1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
5 * Lustre Metadata Server (mds) reintegration routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/obd_support.h>
35 #include <linux/obd_class.h>
36 #include <linux/obd.h>
37 #include <linux/lustre_lib.h>
38 #include <linux/lustre_idl.h>
39 #include <linux/lustre_mds.h>
40 #include <linux/lustre_dlm.h>
41 #include <linux/lustre_log.h>
42 #include <linux/lustre_fsfilt.h>
44 #include "mds_internal.h"
46 void mds_commit_cb(struct obd_device *obd, __u64 transno, void *data,
49 obd_transno_commit_cb(obd, transno, error);
52 struct mds_logcancel_data {
53 struct lov_mds_md *mlcd_lmm;
57 struct llog_cookie mlcd_cookies[0];
61 static void mds_cancel_cookies_cb(struct obd_device *obd, __u64 transno,
62 void *cb_data, int error)
64 struct mds_logcancel_data *mlcd = cb_data;
65 struct lov_stripe_md *lsm = NULL;
66 struct llog_ctxt *ctxt;
69 obd_transno_commit_cb(obd, transno, error);
71 CDEBUG(D_HA, "cancelling %d cookies\n",
72 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
74 rc = obd_unpackmd(obd->u.mds.mds_osc_exp, &lsm, mlcd->mlcd_lmm,
75 mlcd->mlcd_eadatalen);
77 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
78 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
81 ///* XXX 0 normally, SENDNOW for debug */);
82 ctxt = llog_get_context(&obd->obd_llogs,
83 mlcd->mlcd_cookies[0].lgc_subsys + 1);
84 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
85 sizeof(*mlcd->mlcd_cookies),
86 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
88 CERROR("error cancelling %d log cookies: rc %d\n",
89 (int)(mlcd->mlcd_cookielen /
90 sizeof(*mlcd->mlcd_cookies)), rc);
93 OBD_FREE(mlcd, mlcd->mlcd_size);
96 /* Assumes caller has already pushed us into the kernel context. */
97 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
98 struct ptlrpc_request *req, int rc, __u32 op_data)
100 struct mds_export_data *med = &req->rq_export->exp_mds_data;
101 struct mds_client_data *mcd = med->med_mcd;
102 struct obd_device *obd = req->rq_export->exp_obd;
109 /* if the export has already been failed, we have no last_rcvd slot */
110 if (req->rq_export->exp_failed) {
111 CERROR("committing transaction for disconnected client\n");
120 if (handle == NULL) {
121 /* if we're starting our own xaction, use our own inode */
122 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
123 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
124 if (IS_ERR(handle)) {
125 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
126 RETURN(PTR_ERR(handle));
132 transno = req->rq_reqmsg->transno;
134 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
135 } else if (transno == 0) {
136 spin_lock(&mds->mds_transno_lock);
137 transno = ++mds->mds_last_transno;
138 spin_unlock(&mds->mds_transno_lock);
140 spin_lock(&mds->mds_transno_lock);
141 if (transno > mds->mds_last_transno)
142 mds->mds_last_transno = transno;
143 spin_unlock(&mds->mds_transno_lock);
145 req->rq_repmsg->transno = req->rq_transno = transno;
146 mcd->mcd_last_transno = cpu_to_le64(transno);
147 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
148 mcd->mcd_last_result = cpu_to_le32(rc);
149 mcd->mcd_last_data = cpu_to_le32(op_data);
152 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb,
153 transno, handle, mds_commit_cb, NULL);
155 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd, sizeof(*mcd),
164 DEBUG_REQ(log_pri, req,
165 "wrote trans #"LPU64" client %s at idx %u: err = %d",
166 transno, mcd->mcd_uuid, med->med_idx, err);
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_obd *obd = &req->rq_export->exp_obd->u.mds;
350 struct mds_body *body;
352 mds_req_from_mcd(req, med->med_mcd);
354 de = mds_fid2dentry(obd, rec->ur_fid1, NULL);
356 LASSERT(PTR_ERR(de) == req->rq_status);
360 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
361 mds_pack_inode2fid(req2obd(req), &body->fid1, de->d_inode);
362 mds_pack_inode2body(req2obd(req), body, de->d_inode);
364 /* Don't return OST-specific attributes if we didn't just set them */
365 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
366 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
367 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
368 body->valid |= OBD_MD_FLMTIME;
369 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
370 body->valid |= OBD_MD_FLATIME;
375 /* In the raw-setattr case, we lock the child inode.
376 * In the write-back case or if being called from open, the client holds a lock
379 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
380 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
381 struct ptlrpc_request *req,
382 struct lustre_handle *lh)
384 struct mds_obd *mds = mds_req2mds(req);
385 struct obd_device *obd = req->rq_export->exp_obd;
386 struct mds_body *body;
388 struct inode *inode = NULL;
389 struct lustre_handle lockh[2] = {{0}, {0}};
392 struct mds_logcancel_data *mlcd = NULL;
393 int rc = 0, cleanup_phase = 0, err, locked = 0;
396 LASSERT(offset == 0);
398 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x", rec->ur_fid1->id,
399 rec->ur_fid1->generation, rec->ur_iattr.ia_valid);
401 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
403 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
404 de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
406 GOTO(cleanup, rc = PTR_ERR(de));
408 __u64 lockpart = MDS_INODELOCK_UPDATE;
409 if (rec->ur_iattr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID) )
410 lockpart |= MDS_INODELOCK_LOOKUP;
411 de = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW,
412 lockh, &parent_mode, NULL, 0, lockpart);
414 GOTO(cleanup, rc = PTR_ERR(de));
422 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
423 rec->ur_eadata != NULL)
426 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
428 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
430 GOTO(cleanup, rc = PTR_ERR(handle));
432 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
433 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
434 LTIME_S(rec->ur_iattr.ia_mtime),
435 LTIME_S(rec->ur_iattr.ia_ctime));
436 rc = mds_fix_attr(inode, rec);
440 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
441 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
442 (long)&rec->ur_iattr.ia_attr_flags);
444 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
446 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
447 rec->ur_eadata != NULL) {
448 struct lov_stripe_md *lsm = NULL;
450 rc = ll_permission(inode, MAY_WRITE, NULL);
454 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
455 mds->mds_osc_exp, 0, &lsm, rec->ur_eadata);
459 obd_free_memmd(mds->mds_osc_exp, &lsm);
461 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
467 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
468 mds_pack_inode2fid(obd, &body->fid1, inode);
469 mds_pack_inode2body(obd, body, inode);
471 /* Don't return OST-specific attributes if we didn't just set them */
472 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
473 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
474 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
475 body->valid |= OBD_MD_FLMTIME;
476 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
477 body->valid |= OBD_MD_FLATIME;
479 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_osc_obd)) {
480 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
483 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
485 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
486 mlcd->mlcd_cookielen = rec->ur_cookielen;
487 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
488 mlcd->mlcd_cookielen;
489 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
490 mlcd->mlcd_cookielen);
491 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
492 mlcd->mlcd_eadatalen);
494 CERROR("unable to allocate log cancel data\n");
500 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
501 handle, mds_cancel_cookies_cb, mlcd);
502 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
503 switch (cleanup_phase) {
505 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
506 rec->ur_eadata != NULL)
511 if (lockh[1].cookie != 0)
512 ldlm_lock_decref(lockh + 1, parent_mode);
515 ldlm_lock_decref(lockh, LCK_PW);
517 ptlrpc_save_lock (req, lockh, LCK_PW);
532 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
533 struct ptlrpc_request *req)
535 struct mds_export_data *med = &req->rq_export->exp_mds_data;
536 struct mds_obd *obd = &req->rq_export->exp_obd->u.mds;
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_fid2dentry(obd, rec->ur_fid1, NULL);
549 LASSERT(!IS_ERR(parent));
550 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
551 LASSERT(!IS_ERR(child));
552 if ((child->d_flags & DCACHE_CROSS_REF)) {
553 LASSERTF(child->d_inode == NULL, "BUG 3869");
554 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
555 mds_pack_dentry2fid(&body->fid1, child);
556 mds_pack_dentry2body(body, child);
557 body->valid |= OBD_MD_MDS;
558 } else if (child->d_inode == NULL) {
559 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
560 rec->ur_fid1->id, rec->ur_fid1->generation,
561 rec->ur_name, rec->ur_mode);
562 LASSERTF(child->d_inode != NULL, "BUG 3869");
564 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
565 mds_pack_inode2fid(req2obd(req), &body->fid1, child->d_inode);
