1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
5 * Lustre Metadata Server (mds) reintegration routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/jbd.h>
35 #include <linux/ext3_fs.h>
36 #include <linux/obd_support.h>
37 #include <linux/obd_class.h>
38 #include <linux/obd.h>
39 #include <linux/lustre_lib.h>
40 #include <linux/lustre_idl.h>
41 #include <linux/lustre_mds.h>
42 #include <linux/lustre_dlm.h>
43 #include <linux/lustre_log.h>
44 #include <linux/lustre_fsfilt.h>
45 #include "mds_internal.h"
47 struct mds_logcancel_data {
48 struct lov_mds_md *mlcd_lmm;
52 struct llog_cookie mlcd_cookies[0];
55 static void mds_cancel_cookies_cb(struct obd_device *obd,
56 __u64 transno, void *cb_data,
59 struct mds_logcancel_data *mlcd = cb_data;
60 struct lov_stripe_md *lsm = NULL;
61 struct llog_ctxt *ctxt;
64 obd_transno_commit_cb(obd, transno, error);
66 CDEBUG(D_HA, "cancelling %d cookies\n",
67 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
69 rc = obd_unpackmd(obd->u.mds.mds_lov_exp, &lsm, mlcd->mlcd_lmm,
70 mlcd->mlcd_eadatalen);
72 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
73 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
76 ///* XXX 0 normally, SENDNOW for debug */);
77 ctxt = llog_get_context(&obd->obd_llogs,
78 mlcd->mlcd_cookies[0].lgc_subsys + 1);
79 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
80 sizeof(*mlcd->mlcd_cookies),
81 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
83 CERROR("error cancelling %d log cookies: rc %d\n",
84 (int)(mlcd->mlcd_cookielen /
85 sizeof(*mlcd->mlcd_cookies)), rc);
88 OBD_FREE(mlcd, mlcd->mlcd_size);
91 /* Assumes caller has already pushed us into the kernel context. */
92 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
93 struct ptlrpc_request *req, int rc, __u32 op_data)
95 struct mds_export_data *med = &req->rq_export->exp_mds_data;
96 struct obd_device *obd = req->rq_export->exp_obd;
97 struct mds_client_data *mcd = med->med_mcd;
98 int err, log_pri = D_HA;
103 /* if the export has already been failed, we have no last_rcvd slot */
104 if (req->rq_export->exp_failed) {
105 CERROR("committing transaction for disconnected client\n");
107 GOTO(out_commit, rc);
114 if (handle == NULL) {
115 /* if we're starting our own xaction, use our own inode */
116 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
117 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
118 if (IS_ERR(handle)) {
119 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
120 RETURN(PTR_ERR(handle));
126 transno = req->rq_reqmsg->transno;
128 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
129 } else if (transno == 0) {
130 spin_lock(&mds->mds_transno_lock);
131 transno = ++mds->mds_last_transno;
132 spin_unlock(&mds->mds_transno_lock);
134 spin_lock(&mds->mds_transno_lock);
135 if (transno > mds->mds_last_transno)
136 mds->mds_last_transno = transno;
137 spin_unlock(&mds->mds_transno_lock);
139 req->rq_repmsg->transno = req->rq_transno = transno;
140 mcd->mcd_last_transno = cpu_to_le64(transno);
141 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
142 mcd->mcd_last_result = cpu_to_le32(rc);
143 mcd->mcd_last_data = cpu_to_le32(op_data);
145 fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
146 mds_commit_last_transno_cb, NULL);
148 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
149 sizeof(*mcd), &off, 0);
157 DEBUG_REQ(log_pri, req,
158 "wrote trans #"LPU64" client %s at idx %u: err = %d",
159 transno, mcd->mcd_uuid, med->med_idx, err);
161 err = mds_update_last_fid(obd, handle, 0);
168 err = mds_lov_write_objids(obd);
174 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
177 err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
179 CERROR("error committing transaction: %d\n", err);
187 /* this gives the same functionality as the code between
188 * sys_chmod and inode_setattr
189 * chown_common and inode_setattr
190 * utimes and inode_setattr
192 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
194 time_t now = LTIME_S(CURRENT_TIME);
195 struct iattr *attr = &rec->ur_iattr;
196 unsigned int ia_valid = attr->ia_valid;
200 /* only fix up attrs if the client VFS didn't already */
201 if (!(ia_valid & ATTR_RAW))
204 if (!(ia_valid & ATTR_CTIME_SET))
205 LTIME_S(attr->ia_ctime) = now;
206 if (!(ia_valid & ATTR_ATIME_SET))
207 LTIME_S(attr->ia_atime) = now;
208 if (!(ia_valid & ATTR_MTIME_SET))
209 LTIME_S(attr->ia_mtime) = now;
211 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
215 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
216 if (rec->ur_fsuid != inode->i_uid &&
217 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
221 if (ia_valid & ATTR_SIZE) {
222 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
226 if (ia_valid & ATTR_UID) {
229 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
231 if (attr->ia_uid == (uid_t) -1)
232 attr->ia_uid = inode->i_uid;
233 if (attr->ia_gid == (gid_t) -1)
234 attr->ia_gid = inode->i_gid;
235 attr->ia_mode = inode->i_mode;
237 * If the user or group of a non-directory has been
238 * changed by a non-root user, remove the setuid bit.
239 * 19981026 David C Niemi <niemi@tux.org>
241 * Changed this to apply to all users, including root,
242 * to avoid some races. This is the behavior we had in
243 * 2.0. The check for non-root was definitely wrong
244 * for 2.2 anyway, as it should have been using
245 * CAP_FSETID rather than fsuid -- 19990830 SD.
247 if ((inode->i_mode & S_ISUID) == S_ISUID &&
248 !S_ISDIR(inode->i_mode)) {
249 attr->ia_mode &= ~S_ISUID;
250 attr->ia_valid |= ATTR_MODE;
253 * Likewise, if the user or group of a non-directory
254 * has been changed by a non-root user, remove the
255 * setgid bit UNLESS there is no group execute bit
256 * (this would be a file marked for mandatory
257 * locking). 19981026 David C Niemi <niemi@tux.org>
259 * Removed the fsuid check (see the comment above) --
262 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
263 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
264 attr->ia_mode &= ~S_ISGID;
265 attr->ia_valid |= ATTR_MODE;
267 } else if (ia_valid & ATTR_MODE) {
268 int mode = attr->ia_mode;
270 if (attr->ia_mode == (mode_t) -1)
271 attr->ia_mode = inode->i_mode;
273 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
278 void mds_steal_ack_locks(struct ptlrpc_request *req)
280 struct obd_export *exp = req->rq_export;
281 struct list_head *tmp;
282 struct ptlrpc_reply_state *oldrep;
283 struct ptlrpc_service *svc;
284 struct llog_create_locks *lcl;
286 char str[PTL_NALFMT_SIZE];
289 /* CAVEAT EMPTOR: spinlock order */
290 spin_lock_irqsave (&exp->exp_lock, flags);
291 list_for_each (tmp, &exp->exp_outstanding_replies) {
292 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
294 if (oldrep->rs_xid != req->rq_xid)
297 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
298 CERROR ("Resent req xid "LPX64" has mismatched opc: "
299 "new %d old %d\n", req->rq_xid,
300 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
302 svc = oldrep->rs_srv_ni->sni_service;
303 spin_lock (&svc->srv_lock);
305 list_del_init (&oldrep->rs_exp_list);
307 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
309 oldrep->rs_nlocks, oldrep,
310 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
311 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
313 for (i = 0; i < oldrep->rs_nlocks; i++)
314 ptlrpc_save_lock(req,
315 &oldrep->rs_locks[i],
316 oldrep->rs_modes[i]);
317 oldrep->rs_nlocks = 0;
319 lcl = oldrep->rs_llog_locks;
320 oldrep->rs_llog_locks = NULL;
322 ptlrpc_save_llog_lock(req, lcl);
324 DEBUG_REQ(D_HA, req, "stole locks for");
325 ptlrpc_schedule_difficult_reply (oldrep);
327 spin_unlock (&svc->srv_lock);
328 spin_unlock_irqrestore (&exp->exp_lock, flags);
331 spin_unlock_irqrestore (&exp->exp_lock, flags);
334 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
336 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
337 mcd->mcd_last_transno, mcd->mcd_last_result);
338 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
339 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
341 mds_steal_ack_locks(req);
344 static void reconstruct_reint_setattr(struct mds_update_record *rec,
345 int offset, struct ptlrpc_request *req)
347 struct mds_export_data *med = &req->rq_export->exp_mds_data;
348 struct mds_body *body;
351 mds_req_from_mcd(req, med->med_mcd);
353 de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
355 LASSERT(PTR_ERR(de) == req->rq_status);
359 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
360 mds_pack_inode2body(req2obd(req), body, de->d_inode, 0);
362 /* Don't return OST-specific attributes if we didn't just set them */
363 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
364 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
365 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
366 body->valid |= OBD_MD_FLMTIME;
367 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
368 body->valid |= OBD_MD_FLATIME;
373 /* In the raw-setattr case, we lock the child inode.
