4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_reint.c
38 * Lustre Metadata Target (mdt) reintegration routines
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Huang Hua <huanghua@clusterfs.com>
44 * Author: Yury Umanets <umka@clusterfs.com>
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include "mdt_internal.h"
52 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
55 ma->ma_need = MA_INODE;
59 static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
60 struct mdt_object *object,
61 struct mdt_body *repbody)
65 /* for cross-ref mkdir, mds capa has been fetched from remote obj, then
66 * we won't go to below*/
67 if (repbody->valid & OBD_MD_FLMDSCAPA)
70 if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
71 info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
72 struct lustre_capa *capa;
74 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
76 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
77 rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa,
80 repbody->valid |= OBD_MD_FLMDSCAPA;
87 * Get version of object by fid.
89 * Return real version or ENOENT_VERSION if object doesn't exist
91 static void mdt_obj_version_get(struct mdt_thread_info *info,
92 struct mdt_object *o, __u64 *version)
95 LASSERT(mdt_object_exists(o) >= 0);
96 if (mdt_object_exists(o) > 0 && !mdt_object_obf(o))
97 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
99 *version = ENOENT_VERSION;
100 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
101 PFID(mdt_object_fid(o)), *version);
105 * Check version is correct.
107 * Should be called only during replay.
109 static int mdt_version_check(struct ptlrpc_request *req,
110 __u64 version, int idx)
112 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
115 if (!exp_connect_vbr(req->rq_export))
118 LASSERT(req_is_replay(req));
119 /** VBR: version is checked always because costs nothing */
120 LASSERT(idx < PTLRPC_NUM_VERSIONS);
121 /** Sanity check for malformed buffers */
122 if (pre_ver == NULL) {
123 CERROR("No versions in request buffer\n");
124 cfs_spin_lock(&req->rq_export->exp_lock);
125 req->rq_export->exp_vbr_failed = 1;
126 cfs_spin_unlock(&req->rq_export->exp_lock);
128 } else if (pre_ver[idx] != version) {
129 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
130 pre_ver[idx], version);
131 cfs_spin_lock(&req->rq_export->exp_lock);
132 req->rq_export->exp_vbr_failed = 1;
133 cfs_spin_unlock(&req->rq_export->exp_lock);
140 * Save pre-versions in reply.
142 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
147 if (!exp_connect_vbr(req->rq_export))
150 LASSERT(!req_is_replay(req));
151 LASSERT(req->rq_repmsg != NULL);
152 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
154 reply_ver[idx] = version;
158 * Save enoent version, it is needed when it is obvious that object doesn't
159 * exist, e.g. child during create.
161 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
163 /* save version of file name for replay, it must be ENOENT here */
164 if (!req_is_replay(mdt_info_req(info))) {
165 info->mti_ver[idx] = ENOENT_VERSION;
166 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and save in reply buffer.
173 * Versions are saved in reply only during normal operations not replays.
175 void mdt_version_get_save(struct mdt_thread_info *info,
176 struct mdt_object *mto, int idx)
178 /* don't save versions during replay */
179 if (!req_is_replay(mdt_info_req(info))) {
180 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
181 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
186 * Get version from disk and check it, no save in reply.
188 int mdt_version_get_check(struct mdt_thread_info *info,
189 struct mdt_object *mto, int idx)
191 /* only check versions during replay */
192 if (!req_is_replay(mdt_info_req(info)))
195 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
196 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
200 * Get version from disk and check if recovery or just save.
202 int mdt_version_get_check_save(struct mdt_thread_info *info,
203 struct mdt_object *mto, int idx)
207 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
208 if (req_is_replay(mdt_info_req(info)))
209 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
212 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
217 * Lookup with version checking.
219 * This checks version of 'name'. Many reint functions uses 'name' for child not
220 * FID, therefore we need to get object by name and check its version.
222 int mdt_lookup_version_check(struct mdt_thread_info *info,
223 struct mdt_object *p, struct lu_name *lname,
224 struct lu_fid *fid, int idx)
228 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
230 /* Check version only during replay */
231 if (!req_is_replay(mdt_info_req(info)))
234 info->mti_ver[idx] = ENOENT_VERSION;
236 struct mdt_object *child;
237 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
238 if (likely(!IS_ERR(child))) {
239 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
240 mdt_object_put(info->mti_env, child);
243 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
244 return vbrc ? vbrc : rc;
249 * VBR: we save three versions in reply:
250 * 0 - parent. Check that parent version is the same during replay.