566 mds_pack_inode2body(req2obd(req), body, child->d_inode);
573 static int mds_reint_create(struct mds_update_record *rec, int offset,
574 struct ptlrpc_request *req,
575 struct lustre_handle *lh)
577 struct dentry *dparent = NULL;
578 struct mds_obd *mds = mds_req2mds(req);
579 struct obd_device *obd = req->rq_export->exp_obd;
580 struct dentry *dchild = NULL;
581 struct inode *dir = NULL;
583 struct lustre_handle lockh[2] = {{0}, {0}};
585 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
587 struct dentry_params dp;
588 struct mea *mea = NULL;
592 LASSERT(offset == 0);
593 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, LUSTRE_MDS_NAME));
595 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
596 rec->ur_fid1->id, rec->ur_fid1->generation,
597 rec->ur_name, rec->ur_mode);
599 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
601 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
602 GOTO(cleanup, rc = -ESTALE);
604 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, LCK_PW, lockh,
605 &parent_mode, rec->ur_name,
606 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
607 if (IS_ERR(dparent)) {
608 rc = PTR_ERR(dparent);
609 CERROR("parent lookup error %d\n", rc);
612 cleanup_phase = 1; /* locked parent dentry */
613 dir = dparent->d_inode;
616 ldlm_lock_dump_handle(D_OTHER, lockh);
618 /* try to retrieve MEA data for this dir */
619 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
624 /* dir is already splitted, check is requested filename
625 * should live at this MDS or at another one */
627 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
628 if (mea->mea_master != mea->mea_fids[i].mds) {
629 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
630 " should be %d(%d)\n",
631 mea->mea_master, dparent->d_inode->i_ino,
632 dparent->d_inode->i_generation, rec->ur_name,
633 mea->mea_fids[i].mds, i);
634 GOTO(cleanup, rc = -ERESTART);
638 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
639 if (IS_ERR(dchild)) {
640 rc = PTR_ERR(dchild);
641 CERROR("child lookup error %d\n", rc);
645 cleanup_phase = 2; /* child dentry */
647 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
649 if (type == S_IFREG || type == S_IFDIR) {
650 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
651 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
652 obd->obd_name, dparent->d_inode->i_ino,
653 dparent->d_inode->i_generation, rc, parent_mode);
655 /* dir got splitted */
656 GOTO(cleanup, rc = -ERESTART);
658 /* error happened during spitting. */
663 if (dir->i_mode & S_ISGID) {
664 if (S_ISDIR(rec->ur_mode))
665 rec->ur_mode |= S_ISGID;
668 dchild->d_fsdata = (void *)&dp;
669 dp.p_inum = (unsigned long)rec->ur_fid2->id;
674 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
676 GOTO(cleanup, rc = PTR_ERR(handle));
677 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
685 /* as Peter asked, mkdir() should distribute new directories
686 * over the whole cluster in order to distribute namespace
687 * processing load. first, we calculate which MDS to use to put
688 * new directory's inode in. */
689 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
691 if (i == mds->mds_num) {
692 /* inode will be created locally */
693 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
695 GOTO(cleanup, rc = PTR_ERR(handle));
697 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
700 nstripes = *(u16 *)rec->ur_eadata;
702 if (rc == 0 && nstripes) {
703 /* we pass LCK_EX to split routine to
704 * signalthat we have exclusive access
705 * to the directory. simple because
706 * nobody knows it already exists -bzzz */
707 rc = mds_try_to_split_dir(obd, dchild,
711 /* dir got splitted */
714 /* an error occured during
719 } else if (!DENTRY_VALID(dchild)) {
720 /* inode will be created on another MDS */
721 struct obdo *oa = NULL;
722 struct mds_body *body;
724 /* first, create that inode */
729 if (rec->ur_eadata) {
730 /* user asks for creating splitted dir */
731 oa->o_easize = *((u16 *) rec->ur_eadata);
734 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
735 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
736 OBD_MD_FLUID | OBD_MD_FLGID);
737 oa->o_mode = dir->i_mode;
738 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
740 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
741 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
742 obd->obd_name, rec->ur_namelen - 1,
743 rec->ur_name, (unsigned) rec->ur_fid2->id,
744 (unsigned) rec->ur_fid2->generation);
745 oa->o_id = rec->ur_fid2->id;
746 oa->o_generation = rec->ur_fid2->generation;
747 oa->o_flags |= OBD_FL_RECREATE_OBJS;
750 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
756 /* now, add new dir entry for it */
757 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
758 if (IS_ERR(handle)) {
760 GOTO(cleanup, rc = PTR_ERR(handle));
762 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
764 oa->o_id, oa->o_generation,
769 body = lustre_msg_buf(req->rq_repmsg,
770 offset, sizeof (*body));
771 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
772 body->fid1.id = oa->o_id;
774 body->fid1.generation = oa->o_generation;
777 /* requested name exists in the directory */
784 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
786 GOTO(cleanup, rc = PTR_ERR(handle));
787 if (rec->ur_tgt == NULL) /* no target supplied */
788 rc = -EINVAL; /* -EPROTO? */
790 rc = vfs_symlink(dir, dchild, rec->ur_tgt);
798 int rdev = rec->ur_rdev;
799 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
801 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
802 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
807 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
808 dchild->d_fsdata = NULL;
809 GOTO(cleanup, rc = -EINVAL);
812 /* In case we stored the desired inum in here, we want to clean up. */
813 if (dchild->d_fsdata == (void *)(unsigned long)rec->ur_fid2->id)
814 dchild->d_fsdata = NULL;
817 CDEBUG(D_INODE, "error during create: %d\n", rc);
819 } else if (dchild->d_inode) {
821 struct inode *inode = dchild->d_inode;
822 struct mds_body *body;
825 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
826 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
827 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
828 iattr.ia_uid = rec->_ur_fsuid;
829 if (dir->i_mode & S_ISGID)
830 iattr.ia_gid = dir->i_gid;
832 iattr.ia_gid = rec->_ur_fsgid;
833 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
834 ATTR_MTIME | ATTR_CTIME;
836 if (rec->ur_fid2->id) {
837 LASSERT(rec->ur_fid2->id == inode->i_ino);
838 inode->i_generation = rec->ur_fid2->generation;
839 /* Dirtied and committed by the upcoming setattr. */
840 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
841 inode->i_ino, inode->i_generation);
844 struct lustre_handle child_ino_lockh;
847 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
848 inode->i_ino, inode->i_generation);
851 /* The inode we were allocated may have just been freed
852 * by an unlink operation. We take this lock to
853 * synchronize against the matching reply-ack-lock taken
854 * in unlink, to avoid replay problems if this reply
855 * makes it out to the client but the unlink's does not.
856 * See bug 2029 for more detail.*/
857 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
858 if (rc != ELDLM_OK) {
859 CERROR("error locking for unlink/create sync: "
862 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
867 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
869 CERROR("error on child setattr: rc = %d\n", rc);
871 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
872 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
874 CERROR("error on parent setattr: rc = %d\n", rc);
876 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
877 mds_pack_inode2fid(obd, &body->fid1, inode);
878 mds_pack_inode2body(obd, body, inode);
883 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
886 /* Destroy the file we just created. This should not need extra
887 * journal credits, as we have already modified all of the
888 * blocks needed in order to create the file in the first
892 err = vfs_rmdir(dir, dchild);
894 CERROR("rmdir in error path: %d\n", err);
897 err = vfs_unlink(dir, dchild);
899 CERROR("unlink in error path: %d\n", err);
905 switch (cleanup_phase) {
906 case 2: /* child dentry */
908 case 1: /* locked parent dentry */
910 if (lockh[1].cookie != 0)
911 ldlm_lock_decref(lockh + 1, parent_mode);
914 ldlm_lock_decref(lockh, LCK_PW);
916 ptlrpc_save_lock(req, lockh, LCK_PW);
922 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
926 OBD_FREE(mea, mea_size);
931 static int res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
932 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
936 for (i = 0; i < RES_NAME_SIZE; i++) {
937 /* return 1 here, because enqueue_ordered will skip resources
938 * of all zeroes if they're sorted to the end of the list. */
939 if (res1->name[i] == 0 && res2->name[i] != 0)
941 if (res2->name[i] == 0 && res1->name[i] != 0)
944 if (res1->name[i] > res2->name[i])
946 if (res1->name[i] < res2->name[i])
953 if (memcmp(p1, p2, sizeof(*p1)) < 0)
959 /* This function doesn't use ldlm_match_or_enqueue because we're always called
960 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
961 * because they take the place of local semaphores.