374 * In the write-back case or if being called from open, the client holds a lock
377 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
378 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
379 struct ptlrpc_request *req, struct lustre_handle *lh)
381 struct mds_obd *mds = mds_req2mds(req);
382 struct obd_device *obd = req->rq_export->exp_obd;
383 struct mds_body *body;
385 struct inode *inode = NULL;
386 struct lustre_handle lockh[2] = {{0}, {0}};
389 struct mds_logcancel_data *mlcd = NULL;
390 int rc = 0, cleanup_phase = 0, err;
394 LASSERT(offset == 1);
396 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
397 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
398 rec->ur_iattr.ia_valid);
400 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
402 MDS_UPDATE_COUNTER(mds, MDS_SETATTR_COUNT);
404 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
405 de = mds_id2dentry(obd, rec->ur_id1, NULL);
407 GOTO(cleanup, rc = PTR_ERR(de));
409 __u64 lockpart = MDS_INODELOCK_UPDATE;
410 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
411 lockpart |= MDS_INODELOCK_LOOKUP;
412 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
413 lockh, &parent_mode, NULL, 0, lockpart);
415 GOTO(cleanup, rc = PTR_ERR(de));
423 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
424 rec->ur_eadata != NULL)
427 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
429 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
431 GOTO(cleanup, rc = PTR_ERR(handle));
433 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
434 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
435 LTIME_S(rec->ur_iattr.ia_mtime),
436 LTIME_S(rec->ur_iattr.ia_ctime));
437 rc = mds_fix_attr(inode, rec);
441 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
442 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
443 (long)&rec->ur_iattr.ia_attr_flags);
445 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
447 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
448 rec->ur_eadata != NULL) {
449 struct lov_stripe_md *lsm = NULL;
451 rc = ll_permission(inode, MAY_WRITE, NULL);
455 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_lov_exp,
456 0, &lsm, rec->ur_eadata);
460 obd_free_memmd(mds->mds_lov_exp, &lsm);
462 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
468 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
469 mds_pack_inode2body(obd, body, inode, 0);
470 body->id1 = *rec->ur_id1;
472 /* Don't return OST-specific attributes if we didn't just set them */
473 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
474 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
475 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
476 body->valid |= OBD_MD_FLMTIME;
477 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
478 body->valid |= OBD_MD_FLATIME;
480 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_lov_obd)) {
481 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
484 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
486 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
487 mlcd->mlcd_cookielen = rec->ur_cookielen;
488 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
489 mlcd->mlcd_cookielen;
490 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
491 mlcd->mlcd_cookielen);
492 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
493 mlcd->mlcd_eadatalen);
495 CERROR("unable to allocate log cancel data\n");
501 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
502 handle, mds_cancel_cookies_cb, mlcd);
503 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
504 switch (cleanup_phase) {
506 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
507 rec->ur_eadata != NULL)
512 if (lockh[1].cookie != 0)
513 ldlm_lock_decref(lockh + 1, parent_mode);
516 ldlm_lock_decref(lockh, LCK_PW);
518 ptlrpc_save_lock (req, lockh, LCK_PW);
533 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
534 struct ptlrpc_request *req)
536 struct mds_export_data *med = &req->rq_export->exp_mds_data;
537 struct dentry *parent, *child;
538 struct mds_body *body;
541 mds_req_from_mcd(req, med->med_mcd);
543 if (req->rq_status) {
548 parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
549 LASSERT(!IS_ERR(parent));
550 child = ll_lookup_one_len(rec->ur_name, parent,
551 rec->ur_namelen - 1);
552 LASSERT(!IS_ERR(child));
553 if ((child->d_flags & DCACHE_CROSS_REF)) {
554 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
555 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
556 mds_pack_dentry2body(req2obd(req), body, child, 1);
557 } else if (child->d_inode == NULL) {
558 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
559 OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
560 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
562 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
563 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
570 static int mds_reint_create(struct mds_update_record *rec, int offset,
571 struct ptlrpc_request *req,
572 struct lustre_handle *lh)
574 struct dentry *dparent = NULL;
575 struct mds_obd *mds = mds_req2mds(req);
576 struct obd_device *obd = req->rq_export->exp_obd;
577 struct dentry *dchild = NULL;
578 struct inode *dir = NULL;
580 struct lustre_handle lockh[2] = {{0}, {0}};
582 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
584 struct dentry_params dp;
585 struct mea *mea = NULL;
589 LASSERT(offset == 1);
591 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
594 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
595 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
596 rec->ur_name, rec->ur_mode);
598 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
600 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
601 GOTO(cleanup, rc = -ESTALE);
603 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
604 lockh, &parent_mode, rec->ur_name,
605 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
606 if (IS_ERR(dparent)) {
607 rc = PTR_ERR(dparent);
608 CERROR("parent lookup error %d\n", rc);
611 cleanup_phase = 1; /* locked parent dentry */
612 dir = dparent->d_inode;
615 ldlm_lock_dump_handle(D_OTHER, lockh);
617 /* try to retrieve MEA data for this dir */
618 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
624 * dir is already splitted, check is requested filename should
625 * live at this MDS or at another one.
627 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
628 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
629 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
630 " should be %lu(%d)\n",
631 mea->mea_master, dparent->d_inode->i_ino,
632 dparent->d_inode->i_generation, rec->ur_name,
633 (unsigned long)id_group(&mea->mea_ids[i]), i);
634 GOTO(cleanup, rc = -ERESTART);
638 dchild = ll_lookup_one_len(rec->ur_name, dparent,
639 rec->ur_namelen - 1);
640 if (IS_ERR(dchild)) {
641 rc = PTR_ERR(dchild);
642 CERROR("Can't find "DLID4"/%s, error %d\n",
643 OLID4(rec->ur_id1), rec->ur_name, rc);
647 cleanup_phase = 2; /* child dentry */
649 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
651 if (type == S_IFREG || type == S_IFDIR) {
652 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
653 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
654 obd->obd_name, dparent->d_inode->i_ino,
655 dparent->d_inode->i_generation, rc, parent_mode);
657 /* dir got splitted */
658 GOTO(cleanup, rc = -ERESTART);
660 /* error happened during spitting. */
665 if (dir->i_mode & S_ISGID) {
666 if (S_ISDIR(rec->ur_mode))
667 rec->ur_mode |= S_ISGID;
671 * here inode number should be used only in the case of replaying. It is
672 * needed to check if object already created in the case of creating
675 dchild->d_fsdata = (void *)&dp;
676 dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
681 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
683 GOTO(cleanup, rc = PTR_ERR(handle));
684 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
690 struct lustre_id sid;
693 * as Peter asked, mkdir() should distribute new directories
694 * over the whole cluster in order to distribute namespace
695 * processing load. first, we calculate which MDS to use to put
696 * new directory's inode in.
698 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
700 if (i == mds->mds_num) {
701 /* inode will be created locally */
702 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
704 GOTO(cleanup, rc = PTR_ERR(handle));
706 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
708 CERROR("Can't create dir %s, rc = %d\n",
709 dchild->d_name.name, rc);
713 down(&dchild->d_inode->i_sem);
715 rc = mds_update_inode_sid(obd, dchild->d_inode,
716 handle, rec->ur_id2);
718 CERROR("mds_update_inode_sid() failed, inode %lu, "
719 "rc %d\n", dchild->d_inode->i_ino, rc);
723 * make sure, that fid is up-to-date.
725 mds_set_last_fid(obd, id_fid(rec->ur_id2));
727 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
730 CERROR("mds_alloc_inode_sid() failed, inode %lu, "
731 "rc %d\n", dchild->d_inode->i_ino, rc);
734 up(&dchild->d_inode->i_sem);
740 nstripes = *(u16 *)rec->ur_eadata;
742 if (rc == 0 && nstripes) {
743 /* we pass LCK_EX to split routine to signal,
744 * that we have exclusive access to the
745 * directory. Simple because nobody knows it
746 * already exists -bzzz */
747 rc = mds_try_to_split_dir(obd, dchild,
751 /* dir got splitted */
754 /* an error occured during
759 } else if (!DENTRY_VALID(dchild)) {
760 /* inode will be created on another MDS */
761 struct obdo *oa = NULL;
762 struct mds_body *body;
764 /* first, create that inode */
767 GOTO(cleanup, rc = -ENOMEM);
772 if (rec->ur_eadata) {
773 /* user asks for creating splitted dir */
774 oa->o_easize = *((u16 *) rec->ur_eadata);
777 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
778 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
779 OBD_MD_FLUID | OBD_MD_FLGID);
781 oa->o_mode = dir->i_mode;
783 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
786 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
788 * here inode number and generation are
789 * important, as this is replay request and we
790 * need them to check if such an object is
793 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
794 obd->obd_name, rec->ur_namelen - 1,
795 rec->ur_name, (unsigned)id_ino(rec->ur_id2),
796 (unsigned)id_gen(rec->ur_id2));
797 oa->o_id = id_ino(rec->ur_id2);
798 oa->o_fid = id_fid(rec->ur_id2);
799 oa->o_generation = id_gen(rec->ur_id2);
800 oa->o_flags |= OBD_FL_RECREATE_OBJS;
803 /* before obd_create() is called, o_fid is not known. */
804 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
806 CERROR("can't create remote inode: %d\n", rc);
807 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
808 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
809 rec->ur_name, rec->ur_mode);
814 /* now, add new dir entry for it */
815 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
816 if (IS_ERR(handle)) {
818 GOTO(cleanup, rc = PTR_ERR(handle));
821 /* creating local dentry for remote inode. */
822 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
823 rec->ur_namelen - 1, oa->o_id,
824 oa->o_generation, i, oa->o_fid);
827 CERROR("Can't create local entry %*s for "
828 "remote inode.\n", rec->ur_namelen - 1,
834 body = lustre_msg_buf(req->rq_repmsg,
836 body->valid |= OBD_MD_FLID | OBD_MD_MDS |
839 id_group(&body->id1) = i;
840 id_ino(&body->id1) = oa->o_id;
841 id_fid(&body->id1) = oa->o_fid;
842 id_gen(&body->id1) = oa->o_generation;
845 /* requested name exists in the directory */
852 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
854 GOTO(cleanup, rc = PTR_ERR(handle));
855 if (rec->ur_tgt == NULL) /* no target supplied */
856 rc = -EINVAL; /* -EPROTO? */
858 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
866 int rdev = rec->ur_rdev;
867 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
869 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
870 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
875 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
876 dchild->d_fsdata = NULL;
877 GOTO(cleanup, rc = -EINVAL);
880 /* In case we stored the desired inum in here, we want to clean up. */
881 if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
882 dchild->d_fsdata = NULL;
885 CDEBUG(D_INODE, "error during create: %d\n", rc);
887 } else if (dchild->d_inode) {
889 struct mds_body *body;
890 struct inode *inode = dchild->d_inode;
893 iattr.ia_uid = rec->ur_fsuid;
894 LTIME_S(iattr.ia_atime) = rec->ur_time;
895 LTIME_S(iattr.ia_ctime) = rec->ur_time;
896 LTIME_S(iattr.ia_mtime) = rec->ur_time;
898 if (dir->i_mode & S_ISGID)
899 iattr.ia_gid = dir->i_gid;
901 iattr.ia_gid = rec->ur_fsgid;
903 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
904 ATTR_MTIME | ATTR_CTIME;
906 if (id_ino(rec->ur_id2)) {
907 LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
908 inode->i_generation = id_gen(rec->ur_id2);
910 if (type != S_IFDIR) {
912 * updating inode self id, as inode already
913 * exists and we should make sure, its sid will
914 * be the same as we reveived.
