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 LOCK_INODE_MUTEX(inode);
102 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR,
104 if (!IS_ERR(handle)) {
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);
111 UNLOCK_INODE_MUTEX(inode);
118 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
119 void *cb_data, int error)
121 struct mds_logcancel_data *mlcd = cb_data;
122 struct lov_stripe_md *lsm = NULL;
123 struct llog_ctxt *ctxt;
126 obd_transno_commit_cb(obd, transno, NULL, error);
128 CDEBUG(D_RPCTRACE, "cancelling %d cookies\n",
129 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
131 rc = obd_unpackmd(obd->u.mds.mds_lov_exp, &lsm, mlcd->mlcd_lmm,
132 mlcd->mlcd_eadatalen);
134 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
135 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
138 rc = obd_checkmd(obd->u.mds.mds_lov_exp, obd->obd_self_export,
141 CERROR("Can not revalidate lsm %p \n", lsm);
143 ctxt = llog_get_context(obd,mlcd->mlcd_cookies[0].lgc_subsys+1);
144 /* XXX 0 normally, SENDNOW for debug */
145 rc = llog_cancel(ctxt, lsm, mlcd->mlcd_cookielen /
146 sizeof(*mlcd->mlcd_cookies),
147 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW);
151 CERROR("error cancelling %d log cookies: rc %d\n",
152 (int)(mlcd->mlcd_cookielen /
153 sizeof(*mlcd->mlcd_cookies)), rc);
156 OBD_FREE(mlcd, mlcd->mlcd_size);
159 /* fsfilt_set_version return old version. use that here */
160 static void mds_versions_set(struct obd_device *obd,
161 struct inode **inodes, __u64 version)
168 for (i = 0; i < PTLRPC_NUM_VERSIONS; i++)
169 if (inodes[i] != NULL)
170 fsfilt_set_version(obd, inodes[i], version);
173 int mds_version_get_check(struct ptlrpc_request *req, struct inode *inode,
176 /* version recovery */
177 struct obd_device *obd = req->rq_export->exp_obd;
178 __u64 curr_version, *pre_versions;
181 if (inode == NULL || !exp_connect_vbr(req->rq_export))
184 curr_version = fsfilt_get_version(obd, inode);
185 if ((__s64)curr_version == -EOPNOTSUPP)
187 /* VBR: version is checked always because costs nothing */
188 if (lustre_msg_get_transno(req->rq_reqmsg) != 0) {
189 pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
190 LASSERT(index < PTLRPC_NUM_VERSIONS);
191 /* Sanity check for malformed buffers */
192 if (pre_versions == NULL) {
193 CERROR("No versions in request buffer\n");
194 spin_lock(&req->rq_export->exp_lock);
195 req->rq_export->exp_vbr_failed = 1;
196 spin_unlock(&req->rq_export->exp_lock);
198 } else if (pre_versions[index] != curr_version) {
199 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
200 pre_versions[index], curr_version);
201 spin_lock(&req->rq_export->exp_lock);
202 req->rq_export->exp_vbr_failed = 1;
203 spin_unlock(&req->rq_export->exp_lock);
207 /* save pre-versions in reply */
208 LASSERT(req->rq_repmsg != NULL);
209 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
211 pre_versions[index] = curr_version;
215 /* Assumes caller has already pushed us into the kernel context. */
216 int mds_finish_transno(struct mds_obd *mds, struct inode **inodes, void *handle,
217 struct ptlrpc_request *req, int rc, __u32 op_data,
220 struct mds_export_data *med = &req->rq_export->exp_mds_data;
221 struct lsd_client_data *lcd = med->med_lcd;
222 struct obd_device *obd = req->rq_export->exp_obd;
223 __u64 transno, prev_transno;
226 int log_pri = D_RPCTRACE;
227 struct inode *inode = inodes ? inodes[0] : NULL;
228 int version_set = handle ? 1 : 0;
231 if (IS_ERR(handle)) {
236 /* if the export has already been failed, we have no last_rcvd slot */
237 if (req->rq_export->exp_failed || obd->obd_fail) {
238 CWARN("commit transaction for disconnected client %s: rc %d\n",
239 req->rq_export->exp_client_uuid.uuid, rc);
247 if (handle == NULL) {
248 /* if we're starting our own xaction, use our own inode */
249 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
250 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
251 if (IS_ERR(handle)) {
252 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
253 RETURN(PTR_ERR(handle));
257 off = med->med_lr_off;
259 transno = lustre_msg_get_transno(req->rq_reqmsg);
262 CERROR("%s: replay %s transno "LPU64" failed: rc %d\n",
264 libcfs_nid2str(req->rq_export->exp_connection->c_peer.nid),
267 } else if (transno == 0) {
268 spin_lock(&mds->mds_transno_lock);
269 transno = ++mds->mds_last_transno;
270 spin_unlock(&mds->mds_transno_lock);
271 /* VBR: set versions */
272 if (inodes && version_set)
273 mds_versions_set(obd, inodes, transno);
275 spin_lock(&mds->mds_transno_lock);
276 if (transno > mds->mds_last_transno)
277 mds->mds_last_transno = transno;
278 spin_unlock(&mds->mds_transno_lock);
280 /* VBR: replay case. Copy version from replay req and
281 * set new versions */
282 mds_versions_set(obd, inodes, transno);
285 req->rq_transno = transno;
286 lustre_msg_set_transno(req->rq_repmsg, transno);
290 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
292 lcd->lcd_last_close_transno = cpu_to_le64(transno);
293 lcd->lcd_last_close_xid = cpu_to_le64(req->rq_xid);
294 lcd->lcd_last_close_result = cpu_to_le32(rc);
295 lcd->lcd_last_close_data = cpu_to_le32(op_data);
297 prev_transno = le64_to_cpu(lcd->lcd_last_transno);
298 if (((lustre_msg_get_flags(req->rq_reqmsg) &
299 (MSG_RESENT | MSG_REPLAY)) == 0) ||
300 (transno > prev_transno)) {
301 /* VBR: save versions in last_rcvd for reconstruct. */
302 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
304 lcd->lcd_pre_versions[0] = cpu_to_le64(pre_versions[0]);
305 lcd->lcd_pre_versions[1] = cpu_to_le64(pre_versions[1]);
306 lcd->lcd_pre_versions[2] = cpu_to_le64(pre_versions[2]);
307 lcd->lcd_pre_versions[3] = cpu_to_le64(pre_versions[3]);
310 lcd->lcd_last_transno = cpu_to_le64(transno);
311 lcd->lcd_last_xid = cpu_to_le64(req->rq_xid);
312 lcd->lcd_last_result = cpu_to_le32(rc);
313 lcd->lcd_last_data = cpu_to_le32(op_data);
316 /** update trans table */
317 target_trans_table_update(req->rq_export, transno);
320 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
323 struct obd_export *exp = req->rq_export;
325 class_export_get(exp); /* released once the cb is called */
327 force_sync = fsfilt_add_journal_cb(obd, transno,
328 handle, mds_commit_cb,
331 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, lcd,
333 force_sync | exp->exp_need_sync);
335 mds_commit_cb(obd, transno, exp, err);
344 DEBUG_REQ(log_pri, req,
345 "wrote trans #"LPU64" rc %d client %s at idx %u: err = %d",
346 transno, rc, lcd->lcd_uuid, med->med_lr_idx, err);
348 err = mds_lov_write_objids(obd);
354 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
357 err = fsfilt_commit(obd, inode, handle, 0);
359 CERROR("error committing transaction: %d\n", err);
367 /* this gives the same functionality as the code between
368 * sys_chmod and inode_setattr
369 * chown_common and inode_setattr
370 * utimes and inode_setattr
372 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
374 time_t now = CURRENT_SECONDS;
375 struct iattr *attr = &rec->ur_iattr;
376 unsigned int ia_valid = attr->ia_valid;
380 if (ia_valid & ATTR_RAW)
381 attr->ia_valid &= ~ATTR_RAW;
383 if (!(ia_valid & ATTR_CTIME_SET))
384 LTIME_S(attr->ia_ctime) = now;
386 attr->ia_valid &= ~ATTR_CTIME_SET;
387 if (!(ia_valid & ATTR_ATIME_SET))
388 LTIME_S(attr->ia_atime) = now;
389 if (!(ia_valid & ATTR_MTIME_SET))
390 LTIME_S(attr->ia_mtime) = now;
392 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
393 RETURN((attr->ia_valid & ~ATTR_ATTR_FLAG) ? -EPERM : 0);
396 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
397 if (current->fsuid != inode->i_uid &&
398 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
402 if (ia_valid & ATTR_SIZE &&
403 /* NFSD hack for open(O_CREAT|O_TRUNC)=mknod+truncate (bug 5781) */
404 !(rec->ur_uc.luc_fsuid == inode->i_uid &&
405 ia_valid & MDS_OPEN_OWNEROVERRIDE)) {
406 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
410 if (ia_valid & (ATTR_UID | ATTR_GID)) {
413 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
415 if (attr->ia_uid == (uid_t) -1)
416 attr->ia_uid = inode->i_uid;
417 if (attr->ia_gid == (gid_t) -1)
418 attr->ia_gid = inode->i_gid;
419 if (!(ia_valid & ATTR_MODE))
420 attr->ia_mode = inode->i_mode;
422 * If the user or group of a non-directory has been
423 * changed by a non-root user, remove the setuid bit.
424 * 19981026 David C Niemi <niemi@tux.org>
426 * Changed this to apply to all users, including root,
427 * to avoid some races. This is the behavior we had in
428 * 2.0. The check for non-root was definitely wrong
429 * for 2.2 anyway, as it should have been using
430 * CAP_FSETID rather than fsuid -- 19990830 SD.