963 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
964 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
965 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
966 struct lustre_handle *p1_lockh, int p1_lock_mode,
967 ldlm_policy_data_t *p1_policy,
968 struct ldlm_res_id *p2_res_id,
969 struct lustre_handle *p2_lockh, int p2_lock_mode,
970 ldlm_policy_data_t *p2_policy)
972 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
973 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
974 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
975 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
979 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
981 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n", res_id[0]->name[0],
984 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
985 handles[1] = p1_lockh;
986 handles[0] = p2_lockh;
987 res_id[1] = p1_res_id;
988 res_id[0] = p2_res_id;
989 lock_modes[1] = p1_lock_mode;
990 lock_modes[0] = p2_lock_mode;
991 policies[1] = p1_policy;
992 policies[0] = p2_policy;
995 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
996 res_id[0]->name[0], res_id[1]->name[0]);
998 flags = LDLM_FL_LOCAL_ONLY;
999 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1000 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1001 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1002 NULL, 0, NULL, handles[0]);
1005 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1007 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1008 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1009 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1010 ldlm_lock_addref(handles[1], lock_modes[1]);
1011 } else if (res_id[1]->name[0] != 0) {
1012 flags = LDLM_FL_LOCAL_ONLY;
1013 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1014 *res_id[1], LDLM_IBITS, policies[1],
1015 lock_modes[1], &flags, mds_blocking_ast,
1016 ldlm_completion_ast, NULL, NULL, NULL, 0,
1018 if (rc != ELDLM_OK) {
1019 ldlm_lock_decref(handles[0], lock_modes[0]);
1022 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1028 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1029 struct lustre_handle *p1_lockh, int p1_lock_mode,
1030 ldlm_policy_data_t *p1_policy,
1031 struct ldlm_res_id *p2_res_id,
1032 struct lustre_handle *p2_lockh, int p2_lock_mode,
1033 ldlm_policy_data_t *p2_policy,
1034 struct ldlm_res_id *c1_res_id,
1035 struct lustre_handle *c1_lockh, int c1_lock_mode,
1036 ldlm_policy_data_t *c1_policy,
1037 struct ldlm_res_id *c2_res_id,
1038 struct lustre_handle *c2_lockh, int c2_lock_mode,
1039 ldlm_policy_data_t *c2_policy)
1041 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1042 c1_res_id, c2_res_id };
1043 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1044 c1_lockh, c2_lockh };
1045 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1046 c1_lock_mode, c2_lock_mode };
1047 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1048 c1_policy, c2_policy};
1049 int rc, i, j, sorted, flags;
1053 "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1054 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1055 res_id[3]->name[0]);
1057 /* simple insertion sort - we have at most 4 elements */
1058 for (i = 1; i < 4; i++) {
1060 dlm_handles[4] = dlm_handles[i];
1061 res_id[4] = res_id[i];
1062 lock_modes[4] = lock_modes[i];
1063 policies[4] = policies[i];
1067 if (res_gt(res_id[j], res_id[4], policies[j],
1069 dlm_handles[j + 1] = dlm_handles[j];
1070 res_id[j + 1] = res_id[j];
1071 lock_modes[j + 1] = lock_modes[j];
1072 policies[j + 1] = policies[j];
1077 } while (j >= 0 && !sorted);
1079 dlm_handles[j + 1] = dlm_handles[4];
1080 res_id[j + 1] = res_id[4];
1081 lock_modes[j + 1] = lock_modes[4];
1082 policies[j + 1] = policies[4];
1086 "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1087 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1088 res_id[3]->name[0]);
1090 /* XXX we could send ASTs on all these locks first before blocking? */
1091 for (i = 0; i < 4; i++) {
1093 if (res_id[i]->name[0] == 0)
1096 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1097 (policies[i]->l_inodebits.bits &
1098 policies[i-1]->l_inodebits.bits) ) {
1099 memcpy(dlm_handles[i], dlm_handles[i-1],
1100 sizeof(*(dlm_handles[i])));
1101 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1103 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1104 *res_id[i], LDLM_IBITS,
1106 lock_modes[i], &flags,
1108 ldlm_completion_ast, NULL, NULL,
1109 NULL, 0, NULL, dlm_handles[i]);
1111 GOTO(out_err, rc = -EIO);
1112 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1119 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1124 /* In the unlikely case that the child changed while we were waiting
1125 * on the lock, we need to drop the lock on the old child and either:
1126 * - if the child has a lower resource name, then we have to also
1127 * drop the parent lock and regain the locks in the right order
1128 * - in the rename case, if the child has a lower resource name than one of
1129 * the other parent/child resources (maxres) we also need to reget the locks
1130 * - if the child has a higher resource name (this is the common case)
1131 * we can just get the lock on the new child (still in lock order)
1133 * Returns 0 if the child did not change or if it changed but could be locked.
1134 * Returns 1 if the child changed and we need to re-lock (no locks held).
1135 * Returns -ve error with a valid dchild (no locks held). */
1136 static int mds_verify_child(struct obd_device *obd,
1137 struct ldlm_res_id *parent_res_id,
1138 struct lustre_handle *parent_lockh,
1139 struct dentry *dparent, int parent_mode,
1140 struct ldlm_res_id *child_res_id,
1141 struct lustre_handle *child_lockh,
1142 struct dentry **dchildp, int child_mode,
1143 ldlm_policy_data_t *child_policy,
1144 const char *name, int namelen,
1145 struct ldlm_res_id *maxres)
1147 struct dentry *vchild, *dchild = *dchildp;
1148 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1151 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1153 GOTO(cleanup, rc = PTR_ERR(vchild));
1155 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1156 if (child_res_id->name[0] == vchild->d_inum &&
1157 child_res_id->name[1] == vchild->d_generation) {
1166 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1167 (vchild->d_inode != NULL &&
1168 child_res_id->name[0] == vchild->d_inode->i_ino &&
1169 child_res_id->name[1] == vchild->d_inode->i_generation))) {
1178 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1179 vchild->d_inode, dchild ? dchild->d_inode : 0,
1180 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1181 child_res_id->name[0]);
1182 if (child_res_id->name[0] != 0)
1183 ldlm_lock_decref(child_lockh, child_mode);
1187 cleanup_phase = 1; /* parent lock only */
1188 *dchildp = dchild = vchild;
1190 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1192 if (dchild->d_inode) {
1193 child_res_id->name[0] = dchild->d_inode->i_ino;
1194 child_res_id->name[1] = dchild->d_inode->i_generation;
1196 child_res_id->name[0] = dchild->d_inum;
1197 child_res_id->name[1] = dchild->d_generation;
1200 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1201 res_gt(maxres, child_res_id, NULL, NULL)) {
1202 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1203 child_res_id->name[0], parent_res_id->name[0],
1205 GOTO(cleanup, rc = 1);
1208 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1209 *child_res_id, LDLM_IBITS, child_policy,
1210 child_mode, &flags, mds_blocking_ast,
1211 ldlm_completion_ast, NULL, NULL, NULL, 0,
1214 GOTO(cleanup, rc = -EIO);
1217 memset(child_res_id, 0, sizeof(*child_res_id));
1223 switch(cleanup_phase) {
1225 if (child_res_id->name[0] != 0)
1226 ldlm_lock_decref(child_lockh, child_mode);
1228 ldlm_lock_decref(parent_lockh, parent_mode);
1234 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1236 struct lustre_handle *parent_lockh,
1237 struct dentry **dparentp, int parent_mode,
1238 __u64 parent_lockpart, int *update_mode,
1239 char *name, int namelen,
1240 struct lustre_handle *child_lockh,
1241 struct dentry **dchildp, int child_mode,
1242 __u64 child_lockpart, void *clone_info)
1244 struct ldlm_res_id child_res_id = { .name = {0} };
1245 struct ldlm_res_id parent_res_id = { .name = {0} };
1246 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1247 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1248 struct inode *inode;
1249 int rc = 0, cleanup_phase = 0;
1252 /* Step 1: Lookup parent */
1253 *dparentp = mds_fid2dentry(mds, fid, NULL);
1254 if (IS_ERR(*dparentp)) {
1255 rc = PTR_ERR(*dparentp);
1260 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1261 (*dparentp)->d_inode->i_ino, name);
1263 parent_res_id.name[0] = (*dparentp)->d_inode->i_ino;
1264 parent_res_id.name[1] = (*dparentp)->d_inode->i_generation;
1266 parent_lockh[1].cookie = 0;
1267 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1268 struct ldlm_res_id res_id = { .name = {0} };
1269 ldlm_policy_data_t policy;
1271 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1273 res_id.name[0] = (*dparentp)->d_inode->i_ino;
1274 res_id.name[1] = (*dparentp)->d_inode->i_generation;
1275 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1276 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1277 res_id, LDLM_IBITS, &policy,
1278 *update_mode, &flags,
1280 ldlm_completion_ast,
1281 NULL, NULL, NULL, 0, NULL,
1287 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1288 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
1289 (*dparentp)->d_inode->i_ino,
1290 (*dparentp)->d_inode->i_generation,
1291 parent_res_id.name[2]);
1295 cleanup_phase = 1; /* parent dentry */
1296 #ifdef CONFIG_SNAPFS
1298 /*FIXME is there any other FUNC will use d_fsdata,
1299 *excepet creating inode according inum*/
1300 (*dparentp)->d_fsdata = clone_info;
1304 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1305 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1306 if (IS_ERR(*dchildp)) {
1307 rc = PTR_ERR(*dchildp);
1308 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1312 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1313 /* inode lives on another MDS: return * mds/ino/gen
1314 * and LOOKUP lock. drop possible UPDATE lock! */
1315 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1316 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1317 child_res_id.name[0] = (*dchildp)->d_inum;
1318 child_res_id.name[1] = (*dchildp)->d_generation;
1322 inode = (*dchildp)->d_inode;
1324 inode = igrab(inode);
1328 child_res_id.name[0] = inode->i_ino;