917 rc = mds_update_inode_sid(obd, inode,
918 handle, rec->ur_id2);
921 CERROR("Can't update inode self id, "
926 * make sure, that fid is up-to-date.
928 mds_set_last_fid(obd, id_fid(rec->ur_id2));
931 /* dirtied and committed by the upcoming setattr. */
932 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
933 inode->i_ino, inode->i_generation);
935 struct lustre_handle child_ino_lockh;
937 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
938 inode->i_ino, inode->i_generation);
940 if (type != S_IFDIR) {
941 struct lustre_id sid;
944 * allocate new id for @inode if it is not dir,
945 * because for dir it was already done.
948 rc = mds_alloc_inode_sid(obd, inode,
952 CERROR("mds_alloc_inode_sid() failed, "
953 "inode %lu, rc %d\n", inode->i_ino,
960 * the inode we were allocated may have just
961 * been freed by an unlink operation. We take
962 * this lock to synchronize against the matching
963 * reply-ack-lock taken in unlink, to avoid
964 * replay problems if this reply makes it out to
965 * the client but the unlink's does not. See
966 * bug 2029 for more detail.
968 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
969 if (rc != ELDLM_OK) {
970 CERROR("error locking for unlink/create sync: "
973 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
978 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
980 CERROR("error on child setattr: rc = %d\n", rc);
982 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
983 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
985 CERROR("error on parent setattr: rc = %d\n", rc);
987 MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
989 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
990 mds_pack_inode2body(obd, body, inode, 1);
995 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
998 /* Destroy the file we just created. This should not need extra
999 * journal credits, as we have already modified all of the
1000 * blocks needed in order to create the file in the first
1004 err = vfs_rmdir(dir, dchild);
1006 CERROR("rmdir in error path: %d\n", err);
1009 err = vfs_unlink(dir, dchild);
1011 CERROR("unlink in error path: %d\n", err);
1017 switch (cleanup_phase) {
1018 case 2: /* child dentry */
1020 case 1: /* locked parent dentry */
1022 if (lockh[1].cookie != 0)
1023 ldlm_lock_decref(lockh + 1, parent_mode);
1026 ldlm_lock_decref(lockh, LCK_PW);
1028 ptlrpc_save_lock(req, lockh, LCK_PW);
1034 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1038 OBD_FREE(mea, mea_size);
1039 req->rq_status = rc;
1044 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1045 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1049 for (i = 0; i < RES_NAME_SIZE; i++) {
1051 * this is needed to make zeroed res_id entries to be put at the
1052 * end of list in *ordered_locks() .
1054 if (res1->name[i] == 0 && res2->name[i] != 0)
1056 if (res2->name[i] == 0 && res1->name[i] != 0)
1058 if (res1->name[i] > res2->name[i])
1060 if (res1->name[i] < res2->name[i])
1067 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1073 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1074 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1075 * because they take the place of local semaphores.
1077 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1078 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1079 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1080 struct lustre_handle *p1_lockh, int p1_lock_mode,
1081 ldlm_policy_data_t *p1_policy,
1082 struct ldlm_res_id *p2_res_id,
1083 struct lustre_handle *p2_lockh, int p2_lock_mode,
1084 ldlm_policy_data_t *p2_policy)
1086 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1087 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1088 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1089 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1093 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1095 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1096 res_id[0]->name[0], res_id[1]->name[0]);
1098 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1099 handles[1] = p1_lockh;
1100 handles[0] = p2_lockh;
1101 res_id[1] = p1_res_id;
1102 res_id[0] = p2_res_id;
1103 lock_modes[1] = p1_lock_mode;
1104 lock_modes[0] = p2_lock_mode;
1105 policies[1] = p1_policy;
1106 policies[0] = p2_policy;
1109 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1110 res_id[0]->name[0], res_id[1]->name[0]);
1112 flags = LDLM_FL_LOCAL_ONLY;
1113 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1114 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1115 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1116 NULL, 0, NULL, handles[0]);
1119 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1121 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1122 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1123 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1124 ldlm_lock_addref(handles[1], lock_modes[1]);
1125 } else if (res_id[1]->name[0] != 0) {
1126 flags = LDLM_FL_LOCAL_ONLY;
1127 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1128 *res_id[1], LDLM_IBITS, policies[1],
1129 lock_modes[1], &flags, mds_blocking_ast,
1130 ldlm_completion_ast, NULL, NULL, NULL, 0,
1132 if (rc != ELDLM_OK) {
1133 ldlm_lock_decref(handles[0], lock_modes[0]);
1136 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1142 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1143 struct lustre_handle *p1_lockh, int p1_lock_mode,
1144 ldlm_policy_data_t *p1_policy,
1145 struct ldlm_res_id *p2_res_id,
1146 struct lustre_handle *p2_lockh, int p2_lock_mode,
1147 ldlm_policy_data_t *p2_policy,
1148 struct ldlm_res_id *c1_res_id,
1149 struct lustre_handle *c1_lockh, int c1_lock_mode,
1150 ldlm_policy_data_t *c1_policy,
1151 struct ldlm_res_id *c2_res_id,
1152 struct lustre_handle *c2_lockh, int c2_lock_mode,
1153 ldlm_policy_data_t *c2_policy)
1155 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1156 c1_res_id, c2_res_id };
1157 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1158 c1_lockh, c2_lockh };
1159 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1160 c1_lock_mode, c2_lock_mode };
1161 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1162 c1_policy, c2_policy};
1163 int rc, i, j, sorted, flags;
1166 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1167 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1168 res_id[3]->name[0]);
1171 * simple insertion sort - we have at most 4 elements. Note, that zeroed
1172 * res_id should be at the end of list after sorting is finished.
1174 for (i = 1; i < 4; i++) {
1176 dlm_handles[4] = dlm_handles[i];
1177 res_id[4] = res_id[i];
1178 lock_modes[4] = lock_modes[i];
1179 policies[4] = policies[i];
1183 if (res_gt(res_id[j], res_id[4], policies[j],
1185 dlm_handles[j + 1] = dlm_handles[j];
1186 res_id[j + 1] = res_id[j];
1187 lock_modes[j + 1] = lock_modes[j];
1188 policies[j + 1] = policies[j];
1193 } while (j >= 0 && !sorted);
1195 dlm_handles[j + 1] = dlm_handles[4];
1196 res_id[j + 1] = res_id[4];
1197 lock_modes[j + 1] = lock_modes[4];
1198 policies[j + 1] = policies[4];
1201 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1202 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1203 res_id[3]->name[0]);
1205 /* XXX we could send ASTs on all these locks first before blocking? */
1206 for (i = 0; i < 4; i++) {
1210 * nevertheless zeroed res_ids should be at the end of list, and
1211 * could use break here, I think, that it is more correctly for
1212 * clear understanding of code to have continue here, as it
1213 * clearly means, that zeroed res_id should be skipped and does
1214 * not mean, that if we meet zeroed res_id we should stop
1217 if (res_id[i]->name[0] == 0)
1221 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1222 (policies[i]->l_inodebits.bits &
1223 policies[i-1]->l_inodebits.bits) ) {
1224 memcpy(dlm_handles[i], dlm_handles[i-1],
1225 sizeof(*(dlm_handles[i])));
1226 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1228 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1229 *res_id[i], LDLM_IBITS,
1231 lock_modes[i], &flags,
1233 ldlm_completion_ast, NULL, NULL,
1234 NULL, 0, NULL, dlm_handles[i]);
1236 GOTO(out_err, rc = -EIO);
1237 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1244 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1249 /* In the unlikely case that the child changed while we were waiting
1250 * on the lock, we need to drop the lock on the old child and either:
1251 * - if the child has a lower resource name, then we have to also
1252 * drop the parent lock and regain the locks in the right order
1253 * - in the rename case, if the child has a lower resource name than one of
1254 * the other parent/child resources (maxres) we also need to reget the locks
1255 * - if the child has a higher resource name (this is the common case)
1256 * we can just get the lock on the new child (still in lock order)
1258 * Returns 0 if the child did not change or if it changed but could be locked.
1259 * Returns 1 if the child changed and we need to re-lock (no locks held).
