1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mds/mds_reint.c
38 * Lustre Metadata Server (mds) reintegration routines
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
46 # define EXPORT_SYMTAB
48 #define DEBUG_SUBSYSTEM S_MDS
51 #include <obd_support.h>
52 #include <obd_class.h>
54 #include <lustre_lib.h>
55 #include <lustre/lustre_idl.h>
56 #include <lustre_mds.h>
57 #include <lustre_dlm.h>
58 #include <lustre_fsfilt.h>
59 #include <lustre_ucache.h>
60 #include <lustre_net.h>
62 #include "mds_internal.h"
64 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
67 struct obd_export *exp = data;
68 LASSERTF(exp->exp_obd == obd,
69 "%s: bad export (%p), obd (%p) != exp->exp_obd (%p)\n",
70 obd->obd_name, exp, obd, exp->exp_obd);
71 obd_transno_commit_cb(obd, transno, exp, error);
72 class_export_put(exp);
75 struct mds_logcancel_data {
76 struct lov_mds_md *mlcd_lmm;
80 struct llog_cookie mlcd_cookies[0];
83 /** lookup child dentry in parent dentry according to the name.
84 * if dentry is found, delete "lustre_mdt_attrs" EA (with name "lma")
85 * if it exists by checking OBD_INCOMPAT_FID.
87 struct dentry *mds_lookup(struct obd_device *obd, const char *fid_name,
88 struct dentry *dparent, int fid_namelen)
90 struct dentry *dchild;
91 struct lr_server_data *lsd = obd->u.mds.mds_server_data;
94 dchild = ll_lookup_one_len(fid_name, dparent, fid_namelen);
95 if (!IS_ERR(dchild) &&
96 unlikely((lsd->lsd_feature_incompat & OBD_INCOMPAT_FID) ||
97 OBD_FAIL_CHECK(OBD_FAIL_MDS_REMOVE_COMMON_EA))) {
98 struct inode *inode = dchild->d_inode;
101 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR,
103 if (!IS_ERR(handle)) {
104 LOCK_INODE_MUTEX(inode);
105 fsfilt_set_md(obd, inode, handle, NULL, 0,
107 /* result is ignored. */
108 UNLOCK_INODE_MUTEX(inode);
109 fsfilt_commit(obd, inode, handle, 0);
116 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
117 void *cb_data, int error)
119 struct mds_logcancel_data *mlcd = cb_data;
120 struct lov_stripe_md *lsm = NULL;
121 struct llog_ctxt *ctxt;
124 obd_transno_commit_cb(obd, transno, NULL, error);
126 CDEBUG(D_RPCTRACE, "cancelling %d cookies\n",
127 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
129 rc = obd_unpackmd(obd->u.mds.mds_lov_exp, &lsm, mlcd->mlcd_lmm,
130 mlcd->mlcd_eadatalen);
132 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
133 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
136 rc = obd_checkmd(obd->u.mds.mds_lov_exp, obd->obd_self_export,
139 CERROR("Can not revalidate lsm %p \n", lsm);
141 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
142 /* XXX 0 normally, SENDNOW for debug */
143 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
144 sizeof(*mlcd->mlcd_cookies),
145 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
149 CERROR("error cancelling %d log cookies: rc %d\n",
150 (int)(mlcd->mlcd_cookielen /
151 sizeof(*mlcd->mlcd_cookies)), rc);
154 OBD_FREE(mlcd, mlcd->mlcd_size);
157 /* fsfilt_set_version return old version. use that here */
158 static void mds_versions_set(struct obd_device *obd,
159 struct inode **inodes, __u64 version)
166 for (i = 0; i < PTLRPC_NUM_VERSIONS; i++)
167 if (inodes[i] != NULL)
168 fsfilt_set_version(obd, inodes[i], version);
171 int mds_version_get_check(struct ptlrpc_request *req, struct inode *inode,
174 /* version recovery */
175 struct obd_device *obd = req->rq_export->exp_obd;
176 __u64 curr_version, *pre_versions;
179 if (inode == NULL || !exp_connect_vbr(req->rq_export))
182 curr_version = fsfilt_get_version(obd, inode);
183 if ((__s64)curr_version == -EOPNOTSUPP)
185 /* VBR: version is checked always because costs nothing */
186 if (lustre_msg_get_transno(req->rq_reqmsg) != 0) {
187 pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
188 LASSERT(index < PTLRPC_NUM_VERSIONS);
189 /* Sanity check for malformed buffers */
190 if (pre_versions == NULL) {
191 CERROR("No versions in request buffer\n");
192 spin_lock(&req->rq_export->exp_lock);
193 req->rq_export->exp_vbr_failed = 1;
194 spin_unlock(&req->rq_export->exp_lock);
196 } else if (pre_versions[index] != curr_version) {
197 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
198 pre_versions[index], curr_version);
199 spin_lock(&req->rq_export->exp_lock);
200 req->rq_export->exp_vbr_failed = 1;
201 spin_unlock(&req->rq_export->exp_lock);
205 /* save pre-versions in reply */
206 LASSERT(req->rq_repmsg != NULL);
207 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
209 pre_versions[index] = curr_version;
213 /* Assumes caller has already pushed us into the kernel context. */
214 int mds_finish_transno(struct mds_obd *mds, struct inode **inodes, void *handle,
215 struct ptlrpc_request *req, int rc, __u32 op_data,
218 struct mds_export_data *med = &req->rq_export->exp_mds_data;
219 struct lsd_client_data *lcd = med->med_lcd;
220 struct obd_device *obd = req->rq_export->exp_obd;
221 __u64 transno, prev_transno;
224 int log_pri = D_RPCTRACE;
225 struct inode *inode = inodes ? inodes[0] : NULL;
226 int version_set = handle ? 1 : 0;
229 if (IS_ERR(handle)) {
234 /* if the export has already been failed, we have no last_rcvd slot */
235 if (req->rq_export->exp_failed || obd->obd_fail) {
236 CWARN("commit transaction for disconnected client %s: rc %d\n",
237 req->rq_export->exp_client_uuid.uuid, rc);
245 if (handle == NULL) {
246 /* if we're starting our own xaction, use our own inode */
247 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
248 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
249 if (IS_ERR(handle)) {
250 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
251 RETURN(PTR_ERR(handle));
255 off = med->med_lr_off;
257 transno = lustre_msg_get_transno(req->rq_reqmsg);
260 CERROR("%s: replay %s transno "LPU64" failed: rc %d\n",
262 libcfs_nid2str(req->rq_export->exp_connection->c_peer.nid),
265 } else if (transno == 0) {
266 spin_lock(&mds->mds_transno_lock);
267 transno = ++mds->mds_last_transno;
268 spin_unlock(&mds->mds_transno_lock);
269 /* VBR: set versions */
270 if (inodes && version_set)
271 mds_versions_set(obd, inodes, transno);
273 spin_lock(&mds->mds_transno_lock);
274 if (transno > mds->mds_last_transno)
275 mds->mds_last_transno = transno;
276 spin_unlock(&mds->mds_transno_lock);
278 /* VBR: replay case. Copy version from replay req and
279 * set new versions */
280 mds_versions_set(obd, inodes, transno);
283 req->rq_transno = transno;
284 lustre_msg_set_transno(req->rq_repmsg, transno);
288 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
290 lcd->lcd_last_close_transno = cpu_to_le64(transno);
291 lcd->lcd_last_close_xid = cpu_to_le64(req->rq_xid);
292 lcd->lcd_last_close_result = cpu_to_le32(rc);
293 lcd->lcd_last_close_data = cpu_to_le32(op_data);
295 prev_transno = le64_to_cpu(lcd->lcd_last_transno);
296 if (((lustre_msg_get_flags(req->rq_reqmsg) &
297 (MSG_RESENT | MSG_REPLAY)) == 0) ||
298 (transno > prev_transno)) {
299 /* VBR: save versions in last_rcvd for reconstruct. */
300 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
302 lcd->lcd_pre_versions[0] = cpu_to_le64(pre_versions[0]);
303 lcd->lcd_pre_versions[1] = cpu_to_le64(pre_versions[1]);
304 lcd->lcd_pre_versions[2] = cpu_to_le64(pre_versions[2]);
305 lcd->lcd_pre_versions[3] = cpu_to_le64(pre_versions[3]);
308 lcd->lcd_last_transno = cpu_to_le64(transno);
309 lcd->lcd_last_xid = cpu_to_le64(req->rq_xid);
310 lcd->lcd_last_result = cpu_to_le32(rc);
311 lcd->lcd_last_data = cpu_to_le32(op_data);
314 /** update trans table */
315 target_trans_table_update(req->rq_export, transno);
318 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
321 struct obd_export *exp = req->rq_export;
323 class_export_get(exp); /* released once the cb is called */
325 force_sync = fsfilt_add_journal_cb(obd, transno,
326 handle, mds_commit_cb,
329 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, lcd,
331 force_sync | exp->exp_need_sync);
333 mds_commit_cb(obd, transno, exp, err);
342 DEBUG_REQ(log_pri, req,
343 "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
344 transno, rc, lcd->lcd_uuid, med->med_lr_idx, err);
346 err = mds_lov_write_objids(obd);
352 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
355 err = fsfilt_commit(obd, inode, handle, 0);
357 CERROR("error committing transaction: %d\n", err);
365 /* this gives the same functionality as the code between
366 * sys_chmod and inode_setattr
367 * chown_common and inode_setattr
368 * utimes and inode_setattr
370 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
372 time_t now = CURRENT_SECONDS;
373 struct iattr *attr = &rec->ur_iattr;
374 unsigned int ia_valid = attr->ia_valid;
378 if (ia_valid & ATTR_RAW)
379 attr->ia_valid &= ~ATTR_RAW;
381 if (!(ia_valid & ATTR_CTIME_SET))
382 LTIME_S(attr->ia_ctime) = now;
384 attr->ia_valid &= ~ATTR_CTIME_SET;
385 if (!(ia_valid & ATTR_ATIME_SET))
386 LTIME_S(attr->ia_atime) = now;
387 if (!(ia_valid & ATTR_MTIME_SET))
388 LTIME_S(attr->ia_mtime) = now;
390 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
391 RETURN((attr->ia_valid & ~ATTR_ATTR_FLAG) ? -EPERM : 0);
394 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
395 if (current->fsuid != inode->i_uid &&
396 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
400 if (ia_valid & ATTR_SIZE &&
401 /* NFSD hack for open(O_CREAT|O_TRUNC)=mknod+truncate (bug 5781) */
402 !(rec->ur_uc.luc_fsuid == inode->i_uid &&
403 ia_valid & MDS_OPEN_OWNEROVERRIDE)) {
404 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
408 if (ia_valid & (ATTR_UID | ATTR_GID)) {
411 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
413 if (attr->ia_uid == (uid_t) -1)
414 attr->ia_uid = inode->i_uid;
415 if (attr->ia_gid == (gid_t) -1)
416 attr->ia_gid = inode->i_gid;
417 if (!(ia_valid & ATTR_MODE))
418 attr->ia_mode = inode->i_mode;
420 * If the user or group of a non-directory has been
421 * changed by a non-root user, remove the setuid bit.
