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, 2013, Intel Corporation.
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"
50 #include <lustre_lmv.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_lut.lut_mds_capa &&
71 exp_connect_flags(info->mti_exp) & 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 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
96 !fid_is_obf(mdt_object_fid(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 spin_lock(&req->rq_export->exp_lock);
125 req->rq_export->exp_vbr_failed = 1;
126 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 spin_lock(&req->rq_export->exp_lock);
132 req->rq_export->exp_vbr_failed = 1;
133 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, const 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;
268 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
270 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
272 if (!fid_is_md_operative(rr->rr_fid1))
275 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
277 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
279 RETURN(PTR_ERR(parent));
281 if (!mdt_object_exists(parent))
282 GOTO(put_parent, rc = -ENOENT);
284 lh = &info->mti_lh[MDT_LH_PARENT];
285 if (mdt_object_remote(parent)) {
286 mdt_lock_reg_init(lh, LCK_EX);
287 rc = mdt_remote_object_lock(info, parent, &lh->mlh_rreg_lh,
289 MDS_INODELOCK_UPDATE);
291 GOTO(put_parent, rc);
294 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
295 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE,
298 GOTO(put_parent, rc);
300 rc = mdt_version_get_check_save(info, parent, 0);
302 GOTO(unlock_parent, rc);
306 * Check child name version during replay.
307 * During create replay a file may exist with same name.
309 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
310 &info->mti_tmp_fid1, 1);
312 GOTO(unlock_parent, rc = -EEXIST);
314 /* -ENOENT is expected here */
316 GOTO(unlock_parent, rc);
318 /* save version of file name for replay, it must be ENOENT here */
319 mdt_enoent_version_save(info, 1);
321 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
322 if (likely(!IS_ERR(child))) {
323 struct md_object *next = mdt_object_child(parent);
325 if (mdt_object_remote(child)) {
326 struct seq_server_site *ss;
327 struct lu_ucred *uc = mdt_ucred(info);
329 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
331 mdt->mdt_enable_remote_dir_gid &&
332 mdt->mdt_enable_remote_dir_gid != -1) {
333 CERROR("%s: Creating remote dir is only"
334 " permitted for administrator or"
335 " set mdt_enable_remote_dir_gid:"
337 mdt_obd_name(mdt), -EPERM);
338 GOTO(out_put_child, rc = -EPERM);
342 ss = mdt_seq_site(mdt);
343 if (ss->ss_node_id != 0 &&
344 mdt->mdt_enable_remote_dir == 0) {
345 CERROR("%s: remote dir is only permitted on"
347 " mdt.*.enable_remote_dir=1\n",
349 GOTO(out_put_child, rc = -EPERM);
351 if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
352 /* Return -EIO for old client */
353 GOTO(out_put_child, rc = -EIO);
357 ma->ma_need = MA_INODE;
359 /* capa for cross-ref will be stored here */
360 ma->ma_capa = req_capsule_server_get(info->mti_pill,
362 LASSERT(ma->ma_capa);
364 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
365 OBD_FAIL_MDS_REINT_CREATE_WRITE);
367 /* Version of child will be updated on disk. */
368 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
369 rc = mdt_version_get_check_save(info, child, 2);
371 GOTO(out_put_child, rc);
373 /* Let lower layer know current lock mode. */
374 info->mti_spec.sp_cr_mode =
375 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
378 * Do not perform lookup sanity check. We know that name does
381 info->mti_spec.sp_cr_lookup = 0;
382 info->mti_spec.sp_feat = &dt_directory_features;
384 rc = mdo_create(info->mti_env, next, &rr->rr_name,
385 mdt_object_child(child), &info->mti_spec, ma);
387 rc = mdt_attr_get_complex(info, child, ma);
390 /* Return fid & attr to client. */
391 if (ma->ma_valid & MA_INODE)
392 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
393 mdt_object_fid(child));
396 mdt_object_put(info->mti_env, child);
400 mdt_create_pack_capa(info, rc, child, repbody);
402 mdt_object_unlock(info, parent, lh, rc);
404 mdt_object_put(info->mti_env, parent);
408 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
409 struct mdt_object *obj, __u64 ibits,
410 struct ldlm_enqueue_info *einfo)
412 ldlm_policy_data_t *policy = &mti->mti_policy;
416 if (!S_ISDIR(obj->mot_header.loh_attr))
419 memset(policy, 0, sizeof(*policy));
420 policy->l_inodebits.bits = ibits;
422 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
428 * Lock slave stripes if necessary, the lock handles of slave stripes
429 * will be stored in einfo->ei_cbdata.