432 if ((inode->i_mode & S_ISUID) == S_ISUID &&
433 !S_ISDIR(inode->i_mode)) {
434 attr->ia_mode &= ~S_ISUID;
435 attr->ia_valid |= ATTR_MODE;
438 * Likewise, if the user or group of a non-directory
439 * has been changed by a non-root user, remove the
440 * setgid bit UNLESS there is no group execute bit
441 * (this would be a file marked for mandatory
442 * locking). 19981026 David C Niemi <niemi@tux.org>
444 * Removed the fsuid check (see the comment above) --
447 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
448 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
449 attr->ia_mode &= ~S_ISGID;
450 attr->ia_valid |= ATTR_MODE;
452 } else if (ia_valid & ATTR_MODE) {
454 if (!(attr->ia_valid & ATTR_FORCE)) {
455 mode = inode->i_mode;
456 if (((mode & S_ISUID) && (!(attr->ia_mode & S_ISUID))) ||
457 ((mode & S_ISGID) && (mode & S_IXGRP) &&
458 (!(attr->ia_mode & S_ISGID))))
459 attr->ia_valid |= ATTR_FORCE;
461 mode = attr->ia_mode;
463 if (attr->ia_mode == (umode_t)-1)
464 mode = inode->i_mode;
466 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
471 void mds_steal_ack_locks(struct ptlrpc_request *req)
473 struct obd_export *exp = req->rq_export;
474 struct list_head *tmp;
475 struct ptlrpc_reply_state *oldrep;
476 struct ptlrpc_service *svc;
479 /* CAVEAT EMPTOR: spinlock order */
480 spin_lock(&exp->exp_lock);
481 list_for_each (tmp, &exp->exp_outstanding_replies) {
482 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
484 if (oldrep->rs_xid != req->rq_xid)
487 if (lustre_msg_get_opc(oldrep->rs_msg) !=
488 lustre_msg_get_opc(req->rq_reqmsg))
489 CERROR ("Resent req xid "LPU64" has mismatched opc: "
490 "new %d old %d\n", req->rq_xid,
491 lustre_msg_get_opc(req->rq_reqmsg),
492 lustre_msg_get_opc(oldrep->rs_msg));
494 svc = oldrep->rs_service;
495 spin_lock (&svc->srv_lock);
497 list_del_init (&oldrep->rs_exp_list);
499 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
501 oldrep->rs_nlocks, oldrep,
502 oldrep->rs_xid, oldrep->rs_transno,
503 lustre_msg_get_opc(oldrep->rs_msg),
504 libcfs_nid2str(exp->exp_connection->c_peer.nid));
506 for (i = 0; i < oldrep->rs_nlocks; i++)
507 ptlrpc_save_lock(req,
508 &oldrep->rs_locks[i],
509 oldrep->rs_modes[i]);
510 oldrep->rs_nlocks = 0;
512 DEBUG_REQ(D_DLMTRACE, req, "stole locks for");
513 ptlrpc_schedule_difficult_reply (oldrep);
515 spin_unlock (&svc->srv_lock);
518 spin_unlock(&exp->exp_lock);
522 * VBR: restore versions
524 void mds_vbr_reconstruct(struct ptlrpc_request *req,
525 struct lsd_client_data *lcd)
527 __u64 pre_versions[4] = {0};
528 pre_versions[0] = le64_to_cpu(lcd->lcd_pre_versions[0]);
529 pre_versions[1] = le64_to_cpu(lcd->lcd_pre_versions[1]);
530 pre_versions[2] = le64_to_cpu(lcd->lcd_pre_versions[2]);
531 pre_versions[3] = le64_to_cpu(lcd->lcd_pre_versions[3]);
532 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
535 void mds_req_from_lcd(struct ptlrpc_request *req, struct lsd_client_data *lcd)
537 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
538 req->rq_transno = le64_to_cpu(lcd->lcd_last_close_transno);
539 req->rq_status = le32_to_cpu(lcd->lcd_last_close_result);
541 req->rq_transno = le64_to_cpu(lcd->lcd_last_transno);
542 req->rq_status = le32_to_cpu(lcd->lcd_last_result);
543 mds_vbr_reconstruct(req, lcd);
545 if (req->rq_status != 0)
547 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
548 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
549 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
550 req->rq_transno, req->rq_status);
552 mds_steal_ack_locks(req);
555 static void reconstruct_reint_setattr(struct mds_update_record *rec,
556 int offset, struct ptlrpc_request *req)
558 struct obd_export *exp = req->rq_export;
559 struct mds_export_data *med = &exp->exp_mds_data;
560 struct mds_obd *obd = &exp->exp_obd->u.mds;
562 struct mds_body *body;
564 mds_req_from_lcd(req, med->med_lcd);
566 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
570 LCONSOLE_WARN("FID "LPU64"/%u lookup error %d."
571 " Evicting client %s with export %s.\n",
572 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
573 obd_uuid2str(&exp->exp_client_uuid),
574 obd_export_nid2str(exp));
575 mds_export_evict(exp);
580 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
581 mds_pack_inode2body(body, de->d_inode);
583 /* Don't return OST-specific attributes if we didn't just set them */
584 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
585 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
586 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
587 body->valid |= OBD_MD_FLMTIME;
588 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
589 body->valid |= OBD_MD_FLATIME;
594 int mds_osc_setattr_async(struct obd_device *obd, struct inode *inode,
595 struct lov_mds_md *lmm, int lmm_size,
596 struct llog_cookie *logcookies, struct ll_fid *fid)
598 struct mds_obd *mds = &obd->u.mds;
599 struct obd_trans_info oti = { 0 };
600 struct obd_info oinfo = { { { 0 } } };
604 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OST_SETATTR))
607 /* first get memory EA */
608 OBDO_ALLOC(oinfo.oi_oa);
614 rc = obd_unpackmd(mds->mds_lov_exp, &oinfo.oi_md, lmm, lmm_size);
616 CERROR("Error unpack md %p for inode %lu\n", lmm, inode->i_ino);
620 rc = obd_checkmd(mds->mds_lov_exp, obd->obd_self_export, oinfo.oi_md);
622 CERROR("Error revalidate lsm %p \n", oinfo.oi_md);
627 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
628 oinfo.oi_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
629 oinfo.oi_oa->o_id = oinfo.oi_md->lsm_object_id;
630 oinfo.oi_oa->o_gr = 0;
632 oinfo.oi_oa->o_valid |= OBD_MD_FLCOOKIE;
633 oti.oti_logcookies = logcookies;
636 LASSERT(fid != NULL);
637 oinfo.oi_oa->o_fid = fid->id;
638 oinfo.oi_oa->o_generation = fid->generation;
639 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
641 /* do async setattr from mds to ost not waiting for responses. */
642 rc = obd_setattr_async(mds->mds_lov_exp, &oinfo, &oti, NULL);
644 CDEBUG(D_INODE, "mds to ost setattr objid 0x"LPX64
645 " on ost error %d\n", oinfo.oi_md->lsm_object_id, rc);
648 obd_free_memmd(mds->mds_lov_exp, &oinfo.oi_md);
649 OBDO_FREE(oinfo.oi_oa);
653 /* In the raw-setattr case, we lock the child inode.
654 * In the write-back case or if being called from open, the client holds a lock
657 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
658 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
659 struct ptlrpc_request *req,
660 struct lustre_handle *lh)
662 unsigned int ia_valid = rec->ur_iattr.ia_valid;
663 struct mds_obd *mds = mds_req2mds(req);
664 struct obd_device *obd = req->rq_export->exp_obd;
665 struct mds_body *body;
667 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
668 struct inode *inode = NULL;
669 struct lustre_handle lockh;
671 struct mds_logcancel_data *mlcd = NULL;
672 struct lov_mds_md *lmm = NULL;
673 struct llog_cookie *logcookies = NULL;
674 int lmm_size = 0, need_lock = 1, cookie_size = 0;
675 int rc = 0, cleanup_phase = 0, err = 0, locked = 0, sync = 0;
676 int do_vbr = rec->ur_iattr.ia_valid &
677 (ATTR_MODE|ATTR_UID|ATTR_GID|
678 ATTR_FROM_OPEN|ATTR_RAW|ATTR_ATTR_FLAG);
679 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
680 unsigned int qpids[MAXQUOTAS] = { rec->ur_iattr.ia_uid,
681 rec->ur_iattr.ia_gid };
684 LASSERT(offset == REQ_REC_OFF);
685 offset = REPLY_REC_OFF;
687 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
688 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
689 OBD_COUNTER_INCREMENT(obd, setattr);
691 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
694 ldlm_request_cancel(req, rec->ur_dlm, 0);
696 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN ||
697 (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)) {
698 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
700 GOTO(cleanup, rc = PTR_ERR(de));
701 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
702 GOTO(cleanup, rc = -EROFS);
704 __u64 lockpart = MDS_INODELOCK_UPDATE;
705 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
706 lockpart |= MDS_INODELOCK_LOOKUP;
708 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX,
709 &lockh, NULL, 0, lockpart);
711 GOTO(cleanup, rc = PTR_ERR(de));
719 if ((rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) ||
720 (rec->ur_iattr.ia_valid & ATTR_SIZE)) {
721 /* Check write access for the O_TRUNC case */
722 if (mds_query_write_access(inode) < 0)
723 GOTO(cleanup, rc = -ETXTBSY);
726 /* save uid/gid for quota acq/rel */
727 qcids[USRQUOTA] = inode->i_uid;
728 qcids[GRPQUOTA] = inode->i_gid;
730 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
731 rec->ur_eadata != NULL) {
732 LOCK_INODE_MUTEX(inode);
736 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
738 /* VBR: update version if attr changed are important for recovery */
740 rc = mds_version_get_check(req, inode, 0);
742 GOTO(cleanup_no_trans, rc);
744 /* start a log jounal handle if needed */
745 if (S_ISREG(inode->i_mode) &&
746 rec->ur_iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
747 lmm_size = mds->mds_max_mdsize;
748 OBD_ALLOC(lmm, lmm_size);
750 GOTO(cleanup, rc = -ENOMEM);
753 rc = mds_get_md(obd, inode, lmm, &lmm_size, need_lock, 0,
754 req->rq_export->exp_connect_flags);
759 handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL,
760 le32_to_cpu(lmm->lmm_stripe_count));
762 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
765 GOTO(cleanup, rc = PTR_ERR(handle));
767 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
768 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
769 LTIME_S(rec->ur_iattr.ia_mtime),
770 LTIME_S(rec->ur_iattr.ia_ctime));
771 rc = mds_fix_attr(inode, rec);
775 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) { /* ioctl */
776 rc = fsfilt_iocontrol(obd, de, FSFILT_IOC_SETFLAGS,
777 (long)&rec->ur_flags);
778 } else if (rec->ur_iattr.ia_valid) { /* setattr */
779 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
780 /* journal chown/chgrp in llog, just like unlink */
781 if (rc == 0 && lmm_size){
782 cookie_size = mds_get_cookie_size(obd, lmm);
783 OBD_ALLOC(logcookies, cookie_size);
784 if (logcookies == NULL)
785 GOTO(cleanup, rc = -ENOMEM);
787 if (mds_log_op_setattr(obd, inode, lmm, lmm_size,
788 logcookies, cookie_size) <= 0) {
789 OBD_FREE(logcookies, cookie_size);
795 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
796 rec->ur_eadata != NULL) {
797 struct lov_stripe_md *lsm = NULL;
798 struct lov_user_md *lum = NULL;
800 rc = ll_permission(inode, MAY_WRITE, NULL);
804 lum = rec->ur_eadata;
805 /* if { size, offset, count } = { 0, -1, 0 } and no pool
806 * (i.e. all default values specified) then delete default
807 * striping from dir. */
808 if (S_ISDIR(inode->i_mode) &&
809 LOVEA_DELETE_VALUES(lum->lmm_stripe_size,
810 lum->lmm_stripe_count,