422 * 19981026 David C Niemi <niemi@tux.org>
424 * Changed this to apply to all users, including root,
425 * to avoid some races. This is the behavior we had in
426 * 2.0. The check for non-root was definitely wrong
427 * for 2.2 anyway, as it should have been using
428 * CAP_FSETID rather than fsuid -- 19990830 SD.
430 if ((inode->i_mode & S_ISUID) == S_ISUID &&
431 !S_ISDIR(inode->i_mode)) {
432 attr->ia_mode &= ~S_ISUID;
433 attr->ia_valid |= ATTR_MODE;
436 * Likewise, if the user or group of a non-directory
437 * has been changed by a non-root user, remove the
438 * setgid bit UNLESS there is no group execute bit
439 * (this would be a file marked for mandatory
440 * locking). 19981026 David C Niemi <niemi@tux.org>
442 * Removed the fsuid check (see the comment above) --
445 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
446 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
447 attr->ia_mode &= ~S_ISGID;
448 attr->ia_valid |= ATTR_MODE;
450 } else if (ia_valid & ATTR_MODE) {
452 if (!(attr->ia_valid & ATTR_FORCE)) {
453 mode = inode->i_mode;
454 if (((mode & S_ISUID) && (!(attr->ia_mode & S_ISUID))) ||
455 ((mode & S_ISGID) && (mode & S_IXGRP) &&
456 (!(attr->ia_mode & S_ISGID))))
457 attr->ia_valid |= ATTR_FORCE;
459 mode = attr->ia_mode;
461 if (attr->ia_mode == (umode_t)-1)
462 mode = inode->i_mode;
464 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
469 void mds_steal_ack_locks(struct ptlrpc_request *req)
471 struct obd_export *exp = req->rq_export;
472 struct list_head *tmp;
473 struct ptlrpc_reply_state *oldrep;
474 struct ptlrpc_service *svc;
477 /* CAVEAT EMPTOR: spinlock order */
478 spin_lock(&exp->exp_lock);
479 list_for_each (tmp, &exp->exp_outstanding_replies) {
480 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
482 if (oldrep->rs_xid != req->rq_xid)
485 if (lustre_msg_get_opc(oldrep->rs_msg) !=
486 lustre_msg_get_opc(req->rq_reqmsg))
487 CERROR ("Resent req xid "LPU64" has mismatched opc: "
488 "new %d old %d\n", req->rq_xid,
489 lustre_msg_get_opc(req->rq_reqmsg),
490 lustre_msg_get_opc(oldrep->rs_msg));
492 svc = oldrep->rs_service;
493 spin_lock (&svc->srv_lock);
495 list_del_init (&oldrep->rs_exp_list);
497 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
499 oldrep->rs_nlocks, oldrep,
500 oldrep->rs_xid, oldrep->rs_transno,
501 lustre_msg_get_opc(oldrep->rs_msg),
502 libcfs_nid2str(exp->exp_connection->c_peer.nid));
504 for (i = 0; i < oldrep->rs_nlocks; i++)
505 ptlrpc_save_lock(req,
506 &oldrep->rs_locks[i],
507 oldrep->rs_modes[i]);
508 oldrep->rs_nlocks = 0;
510 DEBUG_REQ(D_DLMTRACE, req, "stole locks for");
511 ptlrpc_schedule_difficult_reply (oldrep);
513 spin_unlock (&svc->srv_lock);
516 spin_unlock(&exp->exp_lock);
520 * VBR: restore versions
522 void mds_vbr_reconstruct(struct ptlrpc_request *req,
523 struct lsd_client_data *lcd)
525 __u64 pre_versions[4] = {0};
526 pre_versions[0] = le64_to_cpu(lcd->lcd_pre_versions[0]);
527 pre_versions[1] = le64_to_cpu(lcd->lcd_pre_versions[1]);
528 pre_versions[2] = le64_to_cpu(lcd->lcd_pre_versions[2]);
529 pre_versions[3] = le64_to_cpu(lcd->lcd_pre_versions[3]);
530 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
533 void mds_req_from_lcd(struct ptlrpc_request *req, struct lsd_client_data *lcd)
535 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
536 req->rq_transno = le64_to_cpu(lcd->lcd_last_close_transno);
537 req->rq_status = le32_to_cpu(lcd->lcd_last_close_result);
539 req->rq_transno = le64_to_cpu(lcd->lcd_last_transno);
540 req->rq_status = le32_to_cpu(lcd->lcd_last_result);
541 mds_vbr_reconstruct(req, lcd);
543 if (req->rq_status != 0)
545 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
546 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
547 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
548 req->rq_transno, req->rq_status);
550 mds_steal_ack_locks(req);
553 static void reconstruct_reint_setattr(struct mds_update_record *rec,
554 int offset, struct ptlrpc_request *req)
556 struct obd_export *exp = req->rq_export;
557 struct mds_export_data *med = &exp->exp_mds_data;
558 struct mds_obd *obd = &exp->exp_obd->u.mds;
560 struct mds_body *body;
562 mds_req_from_lcd(req, med->med_lcd);
564 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
568 LCONSOLE_WARN("FID "LPU64"/%u lookup error %d."
569 " Evicting client %s with export %s.\n",
570 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
571 obd_uuid2str(&exp->exp_client_uuid),
572 obd_export_nid2str(exp));
573 mds_export_evict(exp);
578 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
579 mds_pack_inode2body(body, de->d_inode);
581 /* Don't return OST-specific attributes if we didn't just set them */
582 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
583 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
584 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
585 body->valid |= OBD_MD_FLMTIME;
586 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
587 body->valid |= OBD_MD_FLATIME;
592 int mds_osc_setattr_async(struct obd_device *obd, struct inode *inode,
593 struct lov_mds_md *lmm, int lmm_size,
594 struct llog_cookie *logcookies, struct ll_fid *fid)
596 struct mds_obd *mds = &obd->u.mds;
597 struct obd_trans_info oti = { 0 };
598 struct obd_info oinfo = { { { 0 } } };
602 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
605 /* first get memory EA */
606 OBDO_ALLOC(oinfo.oi_oa);
612 rc = obd_unpackmd(mds->mds_lov_exp, &oinfo.oi_md, lmm, lmm_size);
614 CERROR("Error unpack md %p for inode %lu\n", lmm, inode->i_ino);
618 rc = obd_checkmd(mds->mds_lov_exp, obd->obd_self_export, oinfo.oi_md);
620 CERROR("Error revalidate lsm %p \n", oinfo.oi_md);
625 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
626 oinfo.oi_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
627 oinfo.oi_oa->o_id = oinfo.oi_md->lsm_object_id;
628 oinfo.oi_oa->o_gr = 0;
630 oinfo.oi_oa->o_valid |= OBD_MD_FLCOOKIE;
631 oti.oti_logcookies = logcookies;
634 LASSERT(fid != NULL);
635 oinfo.oi_oa->o_fid = fid->id;
636 oinfo.oi_oa->o_generation = fid->generation;
637 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
639 /* do async setattr from mds to ost not waiting for responses. */
640 rc = obd_setattr_async(mds->mds_lov_exp, &oinfo, &oti, NULL);
642 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64
643 " on ost error %d\n", oinfo.oi_md->lsm_object_id, rc);
646 obd_free_memmd(mds->mds_lov_exp, &oinfo.oi_md);
647 OBDO_FREE(oinfo.oi_oa);
651 /* In the raw-setattr case, we lock the child inode.
