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_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
56 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
57 &RMF_MDT_MD, RCL_SERVER);
59 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
61 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
65 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
69 static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
70 struct mdt_object *object,
71 struct mdt_body *repbody)
75 /* for cross-ref mkdir, mds capa has been fetched from remote obj, then
76 * we won't go to below*/
77 if (repbody->valid & OBD_MD_FLMDSCAPA)
80 if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
81 info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
82 struct lustre_capa *capa;
84 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
86 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
87 rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa,
90 repbody->valid |= OBD_MD_FLMDSCAPA;
97 * Get version of object by fid.
99 * Return real version or ENOENT_VERSION if object doesn't exist
101 static void mdt_obj_version_get(struct mdt_thread_info *info,
102 struct mdt_object *o, __u64 *version)
105 LASSERT(mdt_object_exists(o) >= 0);
106 if (mdt_object_exists(o) > 0 && !mdt_object_obf(o))
107 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
109 *version = ENOENT_VERSION;
110 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
111 PFID(mdt_object_fid(o)), *version);
115 * Check version is correct.
117 * Should be called only during replay.
119 static int mdt_version_check(struct ptlrpc_request *req,
120 __u64 version, int idx)
122 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
125 if (!exp_connect_vbr(req->rq_export))
128 LASSERT(req_is_replay(req));
129 /** VBR: version is checked always because costs nothing */
130 LASSERT(idx < PTLRPC_NUM_VERSIONS);
131 /** Sanity check for malformed buffers */
132 if (pre_ver == NULL) {
133 CERROR("No versions in request buffer\n");
134 cfs_spin_lock(&req->rq_export->exp_lock);
135 req->rq_export->exp_vbr_failed = 1;
136 cfs_spin_unlock(&req->rq_export->exp_lock);
138 } else if (pre_ver[idx] != version) {
139 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
140 pre_ver[idx], version);
141 cfs_spin_lock(&req->rq_export->exp_lock);
142 req->rq_export->exp_vbr_failed = 1;
143 cfs_spin_unlock(&req->rq_export->exp_lock);
150 * Save pre-versions in reply.
152 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
157 if (!exp_connect_vbr(req->rq_export))
160 LASSERT(!req_is_replay(req));
161 LASSERT(req->rq_repmsg != NULL);
162 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
164 reply_ver[idx] = version;
168 * Save enoent version, it is needed when it is obvious that object doesn't
169 * exist, e.g. child during create.
171 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
173 /* save version of file name for replay, it must be ENOENT here */
174 if (!req_is_replay(mdt_info_req(info))) {
175 info->mti_ver[idx] = ENOENT_VERSION;
176 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
181 * Get version from disk and save in reply buffer.
183 * Versions are saved in reply only during normal operations not replays.
185 void mdt_version_get_save(struct mdt_thread_info *info,
186 struct mdt_object *mto, int idx)
188 /* don't save versions during replay */
189 if (!req_is_replay(mdt_info_req(info))) {
190 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
191 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
196 * Get version from disk and check it, no save in reply.
198 int mdt_version_get_check(struct mdt_thread_info *info,
199 struct mdt_object *mto, int idx)
201 /* only check versions during replay */
202 if (!req_is_replay(mdt_info_req(info)))
205 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
206 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
210 * Get version from disk and check if recovery or just save.
212 int mdt_version_get_check_save(struct mdt_thread_info *info,
213 struct mdt_object *mto, int idx)
217 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
218 if (req_is_replay(mdt_info_req(info)))
219 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
222 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
227 * Lookup with version checking.
229 * This checks version of 'name'. Many reint functions uses 'name' for child not
230 * FID, therefore we need to get object by name and check its version.
232 int mdt_lookup_version_check(struct mdt_thread_info *info,
233 struct mdt_object *p, struct lu_name *lname,
234 struct lu_fid *fid, int idx)
238 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
240 /* Check version only during replay */
241 if (!req_is_replay(mdt_info_req(info)))
244 info->mti_ver[idx] = ENOENT_VERSION;
246 struct mdt_object *child;
247 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
248 if (likely(!IS_ERR(child))) {
249 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
250 mdt_object_put(info->mti_env, child);
253 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
254 return vbrc ? vbrc : rc;
259 * VBR: we save three versions in reply:
260 * 0 - parent. Check that parent version is the same during replay.