811 lum->lmm_stripe_offset) &&
812 (le32_to_cpu(lum->lmm_magic) != LOV_USER_MAGIC_V3)) {
813 rc = fsfilt_set_md(obd, inode, handle, NULL, 0, "lov");
817 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
819 &lsm, rec->ur_eadata);
823 obd_free_memmd(mds->mds_lov_exp, &lsm);
825 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
826 rec->ur_eadatalen, "lov");
832 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
833 mds_pack_inode2body(body, inode);
835 /* don't return OST-specific attributes if we didn't just set them. */
836 if (ia_valid & ATTR_SIZE)
837 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
838 if (ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
839 body->valid |= OBD_MD_FLMTIME;
840 if (ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
841 body->valid |= OBD_MD_FLATIME;
843 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_lov_obd)) {
844 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
847 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
849 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
850 mlcd->mlcd_cookielen = rec->ur_cookielen;
851 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
852 mlcd->mlcd_cookielen;
853 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
854 mlcd->mlcd_cookielen);
855 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
856 mlcd->mlcd_eadatalen);
858 CERROR("unable to allocate log cancel data\n");
864 sync = fsfilt_add_journal_cb(req->rq_export->exp_obd, 0, handle,
865 mds_cancel_cookies_cb, mlcd);
867 /* permission changes may require sync operation */
868 if (rc == 0 && ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
869 sync |= mds->mds_sync_permission;
871 err = mds_finish_transno(mds, do_vbr ? inodes : NULL, handle, req,
875 /* do mds to ost setattr if needed */
876 if (!rc && !err && lmm_size)
877 mds_osc_setattr_async(obd, inode, lmm, lmm_size,
878 logcookies, rec->ur_fid1);
880 switch (cleanup_phase) {
882 OBD_FREE(lmm, mds->mds_max_mdsize);
884 OBD_FREE(logcookies, cookie_size);
886 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
887 rec->ur_eadata != NULL)
888 UNLOCK_INODE_MUTEX(inode);
892 ldlm_lock_decref(&lockh, LCK_EX);
894 ptlrpc_save_lock (req, &lockh, LCK_EX);
907 /* trigger dqrel/dqacq for original owner and new owner */
908 if (ia_valid & (ATTR_UID | ATTR_GID))
909 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
915 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
916 struct ptlrpc_request *req)
918 struct obd_export *exp = req->rq_export;
919 struct mds_export_data *med = &exp->exp_mds_data;
920 struct mds_obd *obd = &exp->exp_obd->u.mds;
921 struct dentry *parent, *child;
922 struct mds_body *body;
925 mds_req_from_lcd(req, med->med_lcd);
930 parent = mds_fid2dentry(obd, rec->ur_fid1, NULL);
931 if (IS_ERR(parent)) {
932 rc = PTR_ERR(parent);
933 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
934 " Evicting client %s with export %s.\n",
935 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
936 obd_uuid2str(&exp->exp_client_uuid),
937 obd_export_nid2str(exp));
938 mds_export_evict(exp);
942 child = mds_lookup(exp->exp_obd, rec->ur_name, parent,
943 rec->ur_namelen - 1);
946 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
947 " Evicting client %s with export %s.\n",
948 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
949 obd_uuid2str(&exp->exp_client_uuid),
950 obd_export_nid2str(exp));
951 mds_export_evict(exp);
957 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
958 mds_pack_inode2body(body, child->d_inode);
964 static int mds_reint_create(struct mds_update_record *rec, int offset,
965 struct ptlrpc_request *req,
966 struct lustre_handle *lh)
968 struct dentry *dparent = NULL;
969 struct mds_obd *mds = mds_req2mds(req);
970 struct obd_device *obd = req->rq_export->exp_obd;
971 struct dentry *dchild = NULL;
972 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
973 struct inode *dir = NULL;
975 struct lustre_handle lockh;
976 int rc = 0, err = 0, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
978 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
979 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
980 unsigned int ids[MAXQUOTAS] = { 0, 0 };
981 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
982 int quota_pending[2] = {0, 0};
983 unsigned int gid = current->fsgid;
986 LASSERT(offset == REQ_REC_OFF);
987 offset = REPLY_REC_OFF;
989 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
992 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
993 rec->ur_fid1->id, rec->ur_fid1->generation,
994 rec->ur_name, rec->ur_mode);
996 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
998 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
999 GOTO(cleanup, rc = -ESTALE);
1002 ldlm_request_cancel(req, rec->ur_dlm, 0);
1004 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_EX, &lockh,
1005 rec->ur_name, rec->ur_namelen - 1,
1006 MDS_INODELOCK_UPDATE);
1007 if (IS_ERR(dparent)) {
1008 rc = PTR_ERR(dparent);
1009 CDEBUG(D_DENTRY, "parent "LPU64"/%u lookup error %d\n",
1010 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
1013 cleanup_phase = 1; /* locked parent dentry */
1014 dir = dparent->d_inode;
1017 ldlm_lock_dump_handle(D_OTHER, &lockh);
1019 dchild = mds_lookup(obd, rec->ur_name, dparent, rec->ur_namelen - 1);
1020 if (IS_ERR(dchild)) {
1021 rc = PTR_ERR(dchild);
1023 CDEBUG(D_DENTRY, "child lookup error %d\n", rc);
1027 cleanup_phase = 2; /* child dentry */
1029 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
1031 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) {
1032 if (dchild->d_inode)
1033 GOTO(cleanup, rc = -EEXIST);
1034 GOTO(cleanup, rc = -EROFS);
1037 /** check there is no stale orphan with same inode number */
1038 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1042 /* version recovery check */
1043 rc = mds_version_get_check(req, dir, 0);
1045 GOTO(cleanup_no_trans, rc);
1047 if (dir->i_mode & S_ISGID && S_ISDIR(rec->ur_mode))
1048 rec->ur_mode |= S_ISGID;
1050 dchild->d_fsdata = (void *)&dp;
1051 dp.ldp_inum = (unsigned long)rec->ur_fid2->id;
1054 if (dir->i_mode & S_ISGID)
1057 gid = current->fsgid;
1059 /* we try to get enough quota to write here, and let ldiskfs
1060 * decide if it is out of quota or not b=14783
1061 * FIXME: after CMD is used, pointer to obd_trans_info* couldn't
1062 * be NULL, b=14840 */
1063 ids[0] = current->fsuid;
1065 lquota_chkquota(mds_quota_interface_ref, req->rq_export, ids[0], ids[1],
1066 1, quota_pending, NULL, NULL, 0);
1070 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
1072 GOTO(cleanup, rc = PTR_ERR(handle));
1073 LOCK_INODE_MUTEX(dir);
1074 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
1075 UNLOCK_INODE_MUTEX(dir);
1076 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1081 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
1083 GOTO(cleanup, rc = PTR_ERR(handle));
1084 LOCK_INODE_MUTEX(dir);
1085 rc = ll_vfs_mkdir(dir, dchild, mds->mds_vfsmnt, rec->ur_mode);
1086 UNLOCK_INODE_MUTEX(dir);
1087 mds_counter_incr(req->rq_export, LPROC_MDS_MKDIR);
1092 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
1094 GOTO(cleanup, rc = PTR_ERR(handle));
1095 LOCK_INODE_MUTEX(dir);
1096 if (rec->ur_tgt == NULL) /* no target supplied */
1097 rc = -EINVAL; /* -EPROTO? */
1099 rc = ll_vfs_symlink(dir, dchild, mds->mds_vfsmnt,
1100 rec->ur_tgt, S_IALLUGO);
1101 UNLOCK_INODE_MUTEX(dir);
1102 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1110 int rdev = rec->ur_rdev;
1111 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
1113 GOTO(cleanup, rc = PTR_ERR(handle));
1114 LOCK_INODE_MUTEX(dir);
1115 rc = ll_vfs_mknod(dir, dchild, mds->mds_vfsmnt, rec->ur_mode,
1117 UNLOCK_INODE_MUTEX(dir);
1118 mds_counter_incr(req->rq_export, LPROC_MDS_MKNOD);
1123 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
1124 dchild->d_fsdata = NULL;
1125 GOTO(cleanup, rc = -EINVAL);
1128 /* In case we stored the desired inum in here, we want to clean up. */
1129 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
1130 dchild->d_fsdata = NULL;
1133 CDEBUG(D_INODE, "error during create: %d\n", rc);
1137 struct inode *inode = dchild->d_inode;
1138 struct mds_body *body;
1141 LTIME_S(iattr.ia_atime) = rec->ur_time;
1142 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1143 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1144 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1146 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1147 ATTR_MTIME | ATTR_CTIME;
1149 if (rec->ur_fid2->id) {
1150 if (rec->ur_fid2->id != inode->i_ino) {
1151 if (req->rq_export->exp_delayed)
1157 inode->i_generation = rec->ur_fid2->generation;
1158 /* Dirtied and committed by the upcoming setattr. */
1159 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
1160 inode->i_ino, inode->i_generation);
1162 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
1163 inode->i_ino, inode->i_generation);
1166 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1168 CERROR("error on child setattr: rc = %d\n", rc);
1170 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1171 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1173 CERROR("error on parent setattr: rc = %d\n", rc);
1175 if (S_ISDIR(inode->i_mode)) {
1176 struct lov_mds_md_v3 lmm;
1177 int lmm_size = sizeof(lmm);
1178 rc = mds_get_md(obd, dir, &lmm, &lmm_size, 1, 0,
1179 req->rq_export->exp_connect_flags);
1181 LOCK_INODE_MUTEX(inode);
1182 rc = fsfilt_set_md(obd, inode, handle,
1183 &lmm, lmm_size, "lov");
1184 UNLOCK_INODE_MUTEX(inode);
1187 CERROR("error on copy stripe info: rc = %d\n",
1191 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1192 mds_pack_inode2body(body, inode);
1198 inodes[1] = dchild ? dchild->d_inode : NULL;
1199 err = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
1202 if (quota_pending[0] || quota_pending[1])
1203 lquota_pending_commit(mds_quota_interface_ref, obd,
1204 ids[0], ids[1], quota_pending);
1206 if (rc && created) {
1207 /* Destroy the file we just created. This should not need
1208 * extra journal credits, as we have already modified all of
1209 * the blocks needed in order to create the file in the first
1214 LOCK_INODE_MUTEX(dir);
1215 err = ll_vfs_rmdir(dir, dchild, mds->mds_vfsmnt);
1216 UNLOCK_INODE_MUTEX(dir);
1218 CERROR("rmdir in error path: %d\n", err);
1221 LOCK_INODE_MUTEX(dir);
1222 err = ll_vfs_unlink(dir, dchild, mds->mds_vfsmnt);
1223 UNLOCK_INODE_MUTEX(dir);
1225 CERROR("unlink in error path: %d\n", err);
1228 } else if (created) {
1229 /* The inode we were allocated may have just been freed
1230 * by an unlink operation. We take this lock to
1231 * synchronize against the matching reply-ack-lock taken
1232 * in unlink, to avoid replay problems if this reply
1233 * makes it out to the client but the unlink's does not.