652 * In the write-back case or if being called from open, the client holds a lock
655 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
656 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
657 struct ptlrpc_request *req,
658 struct lustre_handle *lh)
660 unsigned int ia_valid = rec->ur_iattr.ia_valid;
661 struct mds_obd *mds = mds_req2mds(req);
662 struct obd_device *obd = req->rq_export->exp_obd;
663 struct mds_body *body;
665 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
666 struct inode *inode = NULL;
667 struct lustre_handle lockh;
669 struct mds_logcancel_data *mlcd = NULL;
670 struct lov_mds_md *lmm = NULL;
671 struct llog_cookie *logcookies = NULL;
672 int lmm_size = 0, need_lock = 1, cookie_size = 0;
673 int rc = 0, cleanup_phase = 0, err = 0, locked = 0, sync = 0;
674 int do_vbr = rec->ur_iattr.ia_valid &
675 (ATTR_MODE|ATTR_UID|ATTR_GID|
676 ATTR_FROM_OPEN|ATTR_RAW|ATTR_ATTR_FLAG);
677 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
678 unsigned int qpids[MAXQUOTAS] = { rec->ur_iattr.ia_uid,
679 rec->ur_iattr.ia_gid };
682 LASSERT(offset == REQ_REC_OFF);
683 offset = REPLY_REC_OFF;
685 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
686 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
687 OBD_COUNTER_INCREMENT(obd, setattr);
689 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
692 ldlm_request_cancel(req, rec->ur_dlm, 0);
694 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
695 (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
696 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
698 GOTO(cleanup, rc = PTR_ERR(de));
699 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
700 GOTO(cleanup, rc = -EROFS);
702 __u64 lockpart = MDS_INODELOCK_UPDATE;
703 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
704 lockpart |= MDS_INODELOCK_LOOKUP;
706 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
707 &lockh, NULL, 0, lockpart);
709 GOTO(cleanup, rc = PTR_ERR(de));
717 if ((rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) ||
718 (rec->ur_iattr.ia_valid & ATTR_SIZE)) {
719 /* Check write access for the O_TRUNC case */
720 if (mds_query_write_access(inode) < 0)
721 GOTO(cleanup, rc = -ETXTBSY);
724 /* save uid/gid for quota acq/rel */
725 qcids[USRQUOTA] = inode->i_uid;
726 qcids[GRPQUOTA] = inode->i_gid;
728 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
729 rec->ur_eadata != NULL) {
730 LOCK_INODE_MUTEX(inode);
734 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
736 /* VBR: update version if attr changed are important for recovery */
738 rc = mds_version_get_check(req, inode, 0);
740 GOTO(cleanup_no_trans, rc);
742 /* start a log jounal handle if needed */
743 if (S_ISREG(inode->i_mode) &&
744 rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
745 lmm_size = mds->mds_max_mdsize;
746 OBD_ALLOC(lmm, lmm_size);
748 GOTO(cleanup, rc = -ENOMEM);
751 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock, 0,
752 req->rq_export->exp_connect_flags);
757 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
758 le32_to_cpu(lmm->lmm_stripe_count));
760 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
763 GOTO(cleanup, rc = PTR_ERR(handle));
765 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
766 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
767 LTIME_S(rec->ur_iattr.ia_mtime),
768 LTIME_S(rec->ur_iattr.ia_ctime));
769 rc = mds_fix_attr(inode, rec);
773 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) { /* ioctl */
774 rc = fsfilt_iocontrol(obd, de, FSFILT_IOC_SETFLAGS,
775 (long)&rec->ur_flags);
776 } else if (rec->ur_iattr.ia_valid) { /* setattr */
777 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
778 /* journal chown/chgrp in llog, just like unlink */
779 if (rc == 0 && lmm_size){
780 cookie_size = mds_get_cookie_size(obd, lmm);
781 OBD_ALLOC(logcookies, cookie_size);
782 if (logcookies == NULL)
783 GOTO(cleanup, rc = -ENOMEM);
785 if (mds_log_op_setattr(obd, inode, lmm, lmm_size,
786 logcookies, cookie_size) <= 0) {
787 OBD_FREE(logcookies, cookie_size);
793 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
794 rec->ur_eadata != NULL) {
795 struct lov_stripe_md *lsm = NULL;
796 struct lov_user_md *lum = NULL;
798 rc = ll_permission(inode, MAY_WRITE, NULL);
802 lum = rec->ur_eadata;
803 /* if { size, offset, count } = { 0, -1, 0 } and no pool
804 * (i.e. all default values specified) then delete default
805 * striping from dir. */
806 if (S_ISDIR(inode->i_mode) &&
807 (lum->lmm_stripe_size == 0 &&
808 lum->lmm_stripe_offset ==
809 (typeof(lum->lmm_stripe_offset))(-1) &&
810 lum->lmm_stripe_count == 0 &&
811 le32_to_cpu(lum->lmm_magic) != LOV_USER_MAGIC_V3)){
812 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, "lov");
816 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
818 &lsm, rec->ur_eadata);
822 obd_free_memmd(mds->mds_lov_exp, &lsm);
824 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
825 rec->ur_eadatalen, "lov");
831 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
832 mds_pack_inode2body(body, inode);
834 /* don't return OST-specific attributes if we didn't just set them. */
835 if (ia_valid & ATTR_SIZE)
836 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
837 if (ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
838 body->valid |= OBD_MD_FLMTIME;
839 if (ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
840 body->valid |= OBD_MD_FLATIME;
842 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_lov_obd)) {
843 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
846 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
848 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
849 mlcd->mlcd_cookielen = rec->ur_cookielen;
850 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
851 mlcd->mlcd_cookielen;
852 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
853 mlcd->mlcd_cookielen);
854 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
855 mlcd->mlcd_eadatalen);
857 CERROR("unable to allocate log cancel data\n");
863 sync = fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
864 mds_cancel_cookies_cb, mlcd);
866 /* permission changes may require sync operation */
867 if (rc == 0 && ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
868 sync |= mds->mds_sync_permission;
870 err = mds_finish_transno(mds, do_vbr ? inodes : NULL, handle, req,
874 /* do mds to ost setattr if needed */
875 if (!rc && !err && lmm_size)
876 mds_osc_setattr_async(obd, inode, lmm, lmm_size,
877 logcookies, rec->ur_fid1);
879 switch (cleanup_phase) {
881 OBD_FREE(lmm, mds->mds_max_mdsize);
883 OBD_FREE(logcookies, cookie_size);
885 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
886 rec->ur_eadata != NULL)
887 UNLOCK_INODE_MUTEX(inode);
891 ldlm_lock_decref(&lockh, LCK_EX);
893 ptlrpc_save_lock (req, &lockh, LCK_EX);
906 /* trigger dqrel/dqacq for original owner and new owner */
907 if (ia_valid & (ATTR_UID | ATTR_GID))
908 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
914 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
915 struct ptlrpc_request *req)
917 struct obd_export *exp = req->rq_export;
918 struct mds_export_data *med = &exp->exp_mds_data;
919 struct mds_obd *obd = &exp->exp_obd->u.mds;
920 struct dentry *parent, *child;
921 struct mds_body *body;
924 mds_req_from_lcd(req, med->med_lcd);
929 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
930 if (IS_ERR(parent)) {
931 rc = PTR_ERR(parent);
932 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
933 " Evicting client %s with export %s.\n",
934 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
935 obd_uuid2str(&exp->exp_client_uuid),
936 obd_export_nid2str(exp));
937 mds_export_evict(exp);
941 child = mds_lookup(exp->exp_obd, rec->ur_name, parent,
942 rec->ur_namelen - 1);
945 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
946 " Evicting client %s with export %s.\n",
947 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
948 obd_uuid2str(&exp->exp_client_uuid),
949 obd_export_nid2str(exp));
950 mds_export_evict(exp);
956 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
957 mds_pack_inode2body(body, child->d_inode);
963 static int mds_reint_create(struct mds_update_record *rec, int offset,
964 struct ptlrpc_request *req,
965 struct lustre_handle *lh)
967 struct dentry *dparent = NULL;
968 struct mds_obd *mds = mds_req2mds(req);
969 struct obd_device *obd = req->rq_export->exp_obd;
970 struct dentry *dchild = NULL;
971 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
972 struct inode *dir = NULL;
974 struct lustre_handle lockh;
975 int rc = 0, err = 0, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
977 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
978 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
979 unsigned int ids[MAXQUOTAS] = { 0, 0 };
980 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
981 int quota_pending[2] = {0, 0};
982 unsigned int gid = current->fsgid;
985 LASSERT(offset == REQ_REC_OFF);
986 offset = REPLY_REC_OFF;
988 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
991 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
992 rec->ur_fid1->id, rec->ur_fid1->generation,
993 rec->ur_name, rec->ur_mode);
995 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
997 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
998 GOTO(cleanup, rc = -ESTALE);
1001 ldlm_request_cancel(req, rec->ur_dlm, 0);
1003 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX, &lockh,
1004 rec->ur_name, rec->ur_namelen - 1,
1005 MDS_INODELOCK_UPDATE);
1006 if (IS_ERR(dparent)) {
1007 rc = PTR_ERR(dparent);
1008 CDEBUG(D_DENTRY, "parent "LPU64"/%u lookup error %d\n",
1009 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
1012 cleanup_phase = 1; /* locked parent dentry */
1013 dir = dparent->d_inode;
1016 ldlm_lock_dump_handle(D_OTHER, &lockh);
1018 dchild = mds_lookup(obd, rec->ur_name, dparent, rec->ur_namelen - 1);
1019 if (IS_ERR(dchild)) {
1020 rc = PTR_ERR(dchild);
1022 CDEBUG(D_DENTRY, "child lookup error %d\n", rc);
1026 cleanup_phase = 2; /* child dentry */
1028 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
1030 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
1031 if (dchild->d_inode)
1032 GOTO(cleanup, rc = -EEXIST);
1033 GOTO(cleanup, rc = -EROFS);
1036 /** check there is no stale orphan with same inode number */
1037 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1041 /* version recovery check */
1042 rc = mds_version_get_check(req, dir, 0);
1044 GOTO(cleanup_no_trans, rc);
1046 if (dir->i_mode & S_ISGID && S_ISDIR(rec->ur_mode))
1047 rec->ur_mode |= S_ISGID;
1049 dchild->d_fsdata = (void *)&dp;
1050 dp.ldp_inum = (unsigned long)rec->ur_fid2->id;
1053 if (dir->i_mode & S_ISGID)
1056 gid = current->fsgid;
1058 /* we try to get enough quota to write here, and let ldiskfs
1059 * decide if it is out of quota or not b=14783
1060 * FIXME: after CMD is used, pointer to obd_trans_info* couldn't
1061 * be NULL, b=14840 */
1062 ids[0] = current->fsuid;
1064 lquota_chkquota(mds_quota_interface_ref, req->rq_export, ids[0], ids[1],
1065 1, quota_pending, NULL, NULL, 0);
1069 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
1071 GOTO(cleanup, rc = PTR_ERR(handle));
1072 LOCK_INODE_MUTEX(dir);
1073 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
1074 UNLOCK_INODE_MUTEX(dir);
1075 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1080 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1082 GOTO(cleanup, rc = PTR_ERR(handle));
