds_internal.h"
49 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
52 obd_transno_commit_cb(obd, transno, error);
55 struct mds_logcancel_data {
56 struct lov_mds_md *mlcd_lmm;
60 struct llog_cookie mlcd_cookies[0];
64 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
65 void *cb_data, int error)
67 struct mds_logcancel_data *mlcd = cb_data;
68 struct lov_stripe_md *lsm = NULL;
69 struct llog_ctxt *ctxt;
72 obd_transno_commit_cb(obd, transno, error);
74 CDEBUG(D_RPCTRACE, "cancelling %d cookies\n",
75 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
77 rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
78 mlcd->mlcd_eadatalen);
80 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
81 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
84 ///* XXX 0 normally, SENDNOW for debug */);
85 rc = obd_checkmd(obd->u.mds.mds_osc_exp, obd->obd_self_export,
88 CERROR("Can not revalidate lsm %p \n", lsm);
90 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
91 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
92 sizeof(*mlcd->mlcd_cookies),
93 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
95 CERROR("error cancelling %d log cookies: rc %d\n",
96 (int)(mlcd->mlcd_cookielen /
97 sizeof(*mlcd->mlcd_cookies)), rc);
100 OBD_FREE(mlcd, mlcd->mlcd_size);
103 /* Assumes caller has already pushed us into the kernel context. */
104 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
105 struct ptlrpc_request *req, int rc, __u32 op_data,
108 struct mds_export_data *med = &req->rq_export->exp_mds_data;
109 struct mds_client_data *mcd = med->med_mcd;
110 struct obd_device *obd = req->rq_export->exp_obd;
112 __u64 transno, prev_transno;
114 int log_pri = D_RPCTRACE;
117 if (IS_ERR(handle)) {
122 /* if the export has already been failed, we have no last_rcvd slot */
123 if (req->rq_export->exp_failed || obd->obd_fail) {
124 CWARN("commit transaction for disconnected client %s: rc %d\n",
125 req->rq_export->exp_client_uuid.uuid, rc);
133 if (handle == NULL) {
134 /* if we're starting our own xaction, use our own inode */
135 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
136 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
137 if (IS_ERR(handle)) {
138 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
139 RETURN(PTR_ERR(handle));
143 off = med->med_lr_off;
145 transno = lustre_msg_get_transno(req->rq_reqmsg);
148 CERROR("%s: replay %s transno "LPU64" failed: rc %d\n",
150 libcfs_nid2str(req->rq_export->exp_connection->c_peer.nid),
154 } else if (transno == 0) {
155 spin_lock(&mds->mds_transno_lock);
156 transno = ++mds->mds_last_transno;
157 spin_unlock(&mds->mds_transno_lock);
159 spin_lock(&mds->mds_transno_lock);
160 if (transno > mds->mds_last_transno)
161 mds->mds_last_transno = transno;
162 spin_unlock(&mds->mds_transno_lock);
165 req->rq_transno = transno;
166 lustre_msg_set_transno(req->rq_repmsg, transno);
167 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
168 prev_transno = le64_to_cpu(mcd->mcd_last_close_transno);
169 mcd->mcd_last_close_transno = cpu_to_le64(transno);
170 mcd->mcd_last_close_xid = cpu_to_le64(req->rq_xid);
171 mcd->mcd_last_close_result = cpu_to_le32(rc);
172 mcd->mcd_last_close_data = cpu_to_le32(op_data);
174 prev_transno = le64_to_cpu(mcd->mcd_last_transno);
175 if (((lustre_msg_get_flags(req->rq_reqmsg) &
176 (MSG_RESENT | MSG_REPLAY)) == 0) ||
177 (transno > prev_transno)) {
178 mcd->mcd_last_transno = cpu_to_le64(transno);
179 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
180 mcd->mcd_last_result = cpu_to_le32(rc);
181 mcd->mcd_last_data = cpu_to_le32(op_data);
184 /* update the server data to not lose the greatest transno. Bug 11125 */
185 if ((transno == 0) && (prev_transno == mds->mds_last_transno))
186 mds_update_server_data(obd, 0);
189 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
192 struct obd_export *exp = req->rq_export;
195 force_sync = fsfilt_add_journal_cb(exp->exp_obd,transno,
196 handle, mds_commit_cb,
199 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
201 force_sync | exp->exp_need_sync);
203 mds_commit_cb(obd, transno, NULL, err);
212 DEBUG_REQ(log_pri, req,
213 "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
214 transno, rc, mcd->mcd_uuid, med->med_lr_idx, err);
216 err = mds_lov_write_objids(obd);
222 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
225 err = fsfilt_commit(obd, inode, handle, 0);
227 CERROR("error committing transaction: %d\n", err);
235 /* this gives the same functionality as the code between
236 * sys_chmod and inode_setattr
237 * chown_common and inode_setattr
238 * utimes and inode_setattr
240 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
242 time_t now = CURRENT_SECONDS;
243 struct iattr *attr = &rec->ur_iattr;
244 unsigned int ia_valid = attr->ia_valid;
248 if (ia_valid & ATTR_RAW)
249 attr->ia_valid &= ~ATTR_RAW;
251 if (!(ia_valid & ATTR_CTIME_SET))
252 LTIME_S(attr->ia_ctime) = now;
254 attr->ia_valid &= ~ATTR_CTIME_SET;
255 if (!(ia_valid & ATTR_ATIME_SET))
256 LTIME_S(attr->ia_atime) = now;
257 if (!(ia_valid & ATTR_MTIME_SET))
258 LTIME_S(attr->ia_mtime) = now;
260 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
261 RETURN((attr->ia_valid & ~ATTR_ATTR_FLAG) ? -EPERM : 0);
264 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
265 if (current->fsuid != inode->i_uid &&
266 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
270 if (ia_valid & ATTR_SIZE &&
271 /* NFSD hack for open(O_CREAT|O_TRUNC)=mknod+truncate (bug 5781) */
272 !(rec->ur_uc.luc_fsuid == inode->i_uid &&
273 ia_valid & MDS_OPEN_OWNEROVERRIDE)) {
274 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
278 if (ia_valid & (ATTR_UID | ATTR_GID)) {
281 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
283 if (attr->ia_uid == (uid_t) -1)
284 attr->ia_uid = inode->i_uid;
285 if (attr->ia_gid == (gid_t) -1)
286 attr->ia_gid = inode->i_gid;
287 if (!(ia_valid & ATTR_MODE))
288 attr->ia_mode = inode->i_mode;
290 * If the user or group of a non-directory has been
291 * changed by a non-root user, remove the setuid bit.
292 * 19981026 David C Niemi <niemi@tux.org>
294 * Changed this to apply to all users, including root,
295 * to avoid some races. This is the behavior we had in
296 * 2.0. The check for non-root was definitely wrong
297 * for 2.2 anyway, as it should have been using
298 * CAP_FSETID rather than fsuid -- 19990830 SD.