251 * 1 - name. Version of 'name' if file exists with the same name or
252 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
253 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
256 static int mdt_md_create(struct mdt_thread_info *info)
258 struct mdt_device *mdt = info->mti_mdt;
259 struct mdt_object *parent;
260 struct mdt_object *child;
261 struct mdt_lock_handle *lh;
262 struct mdt_body *repbody;
263 struct md_attr *ma = &info->mti_attr;
264 struct mdt_reint_record *rr = &info->mti_rr;
265 struct lu_name *lname;
269 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
270 rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
272 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
274 lh = &info->mti_lh[MDT_LH_PARENT];
275 mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
277 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
278 MDS_INODELOCK_UPDATE);
280 RETURN(PTR_ERR(parent));
282 if (mdt_object_obf(parent))
283 GOTO(out_put_parent, rc = -EPERM);
285 rc = mdt_version_get_check_save(info, parent, 0);
287 GOTO(out_put_parent, rc);
290 * Check child name version during replay.
291 * During create replay a file may exist with same name.
293 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
294 rc = mdt_lookup_version_check(info, parent, lname,
295 &info->mti_tmp_fid1, 1);
297 GOTO(out_put_parent, rc = -EEXIST);
299 /* -ENOENT is expected here */
301 GOTO(out_put_parent, rc);
303 /* save version of file name for replay, it must be ENOENT here */
304 mdt_enoent_version_save(info, 1);
306 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
307 if (likely(!IS_ERR(child))) {
308 struct md_object *next = mdt_object_child(parent);
310 ma->ma_need = MA_INODE;
312 /* capa for cross-ref will be stored here */
313 ma->ma_capa = req_capsule_server_get(info->mti_pill,
315 LASSERT(ma->ma_capa);
317 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
318 OBD_FAIL_MDS_REINT_CREATE_WRITE);
320 /* Version of child will be updated on disk. */
321 info->mti_mos = child;
322 rc = mdt_version_get_check_save(info, child, 2);
324 GOTO(out_put_child, rc);
326 /* Let lower layer know current lock mode. */
327 info->mti_spec.sp_cr_mode =
328 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
331 * Do not perform lookup sanity check. We know that name does
334 info->mti_spec.sp_cr_lookup = 0;
335 info->mti_spec.sp_feat = &dt_directory_features;
337 rc = mdo_create(info->mti_env, next, lname,
338 mdt_object_child(child),
339 &info->mti_spec, ma);
341 rc = mdt_attr_get_complex(info, child, ma);
344 /* Return fid & attr to client. */
345 if (ma->ma_valid & MA_INODE)
346 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
347 mdt_object_fid(child));
350 mdt_object_put(info->mti_env, child);
354 mdt_create_pack_capa(info, rc, child, repbody);
356 mdt_object_unlock_put(info, parent, lh, rc);
360 /* Partial request to create object only */
361 static int mdt_md_mkobj(struct mdt_thread_info *info)
363 struct mdt_device *mdt = info->mti_mdt;
364 struct mdt_object *o;
365 struct mdt_body *repbody;
366 struct md_attr *ma = &info->mti_attr;
370 DEBUG_REQ(D_INODE, mdt_info_req(info), "Partial create "DFID"",
371 PFID(info->mti_rr.rr_fid2));
373 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
375 o = mdt_object_find(info->mti_env, mdt, info->mti_rr.rr_fid2);
377 struct md_object *next = mdt_object_child(o);
379 ma->ma_need = MA_INODE;
383 * Cross-ref create can encounter already created obj in case of
384 * recovery, just get attr in that case.
386 if (mdt_object_exists(o) == 1) {
387 rc = mdt_attr_get_complex(info, o, ma);
390 * Here, NO permission check for object_create,
391 * such check has been done on the original MDS.