1083 LOCK_INODE_MUTEX(dir);
1084 rc = ll_vfs_mkdir(dir, dchild, mds->mds_vfsmnt, rec->ur_mode);
1085 UNLOCK_INODE_MUTEX(dir);
1086 mds_counter_incr(req->rq_export, LPROC_MDS_MKDIR);
1091 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
1093 GOTO(cleanup, rc = PTR_ERR(handle));
1094 LOCK_INODE_MUTEX(dir);
1095 if (rec->ur_tgt == NULL) /* no target supplied */
1096 rc = -EINVAL; /* -EPROTO? */
1098 rc = ll_vfs_symlink(dir, dchild, mds->mds_vfsmnt,
1099 rec->ur_tgt, S_IALLUGO);
1100 UNLOCK_INODE_MUTEX(dir);
1101 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1109 int rdev = rec->ur_rdev;
1110 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
1112 GOTO(cleanup, rc = PTR_ERR(handle));
1113 LOCK_INODE_MUTEX(dir);
1114 rc = ll_vfs_mknod(dir, dchild, mds->mds_vfsmnt, rec->ur_mode,
1116 UNLOCK_INODE_MUTEX(dir);
1117 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1122 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
1123 dchild->d_fsdata = NULL;
1124 GOTO(cleanup, rc = -EINVAL);
1127 /* In case we stored the desired inum in here, we want to clean up. */
1128 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
1129 dchild->d_fsdata = NULL;
1132 CDEBUG(D_INODE, "error during create: %d\n", rc);
1136 struct inode *inode = dchild->d_inode;
1137 struct mds_body *body;
1140 LTIME_S(iattr.ia_atime) = rec->ur_time;
1141 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1142 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1143 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1145 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1146 ATTR_MTIME | ATTR_CTIME;
1148 if (rec->ur_fid2->id) {
1149 if (rec->ur_fid2->id != inode->i_ino) {
1150 if (req->rq_export->exp_delayed)
1156 inode->i_generation = rec->ur_fid2->generation;
1157 /* Dirtied and committed by the upcoming setattr. */
1158 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1159 inode->i_ino, inode->i_generation);
1161 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
1162 inode->i_ino, inode->i_generation);
1165 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1167 CERROR("error on child setattr: rc = %d\n", rc);
1169 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1170 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1172 CERROR("error on parent setattr: rc = %d\n", rc);
1174 if (S_ISDIR(inode->i_mode)) {
1175 struct lov_mds_md_v3 lmm;
1176 int lmm_size = sizeof(lmm);
1177 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0,
1178 req->rq_export->exp_connect_flags);
1180 LOCK_INODE_MUTEX(inode);
1181 rc = fsfilt_set_md(obd, inode, handle,
1182 &lmm, lmm_size, "lov");
1183 UNLOCK_INODE_MUTEX(inode);
1186 CERROR("error on copy stripe info: rc = %d\n",
1190 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1191 mds_pack_inode2body(body, inode);
1197 inodes[1] = dchild ? dchild->d_inode : NULL;
1198 err = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
1201 if (quota_pending[0] || quota_pending[1])
1202 lquota_pending_commit(mds_quota_interface_ref, obd,
1203 ids[0], ids[1], quota_pending);
1205 if (rc && created) {
1206 /* Destroy the file we just created. This should not need
1207 * extra journal credits, as we have already modified all of
1208 * the blocks needed in order to create the file in the first
1213 LOCK_INODE_MUTEX(dir);
1214 err = ll_vfs_rmdir(dir, dchild, mds->mds_vfsmnt);
1215 UNLOCK_INODE_MUTEX(dir);
1217 CERROR("rmdir in error path: %d\n", err);
1220 LOCK_INODE_MUTEX(dir);
1221 err = ll_vfs_unlink(dir, dchild, mds->mds_vfsmnt);
1222 UNLOCK_INODE_MUTEX(dir);
1224 CERROR("unlink in error path: %d\n", err);
1227 } else if (created) {
1228 /* The inode we were allocated may have just been freed
1229 * by an unlink operation. We take this lock to
1230 * synchronize against the matching reply-ack-lock taken
1231 * in unlink, to avoid replay problems if this reply
1232 * makes it out to the client but the unlink's does not.
1233 * See bug 2029 for more detail.*/
1234 mds_lock_new_child(obd, dchild->d_inode, NULL);
1235 /* save uid/gid of create inode and parent */
1236 qpids[USRQUOTA] = dir->i_uid;
1237 qpids[GRPQUOTA] = dir->i_gid;
1242 switch (cleanup_phase) {
1243 case 2: /* child dentry */
1245 case 1: /* locked parent dentry */
1247 ldlm_lock_decref(&lockh, LCK_EX);
1249 ptlrpc_save_lock (req, &lockh, LCK_EX);
1255 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1258 req->rq_status = rc;
1260 /* trigger dqacq on the owner of child and parent */
1261 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1266 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1267 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1271 for (i = 0; i < RES_NAME_SIZE; i++) {
1272 /* return 1 here, because enqueue_ordered will skip resources
1273 * of all zeroes if they're sorted to the end of the list. */
1274 if (res1->name[i] == 0 && res2->name[i] != 0)
1276 if (res2->name[i] == 0 && res1->name[i] != 0)
1279 if (res1->name[i] > res2->name[i])
1281 if (res1->name[i] < res2->name[i])
1286 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1291 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1292 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1293 * because they take the place of local semaphores.
1295 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1296 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1297 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1298 struct lustre_handle *p1_lockh, int p1_lock_mode,
1299 ldlm_policy_data_t *p1_policy,
1300 struct ldlm_res_id *p2_res_id,
1301 struct lustre_handle *p2_lockh, int p2_lock_mode,
1302 ldlm_policy_data_t *p2_policy)
1304 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1305 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1306 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1307 ldlm_policy_data_t *policies[2] = {p1_policy, p2_policy};
1311 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1313 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1314 res_id[0]->name[0], res_id[1]->name[0]);
1316 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1317 handles[1] = p1_lockh;
1318 handles[0] = p2_lockh;
1319 res_id[1] = p1_res_id;
1320 res_id[0] = p2_res_id;
1321 lock_modes[1] = p1_lock_mode;
1322 lock_modes[0] = p2_lock_mode;
1323 policies[1] = p1_policy;
1324 policies[0] = p2_policy;
1327 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1328 res_id[0]->name[0], res_id[1]->name[0]);
1330 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1331 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[0],
1332 LDLM_IBITS, policies[0], lock_modes[0],
1333 &flags, ldlm_blocking_ast,
1334 ldlm_completion_ast, NULL, NULL, 0,
1338 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1340 if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0 &&
1341 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1342 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1343 ldlm_lock_addref(handles[1], lock_modes[1]);
1344 } else if (res_id[1]->name[0] != 0) {
1345 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1346 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[1],
1347 LDLM_IBITS, policies[1],
1348 lock_modes[1], &flags,
1350 ldlm_completion_ast, NULL, NULL,
1351 0, NULL, handles[1]);
1352 if (rc != ELDLM_OK) {
1353 ldlm_lock_decref(handles[0], lock_modes[0]);
1356 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1362 static inline int res_eq(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
1364 return !memcmp(res1, res2, sizeof(*res1));
1368 try_to_aggregate_locks(struct ldlm_res_id *res1, ldlm_policy_data_t *p1,
1369 struct ldlm_res_id *res2, ldlm_policy_data_t *p2)
1371 if (!res_eq(res1, res2))
1373 /* XXX: any additional inodebits (to current LOOKUP and UPDATE)
1374 * should be taken with great care here */
1375 p1->l_inodebits.bits |= p2->l_inodebits.bits;
1378 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1379 struct lustre_handle *p1_lockh, int p1_lock_mode,
1380 ldlm_policy_data_t *p1_policy,
1381 struct ldlm_res_id *p2_res_id,
1382 struct lustre_handle *p2_lockh, int p2_lock_mode,
1383 ldlm_policy_data_t *p2_policy,
1384 struct ldlm_res_id *c1_res_id,
1385 struct lustre_handle *c1_lockh, int c1_lock_mode,
1386 ldlm_policy_data_t *c1_policy,
1387 struct ldlm_res_id *c2_res_id,
1388 struct lustre_handle *c2_lockh, int c2_lock_mode,
1389 ldlm_policy_data_t *c2_policy)
1391 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1392 c1_res_id, c2_res_id };
1393 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1394 c1_lockh, c2_lockh };
1395 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1396 c1_lock_mode, c2_lock_mode };
1397 ldlm_policy_data_t *policies[5] = {p1_policy, p2_policy,
1398 c1_policy, c2_policy};
1399 int rc, i, j, sorted, flags;
1402 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1403 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1404 res_id[3]->name[0]);
1406 /* simple insertion sort - we have at most 4 elements */
1407 for (i = 1; i < 4; i++) {
1409 dlm_handles[4] = dlm_handles[i];
1410 res_id[4] = res_id[i];
1411 lock_modes[4] = lock_modes[i];
1412 policies[4] = policies[i];
1416 if (res_gt(res_id[j], res_id[4], policies[j],
1418 dlm_handles[j + 1] = dlm_handles[j];
1419 res_id[j + 1] = res_id[j];
1420 lock_modes[j + 1] = lock_modes[j];
1421 policies[j + 1] = policies[j];
1426 } while (j >= 0 && !sorted);
1428 dlm_handles[j + 1] = dlm_handles[4];
1429 res_id[j + 1] = res_id[4];
1430 lock_modes[j + 1] = lock_modes[4];
1431 policies[j + 1] = policies[4];
1434 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1435 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1436 res_id[3]->name[0]);
1438 /* XXX we could send ASTs on all these locks first before blocking? */
1439 for (i = 0; i < 4; i++) {
1440 flags = LDLM_FL_ATOMIC_CB;
1441 if (res_id[i]->name[0] == 0)
1443 if (i && res_eq(res_id[i], res_id[i-1])) {
1444 memcpy(dlm_handles[i], dlm_handles[i-1],
1445 sizeof(*(dlm_handles[i])));
1446 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1448 /* we need to enqueue locks with different inodebits
1449 * at once, because otherwise concurrent thread can
1450 * hit the windown between these two locks and we'll
1451 * get to deadlock. see bug 10360. note also, that it
1452 * is impossible to have >2 equal res. */
1454 try_to_aggregate_locks(res_id[i], policies[i],
1455 res_id[i+1], policies[i+1]);
1456 rc = ldlm_cli_enqueue_local(obd->obd_namespace,
1457 res_id[i], LDLM_IBITS,
1458 policies[i], lock_modes[i],
1459 &flags, ldlm_blocking_ast,
1460 ldlm_completion_ast, NULL,
1464 GOTO(out_err, rc = -EIO);
1465 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1472 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1477 /* In the unlikely case that the child changed while we were waiting
1478 * on the lock, we need to drop the lock on the old child and either:
1479 * - if the child has a lower resource name, then we have to also
1480 * drop the parent lock and regain the locks in the right order
1481 * - in the rename case, if the child has a lower resource name than one of
1482 * the other parent/child resources (maxres) we also need to reget the locks
1483 * - if the child has a higher resource name (this is the common case)
1484 * we can just get the lock on the new child (still in lock order)
1486 * Returns 0 if the child did not change or if it changed but could be locked.