300 if ((inode->i_mode & S_ISUID) == S_ISUID &&
301 !S_ISDIR(inode->i_mode)) {
302 attr->ia_mode &= ~S_ISUID;
303 attr->ia_valid |= ATTR_MODE;
306 * Likewise, if the user or group of a non-directory
307 * has been changed by a non-root user, remove the
308 * setgid bit UNLESS there is no group execute bit
309 * (this would be a file marked for mandatory
310 * locking). 19981026 David C Niemi <niemi@tux.org>
312 * Removed the fsuid check (see the comment above) --
315 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
316 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
317 attr->ia_mode &= ~S_ISGID;
318 attr->ia_valid |= ATTR_MODE;
320 } else if (ia_valid & ATTR_MODE) {
321 int mode = attr->ia_mode;
323 if (attr->ia_mode == (umode_t)-1)
324 mode = inode->i_mode;
326 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
331 void mds_steal_ack_locks(struct ptlrpc_request *req)
333 struct obd_export *exp = req->rq_export;
334 struct list_head *tmp;
335 struct ptlrpc_reply_state *oldrep;
336 struct ptlrpc_service *svc;
339 /* CAVEAT EMPTOR: spinlock order */
340 spin_lock(&exp->exp_lock);
341 list_for_each (tmp, &exp->exp_outstanding_replies) {
342 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
344 if (oldrep->rs_xid != req->rq_xid)
347 if (lustre_msg_get_opc(oldrep->rs_msg) !=
348 lustre_msg_get_opc(req->rq_reqmsg))
349 CERROR ("Resent req xid "LPX64" has mismatched opc: "
350 "new %d old %d\n", req->rq_xid,
351 lustre_msg_get_opc(req->rq_reqmsg),
352 lustre_msg_get_opc(oldrep->rs_msg));
354 svc = oldrep->rs_service;
355 spin_lock (&svc->srv_lock);
357 list_del_init (&oldrep->rs_exp_list);
359 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
361 oldrep->rs_nlocks, oldrep,
362 oldrep->rs_xid, oldrep->rs_transno,
363 lustre_msg_get_opc(oldrep->rs_msg),
364 libcfs_nid2str(exp->exp_connection->c_peer.nid));
366 for (i = 0; i < oldrep->rs_nlocks; i++)
367 ptlrpc_save_lock(req,
368 &oldrep->rs_locks[i],
369 oldrep->rs_modes[i]);
370 oldrep->rs_nlocks = 0;
372 DEBUG_REQ(D_HA, req, "stole locks for");
373 ptlrpc_schedule_difficult_reply (oldrep);
375 spin_unlock (&svc->srv_lock);
378 spin_unlock(&exp->exp_lock);
380 EXPORT_SYMBOL(mds_steal_ack_locks);
381 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
383 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
384 req->rq_transno = le64_to_cpu(mcd->mcd_last_close_transno);
385 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
386 req->rq_status = le32_to_cpu(mcd->mcd_last_close_result);
387 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
389 req->rq_transno = le64_to_cpu(mcd->mcd_last_transno);
390 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
391 req->rq_status = le32_to_cpu(mcd->mcd_last_result);
392 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
394 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
395 req->rq_transno, req->rq_status);
397 mds_steal_ack_locks(req);
400 static void reconstruct_reint_setattr(struct mds_update_record *rec,
401 int offset, struct ptlrpc_request *req)
403 struct mds_export_data *med = &req->rq_export->exp_mds_data;
404 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
406 struct mds_body *body;
408 mds_req_from_mcd(req, med->med_mcd);
410 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
412 LASSERT(PTR_ERR(de) == req->rq_status);
416 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
417 mds_pack_inode2fid(&body->fid1, de->d_inode);
418 mds_pack_inode2body(body, de->d_inode);
420 /* Don't return OST-specific attributes if we didn't just set them */
421 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
422 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
423 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
424 body->valid |= OBD_MD_FLMTIME;
425 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
426 body->valid |= OBD_MD_FLATIME;
431 int mds_osc_setattr_async(struct obd_device *obd, __u32 uid, __u32 gid,
432 struct lov_mds_md *lmm, int lmm_size,
433 struct llog_cookie *logcookies, __u64 id, __u32 gen,
436 struct mds_obd *mds = &obd->u.mds;
437 struct obd_trans_info oti = { 0 };
438 struct obd_info oinfo = { { { 0 } } };
442 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
445 /* first get memory EA */
446 OBDO_ALLOC(oinfo.oi_oa);
452 rc = obd_unpackmd(mds->mds_osc_exp, &oinfo.oi_md, lmm, lmm_size);
454 CERROR("Error unpack md %p for inode "LPU64"\n", lmm, id);
458 rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, oinfo.oi_md);
460 CERROR("Error revalidate lsm %p \n", oinfo.oi_md);
465 oinfo.oi_oa->o_uid = uid;
466 oinfo.oi_oa->o_gid = gid;
467 oinfo.oi_oa->o_id = oinfo.oi_md->lsm_object_id;
468 oinfo.oi_oa->o_gr = oinfo.oi_md->lsm_object_gr;
469 oinfo.oi_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
470 OBD_MD_FLUID | OBD_MD_FLGID;
472 oinfo.oi_oa->o_valid |= OBD_MD_FLCOOKIE;
473 oti.oti_logcookies = logcookies;
476 oinfo.oi_oa->o_fid = id;
477 oinfo.oi_oa->o_generation = gen;
478 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
481 /* do async setattr from mds to ost not waiting for responses. */
482 rc = obd_setattr_async(mds->mds_osc_exp, &oinfo, &oti, NULL);
484 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64
485 " on ost error %d\n", oinfo.oi_md->lsm_object_id, rc);
488 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
489 OBDO_FREE(oinfo.oi_oa);
492 EXPORT_SYMBOL(mds_osc_setattr_async);
494 /* In the raw-setattr case, we lock the child inode.
495 * In the write-back case or if being called from open, the client holds a lock
498 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
499 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
500 struct ptlrpc_request *req,
501 struct lustre_handle *lh)
503 unsigned int ia_valid = rec->ur_iattr.ia_valid;
504 struct mds_obd *mds = mds_req2mds(req);
505 struct obd_device *obd = req->rq_export->exp_obd;
506 struct mds_body *body;
508 struct inode *inode = NULL;
509 struct lustre_handle lockh;
511 struct mds_logcancel_data *mlcd = NULL;
512 struct lov_mds_md *lmm = NULL;
513 struct llog_cookie *logcookies = NULL;
514 int lmm_size = 0, need_lock = 1, cookie_size = 0;
515 int rc = 0, cleanup_phase = 0, err, locked = 0, sync = 0;
516 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
517 unsigned int qpids[MAXQUOTAS] = { rec->ur_iattr.ia_uid,
518 rec->ur_iattr.ia_gid };
521 LASSERT(offset == REQ_REC_OFF);
522 offset = REPLY_REC_OFF;
524 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
525 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
526 OBD_COUNTER_INCREMENT(obd, setattr);
528 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
531 ldlm_request_cancel(req, rec->ur_dlm, 0);
533 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
534 (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
535 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
537 GOTO(cleanup, rc = PTR_ERR(de));
538 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
539 GOTO(cleanup, rc = -EROFS);
541 __u64 lockpart = MDS_INODELOCK_UPDATE;
542 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
543 lockpart |= MDS_INODELOCK_LOOKUP;
545 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
548 GOTO(cleanup, rc = PTR_ERR(de));
556 if ((rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) ||
557 (rec->ur_iattr.ia_valid & ATTR_SIZE)) {
558 /* Check write access for the O_TRUNC case */
559 if (mds_query_write_access(inode) < 0)
560 GOTO(cleanup, rc = -ETXTBSY);
563 /* save uid/gid for quota acq/rel */
564 qcids[USRQUOTA] = inode->i_uid;
565 qcids[GRPQUOTA] = inode->i_gid;
567 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
568 rec->ur_eadata != NULL) {
569 LOCK_INODE_MUTEX(inode);
573 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
575 /* start a log jounal handle if needed */
576 if (S_ISREG(inode->i_mode) &&
577 rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
578 lmm_size = mds->mds_max_mdsize;
579 OBD_ALLOC(lmm, lmm_size);
581 GOTO(cleanup, rc = -ENOMEM);
584 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock);
589 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
590 le32_to_cpu(lmm->lmm_stripe_count));
592 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
595 GOTO(cleanup, rc = PTR_ERR(handle));
597 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
598 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
599 LTIME_S(rec->ur_iattr.ia_mtime),
600 LTIME_S(rec->ur_iattr.ia_ctime));
601 rc = mds_fix_attr(inode, rec);
605 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) { /* ioctl */
606 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
607 (long)&rec->ur_flags);
608 } else if (rec->ur_iattr.ia_valid) { /* setattr */
609 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
610 /* journal chown/chgrp in llog, just like unlink */
611 if (rc == 0 && lmm_size){
612 cookie_size = mds_get_cookie_size(obd, lmm);
613 OBD_ALLOC(logcookies, cookie_size);
614 if (logcookies == NULL)
615 GOTO(cleanup, rc = -ENOMEM);
617 if (mds_log_op_setattr(obd, inode->i_uid, inode->i_gid,
619 logcookies, cookie_size) <= 0) {
620 OBD_FREE(logcookies, cookie_size);
626 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