261 * 1 - name. Version of 'name' if file exists with the same name or
262 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
263 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
266 static int mdt_md_create(struct mdt_thread_info *info)
268 struct mdt_device *mdt = info->mti_mdt;
269 struct mdt_object *parent;
270 struct mdt_object *child;
271 struct mdt_lock_handle *lh;
272 struct mdt_body *repbody;
273 struct md_attr *ma = &info->mti_attr;
274 struct mdt_reint_record *rr = &info->mti_rr;
275 struct lu_name *lname;
279 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
280 rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
282 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
284 lh = &info->mti_lh[MDT_LH_PARENT];
285 mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
287 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
288 MDS_INODELOCK_UPDATE);
290 RETURN(PTR_ERR(parent));
292 if (mdt_object_obf(parent))
293 GOTO(out_put_parent, rc = -EPERM);
295 rc = mdt_version_get_check_save(info, parent, 0);
297 GOTO(out_put_parent, rc);
300 * Check child name version during replay.
301 * During create replay a file may exist with same name.
303 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
304 rc = mdt_lookup_version_check(info, parent, lname,
305 &info->mti_tmp_fid1, 1);
307 GOTO(out_put_parent, rc = -EEXIST);
309 /* -ENOENT is expected here */
311 GOTO(out_put_parent, rc);
313 /* save version of file name for replay, it must be ENOENT here */
314 mdt_enoent_version_save(info, 1);
316 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
317 if (likely(!IS_ERR(child))) {
318 struct md_object *next = mdt_object_child(parent);
320 ma->ma_need = MA_INODE;
322 /* capa for cross-ref will be stored here */
323 ma->ma_capa = req_capsule_server_get(info->mti_pill,
325 LASSERT(ma->ma_capa);
327 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
328 OBD_FAIL_MDS_REINT_CREATE_WRITE);
330 /* Version of child will be updated on disk. */
331 info->mti_mos = child;
332 rc = mdt_version_get_check_save(info, child, 2);
334 GOTO(out_put_child, rc);
336 /* Let lower layer know current lock mode. */
337 info->mti_spec.sp_cr_mode =
338 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
341 * Do not perform lookup sanity check. We know that name does
344 info->mti_spec.sp_cr_lookup = 0;
345 info->mti_spec.sp_feat = &dt_directory_features;
347 rc = mdo_create(info->mti_env, next, lname,
348 mdt_object_child(child),
349 &info->mti_spec, ma);
351 rc = mdt_attr_get_complex(info, child, ma);
354 /* Return fid & attr to client. */
355 if (ma->ma_valid & MA_INODE)
356 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
357 mdt_object_fid(child));
360 mdt_object_put(info->mti_env, child);
364 mdt_create_pack_capa(info, rc, child, repbody);
366 mdt_object_unlock_put(info, parent, lh, rc);
370 /* Partial request to create object only */
371 static int mdt_md_mkobj(struct mdt_thread_info *info)
373 struct mdt_device *mdt = info->mti_mdt;
374 struct mdt_object *o;
375 struct mdt_body *repbody;
376 struct md_attr *ma = &info->mti_attr;
380 DEBUG_REQ(D_INODE, mdt_info_req(info), "Partial create "DFID"",
381 PFID(info->mti_rr.rr_fid2));
383 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
385 o = mdt_object_find(info->mti_env, mdt, info->mti_rr.rr_fid2);
387 struct md_object *next = mdt_object_child(o);
389 ma->ma_need = MA_INODE;
393 * Cross-ref create can encounter already created obj in case of
394 * recovery, just get attr in that case.
396 if (mdt_object_exists(o) == 1) {
397 rc = mdt_attr_get_complex(info, o, ma);
400 * Here, NO permission check for object_create,
401 * such check has been done on the original MDS.