1329 child_res_id.name[1] = inode->i_generation;
1334 cleanup_phase = 2; /* child dentry */
1336 /* Step 3: Lock parent and child in resource order. If child doesn't
1337 * exist, we still have to lock the parent and re-lookup. */
1338 rc = enqueue_ordered_locks(obd,&parent_res_id,parent_lockh,parent_mode,
1339 &parent_policy, &child_res_id, child_lockh,
1340 child_mode, &child_policy);
1344 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1345 cleanup_phase = 4; /* child lock */
1347 cleanup_phase = 3; /* parent lock */
1349 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1350 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1351 parent_mode, &child_res_id, child_lockh,
1352 dchildp, child_mode, &child_policy,
1353 name, namelen, &parent_res_id);
1363 switch (cleanup_phase) {
1365 ldlm_lock_decref(child_lockh, child_mode);
1367 ldlm_lock_decref(parent_lockh, parent_mode);
1372 if (parent_lockh[1].cookie)
1373 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1382 void mds_reconstruct_generic(struct ptlrpc_request *req)
1384 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1386 mds_req_from_mcd(req, med->med_mcd);
1389 int mds_create_local_dentry(struct mds_update_record *rec,
1390 struct obd_device *obd)
1392 struct mds_obd *mds = &obd->u.mds;
1393 struct inode *fids_dir = mds->mds_fids_dir->d_inode;
1394 int fidlen = 0, rc, cleanup_phase = 0;
1395 struct dentry *new_child = NULL;
1396 char *fidname = rec->ur_name;
1397 struct dentry *child = NULL;
1398 struct lustre_handle lockh[2] = {{0}, {0}};
1402 down(&fids_dir->i_sem);
1403 fidlen = ll_fid2str(fidname, rec->ur_fid1->id, rec->ur_fid1->generation);
1404 CDEBUG(D_OTHER, "look for local dentry '%s' for %u/%u\n",
1405 fidname, (unsigned) rec->ur_fid1->id,
1406 (unsigned) rec->ur_fid1->generation);
1408 new_child = lookup_one_len(fidname, mds->mds_fids_dir, fidlen);
1409 up(&fids_dir->i_sem);
1410 if (IS_ERR(new_child)) {
1411 CERROR("can't lookup %s: %d\n", fidname,
1412 (int) PTR_ERR(new_child));
1413 GOTO(cleanup, rc = PTR_ERR(new_child));
1417 if (new_child->d_inode != NULL) {
1418 /* nice. we've already have local dentry! */
1419 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1420 (unsigned) new_child->d_inode->i_ino,
1421 (unsigned) new_child->d_inode->i_generation);
1422 rec->ur_fid1->id = fids_dir->i_ino;
1423 rec->ur_fid1->generation = fids_dir->i_generation;
1424 rec->ur_namelen = fidlen + 1;
1425 GOTO(cleanup, rc = 0);
1428 /* new, local dentry will be added soon. we need no aliases here */
1431 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1432 child = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1434 child = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1435 LCK_EX, lockh, NULL, NULL, 0,
1436 MDS_INODELOCK_UPDATE);
1439 if (IS_ERR(child)) {
1440 rc = PTR_ERR(child);
1441 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1442 CERROR("can't get victim: %d\n", rc);
1447 handle = fsfilt_start(obd, fids_dir, FSFILT_OP_LINK, NULL);
1449 GOTO(cleanup, rc = PTR_ERR(handle));
1451 rc = fsfilt_add_dir_entry(obd, mds->mds_fids_dir, fidname, fidlen,
1452 rec->ur_fid1->id, rec->ur_fid1->generation,
1455 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1456 (unsigned long) child->d_inode->i_ino,
1457 (unsigned long) child->d_inode->i_generation, rc);
1459 if (S_ISDIR(child->d_inode->i_mode)) {
1460 fids_dir->i_nlink++;
1461 mark_inode_dirty(fids_dir);
1463 mark_inode_dirty(child->d_inode);
1465 fsfilt_commit(obd, mds->mds_sb, fids_dir, handle, 0);
1467 rec->ur_fid1->id = fids_dir->i_ino;
1468 rec->ur_fid1->generation = fids_dir->i_generation;
1469 rec->ur_namelen = fidlen + 1;
1472 switch(cleanup_phase) {
1474 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1475 ldlm_lock_decref(lockh, LCK_EX);
1485 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1486 struct ptlrpc_request *slave)
1488 void *cookie, *cookie2;
1489 struct mds_body *body2;
1490 struct mds_body *body;
1494 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1495 LASSERT(body != NULL);
1497 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1498 LASSERT(body2 != NULL);
1500 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER))) {
1504 memcpy(body2, body, sizeof(*body));
1505 body2->valid &= ~OBD_MD_FLCOOKIE;
1507 if (!(body->valid & OBD_MD_FLEASIZE) &&
1508 !(body->valid & OBD_MD_FLDIREA))
1511 if (body->eadatasize == 0) {
1512 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1516 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1518 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1519 LASSERT(ea != NULL);
1521 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1522 LASSERT(ea2 != NULL);
1524 memcpy(ea2, ea, body->eadatasize);
1526 if (body->valid & OBD_MD_FLCOOKIE) {
1527 LASSERT(master->rq_repmsg->buflens[2] >=
1528 slave->rq_repmsg->buflens[2]);
1529 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1530 slave->rq_repmsg->buflens[2]);
1531 LASSERT(cookie != NULL);
1533 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1534 master->rq_repmsg->buflens[2]);
1535 LASSERT(cookie2 != NULL);
1536 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1537 body2->valid |= OBD_MD_FLCOOKIE;
1542 static int mds_reint_unlink_remote(struct mds_update_record *rec, int offset,
1543 struct ptlrpc_request *req,
1544 struct lustre_handle *parent_lockh,
1546 struct dentry *dparent,
1547 struct lustre_handle *child_lockh,
1548 struct dentry *dchild)
1550 struct obd_device *obd = req->rq_export->exp_obd;
1551 struct mds_obd *mds = mds_req2mds(req);
1552 struct mdc_op_data op_data;
1553 int rc = 0, cleanup_phase = 0;
1554 struct ptlrpc_request *request = NULL;
1558 LASSERT(offset == 0 || offset == 2);
1560 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode %u/%u/%u)",
1561 rec->ur_namelen - 1, rec->ur_name, (unsigned)dchild->d_mdsnum,
1562 (unsigned) dchild->d_inum, (unsigned) dchild->d_generation);
1563 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1564 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode %u/%u/%u)",
1565 rec->ur_namelen - 1, rec->ur_name,
1566 (unsigned)dchild->d_mdsnum,
1567 (unsigned) dchild->d_inum,
1568 (unsigned) dchild->d_generation);
1570 /* time to drop i_nlink on remote MDS */
1571 memset(&op_data, 0, sizeof(op_data));
1572 op_data.fid1.mds = dchild->d_mdsnum;
1573 op_data.fid1.id = dchild->d_inum;
1574 op_data.fid1.generation = dchild->d_generation;
1575 op_data.create_mode = rec->ur_mode;
1576 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1577 op_data.create_mode |= MDS_MODE_REPLAY;
1578 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1582 mds_copy_unlink_reply(req, request);
1583 ptlrpc_req_finished(request);
1586 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1589 GOTO(cleanup, rc = PTR_ERR(handle));
1590 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1591 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1595 req->rq_status = rc;
1598 if (parent_lockh[1].cookie != 0)
1599 ldlm_lock_decref(parent_lockh + 1, update_mode);
1601 ldlm_lock_decref(child_lockh, LCK_EX);
1603 ldlm_lock_decref(parent_lockh, LCK_PW);
1605 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1612 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1613 struct ptlrpc_request *req,
1614 struct lustre_handle *lh)
1616 struct dentry *dparent, *dchild;
1617 struct mds_obd *mds = mds_req2mds(req);
1618 struct obd_device *obd = req->rq_export->exp_obd;
1619 struct mds_body *body = NULL;
1620 struct inode *child_inode;
1621 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1622 struct lustre_handle child_lockh = {0};
1624 struct lustre_handle child_reuse_lockh = {0};
1626 struct lustre_handle * slave_lockh = NULL;
1627 struct llog_create_locks *lcl = NULL;
1628 char fidname[LL_FID_NAMELEN];
1629 void *handle = NULL;
1630 int rc = 0, log_unlink = 0, cleanup_phase = 0;
1631 int unlink_by_fid = 0;
1635 LASSERT(offset == 0 || offset == 2);
1637 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1638 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name);
1640 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1642 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1643 DEBUG_REQ(D_HA, req, "unlink replay\n");
1644 memcpy(lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1645 lustre_msg_buf(req->rq_reqmsg, 2, 0),
1646 req->rq_repmsg->buflens[offset + 2]);
1649 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1650 GOTO(cleanup, rc = -ENOENT);
1652 if (rec->ur_namelen == 1) {
1653 /* this is request to drop i_nlink on local inode */
1655 rec->ur_name = fidname;
1656 rc = mds_create_local_dentry(rec, obd);
1657 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1658 DEBUG_REQ(D_HA, req,
1659 "drop nlink on inode %u/%u/%u (replay)",
1660 (unsigned) rec->ur_fid1->mds,
1661 (unsigned) rec->ur_fid1->id,
1662 (unsigned) rec->ur_fid1->generation);
1668 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1669 /* master mds for directory asks slave removing
1670 * inode is already locked */
1671 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL,
1672 LCK_PW, parent_lockh,
1673 &update_mode, rec->ur_name,
1675 MDS_INODELOCK_UPDATE);
1676 if (IS_ERR(dparent))
1677 GOTO(cleanup, rc = PTR_ERR(dparent));
1678 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1679 rec->ur_namelen - 1);
1681 GOTO(cleanup, rc = PTR_ERR(dchild));
1682 child_lockh.cookie = 0;
1683 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1684 LASSERT(dchild->d_inode != NULL);
1685 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1687 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1,
1688 parent_lockh, &dparent,
1689 LCK_PW, MDS_INODELOCK_UPDATE,
1690 &update_mode, rec->ur_name,
1691 rec->ur_namelen, &child_lockh,
1693 MDS_INODELOCK_LOOKUP |
1694 MDS_INODELOCK_UPDATE, NULL);
1699 if (dchild->d_flags & DCACHE_CROSS_REF) {
1700 /* we should have parent lock only here */
1701 LASSERT(unlink_by_fid == 0);
1702 LASSERT(dchild->d_mdsnum != mds->mds_num);
1703 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1704 update_mode, dparent, &child_lockh, dchild);
1708 cleanup_phase = 1; /* dchild, dparent, locks */
1711 child_inode = dchild->d_inode;
1712 if (child_inode == NULL) {
1713 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1714 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1715 GOTO(cleanup, rc = -ENOENT);
1718 cleanup_phase = 2; /* dchild has a lock */
1720 /* We have to do these checks ourselves, in case we are making an
1721 * orphan. The client tells us whether rmdir() or unlink() was called,
1722 * so we need to return appropriate errors (bug 72).