1234 * See bug 2029 for more detail.*/
1235 mds_lock_new_child(obd, dchild->d_inode, NULL);
1236 /* save uid/gid of create inode and parent */
1237 qpids[USRQUOTA] = dir->i_uid;
1238 qpids[GRPQUOTA] = dir->i_gid;
1243 switch (cleanup_phase) {
1244 case 2: /* child dentry */
1246 case 1: /* locked parent dentry */
1248 ldlm_lock_decref(&lockh, LCK_EX);
1250 ptlrpc_save_lock (req, &lockh, LCK_EX);
1256 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1259 req->rq_status = rc;
1261 /* trigger dqacq on the owner of child and parent */
1262 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1267 int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1268 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1272 for (i = 0; i < RES_NAME_SIZE; i++) {
1273 /* return 1 here, because enqueue_ordered will skip resources
1274 * of all zeroes if they're sorted to the end of the list. */
1275 if (res1->name[i] == 0 && res2->name[i] != 0)
1277 if (res2->name[i] == 0 && res1->name[i] != 0)
1280 if (res1->name[i] > res2->name[i])
1282 if (res1->name[i] < res2->name[i])
1287 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1292 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1293 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1294 * because they take the place of local semaphores.
1296 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1297 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1298 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1299 struct lustre_handle *p1_lockh, int p1_lock_mode,
1300 ldlm_policy_data_t *p1_policy,
1301 struct ldlm_res_id *p2_res_id,
1302 struct lustre_handle *p2_lockh, int p2_lock_mode,
1303 ldlm_policy_data_t *p2_policy)
1305 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1306 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1307 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1308 ldlm_policy_data_t *policies[2] = {p1_policy, p2_policy};
1312 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1314 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1315 res_id[0]->name[0], res_id[1]->name[0]);
1317 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1318 handles[1] = p1_lockh;
1319 handles[0] = p2_lockh;
1320 res_id[1] = p1_res_id;
1321 res_id[0] = p2_res_id;
1322 lock_modes[1] = p1_lock_mode;
1323 lock_modes[0] = p2_lock_mode;
1324 policies[1] = p1_policy;
1325 policies[0] = p2_policy;
1328 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1329 res_id[0]->name[0], res_id[1]->name[0]);
1331 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1332 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[0],
1333 LDLM_IBITS, policies[0], lock_modes[0],
1334 &flags, ldlm_blocking_ast,
1335 ldlm_completion_ast, NULL, NULL, 0,
1339 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1341 if (memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) == 0 &&
1342 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1343 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1344 ldlm_lock_addref(handles[1], lock_modes[1]);
1345 } else if (res_id[1]->name[0] != 0) {
1346 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1347 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id[1],
1348 LDLM_IBITS, policies[1],
1349 lock_modes[1], &flags,
1351 ldlm_completion_ast, NULL, NULL,
1352 0, NULL, handles[1]);
1353 if (rc != ELDLM_OK) {
1354 ldlm_lock_decref(handles[0], lock_modes[0]);
1357 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1363 static inline int res_eq(struct ldlm_res_id *res1, struct ldlm_res_id *res2)
1365 return !memcmp(res1, res2, sizeof(*res1));
1369 try_to_aggregate_locks(struct ldlm_res_id *res1, ldlm_policy_data_t *p1,
1370 struct ldlm_res_id *res2, ldlm_policy_data_t *p2)
1372 if (!res_eq(res1, res2))
1374 /* XXX: any additional inodebits (to current LOOKUP and UPDATE)
1375 * should be taken with great care here */
1376 p1->l_inodebits.bits |= p2->l_inodebits.bits;
1379 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1380 struct lustre_handle *p1_lockh, int p1_lock_mode,
1381 ldlm_policy_data_t *p1_policy,
1382 struct ldlm_res_id *p2_res_id,
1383 struct lustre_handle *p2_lockh, int p2_lock_mode,
1384 ldlm_policy_data_t *p2_policy,
1385 struct ldlm_res_id *c1_res_id,
1386 struct lustre_handle *c1_lockh, int c1_lock_mode,
1387 ldlm_policy_data_t *c1_policy,
1388 struct ldlm_res_id *c2_res_id,
1389 struct lustre_handle *c2_lockh, int c2_lock_mode,
1390 ldlm_policy_data_t *c2_policy)
1392 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1393 c1_res_id, c2_res_id };
1394 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1395 c1_lockh, c2_lockh };
1396 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1397 c1_lock_mode, c2_lock_mode };
1398 ldlm_policy_data_t *policies[5] = {p1_policy, p2_policy,
1399 c1_policy, c2_policy};
1400 int rc, i, j, sorted, flags;
1403 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1404 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1405 res_id[3]->name[0]);
1407 /* simple insertion sort - we have at most 4 elements */
1408 for (i = 1; i < 4; i++) {
1410 dlm_handles[4] = dlm_handles[i];
1411 res_id[4] = res_id[i];
1412 lock_modes[4] = lock_modes[i];
1413 policies[4] = policies[i];
1417 if (res_gt(res_id[j], res_id[4], policies[j],
1419 dlm_handles[j + 1] = dlm_handles[j];
1420 res_id[j + 1] = res_id[j];
1421 lock_modes[j + 1] = lock_modes[j];
1422 policies[j + 1] = policies[j];
1427 } while (j >= 0 && !sorted);
1429 dlm_handles[j + 1] = dlm_handles[4];
1430 res_id[j + 1] = res_id[4];
1431 lock_modes[j + 1] = lock_modes[4];
1432 policies[j + 1] = policies[4];
1435 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1436 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1437 res_id[3]->name[0]);
1439 /* XXX we could send ASTs on all these locks first before blocking? */
1440 for (i = 0; i < 4; i++) {
1441 flags = LDLM_FL_ATOMIC_CB;
1442 if (res_id[i]->name[0] == 0)
1444 if (i && res_eq(res_id[i], res_id[i-1])) {
1445 memcpy(dlm_handles[i], dlm_handles[i-1],
1446 sizeof(*(dlm_handles[i])));
1447 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1449 /* we need to enqueue locks with different inodebits
1450 * at once, because otherwise concurrent thread can
1451 * hit the windown between these two locks and we'll
1452 * get to deadlock. see bug 10360. note also, that it
1453 * is impossible to have >2 equal res. */
1455 try_to_aggregate_locks(res_id[i], policies[i],
1456 res_id[i+1], policies[i+1]);
1457 rc = ldlm_cli_enqueue_local(obd->obd_namespace,
1458 res_id[i], LDLM_IBITS,
1459 policies[i], lock_modes[i],
1460 &flags, ldlm_blocking_ast,
1461 ldlm_completion_ast, NULL,
1465 GOTO(out_err, rc = -EIO);
1466 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1473 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1478 /* In the unlikely case that the child changed while we were waiting
1479 * on the lock, we need to drop the lock on the old child and either:
1480 * - if the child has a lower resource name, then we have to also
1481 * drop the parent lock and regain the locks in the right order
1482 * - in the rename case, if the child has a lower resource name than one of
1483 * the other parent/child resources (maxres) we also need to reget the locks
1484 * - if the child has a higher resource name (this is the common case)
1485 * we can just get the lock on the new child (still in lock order)
1487 * Returns 0 if the child did not change or if it changed but could be locked.
1488 * Returns 1 if the child changed and we need to re-lock (no locks held).