403 rc = mo_object_create(info->mti_env, next,
404 &info->mti_spec, ma);
407 /* Return fid & attr to client. */
408 if (ma->ma_valid & MA_INODE)
409 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
412 mdt_object_put(info->mti_env, o);
416 mdt_create_pack_capa(info, rc, o, repbody);
420 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
421 struct md_attr *ma, int flags)
423 struct mdt_lock_handle *lh;
424 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
425 __u64 lockpart = MDS_INODELOCK_UPDATE;
429 /* attr shouldn't be set on remote object */
430 LASSERT(mdt_object_exists(mo) >= 0);
432 lh = &info->mti_lh[MDT_LH_PARENT];
433 mdt_lock_reg_init(lh, LCK_PW);
435 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
436 lockpart |= MDS_INODELOCK_LOOKUP;
438 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
442 if (mdt_object_exists(mo) == 0)
443 GOTO(out_unlock, rc = -ENOENT);
445 /* all attrs are packed into mti_attr in unpack_setattr */
446 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
447 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
449 /* This is only for set ctime when rename's source is on remote MDS. */
450 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
451 ma->ma_attr_flags |= MDS_VTX_BYPASS;
453 /* VBR: update version if attr changed are important for recovery */
455 /* update on-disk version of changed object */
457 rc = mdt_version_get_check_save(info, mo, 0);
459 GOTO(out_unlock, rc);
462 /* all attrs are packed into mti_attr in unpack_setattr */
463 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
465 GOTO(out_unlock, rc);
469 mdt_object_unlock(info, mo, lh, rc);
473 static int mdt_reint_setattr(struct mdt_thread_info *info,
474 struct mdt_lock_handle *lhc)
476 struct md_attr *ma = &info->mti_attr;
477 struct mdt_reint_record *rr = &info->mti_rr;
478 struct ptlrpc_request *req = mdt_info_req(info);
479 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
480 struct mdt_file_data *mfd;
481 struct mdt_object *mo;
482 struct mdt_body *repbody;
486 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
487 (unsigned int)ma->ma_attr.la_valid);
489 if (info->mti_dlm_req)
490 ldlm_request_cancel(req, info->mti_dlm_req, 0);
492 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
493 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
495 GOTO(out, rc = PTR_ERR(mo));
497 if (mdt_object_obf(mo))
498 GOTO(out_put, rc = -EPERM);
500 /* start a log jounal handle if needed */
501 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
502 if ((ma->ma_attr.la_valid & LA_SIZE) ||
503 (rr->rr_flags & MRF_OPEN_TRUNC)) {
504 /* Check write access for the O_TRUNC case */
505 if (mdt_write_read(mo) < 0)
506 GOTO(out_put, rc = -ETXTBSY);
508 } else if (info->mti_ioepoch &&
509 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
510 /* Truncate case. IOEpoch is opened. */
511 rc = mdt_write_get(mo);
518 GOTO(out_put, rc = -ENOMEM);
521 mdt_ioepoch_open(info, mo, 0);
522 repbody->ioepoch = mo->mot_ioepoch;
524 mdt_object_get(info->mti_env, mo);
525 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
526 mfd->mfd_object = mo;
527 mfd->mfd_xid = req->rq_xid;
529 cfs_spin_lock(&med->med_open_lock);
530 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
531 cfs_spin_unlock(&med->med_open_lock);
532 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
535 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
537 /* SOM Attribute update case. Find the proper mfd and update
538 * SOM attributes on the proper object. */
539 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
540 LASSERT(info->mti_ioepoch);
542 cfs_spin_lock(&med->med_open_lock);
543 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
545 cfs_spin_unlock(&med->med_open_lock);
546 CDEBUG(D_INODE, "no handle for file close: "
547 "fid = "DFID": cookie = "LPX64"\n",
548 PFID(info->mti_rr.rr_fid1),
549 info->mti_ioepoch->handle.cookie);
550 GOTO(out_put, rc = -ESTALE);
552 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
553 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
555 class_handle_unhash(&mfd->mfd_handle);
556 cfs_list_del_init(&mfd->mfd_list);
557 cfs_spin_unlock(&med->med_open_lock);
559 /* Close the found mfd, update attributes. */
560 ma->ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
561 OBD_ALLOC_LARGE(ma->ma_lmm, info->mti_mdt->mdt_max_mdsize);
562 if (ma->ma_lmm == NULL)
563 GOTO(out_put, rc = -ENOMEM);
565 mdt_mfd_close(info, mfd);
567 OBD_FREE_LARGE(ma->ma_lmm, info->mti_mdt->mdt_max_mdsize);
569 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
574 ma->ma_need = MA_INODE;
576 rc = mdt_attr_get_complex(info, mo, ma);