393 rc = mo_object_create(info->mti_env, next,
394 &info->mti_spec, ma);
397 /* Return fid & attr to client. */
398 if (ma->ma_valid & MA_INODE)
399 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
402 mdt_object_put(info->mti_env, o);
406 mdt_create_pack_capa(info, rc, o, repbody);
410 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
411 struct md_attr *ma, int flags)
413 struct mdt_lock_handle *lh;
414 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
415 __u64 lockpart = MDS_INODELOCK_UPDATE;
419 /* attr shouldn't be set on remote object */
420 LASSERT(mdt_object_exists(mo) >= 0);
422 lh = &info->mti_lh[MDT_LH_PARENT];
423 mdt_lock_reg_init(lh, LCK_PW);
425 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
426 lockpart |= MDS_INODELOCK_LOOKUP;
428 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
432 if (mdt_object_exists(mo) == 0)
433 GOTO(out_unlock, rc = -ENOENT);
435 /* all attrs are packed into mti_attr in unpack_setattr */
436 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
437 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
439 /* This is only for set ctime when rename's source is on remote MDS. */
440 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
441 ma->ma_attr_flags |= MDS_VTX_BYPASS;
443 /* VBR: update version if attr changed are important for recovery */
445 /* update on-disk version of changed object */
447 rc = mdt_version_get_check_save(info, mo, 0);
449 GOTO(out_unlock, rc);
452 /* all attrs are packed into mti_attr in unpack_setattr */
453 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
455 GOTO(out_unlock, rc);
459 mdt_object_unlock(info, mo, lh, rc);
463 static int mdt_reint_setattr(struct mdt_thread_info *info,
464 struct mdt_lock_handle *lhc)
466 struct md_attr *ma = &info->mti_attr;
467 struct mdt_reint_record *rr = &info->mti_rr;
468 struct ptlrpc_request *req = mdt_info_req(info);
469 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
470 struct mdt_file_data *mfd;
471 struct mdt_object *mo;
472 struct mdt_body *repbody;
476 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
477 (unsigned int)ma->ma_attr.la_valid);
479 if (info->mti_dlm_req)
480 ldlm_request_cancel(req, info->mti_dlm_req, 0);
482 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
483 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
485 GOTO(out, rc = PTR_ERR(mo));
487 if (mdt_object_obf(mo))
488 GOTO(out_put, rc = -EPERM);
490 /* start a log jounal handle if needed */
491 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
492 if ((ma->ma_attr.la_valid & LA_SIZE) ||
493 (rr->rr_flags & MRF_OPEN_TRUNC)) {
494 /* Check write access for the O_TRUNC case */
495 if (mdt_write_read(mo) < 0)
496 GOTO(out_put, rc = -ETXTBSY);
498 } else if (info->mti_ioepoch &&
499 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
500 /* Truncate case. IOEpoch is opened. */
501 rc = mdt_write_get(mo);
508 GOTO(out_put, rc = -ENOMEM);
511 mdt_ioepoch_open(info, mo, 0);
512 repbody->ioepoch = mo->mot_ioepoch;
514 mdt_object_get(info->mti_env, mo);
515 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
516 mfd->mfd_object = mo;
517 mfd->mfd_xid = req->rq_xid;
519 cfs_spin_lock(&med->med_open_lock);
520 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
521 cfs_spin_unlock(&med->med_open_lock);
522 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
525 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
527 /* SOM Attribute update case. Find the proper mfd and update
528 * SOM attributes on the proper object. */
529 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
530 LASSERT(info->mti_ioepoch);
532 cfs_spin_lock(&med->med_open_lock);
533 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
535 cfs_spin_unlock(&med->med_open_lock);
536 CDEBUG(D_INODE, "no handle for file close: "
537 "fid = "DFID": cookie = "LPX64"\n",
538 PFID(info->mti_rr.rr_fid1),
539 info->mti_ioepoch->handle.cookie);
540 GOTO(out_put, rc = -ESTALE);
542 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
543 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