1489 * Returns -ve error with a valid dchild (no locks held). */
1490 static int mds_verify_child(struct obd_device *obd,
1491 struct ldlm_res_id *parent_res_id,
1492 struct lustre_handle *parent_lockh,
1493 struct dentry *dparent, int parent_mode,
1494 struct ldlm_res_id *child_res_id,
1495 struct lustre_handle *child_lockh,
1496 struct dentry **dchildp, int child_mode,
1497 ldlm_policy_data_t *child_policy,
1498 const char *name, int namelen,
1499 struct ldlm_res_id *maxres)
1501 struct dentry *vchild, *dchild = *dchildp;
1502 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1505 /* not want child - not check it */
1509 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1511 GOTO(cleanup, rc = PTR_ERR(vchild));
1513 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1514 (vchild->d_inode != NULL &&
1515 child_res_id->name[0] == vchild->d_inode->i_ino &&
1516 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1523 /* resouce is changed, but not want child lock, return new child */
1524 if (child_lockh == NULL) {
1527 GOTO(cleanup, rc = 0);
1530 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1531 vchild->d_inode, dchild ? dchild->d_inode : 0,
1532 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1533 child_res_id->name[0]);
1534 if (child_res_id->name[0] != 0)
1535 ldlm_lock_decref(child_lockh, child_mode);
1539 cleanup_phase = 1; /* parent lock only */
1540 *dchildp = dchild = vchild;
1542 if (dchild->d_inode) {
1543 int flags = LDLM_FL_ATOMIC_CB;
1544 child_res_id->name[0] = dchild->d_inode->i_ino;
1545 child_res_id->name[1] = dchild->d_inode->i_generation;
1547 /* Make sure that we don't try to re-enqueue a lock on the
1548 * same resource if it happens that the source is renamed to
1549 * the target by another thread (bug 9974, thanks racer :-) */
1550 if (!res_gt(child_res_id, parent_res_id, NULL, NULL) ||
1551 !res_gt(child_res_id, maxres, NULL, NULL)) {
1552 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1553 child_res_id->name[0], parent_res_id->name[0],
1555 GOTO(cleanup, rc = 1);
1558 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1559 LDLM_IBITS, child_policy,
1562 ldlm_completion_ast, NULL,
1563 NULL, 0, NULL, child_lockh);
1565 GOTO(cleanup, rc = -EIO);
1567 memset(child_res_id, 0, sizeof(*child_res_id));
1568 memset(child_lockh, 0, sizeof(*child_lockh));
1574 switch(cleanup_phase) {
1576 if (child_res_id->name[0] != 0)
1577 ldlm_lock_decref(child_lockh, child_mode);
1579 ldlm_lock_decref(parent_lockh, parent_mode);
1585 #define INODE_CTIME_AGE (10)
1586 #define INODE_CTIME_OLD(inode) (LTIME_S(inode->i_ctime) + \
1587 INODE_CTIME_AGE < CURRENT_SECONDS)
1589 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1591 struct lustre_handle *parent_lockh,
1592 struct dentry **dparentp, int parent_mode,
1593 __u64 parent_lockpart,
1594 char *name, int namelen,
1595 struct lustre_handle *child_lockh,
1596 struct dentry **dchildp, int child_mode,
1597 __u64 child_lockpart)
1599 struct ldlm_res_id child_res_id = { .name = {0} };
1600 struct ldlm_res_id parent_res_id = { .name = {0} };
1601 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1602 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1603 static struct ldlm_res_id child_res_id_nolock = { .name = {0} };
1604 struct inode *inode;
1605 int rc = 0, cleanup_phase = 0;
1608 /* Step 1: Lookup parent */
1609 *dparentp = mds_fid2dentry(mds, fid, NULL);
1610 if (IS_ERR(*dparentp)) {
1611 rc = PTR_ERR(*dparentp);
1616 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1617 (*dparentp)->d_inode->i_ino, name);
1619 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1620 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1622 cleanup_phase = 1; /* parent dentry */
1625 GOTO(retry_locks, rc);
1626 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1627 *dchildp = mds_lookup(obd, name, *dparentp, namelen - 1);
1628 if (IS_ERR(*dchildp)) {
1629 rc = PTR_ERR(*dchildp);
1631 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1635 cleanup_phase = 2; /* child dentry */
1637 inode = (*dchildp)->d_inode;
1638 if (inode != NULL) {
1639 if (is_bad_inode(inode)) {
1640 CERROR("bad inode returned %lu/%u\n",
1641 inode->i_ino, inode->i_generation);
1642 GOTO(cleanup, rc = -ENOENT);
1644 inode = igrab(inode);
1649 child_res_id.name[0] = inode->i_ino;
1650 child_res_id.name[1] = inode->i_generation;
1652 /* If we want a LCK_CR for a directory, and this directory has not been
1653 changed for some time, we return not only a LOOKUP lock, but also an
1654 UPDATE lock to have negative dentry starts working for this dir.
1655 Also we apply same logic to non-directories. If the file is rarely
1656 changed - we return both locks and this might save us RPC on
1658 if ((child_mode & (LCK_CR|LCK_PR|LCK_CW)) && INODE_CTIME_OLD(inode))
1659 child_policy.l_inodebits.bits |= MDS_INODELOCK_UPDATE;
1664 cleanup_phase = 2; /* child dentry */
1666 /* Step 3: Lock parent and child in resource order. If child doesn't
1667 * exist, we still have to lock the parent and re-lookup. */
1668 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1670 child_lockh ? &child_res_id :
1671 &child_res_id_nolock,
1672 child_lockh, child_mode,
1677 if (IS_DEADDIR((*dparentp)->d_inode))
1678 GOTO(cleanup, -ENOENT);
1680 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1681 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1682 parent_mode, &child_res_id, child_lockh, dchildp,
1683 child_mode,&child_policy, name, namelen, &parent_res_id);
1692 switch (cleanup_phase) {
1703 void mds_reconstruct_generic(struct ptlrpc_request *req)
1705 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1707 mds_req_from_lcd(req, med->med_lcd);
1710 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1711 * we instead link the inode into the PENDING directory until it is
1712 * finally released. We can't simply call mds_reint_rename() or some
1713 * part thereof, because we don't have the inode to check for link
1714 * count/open status until after it is locked.
1716 * returns 1 on success
1717 * returns 0 if we lost a race and didn't make a new link
1718 * returns negative on error
1720 static int mds_orphan_add_link(struct mds_update_record *rec,
1721 struct obd_device *obd, struct dentry *dentry)
1723 struct mds_obd *mds = &obd->u.mds;
1724 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1725 struct inode *inode = dentry->d_inode;
1726 struct dentry *pending_child;
1727 char fidname[LL_FID_NAMELEN];
1728 int fidlen = 0, rc, mode;
1731 LASSERT(inode != NULL);
1732 LASSERT(!mds_inode_is_orphan(inode));
1733 LASSERT(TRYLOCK_INODE_MUTEX(pending_dir) == 0);
1735 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1737 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1738 mds_orphan_open_count(inode), inode->i_nlink,
1739 S_ISDIR(inode->i_mode) ? "dir" :
1740 S_ISREG(inode->i_mode) ? "file" : "other",rec->ur_name,fidname);
1742 if (!mds_orphan_needed(obd, inode) || inode->i_nlink != 0)
1745 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
1746 if (IS_ERR(pending_child))
1747 RETURN(PTR_ERR(pending_child));
1749 if (pending_child->d_inode != NULL) {
1750 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1751 LASSERT(pending_child->d_inode == inode);
1752 GOTO(out_dput, rc = 0);
1755 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1756 * for linking and return real mode back then -bzzz */
1757 mode = inode->i_mode;
1758 inode->i_mode = S_IFREG;
1759 /* avoid vfs_link upon 0 nlink inode, inc by 2 instead of 1 because
1760 * ext3_inc_count() can reset i_nlink for indexed directory */
1761 inode->i_nlink += 2;
1762 rc = ll_vfs_link(dentry, mds->mds_vfsmnt, pending_dir, pending_child,
1765 CERROR("error linking orphan %s %s to PENDING: rc = %d\n",
1766 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1769 mds_inode_set_orphan(inode);
1771 /* return mode and correct i_nlink if inode is directory */
1772 inode->i_mode = mode;
1773 LASSERTF(rc || inode->i_nlink == 3, "%s nlink == %d\n",
1774 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1777 if (S_ISDIR(mode)) {
1778 pending_dir->i_nlink++;
1779 if (pending_dir->i_sb->s_op->dirty_inode)
1780 pending_dir->i_sb->s_op->dirty_inode(pending_dir);
1783 inode->i_nlink -= 2;
1784 if (inode->i_sb->s_op->dirty_inode)
1785 inode->i_sb->s_op->dirty_inode(inode);
1790 GOTO(out_dput, rc = 1);
1792 l_dput(pending_child);
1796 int mds_get_cookie_size(struct obd_device *obd, struct lov_mds_md *lmm)
1798 int count = le32_to_cpu(lmm->lmm_stripe_count);
1799 int real_csize = count * sizeof(struct llog_cookie);
1803 static void mds_shrink_reply(struct ptlrpc_request *req,
1804 int reply_mdoff, int have_md, int have_acl)
1806 struct obd_device *obd = req->rq_export->exp_obd;
1807 struct mds_body *reply_body;
1808 int cookie_size = 0, md_size = 0;
1811 /* LSM and cookie is always placed after mds_body */
1812 reply_body = lustre_msg_buf(req->rq_repmsg, reply_mdoff,
1813 sizeof(*reply_body));
1814 if (reply_body == NULL)
1815 /* if there is no mds_body, no point in shrinking the reply */
1820 if (have_md || have_acl) {
1821 if (reply_body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) {
1822 md_size = reply_body->eadatasize;
1823 } else if (reply_body->valid & OBD_MD_LINKNAME)
1824 md_size = reply_body->eadatasize;
1826 if (reply_body->valid & OBD_MD_FLCOOKIE) {
1827 LASSERT(reply_body->valid & OBD_MD_FLEASIZE);
1828 cookie_size = mds_get_cookie_size(obd, lustre_msg_buf(
1831 } else if (reply_body->valid & OBD_MD_FLACL) {
1832 cookie_size = reply_body->aclsize;
1835 CDEBUG(D_INFO, "Shrink %d/%d to md_size %d cookie_size %d \n",
1836 have_md, have_acl, md_size, cookie_size);
1838 if (likely(have_md))
1839 lustre_shrink_reply(req, reply_mdoff, md_size, 1);
1841 if (likely(have_acl))
1842 lustre_shrink_reply(req, reply_mdoff + (md_size > 0),
1846 void mds_shrink_body_reply(struct ptlrpc_request *req,
1850 struct mds_body *rq_body;
1851 const long long have_acl = OBD_MD_FLCOOKIE | OBD_MD_FLACL;
1852 const long have_md = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
1855 /* LSM and cookie is always placed after mds_body */
1856 rq_body = lustre_msg_buf(req->rq_reqmsg, req_mdoff,
1860 /* this check is need for avoid hit asset in case
1861 * OBD_MDS_FLFLAGS */
1862 mds_shrink_reply(req, reply_mdoff,
1863 rq_body->valid & have_md,
1864 rq_body->valid & have_acl);
1867 void mds_shrink_intent_reply(struct ptlrpc_request *req,
1868 int opc, int reply_mdoff)
1873 mds_shrink_reply(req, reply_mdoff, 1, 1);
1876 mds_shrink_reply(req, reply_mdoff, 1, 0);
1883 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1884 struct ptlrpc_request *req,
1885 struct lustre_handle *lh)
1887 struct dentry *dparent = NULL, *dchild;
1888 struct mds_obd *mds = mds_req2mds(req);
1889 struct obd_device *obd = req->rq_export->exp_obd;
1890 struct mds_body *body = NULL;
1891 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
1892 struct inode *child_inode = NULL;
1893 struct lustre_handle parent_lockh, child_lockh, child_reuse_lockh;
1894 void *handle = NULL;
1895 int rc = 0, cleanup_phase = 0;
1896 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
1897 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1900 LASSERT(offset == REQ_REC_OFF); /* || offset == DLM_INTENT_REC_OFF); */
1901 offset = REPLY_REC_OFF;
1903 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1904 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1906 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1908 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1909 GOTO(cleanup, rc = -ENOENT);
1912 ldlm_request_cancel(req, rec->ur_dlm, 0);
1914 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1915 &parent_lockh, &dparent, LCK_EX,
1916 MDS_INODELOCK_UPDATE,
1917 rec->ur_name, rec->ur_namelen,
1918 &child_lockh, &dchild, LCK_EX,
1919 MDS_INODELOCK_FULL);
1923 cleanup_phase = 1; /* dchild, dparent, locks */
1926 child_inode = dchild->d_inode;
1927 if (child_inode == NULL) {
1928 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1929 dparent->d_inode->i_ino, rec->ur_name);