580 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
582 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
583 info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
584 S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
585 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
586 struct lustre_capa *capa;
588 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
590 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
591 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
594 repbody->valid |= OBD_MD_FLOSSCAPA;
599 mdt_object_put(info->mti_env, mo);
602 mdt_counter_incr(req, LPROC_MDT_SETATTR);
604 mdt_client_compatibility(info);
605 rc2 = mdt_fix_reply(info);
611 static int mdt_reint_create(struct mdt_thread_info *info,
612 struct mdt_lock_handle *lhc)
614 struct ptlrpc_request *req = mdt_info_req(info);
618 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
619 RETURN(err_serious(-ESTALE));
621 if (info->mti_dlm_req)
622 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
624 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
626 /* Cross-ref case. */
627 /* TODO: we can add LPROC_MDT_CROSS for cross-ref stats */
628 if (info->mti_cross_ref) {
629 rc = mdt_md_mkobj(info);
631 LASSERT(info->mti_rr.rr_namelen > 0);
632 mdt_counter_incr(req, LPROC_MDT_MKDIR);
633 rc = mdt_md_create(info);
643 /* Special file should stay on the same node as parent. */
644 LASSERT(info->mti_rr.rr_namelen > 0);
645 mdt_counter_incr(req, LPROC_MDT_MKNOD);
646 rc = mdt_md_create(info);
650 rc = err_serious(-EOPNOTSUPP);
656 * VBR: save parent version in reply and child version getting by its name.
657 * Version of child is getting and checking during its lookup. If
659 static int mdt_reint_unlink(struct mdt_thread_info *info,
660 struct mdt_lock_handle *lhc)
662 struct mdt_reint_record *rr = &info->mti_rr;
663 struct ptlrpc_request *req = mdt_info_req(info);
664 struct md_attr *ma = &info->mti_attr;
665 struct lu_fid *child_fid = &info->mti_tmp_fid1;
666 struct mdt_object *mp;
667 struct mdt_object *mc;
668 struct mdt_lock_handle *parent_lh;
669 struct mdt_lock_handle *child_lh;
670 struct lu_name *lname;
674 DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
677 if (info->mti_dlm_req)
678 ldlm_request_cancel(req, info->mti_dlm_req, 0);
680 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
681 RETURN(err_serious(-ENOENT));
684 * step 1: lock the parent. Note, this may be child in case of
685 * remote operation denoted by ->mti_cross_ref flag.
687 parent_lh = &info->mti_lh[MDT_LH_PARENT];
688 if (info->mti_cross_ref) {
690 * Init reg lock for cross ref case when we need to do only
693 mdt_lock_reg_init(parent_lh, LCK_PW);
695 mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
698 mp = mdt_object_find_lock(info, rr->rr_fid1, parent_lh,
699 MDS_INODELOCK_UPDATE);
702 /* errors are possible here in cross-ref cases, see below */
703 if (info->mti_cross_ref)
708 if (mdt_object_obf(mp))
709 GOTO(out_unlock_parent, rc = -EPERM);
711 rc = mdt_version_get_check_save(info, mp, 0);
713 GOTO(out_unlock_parent, rc);
715 mdt_reint_init_ma(info, ma);
716 if (!ma->ma_lmm || !ma->ma_cookie)
717 GOTO(out_unlock_parent, rc = -EINVAL);
719 if (info->mti_cross_ref) {
721 * Remote partial operation. It is possible that replay may
722 * happen on parent MDT and this operation will be repeated.
723 * Therefore the object absense is allowed case and nothing
724 * should be done here.
726 if (mdt_object_exists(mp) > 0) {
727 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
728 rc = mo_ref_del(info->mti_env,
729 mdt_object_child(mp), ma);
731 mdt_handle_last_unlink(info, mp, ma);
734 GOTO(out_unlock_parent, rc);
737 /* step 2: find & lock the child */
738 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
739 /* lookup child object along with version checking */
741 rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
743 GOTO(out_unlock_parent, rc);
745 /* We will lock the child regardless it is local or remote. No harm. */
746 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
748 GOTO(out_unlock_parent, rc = PTR_ERR(mc));
749 child_lh = &info->mti_lh[MDT_LH_CHILD];
750 mdt_lock_reg_init(child_lh, LCK_EX);
751 rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
754 mdt_object_put(info->mti_env, mc);
755 GOTO(out_unlock_parent, rc);
758 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
759 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
760 /* save version when object is locked */
761 mdt_version_get_save(info, mc, 1);
763 * Now we can only make sure we need MA_INODE, in mdd layer, will check
764 * whether need MA_LOV and MA_COOKIE.