431 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
432 ldlm_mode_t mode, __u64 ibits,
433 struct ldlm_enqueue_info *einfo)
435 ldlm_policy_data_t *policy = &mti->mti_policy;
439 if (!S_ISDIR(obj->mot_header.loh_attr))
442 memset(einfo, 0, sizeof(*einfo));
443 einfo->ei_type = LDLM_IBITS;
444 einfo->ei_mode = mode;
445 einfo->ei_cb_bl = mdt_remote_blocking_ast;
446 einfo->ei_cb_cp = ldlm_completion_ast;
447 einfo->ei_enq_slave = 1;
448 memset(policy, 0, sizeof(*policy));
449 policy->l_inodebits.bits = ibits;
451 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
456 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
457 struct md_attr *ma, int flags)
459 struct mdt_lock_handle *lh;
460 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
461 __u64 lockpart = MDS_INODELOCK_UPDATE;
462 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
466 /* attr shouldn't be set on remote object */
467 LASSERT(!mdt_object_remote(mo));
469 lh = &info->mti_lh[MDT_LH_PARENT];
470 mdt_lock_reg_init(lh, LCK_PW);
472 /* Even though the new MDT will grant PERM lock to the old
473 * client, but the old client will almost ignore that during
474 * So it needs to revoke both LOOKUP and PERM lock here, so
475 * both new and old client can cancel the dcache */
476 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
477 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
479 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
483 rc = mdt_lock_slaves(info, mo, LCK_EX, lockpart, einfo);
485 GOTO(out_unlock, rc);
487 if (mdt_object_exists(mo) == 0)
488 GOTO(out_unlock, rc = -ENOENT);
490 /* all attrs are packed into mti_attr in unpack_setattr */
491 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
492 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
494 /* This is only for set ctime when rename's source is on remote MDS. */
495 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
496 ma->ma_attr_flags |= MDS_VTX_BYPASS;
498 /* VBR: update version if attr changed are important for recovery */
500 /* update on-disk version of changed object */
501 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
502 rc = mdt_version_get_check_save(info, mo, 0);
504 GOTO(out_unlock, rc);
507 /* Ensure constant striping during chown(). See LU-2789. */
508 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
509 mutex_lock(&mo->mot_lov_mutex);
511 /* all attrs are packed into mti_attr in unpack_setattr */
512 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
514 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
515 mutex_unlock(&mo->mot_lov_mutex);
518 GOTO(out_unlock, rc);
522 mdt_unlock_slaves(info, mo, lockpart, einfo);
523 mdt_object_unlock(info, mo, lh, rc);
528 * Check HSM flags and add HS_DIRTY flag if relevant.
530 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
531 * and is not RELEASED.