1930 GOTO(cleanup, rc = -ENOENT);
1933 /* save uid/gid for quota acquire/release */
1934 qcids[USRQUOTA] = child_inode->i_uid;
1935 qcids[GRPQUOTA] = child_inode->i_gid;
1936 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1937 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1939 cleanup_phase = 2; /* dchild has a lock */
1941 /* VBR: version recovery check for parent */
1942 rc = mds_version_get_check(req, dparent->d_inode, 0);
1944 GOTO(cleanup_no_trans, rc);
1946 /* version recovery check */
1947 rc = mds_version_get_check(req, child_inode, 1);
1949 GOTO(cleanup_no_trans, rc);
1951 /* We have to do these checks ourselves, in case we are making an
1952 * orphan. The client tells us whether rmdir() or unlink() was called,
1953 * so we need to return appropriate errors (bug 72). */
1954 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1955 if (!S_ISDIR(child_inode->i_mode))
1956 GOTO(cleanup, rc = -ENOTDIR);
1958 if (S_ISDIR(child_inode->i_mode))
1959 GOTO(cleanup, rc = -EISDIR);
1962 /* Check for EROFS after we check ENODENT, ENOTDIR, and EISDIR */
1963 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1964 GOTO(cleanup, rc = -EROFS);
1966 /* Step 3: Get a lock on the ino to sync with creation WRT inode
1967 * reuse (see bug 2029). */
1968 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1972 cleanup_phase = 3; /* child inum lock */
1974 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1976 /* ldlm_reply in buf[0] if called via intent */
1977 if (offset == DLM_INTENT_REC_OFF)
1978 offset = DLM_REPLY_REC_OFF;
1980 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
1981 LASSERT(body != NULL);
1983 /* child orphan sem protects orphan_dec_test && is_orphan race */
1984 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
1985 cleanup_phase = 4; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
1987 /* If this is potentially the last reference to this inode, get the
1988 * OBD EA data first so the client can destroy OST objects. We
1989 * only do the object removal later if no open files/links remain. */
1990 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
1991 child_inode->i_nlink == 1) {
1992 if (mds_orphan_needed(obd, child_inode)) {
1993 /* need to lock pending_dir before transaction */
1994 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
1995 cleanup_phase = 5; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
1996 } else if (S_ISREG(child_inode->i_mode)) {
1997 mds_pack_inode2body(body, child_inode);
1998 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1999 child_inode, MDS_PACK_MD_LOCK, 0,
2000 req->rq_export->exp_connect_flags);
2004 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2005 switch (child_inode->i_mode & S_IFMT) {
2007 /* Drop any lingering child directories before we start our
2008 * transaction, to avoid doing multiple inode dirty/delete
2009 * in our compound transaction (bug 1321). */
2010 shrink_dcache_parent(dchild);
2011 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2014 GOTO(cleanup, rc = PTR_ERR(handle));
2015 LOCK_INODE_MUTEX(dparent->d_inode);
2016 rc = ll_vfs_rmdir(dparent->d_inode, dchild, mds->mds_vfsmnt);
2017 UNLOCK_INODE_MUTEX(dparent->d_inode);
2018 mds_counter_incr(req->rq_export, LPROC_MDS_RMDIR);
2021 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2023 int sz = lustre_msg_buflen(req->rq_repmsg, offset + 1) > 0 ?
2024 le32_to_cpu(lmm->lmm_stripe_count) : 0;
2026 handle = fsfilt_start_log(obd, dparent->d_inode,
2027 FSFILT_OP_UNLINK, NULL, sz);
2029 GOTO(cleanup, rc = PTR_ERR(handle));
2030 LOCK_INODE_MUTEX(dparent->d_inode);
2031 rc = ll_vfs_unlink(dparent->d_inode, dchild, mds->mds_vfsmnt);
2032 UNLOCK_INODE_MUTEX(dparent->d_inode);
2033 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
2041 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2044 GOTO(cleanup, rc = PTR_ERR(handle));
2045 LOCK_INODE_MUTEX(dparent->d_inode);
2046 rc = ll_vfs_unlink(dparent->d_inode, dchild, mds->mds_vfsmnt);
2047 UNLOCK_INODE_MUTEX(dparent->d_inode);
2048 mds_counter_incr(req->rq_export, LPROC_MDS_UNLINK);
2051 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2054 GOTO(cleanup, rc = -EINVAL);
2057 if (rc == 0 && child_inode->i_nlink == 0) {
2058 if (mds_orphan_needed(obd, child_inode)) {
2059 rc = mds_orphan_add_link(rec, obd, dchild);
2061 /* child inode was successfully linked
2063 GOTO(cleanup, rc = 0);
2065 /* we failed to move the file to PENDING,
2066 * really unlink the file as if there were
2067 * no more openers */
2071 if (!S_ISREG(child_inode->i_mode))
2074 if (!(body->valid & OBD_MD_FLEASIZE)) {
2075 body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2076 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2077 } else if (mds_log_op_unlink(obd,
2078 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2079 lustre_msg_buflen(req->rq_repmsg, offset + 1),
2080 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2081 lustre_msg_buflen(req->rq_repmsg, offset+2)) >
2083 body->valid |= OBD_MD_FLCOOKIE;
2093 if (child_inode->i_nlink > 0 ||
2094 mds_orphan_open_count(child_inode) > 0) {
2095 /* update ctime of unlinked file only if it is still
2096 * opened or a link still exists */
2097 iattr.ia_valid = ATTR_CTIME;
2098 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2099 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
2101 CERROR("error on unlinked inode time update: "
2105 /* update mtime and ctime of parent directory*/
2106 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2107 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2108 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2109 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2111 CERROR("error on parent setattr: rc = %d\n", err);
2113 inodes[0] = dparent ? dparent->d_inode : NULL;
2114 inodes[1] = child_inode;
2115 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2118 switch(cleanup_phase) {
2119 case 5: /* pending_dir semaphore */
2120 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2121 case 4: /* child inode semaphore */
2122 MDS_UP_READ_ORPHAN_SEM(child_inode);
2123 case 3: /* child ino-reuse lock */
2124 if (rc && body != NULL) {
2125 // Don't unlink the OST objects if the MDS unlink failed
2129 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2131 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2132 case 2: /* child lock */
2133 ldlm_lock_decref(&child_lockh, LCK_EX);
2134 case 1: /* child and parent dentry, parent lock */
2136 ldlm_lock_decref(&parent_lockh, LCK_EX);
2138 ptlrpc_save_lock(req, &parent_lockh, LCK_EX);
2145 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2148 req->rq_status = rc;
2150 /* trigger dqrel on the owner of child and parent */
2151 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
2156 static int mds_reint_link(struct mds_update_record *rec, int offset,
2157 struct ptlrpc_request *req,
2158 struct lustre_handle *lh)
2160 struct obd_device *obd = req->rq_export->exp_obd;
2161 struct dentry *de_src = NULL;
2162 struct dentry *de_tgt_dir = NULL;
2163 struct dentry *dchild = NULL;
2164 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
2165 struct mds_obd *mds = mds_req2mds(req);
2166 struct lustre_handle *handle = NULL, tgt_dir_lockh, src_lockh;
2167 struct ldlm_res_id src_res_id = { .name = {0} };
2168 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2169 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2170 ldlm_policy_data_t tgt_dir_policy =
2171 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2172 int rc = 0, cleanup_phase = 0;
2175 LASSERT(offset == REQ_REC_OFF);
2177 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2178 rec->ur_fid1->id, rec->ur_fid1->generation,
2179 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2181 mds_counter_incr(req->rq_export, LPROC_MDS_LINK);
2183 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2185 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2186 GOTO(cleanup, rc = -ENOENT);
2189 ldlm_request_cancel(req, rec->ur_dlm, 0);
2191 /* Step 1: Lookup the source inode and target directory by FID */
2192 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2194 GOTO(cleanup, rc = PTR_ERR(de_src));
2196 cleanup_phase = 1; /* source dentry */
2198 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2199 if (IS_ERR(de_tgt_dir)) {
2200 rc = PTR_ERR(de_tgt_dir);
2205 cleanup_phase = 2; /* target directory dentry */
2207 CDEBUG(D_INODE, "linking %.*s/%s to inode %lu\n",
2208 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2209 de_src->d_inode->i_ino);
2211 /* Step 2: Take the two locks */
2212 src_res_id.name[0] = de_src->d_inode->i_ino;
2213 src_res_id.name[1] = de_src->d_inode->i_generation;
2214 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2215 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2217 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2219 &tgt_dir_res_id, &tgt_dir_lockh, LCK_EX,
2224 cleanup_phase = 3; /* locks */
2226 /* version recovery check */
2227 /* directory check */
2228 rc = mds_version_get_check(req, de_tgt_dir->d_inode, 0);
2230 GOTO(cleanup_no_trans, rc);
2231 /* inode version check */
2232 rc = mds_version_get_check(req, de_src->d_inode, 1);
2234 GOTO(cleanup_no_trans, rc);
2236 if (mds_inode_is_orphan(de_src->d_inode)) {
2237 CDEBUG(D_INODE, "an attempt to link an orphan inode %lu/%u\n",
2238 de_src->d_inode->i_ino,
2239 de_src->d_inode->i_generation);
2240 GOTO(cleanup, rc = -ENOENT);
2243 /* Step 3: Lookup the child */
2244 dchild = mds_lookup(obd, rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2245 if (IS_ERR(dchild)) {
2246 rc = PTR_ERR(dchild);
2248 if (rc != -EPERM && rc != -EACCES && rc != -ENAMETOOLONG)
2249 CERROR("child lookup error %d\n", rc);
2253 cleanup_phase = 4; /* child dentry */
2255 if (dchild->d_inode) {
2256 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2257 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2262 /* Step 4: Do it. */
2263 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2265 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2266 GOTO(cleanup, rc = -EROFS);
2268 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2270 GOTO(cleanup, rc = PTR_ERR(handle));
2272 LOCK_INODE_MUTEX(de_tgt_dir->d_inode);
2273 rc = ll_vfs_link(de_src, mds->mds_vfsmnt, de_tgt_dir->d_inode, dchild,
2275 UNLOCK_INODE_MUTEX(de_tgt_dir->d_inode);
2276 if (rc && rc != -EPERM && rc != -EACCES)
2277 CERROR("vfs_link error %d\n", rc);
2282 /* update ctime of old file */
2283 iattr.ia_valid = ATTR_CTIME;
2284 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2285 err = fsfilt_setattr(obd, de_src, handle, &iattr, 0);
2287 CERROR("error on old inode time update: "
2290 /* update mtime and ctime of target directory */
2291 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2292 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2293 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2294 err = fsfilt_setattr(obd, de_tgt_dir, handle, &iattr, 0);
2296 CERROR("error on target dir inode time update: "
2300 inodes[0] = de_tgt_dir ? de_tgt_dir->d_inode : NULL;
2301 inodes[1] = (dchild && !IS_ERR(dchild)) ? dchild->d_inode : NULL;
2302 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2306 switch (cleanup_phase) {
2307 case 4: /* child dentry */
2311 ldlm_lock_decref(&src_lockh, LCK_EX);
2312 ldlm_lock_decref(&tgt_dir_lockh, LCK_EX);
2314 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2315 ptlrpc_save_lock(req, &tgt_dir_lockh, LCK_EX);
2317 case 2: /* target dentry */
2319 case 1: /* source dentry */
2324 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2327 req->rq_status = rc;
2331 /* The idea here is that we need to get four locks in the end:
2332 * one on each parent directory, one on each child. We need to take
2333 * these locks in some kind of order (to avoid deadlocks), and the order
2334 * I selected is "increasing resource number" order. We need to look up
2335 * the children, however, before we know what the resource number(s) are.