1487 * Returns 1 if the child changed and we need to re-lock (no locks held).
1488 * Returns -ve error with a valid dchild (no locks held). */
1489 static int mds_verify_child(struct obd_device *obd,
1490 struct ldlm_res_id *parent_res_id,
1491 struct lustre_handle *parent_lockh,
1492 struct dentry *dparent, int parent_mode,
1493 struct ldlm_res_id *child_res_id,
1494 struct lustre_handle *child_lockh,
1495 struct dentry **dchildp, int child_mode,
1496 ldlm_policy_data_t *child_policy,
1497 const char *name, int namelen,
1498 struct ldlm_res_id *maxres)
1500 struct dentry *vchild, *dchild = *dchildp;
1501 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1504 /* not want child - not check it */
1508 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1510 GOTO(cleanup, rc = PTR_ERR(vchild));
1512 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1513 (vchild->d_inode != NULL &&
1514 child_res_id->name[0] == vchild->d_inode->i_ino &&
1515 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1522 /* resouce is changed, but not want child lock, return new child */
1523 if (child_lockh == NULL) {
1526 GOTO(cleanup, rc = 0);
1529 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1530 vchild->d_inode, dchild ? dchild->d_inode : 0,
1531 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1532 child_res_id->name[0]);
1533 if (child_res_id->name[0] != 0)
1534 ldlm_lock_decref(child_lockh, child_mode);
1538 cleanup_phase = 1; /* parent lock only */
1539 *dchildp = dchild = vchild;
1541 if (dchild->d_inode) {
1542 int flags = LDLM_FL_ATOMIC_CB;
1543 child_res_id->name[0] = dchild->d_inode->i_ino;
1544 child_res_id->name[1] = dchild->d_inode->i_generation;
1546 /* Make sure that we don't try to re-enqueue a lock on the
1547 * same resource if it happens that the source is renamed to
1548 * the target by another thread (bug 9974, thanks racer :-) */
1549 if (!res_gt(child_res_id, parent_res_id, NULL, NULL) ||
1550 !res_gt(child_res_id, maxres, NULL, NULL)) {
1551 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1552 child_res_id->name[0], parent_res_id->name[0],
1554 GOTO(cleanup, rc = 1);
1557 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1558 LDLM_IBITS, child_policy,
1561 ldlm_completion_ast, NULL,
1562 NULL, 0, NULL, child_lockh);
1564 GOTO(cleanup, rc = -EIO);
1566 memset(child_res_id, 0, sizeof(*child_res_id));
1567 memset(child_lockh, 0, sizeof(*child_lockh));
1573 switch(cleanup_phase) {
1575 if (child_res_id->name[0] != 0)
1576 ldlm_lock_decref(child_lockh, child_mode);
1578 ldlm_lock_decref(parent_lockh, parent_mode);
1584 #define INODE_CTIME_AGE (10)
1585 #define INODE_CTIME_OLD(inode) (LTIME_S(inode->i_ctime) + \
1586 INODE_CTIME_AGE < CURRENT_SECONDS)
1588 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1590 struct lustre_handle *parent_lockh,
1591 struct dentry **dparentp, int parent_mode,
1592 __u64 parent_lockpart,
1593 char *name, int namelen,
1594 struct lustre_handle *child_lockh,
1595 struct dentry **dchildp, int child_mode,
1596 __u64 child_lockpart)
1598 struct ldlm_res_id child_res_id = { .name = {0} };
1599 struct ldlm_res_id parent_res_id = { .name = {0} };
1600 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1601 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1602 static struct ldlm_res_id child_res_id_nolock = { .name = {0} };
1603 struct inode *inode;
1604 int rc = 0, cleanup_phase = 0;
1607 /* Step 1: Lookup parent */
1608 *dparentp = mds_fid2dentry(mds, fid, NULL);
1609 if (IS_ERR(*dparentp)) {
1610 rc = PTR_ERR(*dparentp);
1615 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1616 (*dparentp)->d_inode->i_ino, name);
1618 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1619 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1621 cleanup_phase = 1; /* parent dentry */
1624 GOTO(retry_locks, rc);
1625 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1626 *dchildp = mds_lookup(obd, name, *dparentp, namelen - 1);
1627 if (IS_ERR(*dchildp)) {
1628 rc = PTR_ERR(*dchildp);
1630 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1634 cleanup_phase = 2; /* child dentry */
1636 inode = (*dchildp)->d_inode;
1637 if (inode != NULL) {
1638 if (is_bad_inode(inode)) {
1639 CERROR("bad inode returned %lu/%u\n",
1640 inode->i_ino, inode->i_generation);
1641 GOTO(cleanup, rc = -ENOENT);
1643 inode = igrab(inode);
1648 child_res_id.name[0] = inode->i_ino;
1649 child_res_id.name[1] = inode->i_generation;
1651 /* If we want a LCK_CR for a directory, and this directory has not been
1652 changed for some time, we return not only a LOOKUP lock, but also an
1653 UPDATE lock to have negative dentry starts working for this dir.
1654 Also we apply same logic to non-directories. If the file is rarely
1655 changed - we return both locks and this might save us RPC on
1657 if ((child_mode & (LCK_CR|LCK_PR|LCK_CW)) && INODE_CTIME_OLD(inode))
1658 child_policy.l_inodebits.bits |= MDS_INODELOCK_UPDATE;
1663 cleanup_phase = 2; /* child dentry */
1665 /* Step 3: Lock parent and child in resource order. If child doesn't
1666 * exist, we still have to lock the parent and re-lookup. */
1667 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1669 child_lockh ? &child_res_id :
1670 &child_res_id_nolock,
1671 child_lockh, child_mode,
1676 if (IS_DEADDIR((*dparentp)->d_inode))
1677 GOTO(cleanup, -ENOENT);
1679 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1680 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1681 parent_mode, &child_res_id, child_lockh, dchildp,
1682 child_mode,&child_policy, name, namelen, &parent_res_id);
1691 switch (cleanup_phase) {
1702 void mds_reconstruct_generic(struct ptlrpc_request *req)
1704 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1706 mds_req_from_lcd(req, med->med_lcd);
1709 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1710 * we instead link the inode into the PENDING directory until it is
1711 * finally released. We can't simply call mds_reint_rename() or some
1712 * part thereof, because we don't have the inode to check for link
1713 * count/open status until after it is locked.
1715 * returns 1 on success
1716 * returns 0 if we lost a race and didn't make a new link
1717 * returns negative on error
1719 static int mds_orphan_add_link(struct mds_update_record *rec,
1720 struct obd_device *obd, struct dentry *dentry)
1722 struct mds_obd *mds = &obd->u.mds;
1723 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1724 struct inode *inode = dentry->d_inode;
1725 struct dentry *pending_child;
1726 char fidname[LL_FID_NAMELEN];
1727 int fidlen = 0, rc, mode;
1730 LASSERT(inode != NULL);
1731 LASSERT(!mds_inode_is_orphan(inode));
1732 LASSERT(TRYLOCK_INODE_MUTEX(pending_dir) == 0);
1734 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1736 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1737 mds_orphan_open_count(inode), inode->i_nlink,
1738 S_ISDIR(inode->i_mode) ? "dir" :
1739 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1741 if (!mds_orphan_needed(obd, inode) || inode->i_nlink != 0)
1744 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1745 if (IS_ERR(pending_child))
1746 RETURN(PTR_ERR(pending_child));
1748 if (pending_child->d_inode != NULL) {
1749 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1750 LASSERT(pending_child->d_inode == inode);
1751 GOTO(out_dput, rc = 0);
1754 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1755 * for linking and return real mode back then -bzzz */
1756 mode = inode->i_mode;
1757 inode->i_mode = S_IFREG;
1758 /* avoid vfs_link upon 0 nlink inode */
1760 rc = ll_vfs_link(dentry, mds->mds_vfsmnt, pending_dir, pending_child,
1763 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1766 mds_inode_set_orphan(inode);
1768 /* return mode and correct i_nlink if inode is directory */
1769 inode->i_mode = mode;
1770 LASSERTF(inode->i_nlink == 2, "%s nlink == %d\n",
1771 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1773 if (S_ISDIR(mode)) {
1774 pending_dir->i_nlink++;
1775 if (pending_dir->i_sb->s_op->dirty_inode)
1776 pending_dir->i_sb->s_op->dirty_inode(pending_dir);
1779 if (inode->i_sb->s_op->dirty_inode)
1780 inode->i_sb->s_op->dirty_inode(inode);
1783 GOTO(out_dput, rc = 1);
1785 l_dput(pending_child);
1789 int mds_get_cookie_size(struct obd_device *obd, struct lov_mds_md *lmm)
1791 int count = le32_to_cpu(lmm->lmm_stripe_count);
1792 int real_csize = count * sizeof(struct llog_cookie);
1796 static void mds_shrink_reply(struct ptlrpc_request *req,
1797 int reply_mdoff, int have_md, int have_acl)
1799 struct obd_device *obd = req->rq_export->exp_obd;
1800 struct mds_body *reply_body;
1801 int cookie_size = 0, md_size = 0;
1804 /* LSM and cookie is always placed after mds_body */
1805 reply_body = lustre_msg_buf(req->rq_repmsg, reply_mdoff,
1806 sizeof(*reply_body));
1807 if (reply_body == NULL)
1808 /* if there is no mds_body, no point in shrinking the reply */
1813 if (have_md || have_acl) {
1814 if (reply_body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) {
1815 md_size = reply_body->eadatasize;
1816 } else if (reply_body->valid & OBD_MD_LINKNAME)
1817 md_size = reply_body->eadatasize;
1819 if (reply_body->valid & OBD_MD_FLCOOKIE) {
1820 LASSERT(reply_body->valid & OBD_MD_FLEASIZE);
1821 cookie_size = mds_get_cookie_size(obd, lustre_msg_buf(
1824 } else if (reply_body->valid & OBD_MD_FLACL) {
1825 cookie_size = reply_body->aclsize;
1828 CDEBUG(D_INFO, "Shrink %d/%d to md_size %d cookie_size %d \n",
1829 have_md, have_acl, md_size, cookie_size);
1831 if (likely(have_md))
1832 lustre_shrink_reply(req, reply_mdoff, md_size, 1);
1834 if (likely(have_acl))
1835 lustre_shrink_reply(req, reply_mdoff + (md_size > 0),
1839 void mds_shrink_body_reply(struct ptlrpc_request *req,
1843 struct mds_body *rq_body;
1844 const long long have_acl = OBD_MD_FLCOOKIE | OBD_MD_FLACL;
1845 const long have_md = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