627 rec->ur_eadata != NULL) {
628 struct lov_stripe_md *lsm = NULL;
629 struct lov_user_md *lum = NULL;
631 rc = ll_permission(inode, MAY_WRITE, NULL);
635 lum = rec->ur_eadata;
636 /* if { size, offset, count } = { 0, -1, 0 } (i.e. all default
637 * values specified) then delete default striping from dir. */
638 if (S_ISDIR(inode->i_mode) &&
639 ((lum->lmm_stripe_size == 0 &&
640 lum->lmm_stripe_offset ==
641 (typeof(lum->lmm_stripe_offset))(-1) &&
642 lum->lmm_stripe_count == 0) ||
643 /* lmm_stripe_size == -1 is deprecated in 1.4.6 */
644 lum->lmm_stripe_size ==
645 (typeof(lum->lmm_stripe_size))(-1))){
646 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, "lov");
650 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
652 &lsm, rec->ur_eadata);
656 obd_free_memmd(mds->mds_osc_exp, &lsm);
658 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
659 rec->ur_eadatalen, "lov");
665 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
666 mds_pack_inode2fid(&body->fid1, inode);
667 mds_pack_inode2body(body, inode);
669 /* don't return OST-specific attributes if we didn't just set them. */
670 if (ia_valid & ATTR_SIZE)
671 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
672 if (ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
673 body->valid |= OBD_MD_FLMTIME;
674 if (ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
675 body->valid |= OBD_MD_FLATIME;
677 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
678 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
681 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
683 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
684 mlcd->mlcd_cookielen = rec->ur_cookielen;
685 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
686 mlcd->mlcd_cookielen;
687 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
688 mlcd->mlcd_cookielen);
689 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
690 mlcd->mlcd_eadatalen);
692 CERROR("unable to allocate log cancel data\n");
698 sync = fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
699 mds_cancel_cookies_cb, mlcd);
700 err = mds_finish_transno(mds, inode, handle, req, rc, 0, sync);
701 /* do mds to ost setattr if needed */
702 if (!rc && !err && lmm_size)
703 mds_osc_setattr_async(obd, inode->i_ino, inode->i_generation, lmm,
704 lmm_size, logcookies, rec->ur_fid1->id,
705 rec->ur_fid1->generation, NULL);
707 switch (cleanup_phase) {
709 OBD_FREE(lmm, mds->mds_max_mdsize);
711 OBD_FREE(logcookies, cookie_size);
713 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
714 rec->ur_eadata != NULL)
715 UNLOCK_INODE_MUTEX(inode);
719 ldlm_lock_decref(&lockh, LCK_EX);
721 ptlrpc_save_lock (req, &lockh, LCK_EX);
734 /* trigger dqrel/dqacq for original owner and new owner */
735 if (ia_valid & (ATTR_UID | ATTR_GID))
736 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
742 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
743 struct ptlrpc_request *req)
745 struct mds_export_data *med = &req->rq_export->exp_mds_data;
746 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
747 struct dentry *parent, *child;
748 struct mds_body *body;
750 mds_req_from_mcd(req, med->med_mcd);
755 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
756 LASSERT(!IS_ERR(parent));
757 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
758 LASSERT(!IS_ERR(child));
760 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
761 mds_pack_inode2fid(&body->fid1, child->d_inode);
762 mds_pack_inode2body(body, child->d_inode);
768 static int mds_reint_create(struct mds_update_record *rec, int offset,
769 struct ptlrpc_request *req,
770 struct lustre_handle *lh)
772 struct dentry *dparent = NULL;
773 struct mds_obd *mds = mds_req2mds(req);
774 struct obd_device *obd = req->rq_export->exp_obd;
775 struct dentry *dchild = NULL;
776 struct inode *dir = NULL;
778 struct lustre_handle lockh;
779 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
781 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
782 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
783 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
786 LASSERT(offset == REQ_REC_OFF);
787 offset = REPLY_REC_OFF;
789 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
792 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
793 rec->ur_fid1->id, rec->ur_fid1->generation,
794 rec->ur_name, rec->ur_mode);
796 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
798 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
799 GOTO(cleanup, rc = -ESTALE);
802 ldlm_request_cancel(req, rec->ur_dlm, 0);
804 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX, &lockh,
805 MDS_INODELOCK_UPDATE);
806 if (IS_ERR(dparent)) {
807 rc = PTR_ERR(dparent);
809 CERROR("parent "LPU64"/%u lookup error %d\n",
810 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
813 cleanup_phase = 1; /* locked parent dentry */
814 dir = dparent->d_inode;
817 ldlm_lock_dump_handle(D_OTHER, &lockh);
819 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
820 if (IS_ERR(dchild)) {
821 rc = PTR_ERR(dchild);
822 if (rc != -ENAMETOOLONG)
823 CERROR("child lookup error %d\n", rc);
827 cleanup_phase = 2; /* child dentry */
829 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
831 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
833 GOTO(cleanup, rc = -EEXIST);
834 GOTO(cleanup, rc = -EROFS);
837 if (dir->i_mode & S_ISGID && S_ISDIR(rec->ur_mode))
838 rec->ur_mode |= S_ISGID;
840 dchild->d_fsdata = (void *)&dp;
841 dp.ldp_inum = (unsigned long)rec->ur_fid2->id;
846 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
848 GOTO(cleanup, rc = PTR_ERR(handle));
849 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
850 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
855 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
857 GOTO(cleanup, rc = PTR_ERR(handle));
858 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
859 mds_counter_incr(req->rq_export, LPROC_MDS_MKDIR);
864 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
866 GOTO(cleanup, rc = PTR_ERR(handle));
867 if (rec->ur_tgt == NULL) /* no target supplied */
868 rc = -EINVAL; /* -EPROTO? */
870 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
871 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
879 int rdev = rec->ur_rdev;
880 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
882 GOTO(cleanup, rc = PTR_ERR(handle));
883 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
884 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
889 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
890 dchild->d_fsdata = NULL;
891 GOTO(cleanup, rc = -EINVAL);
894 /* In case we stored the desired inum in here, we want to clean up. */
895 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
896 dchild->d_fsdata = NULL;
899 CDEBUG(D_INODE, "error during create: %d\n", rc);
903 struct inode *inode = dchild->d_inode;
904 struct mds_body *body;
907 LTIME_S(iattr.ia_atime) = rec->ur_time;
908 LTIME_S(iattr.ia_ctime) = rec->ur_time;
909 LTIME_S(iattr.ia_mtime) = rec->ur_time;
910 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
911 if (dir->i_mode & S_ISGID)
912 iattr.ia_gid = dir->i_gid;
914 iattr.ia_gid = current->fsgid;
915 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
916 ATTR_MTIME | ATTR_CTIME;
918 if (rec->ur_fid2->id) {
919 LASSERT(rec->ur_fid2->id == inode->i_ino);
920 inode->i_generation = rec->ur_fid2->generation;
921 /* Dirtied and committed by the upcoming setattr. */
922 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
923 inode->i_ino, inode->i_generation);
925 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
926 inode->i_ino, inode->i_generation);
929 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
931 CERROR("error on child setattr: rc = %d\n", rc);
933 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
934 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
936 CERROR("error on parent setattr: rc = %d\n", rc);
938 if (S_ISDIR(inode->i_mode)) {
939 struct lov_mds_md lmm;
940 int lmm_size = sizeof(lmm);
941 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1);
943 LOCK_INODE_MUTEX(inode);
944 rc = fsfilt_set_md(obd, inode, handle,
945 &lmm, lmm_size, "lov");
946 UNLOCK_INODE_MUTEX(inode);
949 CERROR("error on copy stripe info: rc = %d\n",
953 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
954 mds_pack_inode2fid(&body->fid1, inode);
955 mds_pack_inode2body(body, inode);
960 err = mds_finish_transno(mds, dir, handle, req, rc, 0, 0);
963 /* Destroy the file we just created. This should not need
964 * extra journal credits, as we have already modified all of
965 * the blocks needed in order to create the file in the first
970 err = vfs_rmdir(dir, dchild);
972 CERROR("rmdir in error path: %d\n", err);
975 err = vfs_unlink(dir, dchild);
977 CERROR("unlink in error path: %d\n", err);
980 } else if (created) {
981 /* The inode we were allocated may have just been freed
982 * by an unlink operation. We take this lock to
983 * synchronize against the matching reply-ack-lock taken
984 * in unlink, to avoid replay problems if this reply
985 * makes it out to the client but the unlink's does not.