1724 * We don't have to check permissions, because vfs_rename (called from
1725 * mds_open_unlink_rename) also calls may_delete. */
1726 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1727 if (!S_ISDIR(child_inode->i_mode))
1728 GOTO(cleanup, rc = -ENOTDIR);
1730 if (S_ISDIR(child_inode->i_mode))
1731 GOTO(cleanup, rc = -EISDIR);
1734 /* handle splitted dir */
1735 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1740 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1741 * reuse (see bug 2029). */
1742 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1746 cleanup_phase = 3; /* child inum lock */
1748 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
1750 /* ldlm_reply in buf[0] if called via intent */
1754 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
1755 LASSERT(body != NULL);
1757 /* If this is the last reference to this inode, get the OBD EA
1758 * data first so the client can destroy OST objects.
1759 * we only do the object removal if no open files remain.
1760 * Nobody can get at this name anymore because of the locks so
1761 * we make decisions here as to whether to remove the inode */
1762 if (S_ISREG(child_inode->i_mode) && child_inode->i_nlink == 1 &&
1763 mds_open_orphan_count(child_inode) == 0) {
1764 mds_pack_inode2fid(obd, &body->fid1, child_inode);
1765 mds_pack_inode2body(obd, body, child_inode);
1766 mds_pack_md(obd, req->rq_repmsg, offset + 1, body,
1768 if (!(body->valid & OBD_MD_FLEASIZE)) {
1769 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1770 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1776 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
1777 switch (child_inode->i_mode & S_IFMT) {
1779 /* Drop any lingering child directories before we start our
1780 * transaction, to avoid doing multiple inode dirty/delete
1781 * in our compound transaction (bug 1321). */
1782 shrink_dcache_parent(dchild);
1783 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1786 GOTO(cleanup, rc = PTR_ERR(handle));
1787 cleanup_phase = 4; /* transaction */
1788 rc = vfs_rmdir(dparent->d_inode, dchild);
1791 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
1792 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
1794 handle = fsfilt_start_log(obd, dparent->d_inode,
1795 FSFILT_OP_UNLINK, NULL,
1796 le32_to_cpu(lmm->lmm_stripe_count));
1798 GOTO(cleanup, rc = PTR_ERR(handle));
1800 cleanup_phase = 4; /* transaction */
1801 rc = vfs_unlink(dparent->d_inode, dchild);
1803 if (!rc && log_unlink)
1804 if (mds_log_op_unlink(obd, child_inode,
1805 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
1806 req->rq_repmsg->buflens[offset + 1],
1807 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
1808 req->rq_repmsg->buflens[offset + 2], &lcl) > 0)
1809 body->valid |= OBD_MD_FLCOOKIE;
1817 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
1820 GOTO(cleanup, rc = PTR_ERR(handle));
1821 cleanup_phase = 4; /* transaction */
1822 rc = vfs_unlink(dparent->d_inode, dchild);
1825 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
1828 GOTO(cleanup, rc = -EINVAL);
1836 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1837 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1838 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1840 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1842 CERROR("error on parent setattr: rc = %d\n", err);
1845 switch(cleanup_phase) {
1847 LASSERT(dchild != NULL && dchild->d_inode != NULL);
1848 LASSERT(atomic_read(&dchild->d_inode->i_count) > 0);
1849 if (rc == 0 && dchild->d_inode->i_nlink == 0 &&
1850 mds_open_orphan_count(dchild->d_inode) > 0) {
1852 /* filesystem is really going to destroy an inode
1853 * we have to delay this till inode is opened -bzzz */
1854 mds_open_unlink_rename(rec, obd, dparent, dchild, NULL);
1856 /* handle splitted dir */
1858 /* master directory can be non-empty or something else ... */
1859 mds_unlink_slave_objs(obd, dchild);
1861 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1864 (void)obd_set_info(mds->mds_osc_exp, strlen("unlinked"),
1865 "unlinked", 0, NULL);
1867 ptlrpc_save_llog_lock(req, lcl);
1868 case 3: /* child ino-reuse lock */
1870 if (rc && body != NULL) {
1871 // Don't unlink the OST objects if the MDS unlink failed
1875 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
1877 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
1879 case 2: /* child lock */
1880 mds_unlock_slave_objs(obd, dchild, slave_lockh);
1881 if (child_lockh.cookie)
1882 ldlm_lock_decref(&child_lockh, LCK_EX);
1883 case 1: /* child and parent dentry, parent lock */
1885 if (parent_lockh[1].cookie != 0)
1886 ldlm_lock_decref(parent_lockh + 1, update_mode);
1889 ldlm_lock_decref(parent_lockh, LCK_PW);
1891 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1898 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1901 req->rq_status = rc;
1906 * to service requests from remote MDS to increment i_nlink
1908 static int mds_reint_link_acquire(struct mds_update_record *rec,
1909 int offset, struct ptlrpc_request *req,
1910 struct lustre_handle *lh)
1912 struct obd_device *obd = req->rq_export->exp_obd;
1913 struct ldlm_res_id src_res_id = { .name = {0} };
1914 struct lustre_handle *handle = NULL, src_lockh = {0};
1915 struct mds_obd *mds = mds_req2mds(req);
1916 int rc = 0, cleanup_phase = 0;
1917 struct dentry *de_src = NULL;
1918 ldlm_policy_data_t policy;
1922 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks %u/%u/%u\n",
1923 obd->obd_name, (unsigned) rec->ur_fid1->mds,
1924 (unsigned) rec->ur_fid1->id,
1925 (unsigned) rec->ur_fid1->generation);
1927 /* Step 1: Lookup the source inode and target directory by FID */
1928 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
1930 GOTO(cleanup, rc = PTR_ERR(de_src));
1931 cleanup_phase = 1; /* source dentry */
1933 src_res_id.name[0] = de_src->d_inode->i_ino;
1934 src_res_id.name[1] = de_src->d_inode->i_generation;
1935 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1937 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1938 src_res_id, LDLM_IBITS, &policy,
1939 LCK_EX, &flags, mds_blocking_ast,
1940 ldlm_completion_ast, NULL, NULL,
1941 NULL, 0, NULL, &src_lockh);
1943 GOTO(cleanup, rc = -ENOLCK);
1944 cleanup_phase = 2; /* lock */
1946 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
1948 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
1949 if (IS_ERR(handle)) {
1950 rc = PTR_ERR(handle);
1953 de_src->d_inode->i_nlink++;
1954 mark_inode_dirty(de_src->d_inode);
1957 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
1958 handle, req, rc, 0);
1960 switch (cleanup_phase) {
1963 ldlm_lock_decref(&src_lockh, LCK_EX);
1965 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
1971 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1974 req->rq_status = rc;
1979 * request to link to foreign inode:
1980 * - acquire i_nlinks on this inode
1983 static int mds_reint_link_to_remote(struct mds_update_record *rec,
1984 int offset, struct ptlrpc_request *req,
1985 struct lustre_handle *lh)
1987 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
1988 struct obd_device *obd = req->rq_export->exp_obd;
1989 struct dentry *de_tgt_dir = NULL;
1990 struct mds_obd *mds = mds_req2mds(req);
1991 int rc = 0, cleanup_phase = 0;
1992 struct mdc_op_data op_data;
1993 struct ptlrpc_request *request = NULL;
1997 #define fmt "%s: request to link %u/%u/%u:%*s to foreign inode %u/%u/%u\n"
1998 DEBUG_REQ(D_INODE, req, fmt, obd->obd_name,
1999 (unsigned) rec->ur_fid2->mds,