533 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
539 /* If the file was modified, add the dirty flag */
540 ma->ma_need = MA_HSM;
541 rc = mdt_attr_get_complex(info, mo, ma);
543 CERROR("file attribute read error for "DFID": %d.\n",
544 PFID(mdt_object_fid(mo)), rc);
548 /* If an up2date copy exists in the backend, add dirty flag */
549 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
550 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
551 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
553 ma->ma_hsm.mh_flags |= HS_DIRTY;
555 mdt_lock_reg_init(lh, LCK_PW);
556 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
561 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
563 CERROR("file attribute change error for "DFID": %d\n",
564 PFID(mdt_object_fid(mo)), rc);
565 mdt_object_unlock(info, mo, lh, rc);
571 static int mdt_reint_setattr(struct mdt_thread_info *info,
572 struct mdt_lock_handle *lhc)
574 struct md_attr *ma = &info->mti_attr;
575 struct mdt_reint_record *rr = &info->mti_rr;
576 struct ptlrpc_request *req = mdt_info_req(info);
577 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
578 struct mdt_file_data *mfd;
579 struct mdt_object *mo;
580 struct mdt_body *repbody;
584 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
585 (unsigned int)ma->ma_attr.la_valid);
587 if (info->mti_dlm_req)
588 ldlm_request_cancel(req, info->mti_dlm_req, 0);
590 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
591 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
593 GOTO(out, rc = PTR_ERR(mo));
595 /* start a log jounal handle if needed */
596 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
597 if ((ma->ma_attr.la_valid & LA_SIZE) ||
598 (rr->rr_flags & MRF_OPEN_TRUNC)) {
599 /* Check write access for the O_TRUNC case */
600 if (mdt_write_read(mo) < 0)
601 GOTO(out_put, rc = -ETXTBSY);
603 } else if (info->mti_ioepoch &&
604 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
605 /* Truncate case. IOEpoch is opened. */
606 rc = mdt_write_get(mo);
610 mfd = mdt_mfd_new(med);
613 GOTO(out_put, rc = -ENOMEM);
616 mdt_ioepoch_open(info, mo, 0);
617 repbody->ioepoch = mo->mot_ioepoch;
619 mdt_object_get(info->mti_env, mo);
620 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
621 mfd->mfd_object = mo;
622 mfd->mfd_xid = req->rq_xid;
624 spin_lock(&med->med_open_lock);
625 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
626 spin_unlock(&med->med_open_lock);
627 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
630 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
632 /* SOM Attribute update case. Find the proper mfd and update
633 * SOM attributes on the proper object. */
634 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
635 LASSERT(info->mti_ioepoch);
637 spin_lock(&med->med_open_lock);
638 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
641 spin_unlock(&med->med_open_lock);
642 CDEBUG(D_INODE, "no handle for file close: "
643 "fid = "DFID": cookie = "LPX64"\n",
644 PFID(info->mti_rr.rr_fid1),
645 info->mti_ioepoch->handle.cookie);
646 GOTO(out_put, rc = -ESTALE);
648 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
649 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
651 class_handle_unhash(&mfd->mfd_handle);
652 cfs_list_del_init(&mfd->mfd_list);
653 spin_unlock(&med->med_open_lock);
655 mdt_mfd_close(info, mfd);
656 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
657 LASSERT((ma->ma_valid & MA_LOV) == 0);
658 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
661 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
662 struct lu_buf *buf = &info->mti_buf;
663 LASSERT(ma->ma_attr.la_valid == 0);
664 buf->lb_buf = ma->ma_lmm;
665 buf->lb_len = ma->ma_lmm_size;
666 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
667 buf, XATTR_NAME_LOV, 0);
670 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
671 struct lu_buf *buf = &info->mti_buf;
673 LASSERT(ma->ma_attr.la_valid == 0);
674 buf->lb_buf = ma->ma_lmv;
675 buf->lb_len = ma->ma_lmv_size;
676 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
677 buf, XATTR_NAME_DEFAULT_LMV, 0);
683 /* If file data is modified, add the dirty flag */
684 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
685 rc = mdt_add_dirty_flag(info, mo, ma);
687 ma->ma_need = MA_INODE;
689 rc = mdt_attr_get_complex(info, mo, ma);
693 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
695 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
696 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
697 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
698 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
699 struct lustre_capa *capa;
701 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
703 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
704 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
707 repbody->valid |= OBD_MD_FLOSSCAPA;
712 mdt_object_put(info->mti_env, mo);
715 mdt_counter_incr(req, LPROC_MDT_SETATTR);
717 mdt_client_compatibility(info);
718 rc2 = mdt_fix_reply(info);
724 static int mdt_reint_create(struct mdt_thread_info *info,
725 struct mdt_lock_handle *lhc)
727 struct ptlrpc_request *req = mdt_info_req(info);
731 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
732 RETURN(err_serious(-ESTALE));
734 if (info->mti_dlm_req)
735 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
737 if (!lu_name_is_valid(&info->mti_rr.rr_name))
740 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
742 mdt_counter_incr(req, LPROC_MDT_MKDIR);
750 /* Special file should stay on the same node as parent. */
751 mdt_counter_incr(req, LPROC_MDT_MKNOD);
754 CERROR("%s: Unsupported mode %o\n",
755 mdt_obd_name(info->mti_mdt),
756 info->mti_attr.ma_attr.la_mode);
757 RETURN(err_serious(-EOPNOTSUPP));
760 rc = mdt_md_create(info);
765 * VBR: save parent version in reply and child version getting by its name.