986 * See bug 2029 for more detail.*/
987 mds_lock_new_child(obd, dchild->d_inode, NULL);
988 /* save uid/gid of create inode and parent */
989 qpids[USRQUOTA] = dir->i_uid;
990 qpids[GRPQUOTA] = dir->i_gid;
995 switch (cleanup_phase) {
996 case 2: /* child dentry */
998 case 1: /* locked parent dentry */
1000 ldlm_lock_decref(&lockh, LCK_EX);
1002 ptlrpc_save_lock (req, &lockh, LCK_EX);
1008 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1011 req->rq_status = rc;
1013 /* trigger dqacq on the owner of child and parent */
1014 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1019 int res_gt(const struct ldlm_res_id *res1, const struct ldlm_res_id *res2,
1020 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1024 for (i = 0; i < RES_NAME_SIZE; i++) {
1025 /* return 1 here, because enqueue_ordered will skip resources
1026 * of all zeroes if they're sorted to the end of the list. */
1027 if (res1->name[i] == 0 && res2->name[i] != 0)
1029 if (res2->name[i] == 0 && res1->name[i] != 0)
1032 if (res1->name[i] > res2->name[i])
1034 if (res1->name[i] < res2->name[i])
1039 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1044 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1045 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1046 * because they take the place of local semaphores.
1048 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1049 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1050 int enqueue_ordered_locks(struct obd_device *obd,
1051 const struct ldlm_res_id *p1_res_id,
1052 struct lustre_handle *p1_lockh, int p1_lock_mode,
1053 ldlm_policy_data_t *p1_policy,
1054 const struct ldlm_res_id *p2_res_id,
1055 struct lustre_handle *p2_lockh, int p2_lock_mode,
1056 ldlm_policy_data_t *p2_policy)
1058 const struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1059 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1060 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1061 ldlm_policy_data_t *policies[2] = {p1_policy, p2_policy};
1065 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1067 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1068 res_id[0]->name[0], res_id[1]->name[0]);
1070 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1071 handles[1] = p1_lockh;
1072 handles[0] = p2_lockh;
1073 res_id[1] = p1_res_id;
1074 res_id[0] = p2_res_id;
1075 lock_modes[1] = p1_lock_mode;
1076 lock_modes[0] = p2_lock_mode;
1077 policies[1] = p1_policy;
1078 policies[0] = p2_policy;
1081 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1082 res_id[0]->name[0], res_id[1]->name[0]);
1084 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1085 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[0],
1086 LDLM_IBITS, policies[0], lock_modes[0],
1087 &flags, ldlm_blocking_ast,
1088 ldlm_completion_ast, NULL, NULL, 0,
1092 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1094 if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0 &&
1095 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1096 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1097 ldlm_lock_addref(handles[1], lock_modes[1]);
1098 } else if (res_id[1]->name[0] != 0) {
1099 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1100 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[1],
1101 LDLM_IBITS, policies[1],
1102 lock_modes[1], &flags,
1104 ldlm_completion_ast, NULL, NULL,
1105 0, NULL, handles[1]);
1106 if (rc != ELDLM_OK) {
1107 ldlm_lock_decref(handles[0], lock_modes[0]);
1110 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1116 static inline int res_eq(const struct ldlm_res_id *res1,
1117 const struct ldlm_res_id *res2)
1119 return !memcmp(res1, res2, sizeof(*res1));
1123 try_to_aggregate_locks(const struct ldlm_res_id *res1, ldlm_policy_data_t *p1,
1124 const struct ldlm_res_id *res2, ldlm_policy_data_t *p2)
1126 if (!res_eq(res1, res2))
1128 /* XXX: any additional inodebits (to current LOOKUP and UPDATE)
1129 * should be taken with great care here */
1130 p1->l_inodebits.bits |= p2->l_inodebits.bits;
1133 int enqueue_4ordered_locks(struct obd_device *obd,
1134 const struct ldlm_res_id *p1_res_id,
1135 struct lustre_handle *p1_lockh, int p1_lock_mode,
1136 ldlm_policy_data_t *p1_policy,
1137 const struct ldlm_res_id *p2_res_id,
1138 struct lustre_handle *p2_lockh, int p2_lock_mode,
1139 ldlm_policy_data_t *p2_policy,
1140 const struct ldlm_res_id *c1_res_id,
1141 struct lustre_handle *c1_lockh, int c1_lock_mode,
1142 ldlm_policy_data_t *c1_policy,
1143 const struct ldlm_res_id *c2_res_id,
1144 struct lustre_handle *c2_lockh, int c2_lock_mode,
1145 ldlm_policy_data_t *c2_policy)
1147 const struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1148 c1_res_id, c2_res_id };
1149 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1150 c1_lockh, c2_lockh };
1151 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1152 c1_lock_mode, c2_lock_mode };
1153 ldlm_policy_data_t *policies[5] = {p1_policy, p2_policy,
1154 c1_policy, c2_policy};
1155 int rc, i, j, sorted, flags;
1158 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1159 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1160 res_id[3]->name[0]);
1162 /* simple insertion sort - we have at most 4 elements */
1163 for (i = 1; i < 4; i++) {
1165 dlm_handles[4] = dlm_handles[i];
1166 res_id[4] = res_id[i];
1167 lock_modes[4] = lock_modes[i];
1168 policies[4] = policies[i];
1172 if (res_gt(res_id[j], res_id[4], policies[j],
1174 dlm_handles[j + 1] = dlm_handles[j];
1175 res_id[j + 1] = res_id[j];
1176 lock_modes[j + 1] = lock_modes[j];
1177 policies[j + 1] = policies[j];
1182 } while (j >= 0 && !sorted);
1184 dlm_handles[j + 1] = dlm_handles[4];
1185 res_id[j + 1] = res_id[4];
1186 lock_modes[j + 1] = lock_modes[4];
1187 policies[j + 1] = policies[4];
1190 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1191 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1192 res_id[3]->name[0]);
1194 /* XXX we could send ASTs on all these locks first before blocking? */
1195 for (i = 0; i < 4; i++) {
1196 flags = LDLM_FL_ATOMIC_CB;
1197 if (res_id[i]->name[0] == 0)
1199 if (i && res_eq(res_id[i], res_id[i-1])) {
1200 memcpy(dlm_handles[i], dlm_handles[i-1],
1201 sizeof(*(dlm_handles[i])));
1202 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1204 /* we need to enqueue locks with different inodebits
1205 * at once, because otherwise concurrent thread can
1206 * hit the windown between these two locks and we'll
1207 * get to deadlock. see bug 10360. note also, that it
1208 * is impossible to have >2 equal res. */
1210 try_to_aggregate_locks(res_id[i], policies[i],
1211 res_id[i+1], policies[i+1]);
1212 rc = ldlm_cli_enqueue_local(obd->obd_namespace,
1213 res_id[i], LDLM_IBITS,
1214 policies[i], lock_modes[i],
1215 &flags, ldlm_blocking_ast,
1216 ldlm_completion_ast, NULL,
1220 GOTO(out_err, rc = -EIO);
1221 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1228 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1233 /* In the unlikely case that the child changed while we were waiting
1234 * on the lock, we need to drop the lock on the old child and either:
1235 * - if the child has a lower resource name, then we have to also
1236 * drop the parent lock and regain the locks in the right order
1237 * - in the rename case, if the child has a lower resource name than one of
1238 * the other parent/child resources (maxres) we also need to reget the locks
1239 * - if the child has a higher resource name (this is the common case)
1240 * we can just get the lock on the new child (still in lock order)
1242 * Returns 0 if the child did not change or if it changed but could be locked.