1848 /* LSM and cookie is always placed after mds_body */
1849 rq_body = lustre_msg_buf(req->rq_reqmsg, req_mdoff,
1853 /* this check is need for avoid hit asset in case
1854 * OBD_MDS_FLFLAGS */
1855 mds_shrink_reply(req, reply_mdoff,
1856 rq_body->valid & have_md,
1857 rq_body->valid & have_acl);
1860 void mds_shrink_intent_reply(struct ptlrpc_request *req,
1861 int opc, int reply_mdoff)
1866 mds_shrink_reply(req, reply_mdoff, 1, 1);
1869 mds_shrink_reply(req, reply_mdoff, 1, 0);
1876 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1877 struct ptlrpc_request *req,
1878 struct lustre_handle *lh)
1880 struct dentry *dparent = NULL, *dchild;
1881 struct mds_obd *mds = mds_req2mds(req);
1882 struct obd_device *obd = req->rq_export->exp_obd;
1883 struct mds_body *body = NULL;
1884 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
1885 struct inode *child_inode = NULL;
1886 struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1887 void *handle = NULL;
1888 int rc = 0, cleanup_phase = 0;
1889 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
1890 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1893 LASSERT(offset == REQ_REC_OFF); /* || offset == DLM_INTENT_REC_OFF); */
1894 offset = REPLY_REC_OFF;
1896 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1897 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1899 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1901 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1902 GOTO(cleanup, rc = -ENOENT);
1905 ldlm_request_cancel(req, rec->ur_dlm, 0);
1907 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1908 &parent_lockh, &dparent, LCK_EX,
1909 MDS_INODELOCK_UPDATE,
1910 rec->ur_name, rec->ur_namelen,
1911 &child_lockh, &dchild, LCK_EX,
1912 MDS_INODELOCK_FULL);
1916 cleanup_phase = 1; /* dchild, dparent, locks */
1919 child_inode = dchild->d_inode;
1920 if (child_inode == NULL) {
1921 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1922 dparent->d_inode->i_ino, rec->ur_name);
1923 GOTO(cleanup, rc = -ENOENT);
1926 /* save uid/gid for quota acquire/release */
1927 qcids[USRQUOTA] = child_inode->i_uid;
1928 qcids[GRPQUOTA] = child_inode->i_gid;
1929 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1930 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1932 cleanup_phase = 2; /* dchild has a lock */
1934 /* VBR: version recovery check for parent */
1935 rc = mds_version_get_check(req, dparent->d_inode, 0);
1937 GOTO(cleanup_no_trans, rc);
1939 /* version recovery check */
1940 rc = mds_version_get_check(req, child_inode, 1);
1942 GOTO(cleanup_no_trans, rc);
1944 /* We have to do these checks ourselves, in case we are making an
1945 * orphan. The client tells us whether rmdir() or unlink() was called,
1946 * so we need to return appropriate errors (bug 72). */
1947 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1948 if (!S_ISDIR(child_inode->i_mode))
1949 GOTO(cleanup, rc = -ENOTDIR);
1951 if (S_ISDIR(child_inode->i_mode))
1952 GOTO(cleanup, rc = -EISDIR);
1955 /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1956 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1957 GOTO(cleanup, rc = -EROFS);
1959 /* Step 3: Get a lock on the ino to sync with creation WRT inode
1960 * reuse (see bug 2029). */
1961 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1965 cleanup_phase = 3; /* child inum lock */
1967 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1969 /* ldlm_reply in buf[0] if called via intent */
1970 if (offset == DLM_INTENT_REC_OFF)
1971 offset = DLM_REPLY_REC_OFF;
1973 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1974 LASSERT(body != NULL);
1976 /* child orphan sem protects orphan_dec_test && is_orphan race */
1977 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1978 cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1980 /* If this is potentially the last reference to this inode, get the
1981 * OBD EA data first so the client can destroy OST objects. We
1982 * only do the object removal later if no open files/links remain. */
1983 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1984 child_inode->i_nlink == 1) {
1985 if (mds_orphan_needed(obd, child_inode)) {
1986 /* need to lock pending_dir before transaction */
1987 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1988 cleanup_phase = 5; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
1989 } else if (S_ISREG(child_inode->i_mode)) {
1990 mds_pack_inode2body(body, child_inode);
1991 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1992 child_inode, MDS_PACK_MD_LOCK, 0,
1993 req->rq_export->exp_connect_flags);
1997 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1998 switch (child_inode->i_mode & S_IFMT) {
2000 /* Drop any lingering child directories before we start our
2001 * transaction, to avoid doing multiple inode dirty/delete
2002 * in our compound transaction (bug 1321). */
2003 shrink_dcache_parent(dchild);
2004 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2007 GOTO(cleanup, rc = PTR_ERR(handle));
2008 LOCK_INODE_MUTEX(dparent->d_inode);
2009 rc = ll_vfs_rmdir(dparent->d_inode, dchild, mds->mds_vfsmnt);
2010 UNLOCK_INODE_MUTEX(dparent->d_inode);
2011 mds_counter_incr(req->rq_export, LPROC_MDS_RMDIR);
2014 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2016 int sz = lustre_msg_buflen(req->rq_repmsg, offset + 1) > 0 ?
2017 le32_to_cpu(lmm->lmm_stripe_count) : 0;
2019 handle = fsfilt_start_log(obd, dparent->d_inode,
2020 FSFILT_OP_UNLINK, NULL, sz);
2022 GOTO(cleanup, rc = PTR_ERR(handle));
2023 LOCK_INODE_MUTEX(dparent->d_inode);
2024 rc = ll_vfs_unlink(dparent->d_inode, dchild, mds->mds_vfsmnt);
2025 UNLOCK_INODE_MUTEX(dparent->d_inode);
2026 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
2034 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2037 GOTO(cleanup, rc = PTR_ERR(handle));
2038 LOCK_INODE_MUTEX(dparent->d_inode);
2039 rc = ll_vfs_unlink(dparent->d_inode, dchild, mds->mds_vfsmnt);
2040 UNLOCK_INODE_MUTEX(dparent->d_inode);
2041 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
2044 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2047 GOTO(cleanup, rc = -EINVAL);
2050 if (rc == 0 && child_inode->i_nlink == 0) {
2051 if (mds_orphan_needed(obd, child_inode))
2052 rc = mds_orphan_add_link(rec, obd, dchild);
2055 GOTO(cleanup, rc = 0);
2057 if (!S_ISREG(child_inode->i_mode))
2060 if (!(body->valid & OBD_MD_FLEASIZE)) {
2061 body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2062 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2063 } else if (mds_log_op_unlink(obd,
2064 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2065 lustre_msg_buflen(req->rq_repmsg, offset + 1),
2066 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2067 lustre_msg_buflen(req->rq_repmsg, offset+2)) >
2069 body->valid |= OBD_MD_FLCOOKIE;
2079 if (child_inode->i_nlink > 0 ||
2080 mds_orphan_open_count(child_inode) > 0) {
2081 /* update ctime of unlinked file only if it is still
2082 * opened or a link still exists */
2083 iattr.ia_valid = ATTR_CTIME;
2084 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2085 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
2087 CERROR("error on unlinked inode time update: "
2091 /* update mtime and ctime of parent directory*/
2092 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2093 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2094 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2095 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2097 CERROR("error on parent setattr: rc = %d\n", err);
2099 inodes[0] = dparent ? dparent->d_inode : NULL;
2100 inodes[1] = child_inode;
2101 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2104 switch(cleanup_phase) {
2105 case 5: /* pending_dir semaphore */
2106 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2107 case 4: /* child inode semaphore */
2108 MDS_UP_READ_ORPHAN_SEM(child_inode);
2109 case 3: /* child ino-reuse lock */
2110 if (rc && body != NULL) {
2111 // Don't unlink the OST objects if the MDS unlink failed
2115 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2117 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2118 case 2: /* child lock */
2119 ldlm_lock_decref(&child_lockh, LCK_EX);
2120 case 1: /* child and parent dentry, parent lock */
2122 ldlm_lock_decref(&parent_lockh, LCK_EX);
2124 ptlrpc_save_lock(req, &parent_lockh, LCK_EX);
2131 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2134 req->rq_status = rc;
2136 /* trigger dqrel on the owner of child and parent */
2137 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
2142 static int mds_reint_link(struct mds_update_record *rec, int offset,
2143 struct ptlrpc_request *req,
2144 struct lustre_handle *lh)
2146 struct obd_device *obd = req->rq_export->exp_obd;
2147 struct dentry *de_src = NULL;
2148 struct dentry *de_tgt_dir = NULL;
2149 struct dentry *dchild = NULL;
2150 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
2151 struct mds_obd *mds = mds_req2mds(req);
2152 struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
2153 struct ldlm_res_id src_res_id = { .name = {0} };
2154 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2155 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2156 ldlm_policy_data_t tgt_dir_policy =
2157 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2158 int rc = 0, cleanup_phase = 0;
2161 LASSERT(offset == REQ_REC_OFF);
2163 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2164 rec->ur_fid1->id, rec->ur_fid1->generation,
2165 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2167 mds_counter_incr(req->rq_export, LPROC_MDS_LINK);
2169 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2171 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2172 GOTO(cleanup, rc = -ENOENT);
2175 ldlm_request_cancel(req, rec->ur_dlm, 0);
2177 /* Step 1: Lookup the source inode and target directory by FID */