545 class_handle_unhash(&mfd->mfd_handle);
546 cfs_list_del_init(&mfd->mfd_list);
547 cfs_spin_unlock(&med->med_open_lock);
549 mdt_mfd_close(info, mfd);
550 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
551 LASSERT((ma->ma_valid & MA_LOV) == 0);
552 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
555 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
556 struct lu_buf *buf = &info->mti_buf;
557 LASSERT(ma->ma_attr.la_valid == 0);
558 buf->lb_buf = ma->ma_lmm;
559 buf->lb_len = ma->ma_lmm_size;
560 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
561 buf, XATTR_NAME_LOV, 0);
567 ma->ma_need = MA_INODE;
569 rc = mdt_attr_get_complex(info, mo, ma);
573 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
575 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
576 info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
577 S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
578 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
579 struct lustre_capa *capa;
581 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
583 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
584 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
587 repbody->valid |= OBD_MD_FLOSSCAPA;
592 mdt_object_put(info->mti_env, mo);
595 mdt_counter_incr(req, LPROC_MDT_SETATTR);
597 mdt_client_compatibility(info);
598 rc2 = mdt_fix_reply(info);
604 static int mdt_reint_create(struct mdt_thread_info *info,
605 struct mdt_lock_handle *lhc)
607 struct ptlrpc_request *req = mdt_info_req(info);
611 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
612 RETURN(err_serious(-ESTALE));
614 if (info->mti_dlm_req)
615 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
617 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
619 /* Cross-ref case. */
620 /* TODO: we can add LPROC_MDT_CROSS for cross-ref stats */
621 if (info->mti_cross_ref) {
622 rc = mdt_md_mkobj(info);
624 LASSERT(info->mti_rr.rr_namelen > 0);
625 mdt_counter_incr(req, LPROC_MDT_MKDIR);
626 rc = mdt_md_create(info);
636 /* Special file should stay on the same node as parent. */
637 LASSERT(info->mti_rr.rr_namelen > 0);
638 mdt_counter_incr(req, LPROC_MDT_MKNOD);
639 rc = mdt_md_create(info);
643 rc = err_serious(-EOPNOTSUPP);
649 * VBR: save parent version in reply and child version getting by its name.
650 * Version of child is getting and checking during its lookup. If
652 static int mdt_reint_unlink(struct mdt_thread_info *info,
653 struct mdt_lock_handle *lhc)
655 struct mdt_reint_record *rr = &info->mti_rr;
656 struct ptlrpc_request *req = mdt_info_req(info);
657 struct md_attr *ma = &info->mti_attr;
658 struct lu_fid *child_fid = &info->mti_tmp_fid1;
659 struct mdt_object *mp;
660 struct mdt_object *mc;
661 struct mdt_lock_handle *parent_lh;
662 struct mdt_lock_handle *child_lh;
663 struct lu_name *lname;
667 DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
670 if (info->mti_dlm_req)
671 ldlm_request_cancel(req, info->mti_dlm_req, 0);
673 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
674 RETURN(err_serious(-ENOENT));
677 * step 1: lock the parent. Note, this may be child in case of
678 * remote operation denoted by ->mti_cross_ref flag.
680 parent_lh = &info->mti_lh[MDT_LH_PARENT];
681 if (info->mti_cross_ref) {
683 * Init reg lock for cross ref case when we need to do only
686 mdt_lock_reg_init(parent_lh, LCK_PW);
688 mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
691 mp = mdt_object_find_lock(info, rr->rr_fid1, parent_lh,
692 MDS_INODELOCK_UPDATE);
695 /* errors are possible here in cross-ref cases, see below */
696 if (info->mti_cross_ref)
701 if (mdt_object_obf(mp))
702 GOTO(out_unlock_parent, rc = -EPERM);
704 rc = mdt_version_get_check_save(info, mp, 0);
706 GOTO(out_unlock_parent, rc);
708 mdt_reint_init_ma(info, ma);
710 if (info->mti_cross_ref) {
712 * Remote partial operation. It is possible that replay may
713 * happen on parent MDT and this operation will be repeated.
714 * Therefore the object absense is allowed case and nothing
715 * should be done here.