766 ma->ma_need = MA_INODE;
768 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
769 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
770 mdt_object_child(mc), lname, ma);
771 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
772 rc = mdt_attr_get_complex(info, mc, ma);
774 mdt_handle_last_unlink(info, mc, ma);
776 if (ma->ma_valid & MA_INODE) {
777 switch (ma->ma_attr.la_mode & S_IFMT) {
779 mdt_counter_incr(req, LPROC_MDT_RMDIR);
787 mdt_counter_incr(req, LPROC_MDT_UNLINK);
790 LASSERTF(0, "bad file type %o unlinking\n",
791 ma->ma_attr.la_mode);
797 mdt_object_unlock_put(info, mc, child_lh, rc);
799 mdt_object_unlock_put(info, mp, parent_lh, rc);
805 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
808 static int mdt_reint_link(struct mdt_thread_info *info,
809 struct mdt_lock_handle *lhc)
811 struct mdt_reint_record *rr = &info->mti_rr;
812 struct ptlrpc_request *req = mdt_info_req(info);
813 struct md_attr *ma = &info->mti_attr;
814 struct mdt_object *ms;
815 struct mdt_object *mp;
816 struct mdt_lock_handle *lhs;
817 struct mdt_lock_handle *lhp;
818 struct lu_name *lname;
822 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
823 PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
825 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
826 RETURN(err_serious(-ENOENT));
828 if (info->mti_dlm_req)
829 ldlm_request_cancel(req, info->mti_dlm_req, 0);
831 if (info->mti_cross_ref) {
832 /* MDT holding name ask us to add ref. */
833 lhs = &info->mti_lh[MDT_LH_CHILD];
834 mdt_lock_reg_init(lhs, LCK_EX);
835 ms = mdt_object_find_lock(info, rr->rr_fid1, lhs,
836 MDS_INODELOCK_UPDATE);
840 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
841 rc = mo_ref_add(info->mti_env, mdt_object_child(ms), ma);
842 mdt_object_unlock_put(info, ms, lhs, rc);
846 /* Invalid case so return error immediately instead of
848 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
851 /* step 1: find & lock the target parent dir */
852 lhp = &info->mti_lh[MDT_LH_PARENT];
853 mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
855 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
856 MDS_INODELOCK_UPDATE);
860 if (mdt_object_obf(mp))
861 GOTO(out_unlock_parent, rc = -EPERM);
863 rc = mdt_version_get_check_save(info, mp, 0);
865 GOTO(out_unlock_parent, rc);
867 /* step 2: find & lock the source */
868 lhs = &info->mti_lh[MDT_LH_CHILD];
869 mdt_lock_reg_init(lhs, LCK_EX);
871 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
873 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
875 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
878 mdt_object_put(info->mti_env, ms);
879 GOTO(out_unlock_parent, rc);
882 /* step 3: link it */
883 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
884 OBD_FAIL_MDS_REINT_LINK_WRITE);
887 rc = mdt_version_get_check_save(info, ms, 1);
889 GOTO(out_unlock_child, rc);
891 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
892 /** check target version by name during replay */
893 rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
894 if (rc != 0 && rc != -ENOENT)
895 GOTO(out_unlock_child, rc);
896 /* save version of file name for replay, it must be ENOENT here */
897 if (!req_is_replay(mdt_info_req(info))) {
898 info->mti_ver[2] = ENOENT_VERSION;
899 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
902 rc = mdo_link(info->mti_env, mdt_object_child(mp),
903 mdt_object_child(ms), lname, ma);
906 mdt_counter_incr(req, LPROC_MDT_LINK);
910 mdt_object_unlock_put(info, ms, lhs, rc);
912 mdt_object_unlock_put(info, mp, lhp, rc);
916 * lock the part of the directory according to the hash of the name
917 * (lh->mlh_pdo_hash) in parallel directory lock.