2000 (unsigned) rec->ur_fid2->id,
2001 (unsigned) rec->ur_fid2->generation,
2002 rec->ur_namelen - 1, rec->ur_name,
2003 (unsigned) rec->ur_fid1->mds,
2004 (unsigned) rec->ur_fid1->id,
2005 (unsigned)rec->ur_fid1->generation);
2007 de_tgt_dir = mds_fid2locked_dentry(obd, rec->ur_fid2, NULL, LCK_EX,
2008 tgt_dir_lockh, &update_mode,
2009 rec->ur_name, rec->ur_namelen - 1,
2010 MDS_INODELOCK_UPDATE);
2011 if (IS_ERR(de_tgt_dir))
2012 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2015 op_data.fid1 = *(rec->ur_fid1);
2016 op_data.namelen = 0;
2017 op_data.name = NULL;
2018 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2024 ptlrpc_req_finished(request);
2026 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2028 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2029 if (IS_ERR(handle)) {
2030 rc = PTR_ERR(handle);
2034 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2035 rec->ur_namelen - 1, rec->ur_fid1->id,
2036 rec->ur_fid1->generation, rec->ur_fid1->mds);
2040 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2041 handle, req, rc, 0);
2044 switch (cleanup_phase) {
2047 /* FIXME: drop i_nlink on remote inode here */
2048 CERROR("MUST drop drop i_nlink here\n");
2053 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2055 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2058 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2060 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2066 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2069 req->rq_status = rc;
2073 static int mds_reint_link(struct mds_update_record *rec, int offset,
2074 struct ptlrpc_request *req,
2075 struct lustre_handle *lh)
2077 struct obd_device *obd = req->rq_export->exp_obd;
2078 struct dentry *de_src = NULL;
2079 struct dentry *de_tgt_dir = NULL;
2080 struct dentry *dchild = NULL;
2081 struct mds_obd *mds = mds_req2mds(req);
2082 struct lustre_handle *handle = NULL;
2083 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2084 struct ldlm_res_id src_res_id = { .name = {0} };
2085 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2086 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2087 ldlm_policy_data_t tgt_dir_policy =
2088 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2089 int rc = 0, cleanup_phase = 0;
2090 int update_mode = 0;
2093 LASSERT(offset == 0);
2095 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2096 rec->ur_fid1->id, rec->ur_fid1->generation,
2097 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_name);
2099 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2101 // memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2102 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2103 GOTO(cleanup, rc = -ENOENT);
2105 if (rec->ur_fid1->mds != mds->mds_num) {
2106 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2110 if (rec->ur_namelen == 1) {
2111 rc = mds_reint_link_acquire(rec, offset, req, lh);
2115 /* Step 1: Lookup the source inode and target directory by FID */
2116 de_src = mds_fid2dentry(mds, rec->ur_fid1, NULL);
2118 GOTO(cleanup, rc = PTR_ERR(de_src));
2120 cleanup_phase = 1; /* source dentry */
2122 de_tgt_dir = mds_fid2dentry(mds, rec->ur_fid2, NULL);
2123 if (IS_ERR(de_tgt_dir)) {
2124 rc = PTR_ERR(de_tgt_dir);
2129 cleanup_phase = 2; /* target directory dentry */
2131 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2132 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name, rec->ur_name,
2133 de_src->d_inode->i_ino);
2135 /* Step 2: Take the two locks */
2136 src_res_id.name[0] = de_src->d_inode->i_ino;
2137 src_res_id.name[1] = de_src->d_inode->i_generation;
2138 tgt_dir_res_id.name[0] = de_tgt_dir->d_inode->i_ino;
2139 tgt_dir_res_id.name[1] = de_tgt_dir->d_inode->i_generation;
2141 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2143 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2145 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2146 tgt_dir_res_id, LDLM_IBITS,
2147 &src_policy, update_mode, &flags,
2149 ldlm_completion_ast, NULL, NULL,
2150 NULL, 0, NULL, tgt_dir_lockh + 1);
2152 GOTO(cleanup, rc = -ENOLCK);
2155 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2156 rec->ur_namelen - 1);
2157 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
2158 de_tgt_dir->d_inode->i_ino,
2159 de_tgt_dir->d_inode->i_generation,
2160 tgt_dir_res_id.name[2]);
2163 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2165 &tgt_dir_res_id, tgt_dir_lockh, LCK_EX,
2170 cleanup_phase = 3; /* locks */
2172 /* Step 3: Lookup the child */
2173 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir, rec->ur_namelen-1);
2174 if (IS_ERR(dchild)) {
2175 rc = PTR_ERR(dchild);
2176 if (rc != -EPERM && rc != -EACCES)
2177 CERROR("child lookup error %d\n", rc);
2181 cleanup_phase = 4; /* child dentry */
2183 if (dchild->d_inode) {
2184 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2185 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2190 /* Step 4: Do it. */
2191 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2193 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2194 if (IS_ERR(handle)) {
2195 rc = PTR_ERR(handle);
2199 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2200 if (rc && rc != -EPERM && rc != -EACCES)
2201 CERROR("vfs_link error %d\n", rc);
2203 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2204 handle, req, rc, 0);
2207 switch (cleanup_phase) {
2208 case 4: /* child dentry */
2212 ldlm_lock_decref(&src_lockh, LCK_EX);
2213 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2215 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2216 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2218 case 2: /* target dentry */
2220 if (tgt_dir_lockh[1].cookie && update_mode)
2221 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2225 case 1: /* source dentry */
2230 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2233 req->rq_status = rc;
2238 * add a hard link in the PENDING directory, only used by rename()
2240 static int mds_add_link_orphan(struct mds_update_record *rec,
2241 struct obd_device *obd,
2242 struct dentry *dentry)
2244 struct mds_obd *mds = &obd->u.mds;
2245 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
2246 struct dentry *pending_child;
2247 char fidname[LL_FID_NAMELEN];
2251 LASSERT(dentry->d_inode);
2252 LASSERT(!mds_inode_is_orphan(dentry->d_inode));
2254 down(&pending_dir->i_sem);
2255 fidlen = ll_fid2str(fidname, dentry->d_inode->i_ino,
2256 dentry->d_inode->i_generation);
2258 CDEBUG(D_ERROR, "pending destroy of %dx open file %s = %s\n",
2259 mds_open_orphan_count(dentry->d_inode),
2260 rec->ur_name, fidname);
2262 pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
2263 if (IS_ERR(pending_child))
2264 GOTO(out_lock, rc = PTR_ERR(pending_child));
2266 if (pending_child->d_inode != NULL) {
2267 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
2268 LASSERT(pending_child->d_inode == dentry->d_inode);
2269 GOTO(out_dput, rc = 0);
2273 rc = vfs_link(dentry, pending_dir, pending_child);
2276 CERROR("error addlink orphan %s to PENDING: rc = %d\n",
2279 mds_inode_set_orphan(dentry->d_inode);
2281 l_dput(pending_child);
2283 up(&pending_dir->i_sem);
2287 /* The idea here is that we need to get four locks in the end:
2288 * one on each parent directory, one on each child. We need to take
2289 * these locks in some kind of order (to avoid deadlocks), and the order
2290 * I selected is "increasing resource number" order. We need to look up
2291 * the children, however, before we know what the resource number(s) are.