717 if (mdt_object_exists(mp) > 0) {
718 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
719 rc = mo_ref_del(info->mti_env,
720 mdt_object_child(mp), ma);
722 mdt_handle_last_unlink(info, mp, ma);
725 GOTO(out_unlock_parent, rc);
728 /* step 2: find & lock the child */
729 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
730 /* lookup child object along with version checking */
732 rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
734 GOTO(out_unlock_parent, rc);
736 /* We will lock the child regardless it is local or remote. No harm. */
737 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
739 GOTO(out_unlock_parent, rc = PTR_ERR(mc));
740 child_lh = &info->mti_lh[MDT_LH_CHILD];
741 mdt_lock_reg_init(child_lh, LCK_EX);
742 rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
745 mdt_object_put(info->mti_env, mc);
746 GOTO(out_unlock_parent, rc);
749 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
750 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
751 /* save version when object is locked */
752 mdt_version_get_save(info, mc, 1);
754 * Now we can only make sure we need MA_INODE, in mdd layer, will check
755 * whether need MA_LOV and MA_COOKIE.
757 ma->ma_need = MA_INODE;
759 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
760 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
761 mdt_object_child(mc), lname, ma);
762 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
763 rc = mdt_attr_get_complex(info, mc, ma);
765 mdt_handle_last_unlink(info, mc, ma);
767 if (ma->ma_valid & MA_INODE) {
768 switch (ma->ma_attr.la_mode & S_IFMT) {
770 mdt_counter_incr(req, LPROC_MDT_RMDIR);
778 mdt_counter_incr(req, LPROC_MDT_UNLINK);
781 LASSERTF(0, "bad file type %o unlinking\n",
782 ma->ma_attr.la_mode);
788 mdt_object_unlock_put(info, mc, child_lh, rc);
790 mdt_object_unlock_put(info, mp, parent_lh, rc);
796 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
799 static int mdt_reint_link(struct mdt_thread_info *info,
800 struct mdt_lock_handle *lhc)
802 struct mdt_reint_record *rr = &info->mti_rr;
803 struct ptlrpc_request *req = mdt_info_req(info);
804 struct md_attr *ma = &info->mti_attr;
805 struct mdt_object *ms;
806 struct mdt_object *mp;
807 struct mdt_lock_handle *lhs;
808 struct mdt_lock_handle *lhp;
809 struct lu_name *lname;
813 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
814 PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
816 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
817 RETURN(err_serious(-ENOENT));
819 if (info->mti_dlm_req)
820 ldlm_request_cancel(req, info->mti_dlm_req, 0);
822 if (info->mti_cross_ref) {
823 /* MDT holding name ask us to add ref. */
824 lhs = &info->mti_lh[MDT_LH_CHILD];
825 mdt_lock_reg_init(lhs, LCK_EX);
826 ms = mdt_object_find_lock(info, rr->rr_fid1, lhs,
827 MDS_INODELOCK_UPDATE);
831 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
832 rc = mo_ref_add(info->mti_env, mdt_object_child(ms), ma);
833 mdt_object_unlock_put(info, ms, lhs, rc);
837 /* Invalid case so return error immediately instead of
839 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
842 /* step 1: find & lock the target parent dir */
843 lhp = &info->mti_lh[MDT_LH_PARENT];
844 mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
846 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
847 MDS_INODELOCK_UPDATE);
851 if (mdt_object_obf(mp))
852 GOTO(out_unlock_parent, rc = -EPERM);
854 rc = mdt_version_get_check_save(info, mp, 0);
856 GOTO(out_unlock_parent, rc);
858 /* step 2: find & lock the source */
859 lhs = &info->mti_lh[MDT_LH_CHILD];
860 mdt_lock_reg_init(lhs, LCK_EX);
862 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
864 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
866 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
869 mdt_object_put(info->mti_env, ms);
870 GOTO(out_unlock_parent, rc);
873 /* step 3: link it */
874 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
875 OBD_FAIL_MDS_REINT_LINK_WRITE);
878 rc = mdt_version_get_check_save(info, ms, 1);
880 GOTO(out_unlock_child, rc);
882 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
883 /** check target version by name during replay */
884 rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
885 if (rc != 0 && rc != -ENOENT)
886 GOTO(out_unlock_child, rc);
887 /* save version of file name for replay, it must be ENOENT here */
888 if (!req_is_replay(mdt_info_req(info))) {
889 info->mti_ver[2] = ENOENT_VERSION;
890 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
893 rc = mdo_link(info->mti_env, mdt_object_child(mp),
894 mdt_object_child(ms), lname, ma);
897 mdt_counter_incr(req, LPROC_MDT_LINK);
901 mdt_object_unlock_put(info, ms, lhs, rc);
903 mdt_object_unlock_put(info, mp, lhp, rc);
907 * lock the part of the directory according to the hash of the name
908 * (lh->mlh_pdo_hash) in parallel directory lock.