1260 * Returns -ve error with a valid dchild (no locks held). */
1261 static int mds_verify_child(struct obd_device *obd,
1262 struct ldlm_res_id *parent_res_id,
1263 struct lustre_handle *parent_lockh,
1264 struct dentry *dparent, int parent_mode,
1265 struct ldlm_res_id *child_res_id,
1266 struct lustre_handle *child_lockh,
1267 struct dentry **dchildp, int child_mode,
1268 ldlm_policy_data_t *child_policy,
1269 const char *name, int namelen,
1270 struct ldlm_res_id *maxres,
1271 unsigned long child_ino,
1274 struct lustre_id sid;
1275 struct dentry *vchild, *dchild = *dchildp;
1276 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1279 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1281 GOTO(cleanup, rc = PTR_ERR(vchild));
1283 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1284 if (child_gen == vchild->d_generation &&
1285 child_ino == vchild->d_inum) {
1294 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1295 (vchild->d_inode != NULL &&
1296 child_gen == vchild->d_inode->i_generation &&
1297 child_ino == vchild->d_inode->i_ino))) {
1305 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1306 vchild->d_inode, dchild ? dchild->d_inode : 0,
1307 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1308 child_res_id->name[0]);
1310 if (child_res_id->name[0] != 0)
1311 ldlm_lock_decref(child_lockh, child_mode);
1315 cleanup_phase = 1; /* parent lock only */
1316 *dchildp = dchild = vchild;
1318 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1321 if (dchild->d_inode) {
1322 down(&dchild->d_inode->i_sem);
1323 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1324 up(&dchild->d_inode->i_sem);
1326 CERROR("Can't read inode self id, inode %lu,"
1327 " rc %d\n", dchild->d_inode->i_ino, rc);
1330 child_res_id->name[0] = id_fid(&sid);
1331 child_res_id->name[1] = id_group(&sid);
1333 child_res_id->name[0] = dchild->d_fid;
1334 child_res_id->name[1] = dchild->d_mdsnum;
1337 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1338 res_gt(maxres, child_res_id, NULL, NULL)) {
1339 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1340 child_res_id->name[0], parent_res_id->name[0],
1342 GOTO(cleanup, rc = 1);
1345 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1346 *child_res_id, LDLM_IBITS, child_policy,
1347 child_mode, &flags, mds_blocking_ast,
1348 ldlm_completion_ast, NULL, NULL, NULL, 0,
1351 GOTO(cleanup, rc = -EIO);
1354 memset(child_res_id, 0, sizeof(*child_res_id));
1359 switch(cleanup_phase) {
1361 if (child_res_id->name[0] != 0)
1362 ldlm_lock_decref(child_lockh, child_mode);
1364 ldlm_lock_decref(parent_lockh, parent_mode);
1370 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1371 struct lustre_id *id,
1372 struct lustre_handle *parent_lockh,
1373 struct dentry **dparentp, int parent_mode,
1374 __u64 parent_lockpart, int *update_mode,
1375 char *name, int namelen,
1376 struct lustre_handle *child_lockh,
1377 struct dentry **dchildp, int child_mode,
1378 __u64 child_lockpart)
1380 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1381 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1382 struct ldlm_res_id parent_res_id = { .name = {0} };
1383 struct ldlm_res_id child_res_id = { .name = {0} };
1384 int rc = 0, cleanup_phase = 0;
1385 unsigned long child_ino;
1386 struct lustre_id sid;
1387 __u32 child_gen = 0;
1388 struct inode *inode;
1391 /* Step 1: Lookup parent */
1392 *dparentp = mds_id2dentry(obd, id, NULL);
1393 if (IS_ERR(*dparentp)) {
1394 rc = PTR_ERR(*dparentp);
1399 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1400 (*dparentp)->d_inode->i_ino, name);
1402 parent_res_id.name[0] = id_fid(id);
1403 parent_res_id.name[1] = id_group(id);
1406 parent_lockh[1].cookie = 0;
1407 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1408 struct ldlm_res_id res_id = { .name = {0} };
1409 ldlm_policy_data_t policy;
1412 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1414 res_id.name[0] = id_fid(id);
1415 res_id.name[1] = id_group(id);
1416 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1418 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1419 res_id, LDLM_IBITS, &policy,
1420 *update_mode, &flags,
1422 ldlm_completion_ast,
1423 NULL, NULL, NULL, 0, NULL,
1429 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1431 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1432 (unsigned long)id_fid(id), (unsigned long)id_group(id),
1433 parent_res_id.name[2]);
1437 cleanup_phase = 1; /* parent dentry */
1439 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1440 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1441 if (IS_ERR(*dchildp)) {
1442 rc = PTR_ERR(*dchildp);
1443 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1447 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1449 * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1450 * lock. Drop possible UPDATE lock!
1452 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1453 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1455 child_res_id.name[0] = (*dchildp)->d_fid;
1456 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1457 child_gen = (*dchildp)->d_generation;
1458 child_ino = (*dchildp)->d_inum;
1462 inode = (*dchildp)->d_inode;
1464 inode = igrab(inode);
1468 down(&inode->i_sem);
1469 rc = mds_read_inode_sid(obd, inode, &sid);
1472 CERROR("Can't read inode self id, inode %lu, "
1473 "rc %d\n", inode->i_ino, rc);
1478 child_res_id.name[0] = id_fid(&sid);
1479 child_res_id.name[1] = id_group(&sid);
1480 child_gen = inode->i_generation;
1481 child_ino = inode->i_ino;
1485 cleanup_phase = 2; /* child dentry */
1487 /* Step 3: Lock parent and child in resource order. If child doesn't
1488 * exist, we still have to lock the parent and re-lookup. */
1489 rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1490 &parent_policy, &child_res_id, child_lockh,
1491 child_mode, &child_policy);
1495 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1496 cleanup_phase = 4; /* child lock */
1498 cleanup_phase = 3; /* parent lock */
1500 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1501 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1502 parent_mode, &child_res_id, child_lockh,
1503 dchildp, child_mode, &child_policy,
1504 name, namelen, &parent_res_id,
1505 child_ino, child_gen);
1515 switch (cleanup_phase) {
1517 ldlm_lock_decref(child_lockh, child_mode);
1519 ldlm_lock_decref(parent_lockh, parent_mode);
1524 if (parent_lockh[1].cookie)
1525 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1533 void mds_reconstruct_generic(struct ptlrpc_request *req)
1535 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1537 mds_req_from_mcd(req, med->med_mcd);
1539 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1540 * we instead link the inode into the PENDING directory until it is
1541 * finally released. We can't simply call mds_reint_rename() or some
1542 * part thereof, because we don't have the inode to check for link
1543 * count/open status until after it is locked.
1545 * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1546 * before starting journal transaction.
1548 * returns 1 on success
1549 * returns 0 if we lost a race and didn't make a new link
1550 * returns negative on error
1552 static int mds_orphan_add_link(struct mds_update_record *rec,
1553 struct obd_device *obd, struct dentry *dentry)
1555 struct mds_obd *mds = &obd->u.mds;
1556 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1557 struct inode *inode = dentry->d_inode;
1558 struct dentry *pending_child;
1559 char idname[LL_ID_NAMELEN];
1560 int idlen = 0, rc, mode;
1563 LASSERT(inode != NULL);
1564 LASSERT(!mds_inode_is_orphan(inode));
1565 #ifndef HAVE_I_ALLOC_SEM
1566 LASSERT(down_trylock(&inode->i_sem) != 0);
1568 LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1570 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1572 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1573 mds_orphan_open_count(inode), inode->i_nlink,
1574 S_ISDIR(inode->i_mode) ? "dir" :
1575 S_ISREG(inode->i_mode) ? "file" : "other",
1576 rec->ur_name, idname);
1578 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1581 pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1582 if (IS_ERR(pending_child))
1583 RETURN(PTR_ERR(pending_child));
1585 if (pending_child->d_inode != NULL) {
1586 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1587 LASSERT(pending_child->d_inode == inode);
1588 GOTO(out_dput, rc = 0);
1591 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1592 * for linking and return real mode back then -bzzz */
1593 mode = inode->i_mode;
1594 inode->i_mode = S_IFREG;
1595 rc = vfs_link(dentry, pending_dir, pending_child);
1597 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1600 mds_inode_set_orphan(inode);
1602 /* return mode and correct i_nlink if inode is directory */
1603 inode->i_mode = mode;
1604 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1605 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1607 if (S_ISDIR(mode)) {
1609 pending_dir->i_nlink++;
1610 mark_inode_dirty(inode);
1611 mark_inode_dirty(pending_dir);
1614 GOTO(out_dput, rc = 1);
1616 l_dput(pending_child);
1620 int mds_create_local_dentry(struct mds_update_record *rec,
1621 struct obd_device *obd)
1623 struct mds_obd *mds = &obd->u.mds;
1624 struct inode *id_dir = mds->mds_id_dir->d_inode;
1625 int idlen = 0, rc, cleanup_phase = 0;
1626 struct dentry *new_child = NULL;
1627 char *idname = rec->ur_name;
1628 struct dentry *child = NULL;
1629 struct lustre_handle lockh[2] = {{0}, {0}};
1630 struct lustre_id sid;
1634 down(&id_dir->i_sem);
1635 idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1636 id_gen(rec->ur_id1));
1638 CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1639 idname, OLID4(rec->ur_id1));
1641 new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1644 if (IS_ERR(new_child)) {
1645 CERROR("can't lookup %s: %d\n", idname,
1646 (int) PTR_ERR(new_child));
1647 GOTO(cleanup, rc = PTR_ERR(new_child));
1651 down(&id_dir->i_sem);
1652 rc = mds_read_inode_sid(obd, id_dir, &sid);
1655 CERROR("Can't read inode self id, inode %lu, "
1656 "rc %d\n", id_dir->i_ino, rc);
1660 if (new_child->d_inode != NULL) {
1661 /* nice. we've already have local dentry! */
1662 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1663 (unsigned)new_child->d_inode->i_ino,
1664 (unsigned)new_child->d_inode->i_generation);
1666 id_ino(rec->ur_id1) = id_dir->i_ino;
1667 id_gen(rec->ur_id1) = id_dir->i_generation;
1668 rec->ur_namelen = idlen + 1;
1670 id_fid(rec->ur_id1) = id_fid(&sid);
1671 id_group(rec->ur_id1) = id_group(&sid);
1673 GOTO(cleanup, rc = 0);
1676 /* new, local dentry will be added soon. we need no aliases here */
1679 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1680 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1682 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1683 LCK_EX, lockh, NULL, NULL, 0,
1684 MDS_INODELOCK_UPDATE);
1687 if (IS_ERR(child)) {
1688 rc = PTR_ERR(child);
1689 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1690 CERROR("can't get victim: %d\n", rc);
1695 handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1697 GOTO(cleanup, rc = PTR_ERR(handle));
1699 rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
1700 idlen, id_ino(rec->ur_id1),
1701 id_gen(rec->ur_id1), mds->mds_num,
1702 id_fid(rec->ur_id1));
1704 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1705 (unsigned long)child->d_inode->i_ino,
1706 (unsigned long)child->d_inode->i_generation, rc);
1708 if (S_ISDIR(child->d_inode->i_mode)) {
1710 mark_inode_dirty(id_dir);
1712 mark_inode_dirty(child->d_inode);
1714 fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
1716 id_ino(rec->ur_id1) = id_dir->i_ino;
1717 id_gen(rec->ur_id1) = id_dir->i_generation;