2292 * Thus the following plan:
2294 * 1,2. Look up the parents
2295 * 3,4. Look up the children
2296 * 5. Take locks on the parents and children, in order
2297 * 6. Verify that the children haven't changed since they were looked up
2299 * If there was a race and the children changed since they were first looked
2300 * up, it is possible that mds_verify_child() will be able to just grab the
2301 * lock on the new child resource (if it has a higher resource than any other)
2302 * but we need to compare against not only its parent, but also against the
2303 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2305 * We need the fancy igrab() on the child inodes because we aren't holding a
2306 * lock on the parent after the lookup is done, so dentry->d_inode may change
2307 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2309 static int mds_get_parents_children_locked(struct obd_device *obd,
2310 struct mds_obd *mds,
2311 struct ll_fid *p1_fid,
2312 struct dentry **de_srcdirp,
2313 struct ll_fid *p2_fid,
2314 struct dentry **de_tgtdirp,
2316 const char *old_name, int old_len,
2317 struct dentry **de_oldp,
2318 const char *new_name, int new_len,
2319 struct dentry **de_newp,
2320 struct lustre_handle *dlm_handles,
2323 struct ldlm_res_id p1_res_id = { .name = {0} };
2324 struct ldlm_res_id p2_res_id = { .name = {0} };
2325 struct ldlm_res_id c1_res_id = { .name = {0} };
2326 struct ldlm_res_id c2_res_id = { .name = {0} };
2327 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2328 /* Only dentry should change, but the inode itself would be
2330 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2331 /* If something is going to be replaced, both dentry and inode locks are
2333 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2334 MDS_INODELOCK_UPDATE}};
2335 struct ldlm_res_id *maxres_src, *maxres_tgt;
2336 struct inode *inode;
2337 int rc = 0, cleanup_phase = 0;
2340 /* Step 1: Lookup the source directory */
2341 *de_srcdirp = mds_fid2dentry(mds, p1_fid, NULL);
2342 if (IS_ERR(*de_srcdirp))
2343 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2345 cleanup_phase = 1; /* source directory dentry */
2347 p1_res_id.name[0] = (*de_srcdirp)->d_inode->i_ino;
2348 p1_res_id.name[1] = (*de_srcdirp)->d_inode->i_generation;
2350 /* Step 2: Lookup the target directory */
2351 if (memcmp(p1_fid, p2_fid, sizeof(*p1_fid)) == 0) {
2352 *de_tgtdirp = dget(*de_srcdirp);
2354 *de_tgtdirp = mds_fid2dentry(mds, p2_fid, NULL);
2355 if (IS_ERR(*de_tgtdirp)) {
2356 rc = PTR_ERR(*de_tgtdirp);
2362 cleanup_phase = 2; /* target directory dentry */
2364 p2_res_id.name[0] = (*de_tgtdirp)->d_inode->i_ino;
2365 p2_res_id.name[1] = (*de_tgtdirp)->d_inode->i_generation;
2368 dlm_handles[5].cookie = 0;
2369 dlm_handles[6].cookie = 0;
2370 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2371 /* Get a temp lock on just ino, gen to flush client cache and
2372 * to protect dirs from concurrent splitting */
2373 rc = enqueue_ordered_locks(obd, &p1_res_id, &(dlm_handles[5]),
2374 LCK_PW, &p_policy, &p2_res_id,
2375 &(dlm_handles[6]),LCK_PW,&p_policy);
2378 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2379 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2380 CDEBUG(D_INFO, "take locks on %lu:%u:"LPX64", %lu:%u:"LPX64"\n",
2381 (*de_srcdirp)->d_inode->i_ino,
2382 (*de_srcdirp)->d_inode->i_generation, p1_res_id.name[2],
2383 (*de_tgtdirp)->d_inode->i_ino,
2384 (*de_tgtdirp)->d_inode->i_generation, p2_res_id.name[2]);
2389 /* Step 3: Lookup the source child entry */
2390 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp, old_len - 1);
2391 if (IS_ERR(*de_oldp)) {
2392 rc = PTR_ERR(*de_oldp);
2393 CERROR("old child lookup error (%*s): %d\n",
2394 old_len - 1, old_name, rc);
2398 cleanup_phase = 4; /* original name dentry */
2400 inode = (*de_oldp)->d_inode;
2401 if (inode != NULL) {
2402 inode = igrab(inode);
2404 GOTO(cleanup, rc = -ENOENT);
2406 c1_res_id.name[0] = inode->i_ino;
2407 c1_res_id.name[1] = inode->i_generation;
2409 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2410 c1_res_id.name[0] = (*de_oldp)->d_inum;
2411 c1_res_id.name[1] = (*de_oldp)->d_generation;
2414 /* Step 4: Lookup the target child entry */
2415 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp, new_len - 1);
2416 if (IS_ERR(*de_newp)) {
2417 rc = PTR_ERR(*de_newp);
2418 CERROR("new child lookup error (%*s): %d\n",
2419 old_len - 1, old_name, rc);
2423 cleanup_phase = 5; /* target dentry */
2425 inode = (*de_newp)->d_inode;
2426 if (inode != NULL) {
2427 inode = igrab(inode);
2431 c2_res_id.name[0] = inode->i_ino;
2432 c2_res_id.name[1] = inode->i_generation;
2434 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2435 c2_res_id.name[0] = (*de_newp)->d_inum;
2436 c2_res_id.name[1] = (*de_newp)->d_generation;
2440 /* Step 5: Take locks on the parents and child(ren) */
2441 maxres_src = &p1_res_id;
2442 maxres_tgt = &p2_res_id;
2443 cleanup_phase = 5; /* target dentry */
2445 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL,NULL))
2446 maxres_src = &c1_res_id;
2447 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL,NULL))
2448 maxres_tgt = &c2_res_id;
2450 rc = enqueue_4ordered_locks(obd, &p1_res_id,&dlm_handles[0],parent_mode,
2452 &p2_res_id, &dlm_handles[1], parent_mode,
2454 &c1_res_id, &dlm_handles[2], child_mode,
2456 &c2_res_id, &dlm_handles[3], child_mode,
2461 cleanup_phase = 6; /* parent and child(ren) locks */
2463 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2464 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2465 parent_mode, &c1_res_id, &dlm_handles[2],
2466 de_oldp, child_mode, &c1_policy, old_name,old_len,
2469 if (c2_res_id.name[0] != 0)
2470 ldlm_lock_decref(&dlm_handles[3], child_mode);
2471 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2478 if (!DENTRY_VALID(*de_oldp))
2479 GOTO(cleanup, rc = -ENOENT);
2481 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2482 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2483 parent_mode, &c2_res_id, &dlm_handles[3],
2484 de_newp, child_mode, &c2_policy, new_name,
2485 new_len, maxres_src);
2487 ldlm_lock_decref(&dlm_handles[2], child_mode);
2488 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2498 switch (cleanup_phase) {
2499 case 6: /* child lock(s) */
2500 if (c2_res_id.name[0] != 0)
2501 ldlm_lock_decref(&dlm_handles[3], child_mode);
2502 if (c1_res_id.name[0] != 0)
2503 ldlm_lock_decref(&dlm_handles[2], child_mode);
2504 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2505 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2506 case 5: /* target dentry */
2508 case 4: /* source dentry */
2512 if (dlm_handles[5].cookie != 0)
2513 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2514 if (dlm_handles[6].cookie != 0)
2515 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2517 case 2: /* target directory dentry */
2518 l_dput(*de_tgtdirp);
2519 case 1: /* source directry dentry */
2520 l_dput(*de_srcdirp);
2527 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2528 int offset, struct ptlrpc_request *req)
2530 struct obd_device *obd = req->rq_export->exp_obd;
2531 struct dentry *de_srcdir = NULL;
2532 struct dentry *de_new = NULL;
2533 struct mds_obd *mds = mds_req2mds(req);
2534 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2535 struct lustre_handle child_lockh = {0};
2536 int cleanup_phase = 0;
2537 void *handle = NULL;
2538 int update_mode, rc = 0;
2541 /* another MDS executing rename operation has asked us
2542 * to create target name. such a creation should destroy
2543 * existing target name */
2545 CDEBUG(D_OTHER, "%s: request to create name %s for %lu/%lu/%lu\n",
2546 obd->obd_name, rec->ur_tgt,
2547 (unsigned long) rec->ur_fid1->mds,
2548 (unsigned long) rec->ur_fid1->id,
2549 (unsigned long) rec->ur_fid1->generation);
2551 /* first, lookup the target */
2552 child_lockh.cookie = 0;
2553 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid2, parent_lockh,
2554 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2555 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2556 &child_lockh, &de_new, LCK_EX,
2557 MDS_INODELOCK_LOOKUP, NULL);
2564 LASSERT(de_srcdir->d_inode);
2567 if (de_new->d_inode) {
2568 /* name exists and points to local inode
2569 * try to unlink this name and create new one */
2570 CERROR("%s: %s points to local inode %lu/%lu\n",
2571 obd->obd_name, rec->ur_tgt,
2572 (unsigned long) de_new->d_inode->i_ino,
2573 (unsigned long) de_new->d_inode->i_generation);
2574 handle = fsfilt_start(obd, de_srcdir->d_inode,
2575 FSFILT_OP_RENAME, NULL);
2577 GOTO(cleanup, rc = PTR_ERR(handle));
2578 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de_new);
2581 } else if (de_new->d_flags & DCACHE_CROSS_REF) {
2582 /* name exists adn points to remove inode */
2583 CERROR("%s: %s points to remote inode %lu/%lu/%lu\n",
2584 obd->obd_name, rec->ur_tgt,
2585 (unsigned long) de_new->d_mdsnum,
2586 (unsigned long) de_new->d_inum,
2587 (unsigned long) de_new->d_generation);