2336 * Thus the following plan:
2338 * 1,2. Look up the parents
2339 * 3,4. Look up the children
2340 * 5. Take locks on the parents and children, in order
2341 * 6. Verify that the children haven't changed since they were looked up
2343 * If there was a race and the children changed since they were first looked
2344 * up, it is possible that mds_verify_child() will be able to just grab the
2345 * lock on the new child resource (if it has a higher resource than any other)
2346 * but we need to compare against not only its parent, but also against the
2347 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2349 * We need the fancy igrab() on the child inodes because we aren't holding a
2350 * lock on the parent after the lookup is done, so dentry->d_inode may change
2351 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2353 int mds_get_parents_children_locked(struct obd_device *obd,
2354 struct mds_obd *mds,
2355 struct ll_fid *p1_fid,
2356 struct dentry **de_srcdirp,
2357 struct ll_fid *p2_fid,
2358 struct dentry **de_tgtdirp,
2360 const char *old_name, int old_len,
2361 struct dentry **de_oldp,
2362 const char *new_name, int new_len,
2363 struct dentry **de_newp,
2364 struct lustre_handle *dlm_handles,
2367 struct ldlm_res_id p1_res_id = { .name = {0} };
2368 struct ldlm_res_id p2_res_id = { .name = {0} };
2369 struct ldlm_res_id c1_res_id = { .name = {0} };
2370 struct ldlm_res_id c2_res_id = { .name = {0} };
2371 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2372 /* Only dentry should disappear, but the inode itself would be
2373 intact otherwise. */
2374 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2375 /* If something is going to be replaced, both dentry and inode locks are
2377 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_FULL}};
2378 struct ldlm_res_id *maxres_src, *maxres_tgt;
2379 struct inode *inode;
2380 int rc = 0, cleanup_phase = 0;
2383 /* Step 1: Lookup the source directory */
2384 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2385 if (IS_ERR(*de_srcdirp))
2386 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2388 cleanup_phase = 1; /* source directory dentry */
2390 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2391 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2393 /* Step 2: Lookup the target directory */
2394 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2395 *de_tgtdirp = dget(*de_srcdirp);
2397 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2398 if (IS_ERR(*de_tgtdirp)) {
2399 rc = PTR_ERR(*de_tgtdirp);
2405 cleanup_phase = 2; /* target directory dentry */
2407 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2408 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2410 /* Step 3: Lookup the source child entry */
2411 *de_oldp = mds_lookup(obd, old_name, *de_srcdirp, old_len - 1);
2412 if (IS_ERR(*de_oldp)) {
2413 rc = PTR_ERR(*de_oldp);
2414 CDEBUG(D_INODE, "old child lookup error (%.*s): rc %d\n",
2415 old_len - 1, old_name, rc);
2419 cleanup_phase = 3; /* original name dentry */
2421 inode = (*de_oldp)->d_inode;
2423 inode = igrab(inode);
2425 GOTO(cleanup, rc = -ENOENT);
2427 c1_res_id.name[0] = inode->i_ino;
2428 c1_res_id.name[1] = inode->i_generation;
2432 /* Step 4: Lookup the target child entry */
2434 GOTO(retry_locks, rc);
2435 *de_newp = mds_lookup(obd, new_name, *de_tgtdirp, new_len - 1);
2436 if (IS_ERR(*de_newp)) {
2437 rc = PTR_ERR(*de_newp);
2438 CDEBUG(D_DENTRY, "new child lookup error (%.*s): rc %d\n",
2439 old_len - 1, old_name, rc);
2443 cleanup_phase = 4; /* target dentry */
2445 inode = (*de_newp)->d_inode;
2446 if (inode != NULL) {
2447 if (is_bad_inode(inode)) {
2448 CERROR("bad inode returned %lu/%u\n",
2449 inode->i_ino, inode->i_generation);
2450 GOTO(cleanup, rc = -ENOENT);
2452 inode = igrab(inode);
2457 c2_res_id.name[0] = inode->i_ino;
2458 c2_res_id.name[1] = inode->i_generation;
2462 /* Step 5: Take locks on the parents and child(ren) */
2463 maxres_src = &p1_res_id;
2464 maxres_tgt = &p2_res_id;
2465 cleanup_phase = 4; /* target dentry */
2467 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id,NULL,NULL))
2468 maxres_tgt = &c2_res_id;
2470 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2472 &p2_res_id, &dlm_handles[1], parent_mode,
2474 &c1_res_id, &dlm_handles[2], child_mode,
2476 &c2_res_id, &dlm_handles[3], child_mode,
2481 cleanup_phase = 6; /* parent and child(ren) locks */
2483 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2484 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2485 parent_mode, &c1_res_id, &dlm_handles[2], de_oldp,
2486 child_mode, &c1_policy, old_name, old_len,
2489 if (c2_res_id.name[0] != 0)
2490 ldlm_lock_decref(&dlm_handles[3], child_mode);
2491 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2498 if ((*de_oldp)->d_inode == NULL)
2499 GOTO(cleanup, rc = -ENOENT);
2504 /* Safe to skip check for child res being all zero */
2505 if (res_gt(&c1_res_id, maxres_src, NULL, NULL))
2506 maxres_src = &c1_res_id;
2508 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2509 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2510 parent_mode, &c2_res_id, &dlm_handles[3], de_newp,
2511 child_mode, &c2_policy, new_name, new_len,
2514 ldlm_lock_decref(&dlm_handles[2], child_mode);
2515 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2525 switch (cleanup_phase) {
2526 case 6: /* child lock(s) */
2527 if (c2_res_id.name[0] != 0)
2528 ldlm_lock_decref(&dlm_handles[3], child_mode);
2529 if (c1_res_id.name[0] != 0)
2530 ldlm_lock_decref(&dlm_handles[2], child_mode);
2531 case 5: /* parent locks */
2532 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2533 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2534 case 4: /* target dentry */
2536 case 3: /* source dentry */
2538 case 2: /* target directory dentry */
2539 l_dput(*de_tgtdirp);
2540 case 1: /* source directry dentry */
2541 l_dput(*de_srcdirp);
2548 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2549 struct ptlrpc_request *req,
2550 struct lustre_handle *lockh)
2552 struct obd_device *obd = req->rq_export->exp_obd;
2553 struct dentry *de_srcdir = NULL;
2554 struct dentry *de_tgtdir = NULL;
2555 struct dentry *de_old = NULL;
2556 struct dentry *de_new = NULL;
2557 struct dentry *trap;
2558 struct inode *old_inode = NULL, *new_inode = NULL;
2559 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
2560 struct mds_obd *mds = mds_req2mds(req);
2561 struct lustre_handle dlm_handles[4];
2562 struct mds_body *body = NULL;
2563 struct lov_mds_md *lmm = NULL;
2564 int rc = 0, lock_count = 3, cleanup_phase = 0, sz;
2565 void *handle = NULL;
2566 unsigned int qcids[MAXQUOTAS] = { 0, 0 };
2567 unsigned int qpids[4] = { 0, 0, 0, 0 };
2570 LASSERT(offset == REQ_REC_OFF);
2571 offset = REPLY_REC_OFF;
2573 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2574 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2575 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2577 mds_counter_incr(req->rq_export, LPROC_MDS_RENAME);
2579 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2582 ldlm_request_cancel(req, rec->ur_dlm, 0);
2584 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2585 rec->ur_fid2, &de_tgtdir, LCK_EX,
2586 rec->ur_name, rec->ur_namelen,
2587 &de_old, rec->ur_tgt,
2588 rec->ur_tgtlen, &de_new,
2589 dlm_handles, LCK_EX);
2593 cleanup_phase = 1; /* parent(s), children, locks */
2595 old_inode = de_old->d_inode;
2596 new_inode = de_new->d_inode;
2598 if (new_inode != NULL)
2601 /* version recovery check */
2602 rc = mds_version_get_check(req, de_srcdir->d_inode, 0);
2604 GOTO(cleanup_no_trans, rc);
2605 rc = mds_version_get_check(req, old_inode, 1);
2607 GOTO(cleanup_no_trans, rc);
2608 rc = mds_version_get_check(req, de_tgtdir->d_inode, 2);
2610 GOTO(cleanup_no_trans, rc);
2611 rc = mds_version_get_check(req, new_inode, 3);
2613 GOTO(cleanup_no_trans, rc);
2615 /* sanity check for src inode */
2616 if (old_inode->i_ino == de_srcdir->d_inode->i_ino ||
2617 old_inode->i_ino == de_tgtdir->d_inode->i_ino)