1718 rec->ur_namelen = idlen + 1;
1720 id_fid(rec->ur_id1) = id_fid(&sid);
1721 id_group(rec->ur_id1) = id_group(&sid);
1723 switch(cleanup_phase) {
1725 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1726 ldlm_lock_decref(lockh, LCK_EX);
1736 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1737 struct ptlrpc_request *slave)
1739 void *cookie, *cookie2;
1740 struct mds_body *body2;
1741 struct mds_body *body;
1745 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1746 LASSERT(body != NULL);
1748 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1749 LASSERT(body2 != NULL);
1751 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
1754 memcpy(body2, body, sizeof(*body));
1755 body2->valid &= ~OBD_MD_FLCOOKIE;
1757 if (!(body->valid & OBD_MD_FLEASIZE) &&
1758 !(body->valid & OBD_MD_FLDIREA))
1761 if (body->eadatasize == 0) {
1762 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1766 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1768 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1769 LASSERT(ea != NULL);
1771 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1772 LASSERT(ea2 != NULL);
1774 memcpy(ea2, ea, body->eadatasize);
1776 if (body->valid & OBD_MD_FLCOOKIE) {
1777 LASSERT(master->rq_repmsg->buflens[2] >=
1778 slave->rq_repmsg->buflens[2]);
1779 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1780 slave->rq_repmsg->buflens[2]);
1781 LASSERT(cookie != NULL);
1783 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1784 master->rq_repmsg->buflens[2]);
1785 LASSERT(cookie2 != NULL);
1786 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1787 body2->valid |= OBD_MD_FLCOOKIE;
1792 static int mds_reint_unlink_remote(struct mds_update_record *rec,
1793 int offset, struct ptlrpc_request *req,
1794 struct lustre_handle *parent_lockh,
1795 int update_mode, struct dentry *dparent,
1796 struct lustre_handle *child_lockh,
1797 struct dentry *dchild)
1799 struct obd_device *obd = req->rq_export->exp_obd;
1800 struct mds_obd *mds = mds_req2mds(req);
1801 struct ptlrpc_request *request = NULL;
1802 int rc = 0, cleanup_phase = 0;
1803 struct mdc_op_data op_data;
1807 LASSERT(offset == 1 || offset == 3);
1809 /* time to drop i_nlink on remote MDS */
1810 memset(&op_data, 0, sizeof(op_data));
1811 mds_pack_dentry2id(obd, &op_data.id1, dchild, 1);
1812 op_data.create_mode = rec->ur_mode;
1814 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
1815 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data.id1));
1817 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1818 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
1819 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data.id1));
1822 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1823 op_data.create_mode |= MDS_MODE_REPLAY;
1825 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1830 mds_copy_unlink_reply(req, request);
1831 ptlrpc_req_finished(request);
1835 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1838 GOTO(cleanup, rc = PTR_ERR(handle));
1839 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1840 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1845 req->rq_status = rc;
1848 if (parent_lockh[1].cookie != 0)
1849 ldlm_lock_decref(parent_lockh + 1, update_mode);
1851 ldlm_lock_decref(child_lockh, LCK_EX);
1853 ldlm_lock_decref(parent_lockh, LCK_PW);
1855 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1862 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1863 struct ptlrpc_request *req,
1864 struct lustre_handle *lh)
1866 struct dentry *dparent = NULL, *dchild;
1867 struct mds_obd *mds = mds_req2mds(req);
1868 struct obd_device *obd = req->rq_export->exp_obd;
1869 struct mds_body *body = NULL;
1870 struct inode *child_inode = NULL;
1871 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1872 struct lustre_handle child_lockh = {0};
1873 struct lustre_handle child_reuse_lockh = {0};
1874 struct lustre_handle *slave_lockh = NULL;
1875 char idname[LL_ID_NAMELEN];
1876 struct llog_create_locks *lcl = NULL;
1877 void *handle = NULL;
1878 int rc = 0, cleanup_phase = 0;
1879 int unlink_by_id = 0;
1883 LASSERT(offset == 1 || offset == 3);
1885 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1886 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1889 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1891 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1892 DEBUG_REQ(D_HA, req, "unlink replay\n");
1893 LASSERT(offset == 1); /* should not come from intent */
1894 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1895 lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
1896 req->rq_repmsg->buflens[2]);
1899 MDS_UPDATE_COUNTER(mds, MDS_UNLINK_COUNT);
1901 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1902 GOTO(cleanup, rc = -ENOENT);
1904 if (rec->ur_namelen == 1) {
1905 /* this is request to drop i_nlink on local inode */
1907 rec->ur_name = idname;
1908 rc = mds_create_local_dentry(rec, obd);
1909 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1910 DEBUG_REQ(D_HA, req,
1911 "drop nlink on inode "DLID4" (replay)",
1912 OLID4(rec->ur_id1));
1918 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1919 /* master mds for directory asks slave removing inode is already
1921 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1922 LCK_PW, parent_lockh,
1923 &update_mode, rec->ur_name,
1925 MDS_INODELOCK_UPDATE);
1926 if (IS_ERR(dparent))
1927 GOTO(cleanup, rc = PTR_ERR(dparent));
1928 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1929 rec->ur_namelen - 1);
1931 GOTO(cleanup, rc = PTR_ERR(dchild));
1932 child_lockh.cookie = 0;
1933 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1934 LASSERT(dchild->d_inode != NULL);
1935 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1937 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1938 parent_lockh, &dparent,
1939 LCK_PW, MDS_INODELOCK_UPDATE,
1940 &update_mode, rec->ur_name,
1941 rec->ur_namelen, &child_lockh,
1943 MDS_INODELOCK_LOOKUP |
1944 MDS_INODELOCK_UPDATE);
1949 if (dchild->d_flags & DCACHE_CROSS_REF) {
1950 /* we should have parent lock only here */
1951 LASSERT(unlink_by_id == 0);
1952 LASSERT(dchild->d_mdsnum != mds->mds_num);
1953 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1954 update_mode, dparent, &child_lockh, dchild);
1958 cleanup_phase = 1; /* dchild, dparent, locks */
1961 child_inode = dchild->d_inode;
1962 if (child_inode == NULL) {
1963 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1964 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1965 GOTO(cleanup, rc = -ENOENT);
1968 cleanup_phase = 2; /* dchild has a lock */
1970 /* We have to do these checks ourselves, in case we are making an
1971 * orphan. The client tells us whether rmdir() or unlink() was called,
1972 * so we need to return appropriate errors (bug 72).
1974 * We don't have to check permissions, because vfs_rename (called from
1975 * mds_open_unlink_rename) also calls may_delete. */
1976 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1977 if (!S_ISDIR(child_inode->i_mode))
1978 GOTO(cleanup, rc = -ENOTDIR);
1980 if (S_ISDIR(child_inode->i_mode))
1981 GOTO(cleanup, rc = -EISDIR);
1984 /* handle splitted dir */
1985 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1989 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1990 * reuse (see bug 2029). */
1991 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1994 cleanup_phase = 3; /* child inum lock */
1996 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1998 /* ldlm_reply in buf[0] if called via intent */
2004 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2005 LASSERT(body != NULL);
2007 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2008 DOWN_READ_I_ALLOC_SEM(child_inode);
2009 cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2011 /* If this is potentially the last reference to this inode, get the
2012 * OBD EA data first so the client can destroy OST objects. We
2013 * only do the object removal later if no open files/links remain. */
2014 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2015 child_inode->i_nlink == 1) {
2016 if (mds_orphan_open_count(child_inode) > 0) {
2017 /* need to lock pending_dir before transaction */
2018 down(&mds->mds_pending_dir->d_inode->i_sem);
2019 cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2020 } else if (S_ISREG(child_inode->i_mode)) {
2021 mds_pack_inode2body(obd, body, child_inode, 0);
2022 mds_pack_md(obd, req->rq_repmsg, offset + 1,
2023 body, child_inode, MDS_PACK_MD_LOCK);
2027 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2028 switch (child_inode->i_mode & S_IFMT) {
2030 /* Drop any lingering child directories before we start our
2031 * transaction, to avoid doing multiple inode dirty/delete
2032 * in our compound transaction (bug 1321). */
2033 shrink_dcache_parent(dchild);
2034 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2037 GOTO(cleanup, rc = PTR_ERR(handle));
2038 rc = vfs_rmdir(dparent->d_inode, dchild);
2041 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2042 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2044 handle = fsfilt_start_log(obd, dparent->d_inode,
2045 FSFILT_OP_UNLINK, NULL,
2046 le32_to_cpu(lmm->lmm_stripe_count));
2048 GOTO(cleanup, rc = PTR_ERR(handle));
2049 rc = vfs_unlink(dparent->d_inode, dchild);
2057 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2060 GOTO(cleanup, rc = PTR_ERR(handle));
2061 rc = vfs_unlink(dparent->d_inode, dchild);
2064 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2067 GOTO(cleanup, rc = -EINVAL);
2070 if (rc == 0 && child_inode->i_nlink == 0) {
2071 if (mds_orphan_open_count(child_inode) > 0)
2072 rc = mds_orphan_add_link(rec, obd, dchild);
2075 GOTO(cleanup, rc = 0);
2077 if (!S_ISREG(child_inode->i_mode))
2080 if (!(body->valid & OBD_MD_FLEASIZE)) {
2081 body->valid |=(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2082 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2083 } else if (mds_log_op_unlink(obd, child_inode,
2084 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2085 req->rq_repmsg->buflens[offset + 1],
2086 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2087 req->rq_repmsg->buflens[offset+2],
2089 body->valid |= OBD_MD_FLCOOKIE;
2100 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2101 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2102 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2104 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2106 CERROR("error on parent setattr: rc = %d\n", err);
2108 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2109 handle, req, rc, 0);
2111 (void)obd_set_info(mds->mds_lov_exp, strlen("unlinked"),
2112 "unlinked", 0, NULL);
2113 switch(cleanup_phase) {
2114 case 5: /* pending_dir semaphore */
2115 up(&mds->mds_pending_dir->d_inode->i_sem);
2116 case 4: /* child inode semaphore */
2117 UP_READ_I_ALLOC_SEM(child_inode);
2118 /* handle splitted dir */
2120 /* master directory can be non-empty or something else ... */
2121 mds_unlink_slave_objs(obd, dchild);
2124 ptlrpc_save_llog_lock(req, lcl);
2125 case 3: /* child ino-reuse lock */
2126 if (rc && body != NULL) {
2127 // Don't unlink the OST objects if the MDS unlink failed
2131 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2133 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2134 case 2: /* child lock */
2135 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2136 if (child_lockh.cookie)
2137 ldlm_lock_decref(&child_lockh, LCK_EX);
2138 case 1: /* child and parent dentry, parent lock */
2140 if (parent_lockh[1].cookie != 0)
2141 ldlm_lock_decref(parent_lockh + 1, update_mode);
2144 ldlm_lock_decref(parent_lockh, LCK_PW);