910 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
911 struct mdt_lock_handle *lh,
912 struct mdt_object *obj, __u64 ibits)
914 struct ldlm_res_id *res_id = &info->mti_res_id;
915 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
916 ldlm_policy_data_t *policy = &info->mti_policy;
920 * Finish res_id initializing by name hash marking part of
921 * directory which is taking modification.
923 LASSERT(lh->mlh_pdo_hash != 0);
924 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
925 memset(policy, 0, sizeof(*policy));
926 policy->l_inodebits.bits = ibits;
928 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
929 * going to be sent to client. If it is - mdt_intent_policy() path will
930 * fix it up and turn FL_LOCAL flag off.
932 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
933 res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
934 &info->mti_exp->exp_handle.h_cookie);
938 /* partial operation for rename */
939 static int mdt_reint_rename_tgt(struct mdt_thread_info *info)
941 struct mdt_reint_record *rr = &info->mti_rr;
942 struct ptlrpc_request *req = mdt_info_req(info);
943 struct md_attr *ma = &info->mti_attr;
944 struct mdt_object *mtgtdir;
945 struct mdt_object *mtgt = NULL;
946 struct mdt_lock_handle *lh_tgtdir;
947 struct mdt_lock_handle *lh_tgt = NULL;
948 struct lu_fid *tgt_fid = &info->mti_tmp_fid1;
949 struct lu_name *lname;
953 DEBUG_REQ(D_INODE, req, "rename_tgt: insert (%s->"DFID") in "DFID,
954 rr->rr_tgt, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
956 /* step 1: lookup & lock the tgt dir. */
957 lh_tgtdir = &info->mti_lh[MDT_LH_PARENT];
958 mdt_lock_pdo_init(lh_tgtdir, LCK_PW, rr->rr_tgt,
960 mtgtdir = mdt_object_find_lock(info, rr->rr_fid1, lh_tgtdir,
961 MDS_INODELOCK_UPDATE);
963 RETURN(PTR_ERR(mtgtdir));
965 /* step 2: find & lock the target object if exists. */
966 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
967 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
968 rc = mdo_lookup(info->mti_env, mdt_object_child(mtgtdir),
969 lname, tgt_fid, &info->mti_spec);
970 if (rc != 0 && rc != -ENOENT) {
971 GOTO(out_unlock_tgtdir, rc);
972 } else if (rc == 0) {
974 * In case of replay that name can be already inserted, check
975 * that and do nothing if so.
977 if (lu_fid_eq(tgt_fid, rr->rr_fid2))
978 GOTO(out_unlock_tgtdir, rc);
980 lh_tgt = &info->mti_lh[MDT_LH_CHILD];
981 mdt_lock_reg_init(lh_tgt, LCK_EX);
983 mtgt = mdt_object_find_lock(info, tgt_fid, lh_tgt,
984 MDS_INODELOCK_LOOKUP);
986 GOTO(out_unlock_tgtdir, rc = PTR_ERR(mtgt));
988 mdt_reint_init_ma(info, ma);
990 rc = mdo_rename_tgt(info->mti_env, mdt_object_child(mtgtdir),
991 mdt_object_child(mtgt), rr->rr_fid2,
993 } else /* -ENOENT */ {
994 rc = mdo_name_insert(info->mti_env, mdt_object_child(mtgtdir),
995 lname, rr->rr_fid2, ma);
998 /* handle last link of tgt object */
1000 mdt_handle_last_unlink(info, mtgt, ma);
1005 mdt_object_unlock_put(info, mtgt, lh_tgt, rc);
1007 mdt_object_unlock_put(info, mtgtdir, lh_tgtdir, rc);
1011 static int mdt_rename_lock(struct mdt_thread_info *info,
1012 struct lustre_handle *lh)
1014 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1015 ldlm_policy_data_t *policy = &info->mti_policy;
1016 struct ldlm_res_id *res_id = &info->mti_res_id;
1021 ms = mdt_md_site(info->mti_mdt);
1022 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1024 memset(policy, 0, sizeof *policy);
1025 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1027 if (ms->ms_control_exp == NULL) {
1028 int flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1031 * Current node is controller, that is mdt0, where we should
1034 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1035 LCK_EX, &flags, ldlm_blocking_ast,
1036 ldlm_completion_ast, NULL, NULL, 0,
1037 &info->mti_exp->exp_handle.h_cookie,
1040 struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_EX,
1041 ldlm_blocking_ast, ldlm_completion_ast, NULL, NULL, NULL };
1045 * This is the case mdt0 is remote node, issue DLM lock like
1048 rc = ldlm_cli_enqueue(ms->ms_control_exp, NULL, &einfo, res_id,
1049 policy, &flags, NULL, 0, lh, 0);
1055 static void mdt_rename_unlock(struct lustre_handle *lh)
1058 LASSERT(lustre_handle_is_used(lh));
1059 ldlm_lock_decref(lh, LCK_EX);
1064 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1065 * target. Source should not be ancestor of target dir. May be other rename
1066 * checks can be moved here later.