1243 * Returns 1 if the child changed and we need to re-lock (no locks held).
1244 * Returns -ve error with a valid dchild (no locks held). */
1245 static int mds_verify_child(struct obd_device *obd,
1246 const struct ldlm_res_id *parent_res_id,
1247 struct lustre_handle *parent_lockh,
1248 struct dentry *dparent, int parent_mode,
1249 struct ldlm_res_id *child_res_id,
1250 struct lustre_handle *child_lockh,
1251 struct dentry **dchildp, int child_mode,
1252 ldlm_policy_data_t *child_policy,
1253 const char *name, int namelen,
1254 const struct ldlm_res_id *maxres)
1256 struct dentry *vchild, *dchild = *dchildp;
1257 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1260 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1262 GOTO(cleanup, rc = PTR_ERR(vchild));
1264 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1265 (vchild->d_inode != NULL &&
1266 child_res_id->name[0] == vchild->d_inode->i_ino &&
1267 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1275 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1276 vchild->d_inode, dchild ? dchild->d_inode : 0,
1277 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1278 child_res_id->name[0]);
1279 if (child_res_id->name[0] != 0)
1280 ldlm_lock_decref(child_lockh, child_mode);
1284 cleanup_phase = 1; /* parent lock only */
1285 *dchildp = dchild = vchild;
1287 if (dchild->d_inode) {
1288 int flags = LDLM_FL_ATOMIC_CB;
1289 child_res_id->name[0] = dchild->d_inode->i_ino;
1290 child_res_id->name[1] = dchild->d_inode->i_generation;
1292 /* Make sure that we don't try to re-enqueue a lock on the
1293 * same resource if it happens that the source is renamed to
1294 * the target by another thread (bug 9974, thanks racer :-) */
1295 if (!res_gt(child_res_id, parent_res_id, NULL, NULL) ||
1296 !res_gt(child_res_id, maxres, NULL, NULL)) {
1297 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1298 child_res_id->name[0], parent_res_id->name[0],
1300 GOTO(cleanup, rc = 1);
1303 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1304 LDLM_IBITS, child_policy,
1307 ldlm_completion_ast, NULL,
1308 NULL, 0, NULL, child_lockh);
1310 GOTO(cleanup, rc = -EIO);
1312 memset(child_res_id, 0, sizeof(*child_res_id));
1318 switch(cleanup_phase) {
1320 if (child_res_id->name[0] != 0)
1321 ldlm_lock_decref(child_lockh, child_mode);
1323 ldlm_lock_decref(parent_lockh, parent_mode);
1329 #define INODE_CTIME_AGE (10)
1330 #define INODE_CTIME_OLD(inode) (LTIME_S(inode->i_ctime) + \
1331 INODE_CTIME_AGE < CURRENT_SECONDS)
1333 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1335 struct lustre_handle *parent_lockh,
1336 struct dentry **dparentp, int parent_mode,
1337 __u64 parent_lockpart,
1338 char *name, int namelen,
1339 struct lustre_handle *child_lockh,
1340 struct dentry **dchildp, int child_mode,
1341 __u64 child_lockpart)
1343 struct ldlm_res_id child_res_id = { .name = {0} };
1344 struct ldlm_res_id parent_res_id = { .name = {0} };
1345 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1346 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1347 struct inode *inode;
1348 int rc = 0, cleanup_phase = 0;
1351 /* Step 1: Lookup parent */
1352 *dparentp = mds_fid2dentry(mds, fid, NULL);
1353 if (IS_ERR(*dparentp)) {
1354 rc = PTR_ERR(*dparentp);
1359 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1360 (*dparentp)->d_inode->i_ino, name);
1362 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1363 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1365 cleanup_phase = 1; /* parent dentry */
1367 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1368 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1369 if (IS_ERR(*dchildp)) {
1370 rc = PTR_ERR(*dchildp);
1371 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1375 cleanup_phase = 2; /* child dentry */
1376 inode = (*dchildp)->d_inode;
1377 if (inode != NULL) {
1378 if (is_bad_inode(inode)) {
1379 CERROR("bad inode returned %lu/%u\n",
1380 inode->i_ino, inode->i_generation);
1381 GOTO(cleanup, rc = -ENOENT);
1383 inode = igrab(inode);
1388 child_res_id.name[0] = inode->i_ino;
1389 child_res_id.name[1] = inode->i_generation;
1391 /* If we want a LCK_CR for a directory, and this directory has not been
1392 changed for some time, we return not only a LOOKUP lock, but also an
1393 UPDATE lock to have negative dentry starts working for this dir.
1394 Also we apply same logic to non-directories. If the file is rarely
1395 changed - we return both locks and this might save us RPC on
1397 if ((child_mode & (LCK_CR|LCK_PR|LCK_CW)) && INODE_CTIME_OLD(inode))
1398 child_policy.l_inodebits.bits |= MDS_INODELOCK_UPDATE;
1403 cleanup_phase = 2; /* child dentry */
1405 /* Step 3: Lock parent and child in resource order. If child doesn't
1406 * exist, we still have to lock the parent and re-lookup. */
1407 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1409 &child_res_id, child_lockh, child_mode,
1414 if (!(*dchildp)->d_inode)
1415 cleanup_phase = 3; /* parent lock */
1417 cleanup_phase = 4; /* child lock */
1419 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1420 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1421 parent_mode, &child_res_id, child_lockh, dchildp,
1422 child_mode,&child_policy, name, namelen, &parent_res_id);
1432 switch (cleanup_phase) {
1434 ldlm_lock_decref(child_lockh, child_mode);
1436 ldlm_lock_decref(parent_lockh, parent_mode);
1447 void mds_reconstruct_generic(struct ptlrpc_request *req)
1449 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1451 mds_req_from_mcd(req, med->med_mcd);
1454 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1455 * we instead link the inode into the PENDING directory until it is
1456 * finally released. We can't simply call mds_reint_rename() or some
1457 * part thereof, because we don't have the inode to check for link
1458 * count/open status until after it is locked.
1460 * For lock ordering, caller must get child->i_mutex first, then
1461 * pending->i_mutex before starting journal transaction.