2589 /* name doesn't exist. the simplest case */
2590 handle = fsfilt_start(obd, de_srcdir->d_inode,
2591 FSFILT_OP_LINK, NULL);
2593 GOTO(cleanup, rc = PTR_ERR(handle));
2597 rc = fsfilt_add_dir_entry(obd, de_srcdir, rec->ur_tgt,
2598 rec->ur_tgtlen - 1, rec->ur_fid1->id,
2599 rec->ur_fid1->generation, rec->ur_fid1->mds);
2601 CERROR("add_dir_entry() returned error %d\n", rc);
2604 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2605 handle, req, rc, 0);
2606 switch(cleanup_phase) {
2610 if (parent_lockh[1].cookie != 0)
2611 ldlm_lock_decref(&parent_lockh[1], update_mode);
2613 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2614 if (child_lockh.cookie != 0)
2615 ldlm_lock_decref(&child_lockh, LCK_EX);
2623 req->rq_status = rc;
2628 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
2629 struct ptlrpc_request *req)
2631 struct obd_device *obd = req->rq_export->exp_obd;
2632 struct ptlrpc_request *req2 = NULL;
2633 struct dentry *de_srcdir = NULL;
2634 struct dentry *de_old = NULL;
2635 struct mds_obd *mds = mds_req2mds(req);
2636 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2637 struct lustre_handle child_lockh = {0};
2638 struct mdc_op_data opdata;
2639 void *handle = NULL;
2640 int update_mode, rc = 0;
2643 CDEBUG(D_OTHER, "%s: move name %s onto another mds%u\n",
2644 obd->obd_name, rec->ur_name, rec->ur_fid2->mds + 1);
2645 memset(&opdata, 0, sizeof(opdata));
2647 child_lockh.cookie = 0;
2648 rc = mds_get_parent_child_locked(obd, mds, rec->ur_fid1, parent_lockh,
2649 &de_srcdir,LCK_PW,MDS_INODELOCK_UPDATE,
2650 &update_mode, rec->ur_name,
2651 rec->ur_namelen, &child_lockh, &de_old,
2652 LCK_EX, MDS_INODELOCK_LOOKUP, NULL);
2655 LASSERT(de_srcdir->d_inode);
2658 /* we already know the target should be created on another MDS
2659 * so, we have to request that MDS to do it */
2661 /* prepare source fid */
2662 if (de_old->d_flags & DCACHE_CROSS_REF) {
2663 LASSERT(de_old->d_inode == NULL);
2664 CDEBUG(D_OTHER, "request to move remote name\n");
2665 opdata.fid1.mds = de_old->d_mdsnum;
2666 opdata.fid1.id = de_old->d_inum;
2667 opdata.fid1.generation = de_old->d_generation;
2668 } else if (de_old->d_inode == NULL) {
2669 /* oh, source doesn't exist */
2670 GOTO(cleanup, rc = -ENOENT);
2672 LASSERT(de_old->d_inode != NULL);
2673 CDEBUG(D_OTHER, "request to move local name\n");
2674 opdata.fid1.mds = mds->mds_num;
2675 opdata.fid1.id = de_old->d_inode->i_ino;
2676 opdata.fid1.generation = de_old->d_inode->i_generation;
2679 opdata.fid2 = *(rec->ur_fid2);
2680 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0, rec->ur_tgt,
2681 rec->ur_tgtlen - 1, &req2);
2686 handle = fsfilt_start(obd, de_srcdir->d_inode, FSFILT_OP_UNLINK, NULL);
2688 GOTO(cleanup, rc = PTR_ERR(handle));
2689 rc = fsfilt_del_dir_entry(obd, de_old);
2694 rc = mds_finish_transno(mds, de_srcdir ? de_srcdir->d_inode : NULL,
2695 handle, req, rc, 0);
2697 ptlrpc_req_finished(req2);
2700 if (parent_lockh[1].cookie != 0)
2701 ldlm_lock_decref(&parent_lockh[1], update_mode);
2703 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
2704 if (child_lockh.cookie != 0)
2705 ldlm_lock_decref(&child_lockh, LCK_EX);
2710 req->rq_status = rc;
2715 static int mds_reint_rename(struct mds_update_record *rec, int offset,
2716 struct ptlrpc_request *req, struct lustre_handle *lockh)
2718 struct obd_device *obd = req->rq_export->exp_obd;
2719 struct dentry *de_srcdir = NULL;
2720 struct dentry *de_tgtdir = NULL;
2721 struct dentry *de_old = NULL;
2722 struct dentry *de_new = NULL;
2723 struct mds_obd *mds = mds_req2mds(req);
2724 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
2725 struct mds_body *body = NULL;
2726 struct llog_create_locks *lcl = NULL;
2727 int rc = 0, log_unlink = 0, lock_count = 3;
2728 int cleanup_phase = 0;
2729 void *handle = NULL;
2732 LASSERT(offset == 0);
2734 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u %s to "LPU64"/%u %s",
2735 rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name,
2736 rec->ur_fid2->id, rec->ur_fid2->generation, rec->ur_tgt);
2738 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2740 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2741 DEBUG_REQ(D_HA, req, "rename replay\n");
2742 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
2743 lustre_msg_buf(req->rq_reqmsg, 3, 0),
2744 req->rq_repmsg->buflens[2]);
2747 if (rec->ur_namelen == 1) {
2748 rc = mds_reint_rename_create_name(rec, offset, req);
2752 if (rec->ur_fid2->mds != mds->mds_num) {
2753 rc = mds_reint_rename_to_remote(rec, offset, req);
2757 rc = mds_get_parents_children_locked(obd, mds, rec->ur_fid1, &de_srcdir,
2758 rec->ur_fid2, &de_tgtdir, LCK_PW,
2759 rec->ur_name, rec->ur_namelen,
2760 &de_old, rec->ur_tgt,
2761 rec->ur_tgtlen, &de_new,
2762 dlm_handles, LCK_EX);
2766 cleanup_phase = 1; /* parent(s), children, locks */
2768 if (de_new->d_inode)
2771 /* sanity check for src inode */
2772 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2773 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
2774 GOTO(cleanup, rc = -EINVAL);
2776 /* sanity check for dest inode */
2777 if (de_new->d_inode &&
2778 (de_new->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
2779 de_new->d_inode->i_ino == de_tgtdir->d_inode->i_ino))
2780 GOTO(cleanup, rc = -EINVAL);
2782 if (de_old->d_inode == de_new->d_inode)
2783 GOTO(cleanup, rc = 0);
2785 /* if we are about to remove the target at first, pass the EA of
2786 * that inode to client to perform and cleanup on OST */
2787 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2788 LASSERT(body != NULL);
2790 if (de_new->d_inode &&
2791 S_ISREG(de_new->d_inode->i_mode) &&
2792 de_new->d_inode->i_nlink == 1 &&
2793 mds_open_orphan_count(de_new->d_inode) == 0) {
2794 mds_pack_inode2fid(obd, &body->fid1, de_new->d_inode);
2795 mds_pack_inode2body(obd, body, de_new->d_inode);
2796 mds_pack_md(obd, req->rq_repmsg, 1, body, de_new->d_inode, 1);
2797 if (!(body->valid & OBD_MD_FLEASIZE)) {
2798 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2799 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2805 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
2806 de_srcdir->d_inode->i_sb);
2808 handle = fsfilt_start(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME, NULL);
2810 GOTO(cleanup, rc = PTR_ERR(handle));
2812 /* FIXME need adjust the journal block count? */
2813 /* if the target should be moved to PENDING, we at first increase the
2814 * link and later vfs_rename() will decrease the link count again */
2815 if (de_new->d_inode &&
2816 S_ISREG(de_new->d_inode->i_mode) &&
2817 de_new->d_inode->i_nlink == 1 &&
2818 mds_open_orphan_count(de_new->d_inode) > 0) {
2819 rc = mds_add_link_orphan(rec, obd, de_new);
2825 de_old->d_fsdata = req;
2826 de_new->d_fsdata = req;
2827 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
2830 if (!rc && log_unlink) {
2831 if (mds_log_op_unlink(obd, de_tgtdir->d_inode,
2832 lustre_msg_buf(req->rq_repmsg, 1, 0),
2833 req->rq_repmsg->buflens[1],
2834 lustre_msg_buf(req->rq_repmsg, 2, 0),
2835 req->rq_repmsg->buflens[2], &lcl) > 0) {
2836 body->valid |= OBD_MD_FLCOOKIE;
2842 rc = mds_finish_transno(mds, de_tgtdir ? de_tgtdir->d_inode : NULL,
2843 handle, req, rc, 0);
2844 switch (cleanup_phase) {
2847 if (dlm_handles[5].cookie != 0)
2848 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2849 if (dlm_handles[6].cookie != 0)
2850 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2853 ptlrpc_save_llog_lock(req, lcl);
2856 if (lock_count == 4)
2857 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
2858 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
2859 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
2860 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
2862 if (lock_count == 4)
2863 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
2864 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
2865 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
2866 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
2875 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2878 req->rq_status = rc;
2882 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
2883 struct ptlrpc_request *, struct lustre_handle *);
2885 static mds_reinter reinters[REINT_MAX + 1] = {
2886 [REINT_SETATTR] mds_reint_setattr,
2887 [REINT_CREATE] mds_reint_create,
2888 [REINT_LINK] mds_reint_link,
2889 [REINT_UNLINK] mds_reint_unlink,
2890 [REINT_RENAME] mds_reint_rename,
2891 [REINT_OPEN] mds_open
2894 int mds_reint_rec(struct mds_update_record *rec, int offset,
2895 struct ptlrpc_request *req, struct lustre_handle *lockh)
2897 struct obd_device *obd = req->rq_export->exp_obd;
2898 struct lvfs_run_ctxt saved;
2901 /* checked by unpacker */
2902 LASSERT(rec->ur_opcode <= REINT_MAX &&
2903 reinters[rec->ur_opcode] != NULL);
2905 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
2906 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
2907 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);