1068 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1070 struct mdt_reint_record *rr = &info->mti_rr;
1071 struct lu_fid dst_fid = *rr->rr_fid2;
1072 struct mdt_object *dst;
1077 LASSERT(fid_is_sane(&dst_fid));
1078 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1080 rc = mdo_is_subdir(info->mti_env,
1081 mdt_object_child(dst), fid,
1083 mdt_object_put(info->mti_env, dst);
1084 if (rc != -EREMOTE && rc < 0) {
1085 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1087 /* check the found fid */
1088 if (lu_fid_eq(&dst_fid, fid))
1094 } while (rc == -EREMOTE);
1100 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1101 * 2 - src child; 3 - tgt child.
1102 * Update on disk version of src child.
1104 static int mdt_reint_rename(struct mdt_thread_info *info,
1105 struct mdt_lock_handle *lhc)
1107 struct mdt_reint_record *rr = &info->mti_rr;
1108 struct md_attr *ma = &info->mti_attr;
1109 struct ptlrpc_request *req = mdt_info_req(info);
1110 struct mdt_object *msrcdir;
1111 struct mdt_object *mtgtdir;
1112 struct mdt_object *mold;
1113 struct mdt_object *mnew = NULL;
1114 struct mdt_lock_handle *lh_srcdirp;
1115 struct mdt_lock_handle *lh_tgtdirp;
1116 struct mdt_lock_handle *lh_oldp;
1117 struct mdt_lock_handle *lh_newp;
1118 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1119 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1120 struct lustre_handle rename_lh = { 0 };
1121 struct lu_name slname = { 0 };
1122 struct lu_name *lname;
1126 if (info->mti_dlm_req)
1127 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1129 if (info->mti_cross_ref) {
1130 rc = mdt_reint_rename_tgt(info);
1134 DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s",
1135 PFID(rr->rr_fid1), rr->rr_name,
1136 PFID(rr->rr_fid2), rr->rr_tgt);
1138 rc = mdt_rename_lock(info, &rename_lh);
1140 CERROR("Can't lock FS for rename, rc %d\n", rc);
1144 lh_newp = &info->mti_lh[MDT_LH_NEW];
1146 /* step 1: lock the source dir. */
1147 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1148 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name,
1150 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1151 MDS_INODELOCK_UPDATE);
1152 if (IS_ERR(msrcdir))
1153 GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
1155 if (mdt_object_obf(msrcdir))
1156 GOTO(out_unlock_source, rc = -EPERM);
1158 rc = mdt_version_get_check_save(info, msrcdir, 0);
1160 GOTO(out_unlock_source, rc);
1162 /* step 2: find & lock the target dir. */
1163 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1164 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt,
1166 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1167 mdt_object_get(info->mti_env, msrcdir);
1169 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1170 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1171 MDS_INODELOCK_UPDATE);
1173 GOTO(out_unlock_source, rc);
1174 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1177 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1179 if (IS_ERR(mtgtdir))
1180 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1182 if (mdt_object_obf(mtgtdir))
1183 GOTO(out_put_target, rc = -EPERM);
1185 /* check early, the real version will be saved after locking */
1186 rc = mdt_version_get_check(info, mtgtdir, 1);
1188 GOTO(out_put_target, rc);
1190 rc = mdt_object_exists(mtgtdir);
1192 GOTO(out_put_target, rc = -ESTALE);
1193 } else if (rc > 0) {
1194 /* we lock the target dir if it is local */
1195 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1196 MDS_INODELOCK_UPDATE,
1199 GOTO(out_put_target, rc);
1200 /* get and save correct version after locking */
1201 mdt_version_get_save(info, mtgtdir, 1);
1205 /* step 3: find & lock the old object. */