2178 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2180 GOTO(cleanup, rc = PTR_ERR(de_src));
2182 cleanup_phase = 1; /* source dentry */
2184 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2185 if (IS_ERR(de_tgt_dir)) {
2186 rc = PTR_ERR(de_tgt_dir);
2191 cleanup_phase = 2; /* target directory dentry */
2193 CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
2194 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2195 de_src->d_inode->i_ino);
2197 /* Step 2: Take the two locks */
2198 src_res_id.name[0] = de_src->d_inode->i_ino;
2199 src_res_id.name[1] = de_src->d_inode->i_generation;
2200 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2201 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2203 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2205 &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
2210 cleanup_phase = 3; /* locks */
2212 /* version recovery check */
2213 /* directory check */
2214 rc = mds_version_get_check(req, de_tgt_dir->d_inode, 0);
2216 GOTO(cleanup_no_trans, rc);
2217 /* inode version check */
2218 rc = mds_version_get_check(req, de_src->d_inode, 1);
2220 GOTO(cleanup_no_trans, rc);
2222 if (mds_inode_is_orphan(de_src->d_inode)) {
2223 CDEBUG(D_INODE, "an attempt to link an orphan inode %lu/%u\n",
2224 de_src->d_inode->i_ino,
2225 de_src->d_inode->i_generation);
2226 GOTO(cleanup, rc = -ENOENT);
2229 /* Step 3: Lookup the child */
2230 dchild = mds_lookup(obd, rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2231 if (IS_ERR(dchild)) {
2232 rc = PTR_ERR(dchild);
2234 if (rc != -EPERM && rc != -EACCES && rc != -ENAMETOOLONG)
2235 CERROR("child lookup error %d\n", rc);
2239 cleanup_phase = 4; /* child dentry */
2241 if (dchild->d_inode) {
2242 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2243 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2248 /* Step 4: Do it. */
2249 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2251 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2252 GOTO(cleanup, rc = -EROFS);
2254 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2256 GOTO(cleanup, rc = PTR_ERR(handle));
2258 LOCK_INODE_MUTEX(de_tgt_dir->d_inode);
2259 rc = ll_vfs_link(de_src, mds->mds_vfsmnt, de_tgt_dir->d_inode, dchild,
2261 UNLOCK_INODE_MUTEX(de_tgt_dir->d_inode);
2262 if (rc && rc != -EPERM && rc != -EACCES)
2263 CERROR("vfs_link error %d\n", rc);
2268 /* update ctime of old file */
2269 iattr.ia_valid = ATTR_CTIME;
2270 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2271 err = fsfilt_setattr(obd, de_src, handle, &iattr, 0);
2273 CERROR("error on old inode time update: "
2276 /* update mtime and ctime of target directory */
2277 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2278 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2279 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2280 err = fsfilt_setattr(obd, de_tgt_dir, handle, &iattr, 0);
2282 CERROR("error on target dir inode time update: "
2286 inodes[0] = de_tgt_dir ? de_tgt_dir->d_inode : NULL;
2287 inodes[1] = (dchild && !IS_ERR(dchild)) ? dchild->d_inode : NULL;
2288 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2292 switch (cleanup_phase) {
2293 case 4: /* child dentry */
2297 ldlm_lock_decref(&src_lockh, LCK_EX);
2298 ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
2300 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2301 ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
2303 case 2: /* target dentry */
2305 case 1: /* source dentry */
2310 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2313 req->rq_status = rc;
2317 /* The idea here is that we need to get four locks in the end:
2318 * one on each parent directory, one on each child. We need to take
2319 * these locks in some kind of order (to avoid deadlocks), and the order
2320 * I selected is "increasing resource number" order. We need to look up
2321 * the children, however, before we know what the resource number(s) are.
2322 * Thus the following plan:
2324 * 1,2. Look up the parents
2325 * 3,4. Look up the children
2326 * 5. Take locks on the parents and children, in order
2327 * 6. Verify that the children haven't changed since they were looked up
2329 * If there was a race and the children changed since they were first looked
2330 * up, it is possible that mds_verify_child() will be able to just grab the
2331 * lock on the new child resource (if it has a higher resource than any other)
2332 * but we need to compare against not only its parent, but also against the
2333 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2335 * We need the fancy igrab() on the child inodes because we aren't holding a
2336 * lock on the parent after the lookup is done, so dentry->d_inode may change
2337 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2339 int mds_get_parents_children_locked(struct obd_device *obd,
2340 struct mds_obd *mds,
2341 struct ll_fid *p1_fid,
2342 struct dentry **de_srcdirp,
2343 struct ll_fid *p2_fid,
2344 struct dentry **de_tgtdirp,
2346 const char *old_name, int old_len,
2347 struct dentry **de_oldp,
2348 const char *new_name, int new_len,
2349 struct dentry **de_newp,
2350 struct lustre_handle *dlm_handles,
2353 struct ldlm_res_id p1_res_id = { .name = {0} };
2354 struct ldlm_res_id p2_res_id = { .name = {0} };
2355 struct ldlm_res_id c1_res_id = { .name = {0} };
2356 struct ldlm_res_id c2_res_id = { .name = {0} };
2357 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2358 /* Only dentry should disappear, but the inode itself would be
2359 intact otherwise. */
2360 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2361 /* If something is going to be replaced, both dentry and inode locks are
2363 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2364 struct ldlm_res_id *maxres_src, *maxres_tgt;
2365 struct inode *inode;
2366 int rc = 0, cleanup_phase = 0;
2369 /* Step 1: Lookup the source directory */
2370 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2371 if (IS_ERR(*de_srcdirp))
2372 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2374 cleanup_phase = 1; /* source directory dentry */
2376 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2377 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2379 /* Step 2: Lookup the target directory */
2380 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2381 *de_tgtdirp = dget(*de_srcdirp);
2383 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2384 if (IS_ERR(*de_tgtdirp)) {
2385 rc = PTR_ERR(*de_tgtdirp);
2391 cleanup_phase = 2; /* target directory dentry */
2393 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2394 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2396 /* Step 3: Lookup the source child entry */
2397 *de_oldp = mds_lookup(obd, old_name, *de_srcdirp, old_len - 1);
2398 if (IS_ERR(*de_oldp)) {
2399 rc = PTR_ERR(*de_oldp);
2400 CDEBUG(D_INODE, "old child lookup error (%.*s): rc %d\n",
2401 old_len - 1, old_name, rc);
2405 cleanup_phase = 3; /* original name dentry */
2407 inode = (*de_oldp)->d_inode;
2409 inode = igrab(inode);
2411 GOTO(cleanup, rc = -ENOENT);
2413 c1_res_id.name[0] = inode->i_ino;
2414 c1_res_id.name[1] = inode->i_generation;
2418 /* Step 4: Lookup the target child entry */
2420 GOTO(retry_locks, rc);
2421 *de_newp = mds_lookup(obd, new_name, *de_tgtdirp, new_len - 1);
2422 if (IS_ERR(*de_newp)) {
2423 rc = PTR_ERR(*de_newp);
2424 CDEBUG(D_DENTRY, "new child lookup error (%.*s): rc %d\n",
2425 old_len - 1, old_name, rc);
2429 cleanup_phase = 4; /* target dentry */
2431 inode = (*de_newp)->d_inode;
2432 if (inode != NULL) {
2433 if (is_bad_inode(inode)) {
2434 CERROR("bad inode returned %lu/%u\n",
2435 inode->i_ino, inode->i_generation);
2436 GOTO(cleanup, rc = -ENOENT);
2438 inode = igrab(inode);
2443 c2_res_id.name[0] = inode->i_ino;
2444 c2_res_id.name[1] = inode->i_generation;
2448 /* Step 5: Take locks on the parents and child(ren) */
2449 maxres_src = &p1_res_id;
2450 maxres_tgt = &p2_res_id;
2451 cleanup_phase = 4; /* target dentry */
2453 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id,NULL,NULL))
2454 maxres_tgt = &c2_res_id;
2456 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2458 &p2_res_id, &dlm_handles[1], parent_mode,
2460 &c1_res_id, &dlm_handles[2], child_mode,
2462 &c2_res_id, &dlm_handles[3], child_mode,
2467 cleanup_phase = 6; /* parent and child(ren) locks */
2469 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2470 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2471 parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
2472 child_mode, &c1_policy, old_name, old_len,
2475 if (c2_res_id.name[0] != 0)
2476 ldlm_lock_decref(&dlm_handles[3], child_mode);
2477 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2484 if ((*de_oldp)->d_inode == NULL)
2485 GOTO(cleanup, rc = -ENOENT);
2490 /* Safe to skip check for child res being all zero */
2491 if (res_gt(&c1_res_id, maxres_src, NULL, NULL))
2492 maxres_src = &c1_res_id;
2494 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2495 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2496 parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
2497 child_mode, &c2_policy, new_name, new_len,
2500 ldlm_lock_decref(&dlm_handles[2], child_mode);
2501 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2511 switch (cleanup_phase) {
2512 case 6: /* child lock(s) */
2513 if (c2_res_id.name[0] != 0)
2514 ldlm_lock_decref(&dlm_handles[3], child_mode);
2515 if (c1_res_id.name[0] != 0)
2516 ldlm_lock_decref(&dlm_handles[2], child_mode);
2517 case 5: /* parent locks */
2518 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2519 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2520 case 4: /* target dentry */
2522 case 3: /* source dentry */
2524 case 2: /* target directory dentry */
2525 l_dput(*de_tgtdirp);
2526 case 1: /* source directry dentry */
2527 l_dput(*de_srcdirp);
2534 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2535 struct ptlrpc_request *req,