1463 * returns 1 on success
1464 * returns 0 if we lost a race and didn't make a new link
1465 * returns negative on error
1467 static int mds_orphan_add_link(struct mds_update_record *rec,
1468 struct obd_device *obd, struct dentry *dentry)
1470 struct mds_obd *mds = &obd->u.mds;
1471 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1472 struct inode *inode = dentry->d_inode;
1473 struct dentry *pending_child;
1474 char fidname[LL_FID_NAMELEN];
1475 int fidlen = 0, rc, mode;
1478 LASSERT(inode != NULL);
1479 LASSERT(!mds_inode_is_orphan(inode));
1480 #ifndef HAVE_I_ALLOC_SEM
1481 LASSERT(TRYLOCK_INODE_MUTEX(inode) == 0);
1483 LASSERT(TRYLOCK_INODE_MUTEX(pending_dir) == 0);
1485 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1487 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1488 mds_orphan_open_count(inode), inode->i_nlink,
1489 S_ISDIR(inode->i_mode) ? "dir" :
1490 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1492 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1495 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1496 if (IS_ERR(pending_child))
1497 RETURN(PTR_ERR(pending_child));
1499 if (pending_child->d_inode != NULL) {
1500 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1501 LASSERT(pending_child->d_inode == inode);
1502 GOTO(out_dput, rc = 0);
1505 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1506 * for linking and return real mode back then -bzzz */
1507 mode = inode->i_mode;
1508 inode->i_mode = S_IFREG;
1509 rc = vfs_link(dentry, pending_dir, pending_child);
1511 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1514 mds_inode_set_orphan(inode);
1516 /* return mode and correct i_nlink if inode is directory */
1517 inode->i_mode = mode;
1518 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1519 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1521 if (S_ISDIR(mode)) {
1523 pending_dir->i_nlink++;
1524 mark_inode_dirty(inode);
1525 mark_inode_dirty(pending_dir);
1528 GOTO(out_dput, rc = 1);
1530 l_dput(pending_child);
1534 int mds_get_cookie_size(struct obd_device *obd, struct lov_mds_md *lmm)
1536 int count = le32_to_cpu(lmm->lmm_stripe_count);
1537 int real_csize = count * sizeof(struct llog_cookie);
1541 void mds_shrink_reply(struct obd_device *obd, struct ptlrpc_request *req,
1542 struct mds_body *body, int md_off)
1544 int cookie_size = 0, md_size = 0;
1546 if (body && body->valid & OBD_MD_FLEASIZE) {
1547 md_size = body->eadatasize;
1549 if (body && body->valid & OBD_MD_FLCOOKIE) {
1550 LASSERT(body->valid & OBD_MD_FLEASIZE);
1551 cookie_size = mds_get_cookie_size(obd, lustre_msg_buf(
1552 req->rq_repmsg, md_off, 0));
1555 CDEBUG(D_INFO, "Shrink to md_size %d cookie_size %d \n", md_size,
1558 lustre_shrink_reply(req, md_off, md_size, 1);
1560 lustre_shrink_reply(req, md_off + (md_size > 0), cookie_size, 0);
1563 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1564 struct ptlrpc_request *req,
1565 struct lustre_handle *lh)
1567 struct dentry *dparent = NULL, *dchild;
1568 struct mds_obd *mds = mds_req2mds(req);
1569 struct obd_device *obd = req->rq_export->exp_obd;
1570 struct mds_body *body = NULL;
1571 struct inode *child_inode = NULL;
1572 struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1573 void *handle = NULL;
1574 int rc = 0, cleanup_phase = 0;
1575 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
1576 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1579 LASSERT(offset == REQ_REC_OFF); /* || offset == DLM_INTENT_REC_OFF); */
1580 offset = REPLY_REC_OFF;
1582 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1583 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1585 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1587 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1588 GOTO(cleanup, rc = -ENOENT);
1591 ldlm_request_cancel(req, rec->ur_dlm, 0);
1593 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1594 &parent_lockh, &dparent, LCK_EX,
1595 MDS_INODELOCK_UPDATE,
1596 rec->ur_name, rec->ur_namelen,
1597 &child_lockh, &dchild, LCK_EX,
1598 MDS_INODELOCK_FULL);
1602 cleanup_phase = 1; /* dchild, dparent, locks */
1605 child_inode = dchild->d_inode;
1606 if (child_inode == NULL) {
1607 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1608 dparent->d_inode->i_ino, rec->ur_name);
1609 GOTO(cleanup, rc = -ENOENT);
1612 /* save uid/gid for quota acquire/release */
1613 qcids[USRQUOTA] = child_inode->i_uid;
1614 qcids[GRPQUOTA] = child_inode->i_gid;
1615 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1616 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1618 cleanup_phase = 2; /* dchild has a lock */
1620 /* We have to do these checks ourselves, in case we are making an
1621 * orphan. The client tells us whether rmdir() or unlink() was called,
1622 * so we need to return appropriate errors (bug 72). */
1623 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1624 if (!S_ISDIR(child_inode->i_mode))
1625 GOTO(cleanup, rc = -ENOTDIR);
1627 if (S_ISDIR(child_inode->i_mode))
1628 GOTO(cleanup, rc = -EISDIR);
1631 /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1632 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1633 GOTO(cleanup, rc = -EROFS);
1635 /* Step 3: Get a lock on the ino to sync with creation WRT inode
1636 * reuse (see bug 2029). */
1637 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1641 cleanup_phase = 3; /* child inum lock */
1643 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1645 /* ldlm_reply in buf[0] if called via intent */
1646 if (offset == DLM_INTENT_REC_OFF)
1647 offset = DLM_REPLY_REC_OFF;
1649 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1650 LASSERT(body != NULL);
1652 /* child orphan sem protects orphan_dec_test && is_orphan race */
1653 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1654 cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1656 /* If this is potentially the last reference to this inode, get the
1657 * OBD EA data first so the client can destroy OST objects. We
1658 * only do the object removal later if no open files/links remain. */
1659 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1660 child_inode->i_nlink == 1) {
1661 if (mds_orphan_open_count(child_inode) > 0) {
1662 /* need to lock pending_dir before transaction */
1663 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1664 cleanup_phase = 5; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
1665 } else if (S_ISREG(child_inode->i_mode)) {
1666 mds_pack_inode2fid(&body->fid1, child_inode);
1667 mds_pack_inode2body(body, child_inode);
1668 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1669 child_inode, MDS_PACK_MD_LOCK);
1673 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1674 switch (child_inode->i_mode & S_IFMT) {
1676 /* Drop any lingering child directories before we start our
1677 * transaction, to avoid doing multiple inode dirty/delete
1678 * in our compound transaction (bug 1321). */
1679 shrink_dcache_parent(dchild);
1680 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1683 GOTO(cleanup, rc = PTR_ERR(handle));
1684 rc = vfs_rmdir(dparent->d_inode, dchild);
1685 mds_counter_incr(req->rq_export, LPROC_MDS_RMDIR);
1688 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1690 handle = fsfilt_start_log(obd, dparent->d_inode,
1691 FSFILT_OP_UNLINK, NULL,
1692 le32_to_cpu(lmm->lmm_stripe_count));
1694 GOTO(cleanup, rc = PTR_ERR(handle));
1695 rc = vfs_unlink(dparent->d_inode, dchild);
1696 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
1704 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1707 GOTO(cleanup, rc = PTR_ERR(handle));
1708 rc = vfs_unlink(dparent->d_inode, dchild);
1709 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
1712 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1715 GOTO(cleanup, rc = -EINVAL);
1718 if (rc == 0 && child_inode->i_nlink == 0) {
1719 if (mds_orphan_open_count(child_inode) > 0)
1720 rc = mds_orphan_add_link(rec, obd, dchild);
1723 GOTO(cleanup, rc = 0);
1725 if (!S_ISREG(child_inode->i_mode))
1728 if (!(body->valid & OBD_MD_FLEASIZE)) {
1729 body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1730 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1731 } else if (mds_log_op_unlink(obd,
1732 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1733 lustre_msg_buflen(req->rq_repmsg, offset + 1),
1734 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1735 lustre_msg_buflen(req->rq_repmsg, offset+2)) >
1737 body->valid |= OBD_MD_FLCOOKIE;
1747 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1748 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1749 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1751 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1753 CERROR("error on parent setattr: rc = %d\n", err);
1756 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
1757 handle, req, rc, 0, 0);
1759 (void)obd_set_info_async(mds->mds_osc_exp, strlen("unlinked"),
1760 "unlinked", 0, NULL, NULL);
1761 switch(cleanup_phase) {
1762 case 5: /* pending_dir semaphore */