2146 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2153 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2156 req->rq_status = rc;
2161 * to service requests from remote MDS to increment i_nlink
2163 static int mds_reint_link_acquire(struct mds_update_record *rec,
2164 int offset, struct ptlrpc_request *req,
2165 struct lustre_handle *lh)
2167 struct obd_device *obd = req->rq_export->exp_obd;
2168 struct ldlm_res_id src_res_id = { .name = {0} };
2169 struct lustre_handle *handle = NULL, src_lockh = {0};
2170 struct mds_obd *mds = mds_req2mds(req);
2171 int rc = 0, cleanup_phase = 0;
2172 struct dentry *de_src = NULL;
2173 ldlm_policy_data_t policy;
2177 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2178 obd->obd_name, OLID4(rec->ur_id1));
2180 /* Step 1: Lookup the source inode and target directory by ID */
2181 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2183 GOTO(cleanup, rc = PTR_ERR(de_src));
2184 cleanup_phase = 1; /* source dentry */
2186 src_res_id.name[0] = id_fid(rec->ur_id1);
2187 src_res_id.name[1] = id_group(rec->ur_id1);
2188 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2190 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2191 src_res_id, LDLM_IBITS, &policy,
2192 LCK_EX, &flags, mds_blocking_ast,
2193 ldlm_completion_ast, NULL, NULL,
2194 NULL, 0, NULL, &src_lockh);
2196 GOTO(cleanup, rc = -ENOLCK);
2197 cleanup_phase = 2; /* lock */
2199 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2201 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2202 if (IS_ERR(handle)) {
2203 rc = PTR_ERR(handle);
2206 de_src->d_inode->i_nlink++;
2207 mark_inode_dirty(de_src->d_inode);
2210 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2211 handle, req, rc, 0);
2213 switch (cleanup_phase) {
2216 ldlm_lock_decref(&src_lockh, LCK_EX);
2218 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2224 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2227 req->rq_status = rc;
2232 * request to link to foreign inode:
2233 * - acquire i_nlinks on this inode
2236 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2237 int offset, struct ptlrpc_request *req,
2238 struct lustre_handle *lh)
2240 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2241 struct obd_device *obd = req->rq_export->exp_obd;
2242 struct dentry *de_tgt_dir = NULL;
2243 struct mds_obd *mds = mds_req2mds(req);
2244 int rc = 0, cleanup_phase = 0;
2245 struct mdc_op_data op_data;
2246 struct ptlrpc_request *request = NULL;
2250 DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2251 ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2252 OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2253 OLID4(rec->ur_id1));
2255 de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2256 tgt_dir_lockh, &update_mode,
2257 rec->ur_name, rec->ur_namelen - 1,
2258 MDS_INODELOCK_UPDATE);
2259 if (IS_ERR(de_tgt_dir))
2260 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2263 op_data.id1 = *(rec->ur_id1);
2264 op_data.namelen = 0;
2265 op_data.name = NULL;
2266 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2272 ptlrpc_req_finished(request);
2274 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2276 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2277 if (IS_ERR(handle)) {
2278 rc = PTR_ERR(handle);
2282 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2283 rec->ur_namelen - 1, id_ino(rec->ur_id1),
2284 id_gen(rec->ur_id1), id_group(rec->ur_id1),
2285 id_fid(rec->ur_id1));
2289 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2290 handle, req, rc, 0);
2293 switch (cleanup_phase) {
2296 /* FIXME: drop i_nlink on remote inode here */
2297 CERROR("MUST drop drop i_nlink here\n");
2302 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2304 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2307 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2309 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2315 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2318 req->rq_status = rc;
2322 static int mds_reint_link(struct mds_update_record *rec, int offset,
2323 struct ptlrpc_request *req,
2324 struct lustre_handle *lh)
2326 struct obd_device *obd = req->rq_export->exp_obd;
2327 struct dentry *de_src = NULL;
2328 struct dentry *de_tgt_dir = NULL;
2329 struct dentry *dchild = NULL;
2330 struct mds_obd *mds = mds_req2mds(req);
2331 struct lustre_handle *handle = NULL;
2332 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2333 struct ldlm_res_id src_res_id = { .name = {0} };
2334 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2335 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2336 ldlm_policy_data_t tgt_dir_policy =
2337 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2338 int rc = 0, cleanup_phase = 0;
2340 int update_mode = 0;
2344 LASSERT(offset == 1);
2346 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2347 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2348 id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2351 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2352 MDS_UPDATE_COUNTER(mds, MDS_LINK_COUNT);
2354 // memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2355 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2356 GOTO(cleanup, rc = -ENOENT);
2358 if (id_group(rec->ur_id1) != mds->mds_num) {
2359 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2363 if (rec->ur_namelen == 1) {
2364 rc = mds_reint_link_acquire(rec, offset, req, lh);
2368 /* Step 1: Lookup the source inode and target directory by ID */
2369 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2371 GOTO(cleanup, rc = PTR_ERR(de_src));
2373 cleanup_phase = 1; /* source dentry */
2375 de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2376 if (IS_ERR(de_tgt_dir)) {
2377 rc = PTR_ERR(de_tgt_dir);
2382 cleanup_phase = 2; /* target directory dentry */
2384 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2385 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2386 rec->ur_name, de_src->d_inode->i_ino);
2388 /* Step 2: Take the two locks */
2389 src_res_id.name[0] = id_fid(rec->ur_id1);
2390 src_res_id.name[1] = id_group(rec->ur_id1);
2391 tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2392 tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2395 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2397 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2399 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2400 tgt_dir_res_id, LDLM_IBITS,
2401 &src_policy, update_mode, &flags,
2403 ldlm_completion_ast, NULL, NULL,
2404 NULL, 0, NULL, tgt_dir_lockh + 1);
2406 GOTO(cleanup, rc = -ENOLCK);
2409 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2410 rec->ur_namelen - 1);
2411 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2412 (unsigned long)id_fid(rec->ur_id2),
2413 (unsigned long)id_group(rec->ur_id2),
2414 tgt_dir_res_id.name[2]);
2417 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2418 &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2419 LCK_EX, &tgt_dir_policy);
2423 cleanup_phase = 3; /* locks */
2425 /* Step 3: Lookup the child */
2426 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2427 rec->ur_namelen - 1);
2428 if (IS_ERR(dchild)) {
2429 rc = PTR_ERR(dchild);
2430 if (rc != -EPERM && rc != -EACCES)
2431 CERROR("child lookup error %d\n", rc);
2435 cleanup_phase = 4; /* child dentry */
2437 if (dchild->d_inode) {
2438 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2439 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2444 /* Step 4: Do it. */
2445 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2447 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2448 if (IS_ERR(handle)) {
2449 rc = PTR_ERR(handle);
2453 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2454 if (rc && rc != -EPERM && rc != -EACCES)
2455 CERROR("vfs_link error %d\n", rc);
2457 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2458 handle, req, rc, 0);
2461 switch (cleanup_phase) {
2462 case 4: /* child dentry */
2466 ldlm_lock_decref(&src_lockh, LCK_EX);
2467 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2469 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2470 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2472 case 2: /* target dentry */
2474 if (tgt_dir_lockh[1].cookie && update_mode)
2475 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2479 case 1: /* source dentry */
2484 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2487 req->rq_status = rc;
2491 /* The idea here is that we need to get four locks in the end:
2492 * one on each parent directory, one on each child. We need to take
2493 * these locks in some kind of order (to avoid deadlocks), and the order
2494 * I selected is "increasing resource number" order. We need to look up
2495 * the children, however, before we know what the resource number(s) are.
2496 * Thus the following plan:
2498 * 1,2. Look up the parents
2499 * 3,4. Look up the children
2500 * 5. Take locks on the parents and children, in order
2501 * 6. Verify that the children haven't changed since they were looked up
2503 * If there was a race and the children changed since they were first looked
2504 * up, it is possible that mds_verify_child() will be able to just grab the
2505 * lock on the new child resource (if it has a higher resource than any other)
2506 * but we need to compare against not only its parent, but also against the
2507 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2509 * We need the fancy igrab() on the child inodes because we aren't holding a
2510 * lock on the parent after the lookup is done, so dentry->d_inode may change
2511 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2513 static int mds_get_parents_children_locked(struct obd_device *obd,
2514 struct mds_obd *mds,
2515 struct lustre_id *p1_id,
2516 struct dentry **de_srcdirp,
2517 struct lustre_id *p2_id,
2518 struct dentry **de_tgtdirp,
2520 const char *old_name, int old_len,
2521 struct dentry **de_oldp,
2522 const char *new_name, int new_len,
2523 struct dentry **de_newp,
2524 struct lustre_handle *dlm_handles,
2527 struct ldlm_res_id p1_res_id = { .name = {0} };
2528 struct ldlm_res_id p2_res_id = { .name = {0} };
2529 struct ldlm_res_id c1_res_id = { .name = {0} };
2530 struct ldlm_res_id c2_res_id = { .name = {0} };
2531 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2532 /* Only dentry should change, but the inode itself would be
2534 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2535 /* If something is going to be replaced, both dentry and inode locks are
2537 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2538 MDS_INODELOCK_UPDATE}};
2539 struct ldlm_res_id *maxres_src, *maxres_tgt;
2540 struct inode *inode;
2541 __u32 child1_gen = 0;
2542 __u32 child2_gen = 0;
2543 unsigned long child1_ino;
2544 unsigned long child2_ino;
2545 int rc = 0, cleanup_phase = 0;
2548 /* Step 1: Lookup the source directory */
2549 *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2550 if (IS_ERR(*de_srcdirp))
2551 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2553 cleanup_phase = 1; /* source directory dentry */
2555 p1_res_id.name[0] = id_fid(p1_id);
2556 p1_res_id.name[1] = id_group(p1_id);
2558 /* Step 2: Lookup the target directory */
2559 if (id_equal_stc(p1_id, p2_id)) {
2560 *de_tgtdirp = dget(*de_srcdirp);
2562 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2563 if (IS_ERR(*de_tgtdirp)) {
2564 rc = PTR_ERR(*de_tgtdirp);
2570 cleanup_phase = 2; /* target directory dentry */
2572 p2_res_id.name[0] = id_fid(p2_id);
2573 p2_res_id.name[1] = id_group(p2_id);
2576 dlm_handles[5].cookie = 0;
2577 dlm_handles[6].cookie = 0;
2579 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2581 * get a temp lock on just fid, group to flush client cache and
2582 * to protect dirs from concurrent splitting.