919 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
920 struct mdt_lock_handle *lh,
921 struct mdt_object *obj, __u64 ibits)
923 struct ldlm_res_id *res_id = &info->mti_res_id;
924 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
925 ldlm_policy_data_t *policy = &info->mti_policy;
929 * Finish res_id initializing by name hash marking part of
930 * directory which is taking modification.
932 LASSERT(lh->mlh_pdo_hash != 0);
933 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
934 memset(policy, 0, sizeof(*policy));
935 policy->l_inodebits.bits = ibits;
937 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
938 * going to be sent to client. If it is - mdt_intent_policy() path will
939 * fix it up and turn FL_LOCAL flag off.
941 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
942 res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
943 &info->mti_exp->exp_handle.h_cookie);
947 /* partial operation for rename */
948 static int mdt_reint_rename_tgt(struct mdt_thread_info *info)
950 struct mdt_reint_record *rr = &info->mti_rr;
951 struct ptlrpc_request *req = mdt_info_req(info);
952 struct md_attr *ma = &info->mti_attr;
953 struct mdt_object *mtgtdir;
954 struct mdt_object *mtgt = NULL;
955 struct mdt_lock_handle *lh_tgtdir;
956 struct mdt_lock_handle *lh_tgt = NULL;
957 struct lu_fid *tgt_fid = &info->mti_tmp_fid1;
958 struct lu_name *lname;
962 DEBUG_REQ(D_INODE, req, "rename_tgt: insert (%s->"DFID") in "DFID,
963 rr->rr_tgt, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
965 /* step 1: lookup & lock the tgt dir. */
966 lh_tgtdir = &info->mti_lh[MDT_LH_PARENT];
967 mdt_lock_pdo_init(lh_tgtdir, LCK_PW, rr->rr_tgt,
969 mtgtdir = mdt_object_find_lock(info, rr->rr_fid1, lh_tgtdir,
970 MDS_INODELOCK_UPDATE);
972 RETURN(PTR_ERR(mtgtdir));
974 /* step 2: find & lock the target object if exists. */
975 mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
976 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
977 rc = mdo_lookup(info->mti_env, mdt_object_child(mtgtdir),
978 lname, tgt_fid, &info->mti_spec);
979 if (rc != 0 && rc != -ENOENT) {
980 GOTO(out_unlock_tgtdir, rc);
981 } else if (rc == 0) {
983 * In case of replay that name can be already inserted, check
984 * that and do nothing if so.
986 if (lu_fid_eq(tgt_fid, rr->rr_fid2))
987 GOTO(out_unlock_tgtdir, rc);
989 lh_tgt = &info->mti_lh[MDT_LH_CHILD];
990 mdt_lock_reg_init(lh_tgt, LCK_EX);
992 mtgt = mdt_object_find_lock(info, tgt_fid, lh_tgt,
993 MDS_INODELOCK_LOOKUP);
995 GOTO(out_unlock_tgtdir, rc = PTR_ERR(mtgt));
997 mdt_reint_init_ma(info, ma);
998 if (!ma->ma_lmm || !ma->ma_cookie)
999 GOTO(out_unlock_tgt, rc = -EINVAL);
1001 rc = mdo_rename_tgt(info->mti_env, mdt_object_child(mtgtdir),
1002 mdt_object_child(mtgt), rr->rr_fid2,
1004 } else /* -ENOENT */ {
1005 rc = mdo_name_insert(info->mti_env, mdt_object_child(mtgtdir),
1006 lname, rr->rr_fid2, ma);
1009 /* handle last link of tgt object */
1010 if (rc == 0 && mtgt)
1011 mdt_handle_last_unlink(info, mtgt, ma);
1016 mdt_object_unlock_put(info, mtgt, lh_tgt, rc);
1018 mdt_object_unlock_put(info, mtgtdir, lh_tgtdir, rc);
1022 static int mdt_rename_lock(struct mdt_thread_info *info,
1023 struct lustre_handle *lh)
1025 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1026 ldlm_policy_data_t *policy = &info->mti_policy;
1027 struct ldlm_res_id *res_id = &info->mti_res_id;
1032 ms = mdt_md_site(info->mti_mdt);
1033 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1035 memset(policy, 0, sizeof *policy);
1036 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1038 if (ms->ms_control_exp == NULL) {
1039 int flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1042 * Current node is controller, that is mdt0, where we should
1045 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1046 LCK_EX, &flags, ldlm_blocking_ast,
1047 ldlm_completion_ast, NULL, NULL, 0,
1048 &info->mti_exp->exp_handle.h_cookie,
1051 struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_EX,
1052 ldlm_blocking_ast, ldlm_completion_ast, NULL, NULL, NULL };
1056 * This is the case mdt0 is remote node, issue DLM lock like
1059 rc = ldlm_cli_enqueue(ms->ms_control_exp, NULL, &einfo, res_id,
1060 policy, &flags, NULL, 0, lh, 0);
1066 static void mdt_rename_unlock(struct lustre_handle *lh)
1069 LASSERT(lustre_handle_is_used(lh));
1070 ldlm_lock_decref(lh, LCK_EX);
1075 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1076 * target. Source should not be ancestor of target dir. May be other rename
1077 * checks can be moved here later.