1763 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1764 case 4: /* child inode semaphore */
1765 MDS_UP_READ_ORPHAN_SEM(child_inode);
1766 case 3: /* child ino-reuse lock */
1767 if (rc && body != NULL) {
1768 // Don't unlink the OST objects if the MDS unlink failed
1772 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1774 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1775 case 2: /* child lock */
1776 ldlm_lock_decref(&child_lockh, LCK_EX);
1777 case 1: /* child and parent dentry, parent lock */
1779 ldlm_lock_decref(&parent_lockh, LCK_EX);
1781 ptlrpc_save_lock(req, &parent_lockh, LCK_EX);
1788 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1791 req->rq_status = rc;
1793 mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1795 /* trigger dqrel on the owner of child and parent */
1796 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc, FSFILT_OP_UNLINK);
1800 static int mds_reint_link(struct mds_update_record *rec, int offset,
1801 struct ptlrpc_request *req,
1802 struct lustre_handle *lh)
1804 struct obd_device *obd = req->rq_export->exp_obd;
1805 struct dentry *de_src = NULL;
1806 struct dentry *de_tgt_dir = NULL;
1807 struct dentry *dchild = NULL;
1808 struct mds_obd *mds = mds_req2mds(req);
1809 struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
1810 struct ldlm_res_id src_res_id = { .name = {0} };
1811 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
1812 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
1813 ldlm_policy_data_t tgt_dir_policy =
1814 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1815 int rc = 0, cleanup_phase = 0;
1818 LASSERT(offset == REQ_REC_OFF);
1820 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
1821 rec->ur_fid1->id, rec->ur_fid1->generation,
1822 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
1823 mds_counter_incr(req->rq_export, LPROC_MDS_LINK);
1825 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1827 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1828 GOTO(cleanup, rc = -ENOENT);
1831 ldlm_request_cancel(req, rec->ur_dlm, 0);
1833 /* Step 1: Lookup the source inode and target directory by FID */
1834 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1836 GOTO(cleanup, rc = PTR_ERR(de_src));
1838 cleanup_phase = 1; /* source dentry */
1840 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
1841 if (IS_ERR(de_tgt_dir)) {
1842 rc = PTR_ERR(de_tgt_dir);
1847 cleanup_phase = 2; /* target directory dentry */
1849 CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
1850 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
1851 de_src->d_inode->i_ino);
1853 /* Step 2: Take the two locks */
1854 src_res_id.name[0] = de_src->d_inode->i_ino;
1855 src_res_id.name[1] = de_src->d_inode->i_generation;
1856 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
1857 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
1859 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
1861 &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
1866 cleanup_phase = 3; /* locks */
1868 if (mds_inode_is_orphan(de_src->d_inode)) {
1869 CDEBUG(D_INODE, "an attempt to link an orphan inode %lu/%u\n",
1870 de_src->d_inode->i_ino,
1871 de_src->d_inode->i_generation);
1872 GOTO(cleanup, rc = -ENOENT);
1875 /* Step 3: Lookup the child */
1876 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
1877 if (IS_ERR(dchild)) {
1878 rc = PTR_ERR(dchild);
1879 if (rc != -EPERM && rc != -EACCES && rc != -ENAMETOOLONG)
1880 CERROR("child lookup error %d\n", rc);
1884 cleanup_phase = 4; /* child dentry */
1886 if (dchild->d_inode) {
1887 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
1888 de_tgt_dir->d_inode->i_ino, rec->ur_name);
1893 /* Step 4: Do it. */
1894 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1896 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1897 GOTO(cleanup, rc = -EROFS);
1899 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
1901 GOTO(cleanup, rc = PTR_ERR(handle));
1903 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
1904 if (rc && rc != -EPERM && rc != -EACCES)
1905 CERROR("vfs_link error %d\n", rc);
1907 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
1908 handle, req, rc, 0, 0);
1911 switch (cleanup_phase) {
1912 case 4: /* child dentry */
1916 ldlm_lock_decref(&src_lockh, LCK_EX);
1917 ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
1919 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1920 ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
1922 case 2: /* target dentry */
1924 case 1: /* source dentry */
1929 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1932 req->rq_status = rc;
1936 /* The idea here is that we need to get four locks in the end:
1937 * one on each parent directory, one on each child. We need to take
1938 * these locks in some kind of order (to avoid deadlocks), and the order
1939 * I selected is "increasing resource number" order. We need to look up
1940 * the children, however, before we know what the resource number(s) are.
1941 * Thus the following plan:
1943 * 1,2. Look up the parents
1944 * 3,4. Look up the children
1945 * 5. Take locks on the parents and children, in order
1946 * 6. Verify that the children haven't changed since they were looked up
1948 * If there was a race and the children changed since they were first looked
1949 * up, it is possible that mds_verify_child() will be able to just grab the
1950 * lock on the new child resource (if it has a higher resource than any other)
1951 * but we need to compare against not only its parent, but also against the
1952 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
1954 * We need the fancy igrab() on the child inodes because we aren't holding a
1955 * lock on the parent after the lookup is done, so dentry->d_inode may change
1956 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
1958 int mds_get_parents_children_locked(struct obd_device *obd,
1959 struct mds_obd *mds,
1960 struct ll_fid *p1_fid,
1961 struct dentry **de_srcdirp,
1962 struct ll_fid *p2_fid,
1963 struct dentry **de_tgtdirp,
1965 const char *old_name, int old_len,
1966 struct dentry **de_oldp,
1967 const char *new_name, int new_len,
1968 struct dentry **de_newp,
1969 struct lustre_handle *dlm_handles,
1972 struct ldlm_res_id p1_res_id = { .name = {0} };
1973 struct ldlm_res_id p2_res_id = { .name = {0} };
1974 struct ldlm_res_id c1_res_id = { .name = {0} };
1975 struct ldlm_res_id c2_res_id = { .name = {0} };
1976 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
1977 /* Only dentry should disappear, but the inode itself would be
1978 intact otherwise. */
1979 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
1980 /* If something is going to be replaced, both dentry and inode locks are needed */
1981 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
1982 struct ldlm_res_id *maxres_src, *maxres_tgt;
1983 struct inode *inode;
1984 int rc = 0, cleanup_phase = 0;
1987 /* Step 1: Lookup the source directory */
1988 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
1989 if (IS_ERR(*de_srcdirp))
1990 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
1992 cleanup_phase = 1; /* source directory dentry */
1994 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
1995 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
1997 /* Step 2: Lookup the target directory */
1998 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
1999 *de_tgtdirp = dget(*de_srcdirp);
2001 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2002 if (IS_ERR(*de_tgtdirp)) {
2003 rc = PTR_ERR(*de_tgtdirp);
2009 cleanup_phase = 2; /* target directory dentry */
2011 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2012 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2014 /* Step 3: Lookup the source child entry */
2015 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2016 if (IS_ERR(*de_oldp)) {
2017 rc = PTR_ERR(*de_oldp);
2018 CDEBUG(D_INODE, "old child lookup error (%.*s): %d\n",
2019 old_len - 1, old_name, rc);
2023 cleanup_phase = 3; /* original name dentry */
2025 inode = (*de_oldp)->d_inode;
2027 inode = igrab(inode);
2029 GOTO(cleanup, rc = -ENOENT);
2031 c1_res_id.name[0] = inode->i_ino;
2032 c1_res_id.name[1] = inode->i_generation;
2036 /* Step 4: Lookup the target child entry */
2038 GOTO(retry_locks, rc);
2039 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2040 if (IS_ERR(*de_newp)) {
2041 rc = PTR_ERR(*de_newp);
2042 if (rc != -ENAMETOOLONG)
2043 CERROR("new child lookup error (%.*s): %d\n",
2044 old_len - 1, old_name, rc);
2048 cleanup_phase = 4; /* target dentry */
2050 inode = (*de_newp)->d_inode;
2052 inode = igrab(inode);
2056 c2_res_id.name[0] = inode->i_ino;
2057 c2_res_id.name[1] = inode->i_generation;
2061 /* Step 5: Take locks on the parents and child(ren) */
2062 maxres_src = &p1_res_id;
2063 maxres_tgt = &p2_res_id;
2064 cleanup_phase = 4; /* target dentry */
2066 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id,NULL,NULL))
2067 maxres_src = &c1_res_id;
2068 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id,NULL,NULL))