2584 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2585 LCK_PW, &p_policy, &p2_res_id,
2586 &dlm_handles[6], LCK_PW, &p_policy);
2590 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2591 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2593 CDEBUG(D_INFO, "take locks on "
2594 LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2595 p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2596 p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2601 /* Step 3: Lookup the source child entry */
2602 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2604 if (IS_ERR(*de_oldp)) {
2605 rc = PTR_ERR(*de_oldp);
2606 CERROR("old child lookup error (%*s): %d\n",
2607 old_len - 1, old_name, rc);
2611 cleanup_phase = 4; /* original name dentry */
2613 inode = (*de_oldp)->d_inode;
2614 if (inode != NULL) {
2615 struct lustre_id sid;
2617 inode = igrab(inode);
2619 GOTO(cleanup, rc = -ENOENT);
2621 down(&inode->i_sem);
2622 rc = mds_read_inode_sid(obd, inode, &sid);
2625 CERROR("Can't read inode self id, inode %lu, "
2626 "rc %d\n", inode->i_ino, rc);
2631 c1_res_id.name[0] = id_fid(&sid);
2632 c1_res_id.name[1] = id_group(&sid);
2633 child1_gen = inode->i_generation;
2634 child1_ino = inode->i_ino;
2636 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2637 c1_res_id.name[0] = (*de_oldp)->d_fid;
2638 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2639 child1_gen = (*de_oldp)->d_generation;
2640 child1_ino = (*de_oldp)->d_inum;
2642 GOTO(cleanup, rc = -ENOENT);
2645 /* Step 4: Lookup the target child entry */
2646 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2648 if (IS_ERR(*de_newp)) {
2649 rc = PTR_ERR(*de_newp);
2650 CERROR("new child lookup error (%*s): %d\n",
2651 old_len - 1, old_name, rc);
2655 cleanup_phase = 5; /* target dentry */
2657 inode = (*de_newp)->d_inode;
2658 if (inode != NULL) {
2659 struct lustre_id sid;
2661 inode = igrab(inode);
2665 down(&inode->i_sem);
2666 rc = mds_read_inode_sid(obd, inode, &sid);
2669 CERROR("Can't read inode self id, inode %lu, "
2670 "rc %d\n", inode->i_ino, rc);
2674 c2_res_id.name[0] = id_fid(&sid);
2675 c2_res_id.name[1] = id_group(&sid);
2676 child2_gen = inode->i_generation;
2677 child2_ino = inode->i_ino;
2679 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2680 c2_res_id.name[0] = (*de_newp)->d_fid;
2681 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
2682 child2_gen = (*de_newp)->d_generation;
2683 child2_ino = (*de_newp)->d_inum;
2687 /* Step 5: Take locks on the parents and child(ren) */
2688 maxres_src = &p1_res_id;
2689 maxres_tgt = &p2_res_id;
2690 cleanup_phase = 5; /* target dentry */
2692 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
2693 maxres_src = &c1_res_id;
2694 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
2695 maxres_tgt = &c2_res_id;
2697 rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
2699 &p2_res_id, &dlm_handles[1], parent_mode,
2701 &c1_res_id, &dlm_handles[2], child_mode,
2703 &c2_res_id, &dlm_handles[3], child_mode,
2708 cleanup_phase = 6; /* parent and child(ren) locks */
2710 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2711 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2712 parent_mode, &c1_res_id, &dlm_handles[2],
2713 de_oldp, child_mode, &c1_policy, old_name,old_len,
2714 maxres_tgt, child1_ino, child1_gen);
2716 if (c2_res_id.name[0] != 0)
2717 ldlm_lock_decref(&dlm_handles[3], child_mode);
2718 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2725 if (!DENTRY_VALID(*de_oldp))
2726 GOTO(cleanup, rc = -ENOENT);
2728 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2729 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2730 parent_mode, &c2_res_id, &dlm_handles[3],
2731 de_newp, child_mode, &c2_policy, new_name,
2732 new_len, maxres_src, child2_ino, child2_gen);
2734 ldlm_lock_decref(&dlm_handles[2], child_mode);
2735 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2745 switch (cleanup_phase) {
2746 case 6: /* child lock(s) */
2747 if (c2_res_id.name[0] != 0)
2748 ldlm_lock_decref(&dlm_handles[3], child_mode);
2749 if (c1_res_id.name[0] != 0)
2750 ldlm_lock_decref(&dlm_handles[2], child_mode);
2751 if (dlm_handles[1].cookie != 0)
2752 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2753 if (dlm_handles[0].cookie != 0)
2754 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2755 case 5: /* target dentry */
2757 case 4: /* source dentry */
2761 if (dlm_handles[5].cookie != 0)
2762 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2763 if (dlm_handles[6].cookie != 0)
2764 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2766 case 2: /* target directory dentry */
2767 l_dput(*de_tgtdirp);
2768 case 1: /* source directry dentry */
2769 l_dput(*de_srcdirp);
2777 * checks if dentry can be removed. This function also handles cross-ref
2780 static int mds_check_for_rename(struct obd_device *obd,
2781 struct dentry *dentry)
2783 struct mds_obd *mds = &obd->u.mds;
2784 struct lustre_handle *rlockh;
2785 struct ptlrpc_request *req;
2786 struct mdc_op_data op_data;
2787 struct lookup_intent it;
2788 int handle_size, rc = 0;
2791 LASSERT(dentry != NULL);
2793 if (dentry->d_inode) {
2794 if (S_ISDIR(dentry->d_inode->i_mode) &&
2795 !mds_is_dir_empty(obd, dentry))
2798 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
2799 handle_size = sizeof(struct lustre_handle);
2801 OBD_ALLOC(rlockh, handle_size);
2805 memset(rlockh, 0, handle_size);
2806 memset(&op_data, 0, sizeof(op_data));
2807 mds_pack_dentry2id(obd, &op_data.id1, dentry, 1);
2809 it.it_op = IT_UNLINK;
2810 rc = md_enqueue(mds->mds_lmv_exp, LDLM_IBITS, &it, LCK_EX,
2811 &op_data, rlockh, NULL, 0, ldlm_completion_ast,
2812 mds_blocking_ast, NULL);
2817 if (rlockh->cookie != 0)
2818 ldlm_lock_decref(rlockh, LCK_EX);
2820 if (it.d.lustre.it_data) {
2821 req = (struct ptlrpc_request *)it.d.lustre.it_data;
2822 ptlrpc_req_finished(req);
2825 if (it.d.lustre.it_status)
2826 rc = it.d.lustre.it_status;
2827 OBD_FREE(rlockh, handle_size);
2832 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
2833 struct ptlrpc_request *req, struct lustre_id *id,
2834 struct dentry *de_dir, struct dentry *de,
2837 struct obd_device *obd = req->rq_export->exp_obd;
2838 struct mds_obd *mds = mds_req2mds(req);
2839 void *handle = NULL;
2845 * name exists and points to local inode try to unlink this name
2846 * and create new one.
2848 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
2849 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
2850 (unsigned long)de->d_inode->i_generation);
2852 /* checking if we can remove local dentry. */
2853 rc = mds_check_for_rename(obd, de);
2857 handle = fsfilt_start(obd, de_dir->d_inode,
2858 FSFILT_OP_RENAME, NULL);
2860 GOTO(cleanup, rc = PTR_ERR(handle));
2861 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2864 } else if (de->d_flags & DCACHE_CROSS_REF) {
2865 struct lustre_id de_id;
2867 /* name exists and points to remote inode */
2868 mds_pack_dentry2id(obd, &de_id, de, 1);
2870 CDEBUG(D_OTHER, "%s: %s points to remote inode "DLID4"\n",
2871 obd->obd_name, rec->ur_tgt, OLID4(&de_id));
2873 /* checking if we can remove local dentry. */
2874 rc = mds_check_for_rename(obd, de);
2879 * to be fully POSIX compatible, we should add one more check:
2881 * if de_new is subdir of dir rec->ur_id1. If so - return
2884 * I do not know how to implement it right now, because
2885 * inodes/dentries for new and old names lie on different MDS,
2886 * so add this notice here just to make it visible for the rest
2887 * of developers and do not forget about. And when this check
2888 * will be added, del_cross_ref should gone, that is local
2889 * dentry is able to be removed if all checks passed.
2891 * Currently -EEXISTS is returned by fsfilt_add_dir_entry() what
2892 * is not fully correct. --umka
2895 if (del_cross_ref) {
2896 handle = fsfilt_start(obd, de_dir->d_inode,
2897 FSFILT_OP_RENAME, NULL);
2899 GOTO(cleanup, rc = PTR_ERR(handle));
2900 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2906 /* name doesn't exist. the simplest case. */
2907 handle = fsfilt_start(obd, de_dir->d_inode,
2908 FSFILT_OP_LINK, NULL);
2910 GOTO(cleanup, rc = PTR_ERR(handle));
2913 rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
2914 rec->ur_tgtlen - 1, id_ino(id),
2915 id_gen(id), id_group(id), id_fid(id));
2917 CERROR("add_dir_entry() returned error %d\n", rc);
2923 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2924 handle, req, rc, 0);
2929 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
2930 struct ptlrpc_request *req, struct dentry *de_dir,
2933 struct obd_device *obd = req->rq_export->exp_obd;
2934 struct mds_obd *mds = mds_req2mds(req);
2935 void *handle = NULL;
2939 handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
2941 GOTO(cleanup, rc = PTR_ERR(handle));
2942 rc = fsfilt_del_dir_entry(obd, de);
2947 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2948 handle, req, rc, 0);
2952 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2953 int offset, struct ptlrpc_request *req)
2955 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2956 struct obd_device *obd = req->rq_export->exp_obd;
2957 struct mds_obd *mds = mds_req2mds(req);
2958 struct lustre_handle child_lockh = {0};
2959 struct dentry *de_tgtdir = NULL;
2960 struct dentry *de_new = NULL;
2961 int cleanup_phase = 0;
2962 int update_mode, rc = 0;
2966 * another MDS executing rename operation has asked us to create target
2967 * name. such a creation should destroy existing target name.