1079 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1081 struct mdt_reint_record *rr = &info->mti_rr;
1082 struct lu_fid dst_fid = *rr->rr_fid2;
1083 struct mdt_object *dst;
1088 LASSERT(fid_is_sane(&dst_fid));
1089 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1091 rc = mdo_is_subdir(info->mti_env,
1092 mdt_object_child(dst), fid,
1094 mdt_object_put(info->mti_env, dst);
1095 if (rc != -EREMOTE && rc < 0) {
1096 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1098 /* check the found fid */
1099 if (lu_fid_eq(&dst_fid, fid))
1105 } while (rc == -EREMOTE);
1111 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1112 * 2 - src child; 3 - tgt child.
1113 * Update on disk version of src child.
1115 static int mdt_reint_rename(struct mdt_thread_info *info,
1116 struct mdt_lock_handle *lhc)
1118 struct mdt_reint_record *rr = &info->mti_rr;
1119 struct md_attr *ma = &info->mti_attr;
1120 struct ptlrpc_request *req = mdt_info_req(info);
1121 struct mdt_object *msrcdir;
1122 struct mdt_object *mtgtdir;
1123 struct mdt_object *mold;
1124 struct mdt_object *mnew = NULL;
1125 struct mdt_lock_handle *lh_srcdirp;
1126 struct mdt_lock_handle *lh_tgtdirp;
1127 struct mdt_lock_handle *lh_oldp;
1128 struct mdt_lock_handle *lh_newp;
1129 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1130 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1131 struct lustre_handle rename_lh = { 0 };
1132 struct lu_name slname = { 0 };
1133 struct lu_name *lname;
1137 if (info->mti_dlm_req)
1138 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1140 if (info->mti_cross_ref) {
1141 rc = mdt_reint_rename_tgt(info);
1145 DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s",
1146 PFID(rr->rr_fid1), rr->rr_name,
1147 PFID(rr->rr_fid2), rr->rr_tgt);
1149 rc = mdt_rename_lock(info, &rename_lh);
1151 CERROR("Can't lock FS for rename, rc %d\n", rc);
1155 lh_newp = &info->mti_lh[MDT_LH_NEW];
1157 /* step 1: lock the source dir. */
1158 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1159 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name,
1161 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1162 MDS_INODELOCK_UPDATE);
1163 if (IS_ERR(msrcdir))
1164 GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
1166 if (mdt_object_obf(msrcdir))
1167 GOTO(out_unlock_source, rc = -EPERM);
1169 rc = mdt_version_get_check_save(info, msrcdir, 0);
1171 GOTO(out_unlock_source, rc);
1173 /* step 2: find & lock the target dir. */
1174 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1175 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt,
1177 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1178 mdt_object_get(info->mti_env, msrcdir);
1180 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1181 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1182 MDS_INODELOCK_UPDATE);
1184 GOTO(out_unlock_source, rc);
1185 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1188 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1190 if (IS_ERR(mtgtdir))
1191 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1193 if (mdt_object_obf(mtgtdir))
1194 GOTO(out_put_target, rc = -EPERM);
1196 /* check early, the real version will be saved after locking */
1197 rc = mdt_version_get_check(info, mtgtdir, 1);
1199 GOTO(out_put_target, rc);
1201 rc = mdt_object_exists(mtgtdir);
1203 GOTO(out_put_target, rc = -ESTALE);
1204 } else if (rc > 0) {
1205 /* we lock the target dir if it is local */
1206 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1207 MDS_INODELOCK_UPDATE,
1210 GOTO(out_put_target, rc);
1211 /* get and save correct version after locking */
1212 mdt_version_get_save(info, mtgtdir, 1);
1216 /* step 3: find & lock the old object. */