2069 maxres_tgt = &c2_res_id;
2071 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2073 &p2_res_id, &dlm_handles[1], parent_mode,
2075 &c1_res_id, &dlm_handles[2], child_mode,
2077 &c2_res_id, &dlm_handles[3], child_mode,
2082 cleanup_phase = 6; /* parent and child(ren) locks */
2084 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2085 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2086 parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
2087 child_mode, &c1_policy, old_name, old_len,
2090 if (c2_res_id.name[0] != 0)
2091 ldlm_lock_decref(&dlm_handles[3], child_mode);
2092 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2099 if ((*de_oldp)->d_inode == NULL)
2100 GOTO(cleanup, rc = -ENOENT);
2104 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2105 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2106 parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
2107 child_mode, &c2_policy, new_name, new_len,
2110 ldlm_lock_decref(&dlm_handles[2], child_mode);
2111 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2121 switch (cleanup_phase) {
2122 case 6: /* child lock(s) */
2123 if (c2_res_id.name[0] != 0)
2124 ldlm_lock_decref(&dlm_handles[3], child_mode);
2125 if (c1_res_id.name[0] != 0)
2126 ldlm_lock_decref(&dlm_handles[2], child_mode);
2127 case 5: /* parent locks */
2128 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2129 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2130 case 4: /* target dentry */
2132 case 3: /* source dentry */
2134 case 2: /* target directory dentry */
2135 l_dput(*de_tgtdirp);
2136 case 1: /* source directry dentry */
2137 l_dput(*de_srcdirp);
2144 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2145 struct ptlrpc_request *req,
2146 struct lustre_handle *lockh)
2148 struct obd_device *obd = req->rq_export->exp_obd;
2149 struct dentry *de_srcdir = NULL;
2150 struct dentry *de_tgtdir = NULL;
2151 struct dentry *de_old = NULL;
2152 struct dentry *de_new = NULL;
2153 struct inode *old_inode = NULL, *new_inode = NULL;
2154 struct mds_obd *mds = mds_req2mds(req);
2155 struct lustre_handle dlm_handles[4];
2156 struct mds_body *body = NULL;
2157 struct lov_mds_md *lmm = NULL;
2158 int rc = 0, lock_count = 3, cleanup_phase = 0;
2159 void *handle = NULL;
2160 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
2161 unsigned int qpids[4] = { 0, 0, 0, 0 };
2164 LASSERT(offset == REQ_REC_OFF);
2165 offset = REPLY_REC_OFF;
2167 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2168 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2169 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2170 mds_counter_incr(req->rq_export, LPROC_MDS_RENAME);
2172 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2175 ldlm_request_cancel(req, rec->ur_dlm, 0);
2177 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2178 rec->ur_fid2, &de_tgtdir, LCK_EX,
2179 rec->ur_name, rec->ur_namelen,
2180 &de_old, rec->ur_tgt,
2181 rec->ur_tgtlen, &de_new,
2182 dlm_handles, LCK_EX);
2186 cleanup_phase = 1; /* parent(s), children, locks */
2188 old_inode = de_old->d_inode;
2189 new_inode = de_new->d_inode;
2191 if (new_inode != NULL)
2194 /* sanity check for src inode */
2195 if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
2196 old_inode->i_ino == de_tgtdir->d_inode->i_ino)
2197 GOTO(cleanup, rc = -EINVAL);
2199 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2200 GOTO(cleanup, rc = -EROFS);
2202 if (new_inode == NULL)
2206 cleanup_phase = 2; /* iput(new_inode) when finished */
2208 /* sanity check for dest inode */
2209 if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
2210 new_inode->i_ino == de_tgtdir->d_inode->i_ino)
2211 GOTO(cleanup, rc = -EINVAL);
2213 if (old_inode == new_inode)
2214 GOTO(cleanup, rc = 0);
2216 /* save uids/gids for qunit acquire/release */
2217 qcids[USRQUOTA] = old_inode->i_uid;
2218 qcids[GRPQUOTA] = old_inode->i_gid;
2219 qpids[USRQUOTA] = de_tgtdir->d_inode->i_uid;
2220 qpids[GRPQUOTA] = de_tgtdir->d_inode->i_gid;
2221 qpids[2] = de_srcdir->d_inode->i_uid;
2222 qpids[3] = de_srcdir->d_inode->i_gid;
2224 /* if we are about to remove the target at first, pass the EA of
2225 * that inode to client to perform and cleanup on OST */
2226 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
2227 LASSERT(body != NULL);
2229 /* child orphan sem protects orphan_dec_test && is_orphan race */
2230 MDS_DOWN_READ_ORPHAN_SEM(new_inode);
2231 cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
2233 if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
2234 new_inode->i_nlink == 1) {
2235 if (mds_orphan_open_count(new_inode) > 0) {
2236 /* need to lock pending_dir before transaction */
2237 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2238 cleanup_phase = 4; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
2239 } else if (S_ISREG(new_inode->i_mode)) {
2240 mds_pack_inode2fid(&body->fid1, new_inode);
2241 mds_pack_inode2body(body, new_inode);
2242 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
2243 new_inode, MDS_PACK_MD_LOCK);
2248 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2249 de_srcdir->d_inode->i_sb);
2251 /* Check if we are moving old entry into its child. 2.6 does not
2252 check for this in vfs_rename() anymore */
2253 if (is_subdir(de_new, de_old))
2254 GOTO(cleanup, rc = -EINVAL);
2256 lmm = lustre_msg_buf(req->rq_repmsg, offset + 1, 0);
2257 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
2258 NULL, le32_to_cpu(lmm->lmm_stripe_count));
2261 GOTO(cleanup, rc = PTR_ERR(handle));
2264 de_old->d_fsdata = req;
2265 de_new->d_fsdata = req;
2267 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2270 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2271 if (mds_orphan_open_count(new_inode) > 0)
2272 rc = mds_orphan_add_link(rec, obd, de_new);
2275 GOTO(cleanup, rc = 0);
2277 if (!S_ISREG(new_inode->i_mode))
2280 if (!(body->valid & OBD_MD_FLEASIZE)) {
2281 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2282 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2283 } else if (mds_log_op_unlink(obd,
2284 lustre_msg_buf(req->rq_repmsg,
2286 lustre_msg_buflen(req->rq_repmsg,
2288 lustre_msg_buf(req->rq_repmsg,
2290 lustre_msg_buflen(req->rq_repmsg,
2293 body->valid |= OBD_MD_FLCOOKIE;
2299 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2300 handle, req, rc, 0, 0);
2302 switch (cleanup_phase) {
2304 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2306 MDS_UP_READ_ORPHAN_SEM(new_inode);
2311 if (lock_count == 4)
2312 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2313 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2314 ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
2315 ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
2317 if (lock_count == 4)
2318 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2319 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2320 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_EX);
2321 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_EX);
2330 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2333 req->rq_status = rc;
2335 /* acquire/release qunit */
2336 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc, FSFILT_OP_RENAME);
2340 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2341 struct ptlrpc_request *, struct lustre_handle *);
2343 static mds_reinter reinters[REINT_MAX] = {
2344 [REINT_SETATTR] mds_reint_setattr,
2345 [REINT_CREATE] mds_reint_create,
2346 [REINT_LINK] mds_reint_link,
2347 [REINT_UNLINK] mds_reint_unlink,
2348 [REINT_RENAME] mds_reint_rename,
2349 [REINT_OPEN] mds_open
2352 int mds_reint_rec(struct mds_update_record *rec, int offset,
2353 struct ptlrpc_request *req, struct lustre_handle *lockh)
2355 struct obd_device *obd = req->rq_export->exp_obd;
2357 struct mds_obd *mds = &obd->u.mds;
2359 gid_t fsgid = rec->ur_uc.luc_fsgid;
2362 struct lvfs_run_ctxt saved;
2367 if (req->rq_uid != LNET_UID_ANY) {
2368 /* non-root local cluster client
2369 * NB root's creds are believed... */
2370 LASSERT (req->rq_uid != 0);
2371 rec->ur_uc.luc_fsuid = req->rq_uid;
2372 rec->ur_uc.luc_cap = 0;
2377 /* get group info of this user */
2378 rec->ur_uc.luc_uce = upcall_cache_get_entry(mds->mds_group_hash,
2379 rec->ur_uc.luc_fsuid,
2380 rec->ur_uc.luc_fsgid, 2,
2381 &rec->ur_uc.luc_suppgid1);
2383 if (IS_ERR(rec->ur_uc.luc_uce)) {
2384 rc = PTR_ERR(rec->ur_uc.luc_uce);
2385 rec->ur_uc.luc_uce = NULL;
2389 /* checked by unpacker */
2390 LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2393 if (rec->ur_uc.luc_uce)
2394 rec->ur_uc.luc_fsgid = rec->ur_uc.luc_uce->ue_primary;
2398 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2402 if (rec->ur_uc.luc_uce && fsgid != rec->ur_uc.luc_fsgid &&
2403 in_group_p(fsgid)) {
2404 rec->ur_uc.luc_fsgid = fsgid;
2405 current->fsgid = saved.luc.luc_fsgid = fsgid;
2410 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2411 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2414 upcall_cache_put_entry(mds->mds_group_hash, rec->ur_uc.luc_uce);