2969 CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
2970 obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
2972 /* first, lookup the target */
2973 child_lockh.cookie = 0;
2975 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
2976 &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
2977 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2978 &child_lockh, &de_new, LCK_EX,
2979 MDS_INODELOCK_LOOKUP);
2986 LASSERT(de_tgtdir->d_inode);
2989 rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
2990 de_tgtdir, de_new, 0);
2996 if (cleanup_phase == 1) {
2998 if (parent_lockh[1].cookie != 0)
2999 ldlm_lock_decref(&parent_lockh[1], update_mode);
3001 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
3002 if (child_lockh.cookie != 0)
3003 ldlm_lock_decref(&child_lockh, LCK_EX);
3008 req->rq_status = rc;
3012 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3013 struct ptlrpc_request *req)
3015 struct obd_device *obd = req->rq_export->exp_obd;
3016 struct ptlrpc_request *req2 = NULL;
3017 struct dentry *de_srcdir = NULL;
3018 struct dentry *de_old = NULL;
3019 struct mds_obd *mds = mds_req2mds(req);
3020 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3021 struct lustre_handle child_lockh = {0};
3022 struct mdc_op_data opdata;
3023 int update_mode, rc = 0;
3026 CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3027 obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3028 memset(&opdata, 0, sizeof(opdata));
3030 child_lockh.cookie = 0;
3031 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3032 &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3033 &update_mode, rec->ur_name,
3034 rec->ur_namelen, &child_lockh, &de_old,
3035 LCK_EX, MDS_INODELOCK_LOOKUP);
3038 LASSERT(de_srcdir->d_inode);
3042 * we already know the target should be created on another MDS so, we
3043 * have to request that MDS to do it.
3046 /* prepare source id */
3047 if (de_old->d_flags & DCACHE_CROSS_REF) {
3048 LASSERT(de_old->d_inode == NULL);
3049 CDEBUG(D_OTHER, "request to move remote name\n");
3050 mds_pack_dentry2id(obd, &opdata.id1, de_old, 1);
3051 } else if (de_old->d_inode == NULL) {
3052 /* oh, source doesn't exist */
3053 GOTO(cleanup, rc = -ENOENT);
3055 struct lustre_id sid;
3056 struct inode *inode = de_old->d_inode;
3058 LASSERT(inode != NULL);
3059 CDEBUG(D_OTHER, "request to move local name\n");
3060 id_ino(&opdata.id1) = inode->i_ino;
3061 id_group(&opdata.id1) = mds->mds_num;
3062 id_gen(&opdata.id1) = inode->i_generation;
3064 down(&inode->i_sem);
3065 rc = mds_read_inode_sid(obd, inode, &sid);
3068 CERROR("Can't read inode self id, "
3069 "inode %lu, rc = %d\n",
3074 id_fid(&opdata.id1) = id_fid(&sid);
3077 opdata.id2 = *rec->ur_id2;
3078 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0,
3079 rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3084 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3088 ptlrpc_req_finished(req2);
3091 if (parent_lockh[1].cookie != 0)
3092 ldlm_lock_decref(&parent_lockh[1], update_mode);
3094 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
3095 if (child_lockh.cookie != 0)
3096 ldlm_lock_decref(&child_lockh, LCK_EX);
3101 req->rq_status = rc;
3105 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3106 struct ptlrpc_request *req, struct lustre_handle *lockh)
3108 struct obd_device *obd = req->rq_export->exp_obd;
3109 struct dentry *de_srcdir = NULL;
3110 struct dentry *de_tgtdir = NULL;
3111 struct dentry *de_old = NULL;
3112 struct dentry *de_new = NULL;
3113 struct inode *old_inode = NULL, *new_inode = NULL;
3114 struct mds_obd *mds = mds_req2mds(req);
3115 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3116 struct mds_body *body = NULL;
3117 struct llog_create_locks *lcl = NULL;
3118 struct lov_mds_md *lmm = NULL;
3119 int rc = 0, cleanup_phase = 0;
3120 void *handle = NULL;
3123 LASSERT(offset == 1);
3125 DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3126 OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3129 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3131 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3132 DEBUG_REQ(D_HA, req, "rename replay\n");
3133 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3134 lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3135 req->rq_repmsg->buflens[2]);
3138 MDS_UPDATE_COUNTER(mds, MDS_RENAME_COUNT);
3140 if (rec->ur_namelen == 1) {
3141 rc = mds_reint_rename_create_name(rec, offset, req);
3145 /* check if new name should be located on remote target. */
3146 if (id_group(rec->ur_id2) != mds->mds_num) {
3147 rc = mds_reint_rename_to_remote(rec, offset, req);
3151 rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3152 rec->ur_id2, &de_tgtdir, LCK_PW,
3153 rec->ur_name, rec->ur_namelen,
3154 &de_old, rec->ur_tgt,
3155 rec->ur_tgtlen, &de_new,
3156 dlm_handles, LCK_EX);
3160 cleanup_phase = 1; /* parent(s), children, locks */
3161 old_inode = de_old->d_inode;
3162 new_inode = de_new->d_inode;
3164 /* sanity check for src inode */
3165 if (de_old->d_flags & DCACHE_CROSS_REF) {
3166 LASSERT(de_old->d_inode == NULL);
3169 * in the case of cross-ref dir, we can perform this check only
3170 * if child and parent lie on the same mds. This is because
3171 * otherwise they can have the same inode numbers.
3173 if (de_old->d_mdsnum == mds->mds_num) {
3174 if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3175 de_old->d_inum == de_tgtdir->d_inode->i_ino)
3176 GOTO(cleanup, rc = -EINVAL);
3179 LASSERT(de_old->d_inode != NULL);
3180 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3181 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3182 GOTO(cleanup, rc = -EINVAL);
3185 /* sanity check for dest inode */
3186 if (de_new->d_flags & DCACHE_CROSS_REF) {
3187 LASSERT(new_inode == NULL);
3189 /* the same check about target dentry. */
3190 if (de_new->d_mdsnum == mds->mds_num) {
3191 if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3192 de_new->d_inum == de_tgtdir->d_inode->i_ino)
3193 GOTO(cleanup, rc = -EINVAL);
3197 * regular files usualy do not have ->rename() implemented. But
3198 * we handle only this case when @de_new is cross-ref entry,
3199 * because in other cases it will be handled by vfs_rename().
3201 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3202 !de_old->d_inode->i_op->rename))
3203 GOTO(cleanup, rc = -EPERM);
3206 (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3207 new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3208 GOTO(cleanup, rc = -EINVAL);
3212 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3214 * check if we are moving old entry into its child. 2.6 does not check
3215 * for this in vfs_rename() anymore.
3217 if (is_subdir(de_new, de_old))
3218 GOTO(cleanup, rc = -EINVAL);
3221 /* check if inodes point to each other. */
3222 if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3223 !(de_new->d_flags & DCACHE_CROSS_REF) &&
3224 old_inode == new_inode)
3225 GOTO(cleanup, rc = 0);
3228 * if we are about to remove the target at first, pass the EA of that
3229 * inode to client to perform and cleanup on OST.
3231 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3232 LASSERT(body != NULL);
3234 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3236 DOWN_READ_I_ALLOC_SEM(new_inode);
3238 cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3240 if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3241 new_inode->i_nlink == 2) ||
3242 new_inode->i_nlink == 1)) {
3243 if (mds_orphan_open_count(new_inode) > 0) {
3244 /* need to lock pending_dir before transaction */
3245 down(&mds->mds_pending_dir->d_inode->i_sem);
3246 cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3247 } else if (S_ISREG(new_inode->i_mode)) {
3248 mds_pack_inode2body(obd, body, new_inode, 0);
3249 mds_pack_md(obd, req->rq_repmsg, 1, body,
3250 new_inode, MDS_PACK_MD_LOCK);
3254 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3255 de_srcdir->d_inode->i_sb);
3257 if (de_old->d_flags & DCACHE_CROSS_REF) {
3258 struct lustre_id old_id;
3260 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3262 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3263 de_tgtdir, de_new, 1);
3267 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3272 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3273 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3274 NULL, le32_to_cpu(lmm->lmm_stripe_count));
3277 GOTO(cleanup, rc = PTR_ERR(handle));
3280 de_old->d_fsdata = req;
3281 de_new->d_fsdata = req;
3282 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3285 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3286 if (mds_orphan_open_count(new_inode) > 0)
3287 rc = mds_orphan_add_link(rec, obd, de_new);
3290 GOTO(cleanup, rc = 0);
3292 if (!S_ISREG(new_inode->i_mode))
3295 if (!(body->valid & OBD_MD_FLEASIZE)) {
3296 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3297 OBD_MD_FLATIME | OBD_MD_FLMTIME);
3298 } else if (mds_log_op_unlink(obd, new_inode,
3299 lustre_msg_buf(req->rq_repmsg,1,0),
3300 req->rq_repmsg->buflens[1],
3301 lustre_msg_buf(req->rq_repmsg,2,0),
3302 req->rq_repmsg->buflens[2],
3304 body->valid |= OBD_MD_FLCOOKIE;
3311 rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3312 handle, req, rc, 0);
3314 switch (cleanup_phase) {
3316 up(&mds->mds_pending_dir->d_inode->i_sem);
3319 UP_READ_I_ALLOC_SEM(new_inode);
3322 if (dlm_handles[5].cookie != 0)
3323 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3324 if (dlm_handles[6].cookie != 0)
3325 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3328 ptlrpc_save_llog_lock(req, lcl);
3331 if (dlm_handles[3].cookie != 0)
3332 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3333 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3334 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3335 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3337 if (dlm_handles[3].cookie != 0)
3338 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3339 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3340 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3341 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3350 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3353 req->rq_status = rc;
3357 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3358 struct ptlrpc_request *, struct lustre_handle *);
3360 static mds_reinter reinters[REINT_MAX + 1] = {
3361 [REINT_SETATTR] mds_reint_setattr,
3362 [REINT_CREATE] mds_reint_create,
3363 [REINT_LINK] mds_reint_link,
3364 [REINT_UNLINK] mds_reint_unlink,
3365 [REINT_RENAME] mds_reint_rename,
3366 [REINT_OPEN] mds_open
3369 int mds_reint_rec(struct mds_update_record *rec, int offset,
3370 struct ptlrpc_request *req, struct lustre_handle *lockh)
3372 struct obd_device *obd = req->rq_export->exp_obd;
3373 struct lvfs_run_ctxt saved;
3376 /* checked by unpacker */
3377 LASSERT(rec->ur_opcode <= REINT_MAX &&
3378 reinters[rec->ur_opcode] != NULL);
3380 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3381 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3382 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);