1217 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1218 mdt_name_copy(&slname, lname);
1220 rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2);
1222 GOTO(out_unlock_target, rc);
1224 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1225 GOTO(out_unlock_target, rc = -EINVAL);
1227 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1229 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1231 if (mdt_object_obf(mold)) {
1232 mdt_object_put(info->mti_env, mold);
1233 GOTO(out_unlock_target, rc = -EPERM);
1236 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1237 mdt_lock_reg_init(lh_oldp, LCK_EX);
1238 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
1241 mdt_object_put(info->mti_env, mold);
1242 GOTO(out_unlock_target, rc);
1245 info->mti_mos = mold;
1246 /* save version after locking */
1247 mdt_version_get_save(info, mold, 2);
1248 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1250 /* step 4: find & lock the new object. */
1251 /* new target object may not exist now */
1252 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
1253 /* lookup with version checking */
1255 rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3);
1257 /* the new_fid should have been filled at this moment */
1258 if (lu_fid_eq(old_fid, new_fid))
1259 GOTO(out_unlock_old, rc);
1261 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1262 lu_fid_eq(new_fid, rr->rr_fid2))
1263 GOTO(out_unlock_old, rc = -EINVAL);
1265 mdt_lock_reg_init(lh_newp, LCK_EX);
1266 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1268 GOTO(out_unlock_old, rc = PTR_ERR(mnew));
1270 if (mdt_object_obf(mnew)) {
1271 mdt_object_put(info->mti_env, mnew);
1272 GOTO(out_unlock_old, rc = -EPERM);
1275 rc = mdt_object_lock(info, mnew, lh_newp,
1276 MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
1278 mdt_object_put(info->mti_env, mnew);
1279 GOTO(out_unlock_old, rc);
1281 /* get and save version after locking */
1282 mdt_version_get_save(info, mnew, 3);
1283 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1284 } else if (rc != -EREMOTE && rc != -ENOENT) {
1285 GOTO(out_unlock_old, rc);
1287 mdt_enoent_version_save(info, 3);
1290 /* step 5: rename it */
1291 mdt_reint_init_ma(info, ma);
1292 if (!ma->ma_lmm || !ma->ma_cookie)
1293 GOTO(out_unlock_new, rc = -EINVAL);
1295 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1296 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1299 /* Check if @dst is subdir of @src. */
1300 rc = mdt_rename_sanity(info, old_fid);
1302 GOTO(out_unlock_new, rc);
1304 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1305 mdt_object_child(mtgtdir), old_fid, &slname,
1306 (mnew ? mdt_object_child(mnew) : NULL),
1309 /* handle last link of tgt object */
1311 mdt_counter_incr(req, LPROC_MDT_RENAME);
1313 mdt_handle_last_unlink(info, mnew, ma);
1315 mdt_rename_counter_tally(info, info->mti_mdt, req,
1322 mdt_object_unlock_put(info, mnew, lh_newp, rc);
1324 mdt_object_unlock_put(info, mold, lh_oldp, rc);
1326 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1328 mdt_object_put(info->mti_env, mtgtdir);
1330 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1332 mdt_rename_unlock(&rename_lh);
1336 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1337 struct mdt_lock_handle *lhc);
1339 static mdt_reinter reinters[REINT_MAX] = {
1340 [REINT_SETATTR] = mdt_reint_setattr,
1341 [REINT_CREATE] = mdt_reint_create,
1342 [REINT_LINK] = mdt_reint_link,
1343 [REINT_UNLINK] = mdt_reint_unlink,
1344 [REINT_RENAME] = mdt_reint_rename,
1345 [REINT_OPEN] = mdt_reint_open,
1346 [REINT_SETXATTR] = mdt_reint_setxattr
1349 int mdt_reint_rec(struct mdt_thread_info *info,
1350 struct mdt_lock_handle *lhc)
1355 rc = reinters[info->mti_rr.rr_opcode](info, lhc);