2536 struct lustre_handle *lockh)
2538 struct obd_device *obd = req->rq_export->exp_obd;
2539 struct dentry *de_srcdir = NULL;
2540 struct dentry *de_tgtdir = NULL;
2541 struct dentry *de_old = NULL;
2542 struct dentry *de_new = NULL;
2543 struct dentry *trap;
2544 struct inode *old_inode = NULL, *new_inode = NULL;
2545 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
2546 struct mds_obd *mds = mds_req2mds(req);
2547 struct lustre_handle dlm_handles[4];
2548 struct mds_body *body = NULL;
2549 struct lov_mds_md *lmm = NULL;
2550 int rc = 0, lock_count = 3, cleanup_phase = 0, sz;
2551 void *handle = NULL;
2552 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
2553 unsigned int qpids[4] = { 0, 0, 0, 0 };
2556 LASSERT(offset == REQ_REC_OFF);
2557 offset = REPLY_REC_OFF;
2559 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2560 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2561 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2563 mds_counter_incr(req->rq_export, LPROC_MDS_RENAME);
2565 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2568 ldlm_request_cancel(req, rec->ur_dlm, 0);
2570 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2571 rec->ur_fid2, &de_tgtdir, LCK_EX,
2572 rec->ur_name, rec->ur_namelen,
2573 &de_old, rec->ur_tgt,
2574 rec->ur_tgtlen, &de_new,
2575 dlm_handles, LCK_EX);
2579 cleanup_phase = 1; /* parent(s), children, locks */
2581 old_inode = de_old->d_inode;
2582 new_inode = de_new->d_inode;
2584 if (new_inode != NULL)
2587 /* version recovery check */
2588 rc = mds_version_get_check(req, de_srcdir->d_inode, 0);
2590 GOTO(cleanup_no_trans, rc);
2591 rc = mds_version_get_check(req, old_inode, 1);
2593 GOTO(cleanup_no_trans, rc);
2594 rc = mds_version_get_check(req, de_tgtdir->d_inode, 2);
2596 GOTO(cleanup_no_trans, rc);
2597 rc = mds_version_get_check(req, new_inode, 3);
2599 GOTO(cleanup_no_trans, rc);
2601 /* sanity check for src inode */
2602 if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
2603 old_inode->i_ino == de_tgtdir->d_inode->i_ino)
2604 GOTO(cleanup, rc = -EINVAL);
2606 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2607 GOTO(cleanup, rc = -EROFS);
2609 if (new_inode == NULL)
2613 cleanup_phase = 2; /* iput(new_inode) when finished */
2615 /* sanity check for dest inode */
2616 if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
2617 new_inode->i_ino == de_tgtdir->d_inode->i_ino)
2618 GOTO(cleanup, rc = -EINVAL);
2620 if (old_inode == new_inode)
2621 GOTO(cleanup, rc = 0);
2623 /* save uids/gids for qunit acquire/release */
2624 qcids[USRQUOTA] = old_inode->i_uid;
2625 qcids[GRPQUOTA] = old_inode->i_gid;
2626 qpids[USRQUOTA] = de_tgtdir->d_inode->i_uid;
2627 qpids[GRPQUOTA] = de_tgtdir->d_inode->i_gid;
2628 qpids[2] = de_srcdir->d_inode->i_uid;
2629 qpids[3] = de_srcdir->d_inode->i_gid;
2631 /* if we are about to remove the target at first, pass the EA of
2632 * that inode to client to perform and cleanup on OST */
2633 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
2634 LASSERT(body != NULL);
2636 /* child orphan sem protects orphan_dec_test && is_orphan race */
2637 MDS_DOWN_READ_ORPHAN_SEM(new_inode);
2638 cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
2640 if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
2641 new_inode->i_nlink == 1) {
2642 if (mds_orphan_needed(obd, new_inode)) {
2643 /* need to lock pending_dir before transaction */
2644 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2645 cleanup_phase = 4; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
2646 } else if (S_ISREG(new_inode->i_mode)) {
2647 mds_pack_inode2body(body, new_inode);
2648 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
2649 new_inode, MDS_PACK_MD_LOCK, 0,
2650 req->rq_export->exp_connect_flags);
2655 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_RENAME_WRITE,
2656 de_srcdir->d_inode->i_sb);
2658 lmm = lustre_msg_buf(req->rq_repmsg, offset + 1, 0);
2659 /* check that lmm size is not 0 */
2660 sz = lustre_msg_buflen(req->rq_repmsg, offset + 1) > 0 ?
2661 le32_to_cpu(lmm->lmm_stripe_count) : 0;
2663 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
2666 GOTO(cleanup, rc = PTR_ERR(handle));
2668 trap = lock_rename(de_tgtdir, de_srcdir);
2669 /* source should not be ancestor of target */
2670 if (de_old == trap) {
2671 unlock_rename(de_tgtdir, de_srcdir);
2672 GOTO(cleanup, rc = -EINVAL);
2674 /* target should not be an ancestor of source */
2675 if (de_new == trap) {
2676 unlock_rename(de_tgtdir, de_srcdir);
2677 GOTO(cleanup, rc = -ENOTEMPTY);
2680 de_old->d_fsdata = req;
2681 de_new->d_fsdata = req;
2683 rc = ll_vfs_rename(de_srcdir->d_inode, de_old, mds->mds_vfsmnt,
2684 de_tgtdir->d_inode, de_new, mds->mds_vfsmnt);
2686 unlock_rename(de_tgtdir, de_srcdir);
2692 /* update ctime of renamed file */
2693 iattr.ia_valid = ATTR_CTIME;
2694 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2695 if (S_ISDIR(de_old->d_inode->i_mode) &&
2696 de_srcdir->d_inode != de_tgtdir->d_inode) {
2697 /* cross directory rename of a directory, ".."
2698 changed, update mtime also */
2699 iattr.ia_valid |= ATTR_MTIME;
2700 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2702 err = fsfilt_setattr(obd, de_old, handle, &iattr, 0);
2704 CERROR("error on old inode time update: "
2707 if (de_new->d_inode) {
2708 /* target file exists, update its ctime as it
2710 iattr.ia_valid = ATTR_CTIME;
2711 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2712 err = fsfilt_setattr(obd, de_new, handle, &iattr, 0);
2714 CERROR("error on target inode time update: "
2718 /* update mtime and ctime of old directory */
2719 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2720 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2721 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2722 err = fsfilt_setattr(obd, de_srcdir, handle, &iattr, 0);
2724 CERROR("error on old dir inode update: "
2727 if (de_srcdir->d_inode != de_tgtdir->d_inode) {
2728 /* cross directory rename, update
2729 mtime and ctime of new directory */
2730 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2731 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2732 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2733 err = fsfilt_setattr(obd, de_tgtdir, handle, &iattr, 0);
2735 CERROR("error on new dir inode time update: "
2740 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2741 if (mds_orphan_needed(obd, new_inode))
2742 rc = mds_orphan_add_link(rec, obd, de_new);
2745 GOTO(cleanup, rc = 0);
2747 if (!S_ISREG(new_inode->i_mode))
2750 if (!(body->valid & OBD_MD_FLEASIZE)) {
2751 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2752 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2753 } else if (mds_log_op_unlink(obd,
2754 lustre_msg_buf(req->rq_repmsg,
2756 lustre_msg_buflen(req->rq_repmsg,
2758 lustre_msg_buf(req->rq_repmsg,
2760 lustre_msg_buflen(req->rq_repmsg,
2763 body->valid |= OBD_MD_FLCOOKIE;
2769 inodes[0] = de_srcdir && !IS_ERR(de_srcdir) ? de_srcdir->d_inode : NULL;
2770 inodes[1] = old_inode;
2771 inodes[2] = de_tgtdir && !IS_ERR(de_tgtdir) ? de_tgtdir->d_inode : NULL;
2772 inodes[3] = new_inode;
2773 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2776 switch (cleanup_phase) {
2778 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2780 MDS_UP_READ_ORPHAN_SEM(new_inode);
2785 if (lock_count == 4)
2786 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2787 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2788 ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
2789 ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
2791 if (lock_count == 4)
2792 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2793 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2794 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_EX);
2795 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_EX);
2804 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2807 req->rq_status = rc;
2809 /* acquire/release qunit */
2810 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
2815 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2816 struct ptlrpc_request *, struct lustre_handle *);
2818 static mds_reinter reinters[REINT_MAX] = {
2819 [REINT_SETATTR] mds_reint_setattr,
2820 [REINT_CREATE] mds_reint_create,
2821 [REINT_LINK] mds_reint_link,
2822 [REINT_UNLINK] mds_reint_unlink,
2823 [REINT_RENAME] mds_reint_rename,
2824 [REINT_OPEN] mds_open
2827 int mds_reint_rec(struct mds_update_record *rec, int offset,
2828 struct ptlrpc_request *req, struct lustre_handle *lockh)
2830 struct obd_device *obd = req->rq_export->exp_obd;
2831 struct mds_obd *mds = &obd->u.mds;
2832 struct lvfs_run_ctxt saved;
2835 gid_t fsgid = rec->ur_uc.luc_fsgid;
2840 if (req->rq_uid != LNET_UID_ANY) {
2841 /* non-root local cluster client
2842 * NB root's creds are believed... */
2843 LASSERT (req->rq_uid != 0);
2844 rec->ur_uc.luc_fsuid = req->rq_uid;
2845 cfs_kernel_cap_unpack(&rec->ur_uc.luc_cap, 0);
2849 /* get group info of this user */
2850 rec->ur_uc.luc_uce = upcall_cache_get_entry(mds->mds_group_hash,
2851 rec->ur_uc.luc_fsuid,
2852 rec->ur_uc.luc_fsgid, 2,
2853 &rec->ur_uc.luc_suppgid1);
2855 if (IS_ERR(rec->ur_uc.luc_uce)) {
2856 rc = PTR_ERR(rec->ur_uc.luc_uce);
2857 rec->ur_uc.luc_uce = NULL;
2861 /* checked by unpacker */
2862 LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2865 if (rec->ur_uc.luc_uce)
2866 rec->ur_uc.luc_fsgid = rec->ur_uc.luc_uce->ue_primary;
2869 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2872 if (rec->ur_uc.luc_uce && fsgid != rec->ur_uc.luc_fsgid &&
2873 in_group_p(fsgid)) {
2874 rec->ur_uc.luc_fsgid = fsgid;
2875 current->fsgid = saved.luc.luc_fsgid = fsgid;
2879 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2880 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2882 upcall_cache_put_entry(mds->mds_group_hash, rec->ur_uc.luc_uce);