1206 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1207 mdt_name_copy(&slname, lname);
1209 rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2);
1211 GOTO(out_unlock_target, rc);
1213 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1214 GOTO(out_unlock_target, rc = -EINVAL);
1216 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1218 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1220 if (mdt_object_obf(mold)) {
1221 mdt_object_put(info->mti_env, mold);
1222 GOTO(out_unlock_target, rc = -EPERM);
1225 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1226 mdt_lock_reg_init(lh_oldp, LCK_EX);
1227 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
1230 mdt_object_put(info->mti_env, mold);
1231 GOTO(out_unlock_target, rc);
1234 info->mti_mos = mold;
1235 /* save version after locking */
1236 mdt_version_get_save(info, mold, 2);
1237 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1239 /* step 4: find & lock the new object. */
1240 /* new target object may not exist now */
1241 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
1242 /* lookup with version checking */
1244 rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3);
1246 /* the new_fid should have been filled at this moment */
1247 if (lu_fid_eq(old_fid, new_fid))
1248 GOTO(out_unlock_old, rc);
1250 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1251 lu_fid_eq(new_fid, rr->rr_fid2))
1252 GOTO(out_unlock_old, rc = -EINVAL);
1254 mdt_lock_reg_init(lh_newp, LCK_EX);
1255 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1257 GOTO(out_unlock_old, rc = PTR_ERR(mnew));
1259 if (mdt_object_obf(mnew)) {
1260 mdt_object_put(info->mti_env, mnew);
1261 GOTO(out_unlock_old, rc = -EPERM);
1264 rc = mdt_object_lock(info, mnew, lh_newp,
1265 MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
1267 mdt_object_put(info->mti_env, mnew);
1268 GOTO(out_unlock_old, rc);
1270 /* get and save version after locking */
1271 mdt_version_get_save(info, mnew, 3);
1272 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1273 } else if (rc != -EREMOTE && rc != -ENOENT) {
1274 GOTO(out_unlock_old, rc);
1276 mdt_enoent_version_save(info, 3);
1279 /* step 5: rename it */
1280 mdt_reint_init_ma(info, ma);
1282 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1283 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1286 /* Check if @dst is subdir of @src. */
1287 rc = mdt_rename_sanity(info, old_fid);
1289 GOTO(out_unlock_new, rc);
1291 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1292 mdt_object_child(mtgtdir), old_fid, &slname,
1293 (mnew ? mdt_object_child(mnew) : NULL),
1296 /* handle last link of tgt object */
1298 mdt_counter_incr(req, LPROC_MDT_RENAME);
1300 mdt_handle_last_unlink(info, mnew, ma);
1302 mdt_rename_counter_tally(info, info->mti_mdt, req,
1309 mdt_object_unlock_put(info, mnew, lh_newp, rc);
1311 mdt_object_unlock_put(info, mold, lh_oldp, rc);
1313 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1315 mdt_object_put(info->mti_env, mtgtdir);
1317 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1319 mdt_rename_unlock(&rename_lh);
1323 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1324 struct mdt_lock_handle *lhc);
1326 static mdt_reinter reinters[REINT_MAX] = {
1327 [REINT_SETATTR] = mdt_reint_setattr,
1328 [REINT_CREATE] = mdt_reint_create,
1329 [REINT_LINK] = mdt_reint_link,
1330 [REINT_UNLINK] = mdt_reint_unlink,
1331 [REINT_RENAME] = mdt_reint_rename,
1332 [REINT_OPEN] = mdt_reint_open,
1333 [REINT_SETXATTR] = mdt_reint_setxattr
1336 int mdt_reint_rec(struct mdt_thread_info *info,
1337 struct mdt_lock_handle *lhc)
1342 rc = reinters[info->mti_rr.rr_opcode](info, lhc);