2618 GOTO(cleanup, rc = -EINVAL);
2620 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2621 GOTO(cleanup, rc = -EROFS);
2623 if (new_inode == NULL)
2627 cleanup_phase = 2; /* iput(new_inode) when finished */
2629 /* sanity check for dest inode */
2630 if (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
2631 new_inode->i_ino == de_tgtdir->d_inode->i_ino)
2632 GOTO(cleanup, rc = -EINVAL);
2634 if (old_inode == new_inode)
2635 GOTO(cleanup, rc = 0);
2637 /* save uids/gids for qunit acquire/release */
2638 qcids[USRQUOTA] = old_inode->i_uid;
2639 qcids[GRPQUOTA] = old_inode->i_gid;
2640 qpids[USRQUOTA] = de_tgtdir->d_inode->i_uid;
2641 qpids[GRPQUOTA] = de_tgtdir->d_inode->i_gid;
2642 qpids[2] = de_srcdir->d_inode->i_uid;
2643 qpids[3] = de_srcdir->d_inode->i_gid;
2645 /* if we are about to remove the target at first, pass the EA of
2646 * that inode to client to perform and cleanup on OST */
2647 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof(*body));
2648 LASSERT(body != NULL);
2650 /* child orphan sem protects orphan_dec_test && is_orphan race */
2651 MDS_DOWN_READ_ORPHAN_SEM(new_inode);
2652 cleanup_phase = 3; /* MDS_UP_READ_ORPHAN_SEM(new_inode) when finished */
2654 if ((S_ISDIR(new_inode->i_mode) && new_inode->i_nlink == 2) ||
2655 new_inode->i_nlink == 1) {
2656 if (mds_orphan_needed(obd, new_inode)) {
2657 /* need to lock pending_dir before transaction */
2658 LOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2659 cleanup_phase = 4; /* UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode); */
2660 } else if (S_ISREG(new_inode->i_mode)) {
2661 mds_pack_inode2body(body, new_inode);
2662 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
2663 new_inode, MDS_PACK_MD_LOCK, 0,
2664 req->rq_export->exp_connect_flags);
2669 OBD_FAIL_WRITE(obd, OBD_FAIL_MDS_REINT_RENAME_WRITE,
2670 de_srcdir->d_inode->i_sb);
2672 lmm = lustre_msg_buf(req->rq_repmsg, offset + 1, 0);
2673 /* check that lmm size is not 0 */
2674 sz = lustre_msg_buflen(req->rq_repmsg, offset + 1) > 0 ?
2675 le32_to_cpu(lmm->lmm_stripe_count) : 0;
2677 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
2680 GOTO(cleanup, rc = PTR_ERR(handle));
2682 trap = lock_rename(de_tgtdir, de_srcdir);
2683 /* source should not be ancestor of target */
2684 if (de_old == trap) {
2685 unlock_rename(de_tgtdir, de_srcdir);
2686 GOTO(cleanup, rc = -EINVAL);
2688 /* target should not be an ancestor of source */
2689 if (de_new == trap) {
2690 unlock_rename(de_tgtdir, de_srcdir);
2691 GOTO(cleanup, rc = -ENOTEMPTY);
2694 de_old->d_fsdata = req;
2695 de_new->d_fsdata = req;
2697 rc = ll_vfs_rename(de_srcdir->d_inode, de_old, mds->mds_vfsmnt,
2698 de_tgtdir->d_inode, de_new, mds->mds_vfsmnt);
2700 unlock_rename(de_tgtdir, de_srcdir);
2706 /* update ctime of renamed file */
2707 iattr.ia_valid = ATTR_CTIME;
2708 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2709 if (S_ISDIR(de_old->d_inode->i_mode) &&
2710 de_srcdir->d_inode != de_tgtdir->d_inode) {
2711 /* cross directory rename of a directory, ".."
2712 changed, update mtime also */
2713 iattr.ia_valid |= ATTR_MTIME;
2714 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2716 err = fsfilt_setattr(obd, de_old, handle, &iattr, 0);
2718 CERROR("error on old inode time update: "
2721 if (de_new->d_inode) {
2722 /* target file exists, update its ctime as it
2724 iattr.ia_valid = ATTR_CTIME;
2725 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2726 err = fsfilt_setattr(obd, de_new, handle, &iattr, 0);
2728 CERROR("error on target inode time update: "
2732 /* update mtime and ctime of old directory */
2733 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2734 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2735 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2736 err = fsfilt_setattr(obd, de_srcdir, handle, &iattr, 0);
2738 CERROR("error on old dir inode update: "
2741 if (de_srcdir->d_inode != de_tgtdir->d_inode) {
2742 /* cross directory rename, update
2743 mtime and ctime of new directory */
2744 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2745 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2746 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2747 err = fsfilt_setattr(obd, de_tgtdir, handle, &iattr, 0);
2749 CERROR("error on new dir inode time update: "
2754 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
2755 if (mds_orphan_needed(obd, new_inode)) {
2756 rc = mds_orphan_add_link(rec, obd, de_new);
2759 /* inode successfully linked to PENDING */
2760 GOTO(cleanup, rc = 0);
2762 /* we failed to move the file to PENDING,
2763 * really unlink the file as if there were
2764 * no more openers */
2768 if (!S_ISREG(new_inode->i_mode))
2771 if (!(body->valid & OBD_MD_FLEASIZE)) {
2772 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2773 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2774 } else if (mds_log_op_unlink(obd,
2775 lustre_msg_buf(req->rq_repmsg,
2777 lustre_msg_buflen(req->rq_repmsg,
2779 lustre_msg_buf(req->rq_repmsg,
2781 lustre_msg_buflen(req->rq_repmsg,
2784 body->valid |= OBD_MD_FLCOOKIE;
2790 inodes[0] = de_srcdir && !IS_ERR(de_srcdir) ? de_srcdir->d_inode : NULL;
2791 inodes[1] = old_inode;
2792 inodes[2] = de_tgtdir && !IS_ERR(de_tgtdir) ? de_tgtdir->d_inode : NULL;
2793 inodes[3] = new_inode;
2794 rc = mds_finish_transno(mds, inodes, handle, req, rc, 0, 0);
2797 switch (cleanup_phase) {
2799 UNLOCK_INODE_MUTEX(mds->mds_pending_dir->d_inode);
2801 MDS_UP_READ_ORPHAN_SEM(new_inode);
2806 if (lock_count == 4)
2807 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2808 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2809 ldlm_lock_decref(&(dlm_handles[1]), LCK_EX);
2810 ldlm_lock_decref(&(dlm_handles[0]), LCK_EX);
2812 if (lock_count == 4)
2813 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2814 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2815 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_EX);
2816 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_EX);
2825 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2828 req->rq_status = rc;
2830 /* acquire/release qunit */
2831 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
2836 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2837 struct ptlrpc_request *, struct lustre_handle *);
2839 static mds_reinter reinters[REINT_MAX] = {
2840 [REINT_SETATTR] mds_reint_setattr,
2841 [REINT_CREATE] mds_reint_create,
2842 [REINT_LINK] mds_reint_link,
2843 [REINT_UNLINK] mds_reint_unlink,
2844 [REINT_RENAME] mds_reint_rename,
2845 [REINT_OPEN] mds_open
2848 int mds_reint_rec(struct mds_update_record *rec, int offset,
2849 struct ptlrpc_request *req, struct lustre_handle *lockh)
2851 struct obd_device *obd = req->rq_export->exp_obd;
2852 struct mds_obd *mds = &obd->u.mds;
2853 struct lvfs_run_ctxt saved;
2856 gid_t fsgid = rec->ur_uc.luc_fsgid;
2861 if (req->rq_uid != LNET_UID_ANY) {
2862 /* non-root local cluster client
2863 * NB root's creds are believed... */
2864 LASSERT (req->rq_uid != 0);
2865 rec->ur_uc.luc_fsuid = req->rq_uid;
2866 cfs_kernel_cap_unpack(&rec->ur_uc.luc_cap, 0);
2870 /* get group info of this user */
2871 rec->ur_uc.luc_uce = upcall_cache_get_entry(mds->mds_group_hash,
2872 rec->ur_uc.luc_fsuid,
2873 rec->ur_uc.luc_fsgid, 2,
2874 &rec->ur_uc.luc_suppgid1);
2876 if (IS_ERR(rec->ur_uc.luc_uce)) {
2877 rc = PTR_ERR(rec->ur_uc.luc_uce);
2878 rec->ur_uc.luc_uce = NULL;
2882 /* checked by unpacker */
2883 LASSERT(rec->ur_opcode < REINT_MAX && reinters[rec->ur_opcode] != NULL);
2886 if (rec->ur_uc.luc_uce)
2887 rec->ur_uc.luc_fsgid = rec->ur_uc.luc_uce->ue_primary;
2890 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2893 if (rec->ur_uc.luc_uce && fsgid != rec->ur_uc.luc_fsgid &&
2894 in_group_p(fsgid)) {
2895 rec->ur_uc.luc_fsgid = fsgid;
2896 current->fsgid = saved.luc.luc_fsgid = fsgid;
2900 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2901 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2903 upcall_cache_put_entry(mds->mds_group_hash, rec->ur_uc.luc_uce);