766 * Version of child is getting and checking during its lookup. If
768 static int mdt_reint_unlink(struct mdt_thread_info *info,
769 struct mdt_lock_handle *lhc)
771 struct mdt_reint_record *rr = &info->mti_rr;
772 struct ptlrpc_request *req = mdt_info_req(info);
773 struct md_attr *ma = &info->mti_attr;
774 struct lu_fid *child_fid = &info->mti_tmp_fid1;
775 struct mdt_object *mp;
776 struct mdt_object *mc;
777 struct mdt_lock_handle *parent_lh;
778 struct mdt_lock_handle *child_lh;
779 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
784 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
785 PNAME(&rr->rr_name));
787 if (info->mti_dlm_req)
788 ldlm_request_cancel(req, info->mti_dlm_req, 0);
790 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
791 RETURN(err_serious(-ENOENT));
793 if (!fid_is_md_operative(rr->rr_fid1))
797 * step 1: Found the parent.
799 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
805 parent_lh = &info->mti_lh[MDT_LH_PARENT];
807 if (mdt_object_remote(mp)) {
808 mdt_lock_reg_init(parent_lh, LCK_EX);
809 rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
810 parent_lh->mlh_rreg_mode,
811 MDS_INODELOCK_UPDATE);
813 GOTO(put_parent, rc);
816 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
817 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
820 GOTO(put_parent, rc);
822 rc = mdt_version_get_check_save(info, mp, 0);
824 GOTO(unlock_parent, rc);
827 /* step 2: find & lock the child */
828 /* lookup child object along with version checking */
830 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
832 /* Name might not be able to find during resend of
833 * remote unlink, considering following case.
834 * dir_A is a remote directory, the name entry of
835 * dir_A is on MDT0, the directory is on MDT1,
837 * 1. client sends unlink req to MDT1.
838 * 2. MDT1 sends name delete update to MDT0.
839 * 3. name entry is being deleted in MDT0 synchronously.
840 * 4. MDT1 is restarted.
841 * 5. client resends unlink req to MDT1. So it can not
842 * find the name entry on MDT0 anymore.
843 * In this case, MDT1 only needs to destory the local
846 if (mdt_object_remote(mp) && rc == -ENOENT &&
847 !fid_is_zero(rr->rr_fid2) &&
848 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
850 *child_fid = *rr->rr_fid2;
852 GOTO(unlock_parent, rc);
856 if (!fid_is_md_operative(child_fid))
857 GOTO(unlock_parent, rc = -EPERM);
859 /* We will lock the child regardless it is local or remote. No harm. */
860 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
862 GOTO(unlock_parent, rc = PTR_ERR(mc));
864 child_lh = &info->mti_lh[MDT_LH_CHILD];
865 mdt_lock_reg_init(child_lh, LCK_EX);
866 if (mdt_object_remote(mc)) {
867 struct mdt_body *repbody;
869 if (!fid_is_zero(rr->rr_fid2)) {
870 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
871 mdt_obd_name(info->mti_mdt),
872 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
873 GOTO(put_child, rc = -ENOENT);
875 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
876 mdt_obd_name(info->mti_mdt),
877 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
879 if (!mdt_is_dne_client(req->rq_export))
880 /* Return -EIO for old client */
881 GOTO(put_child, rc = -EIO);
883 if (info->mti_spec.sp_rm_entry) {
884 struct lu_ucred *uc = mdt_ucred(info);
886 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
887 CERROR("%s: unlink remote entry is only "
888 "permitted for administrator: rc = %d\n",
889 mdt_obd_name(info->mti_mdt),
891 GOTO(put_child, rc = -EPERM);
894 ma->ma_need = MA_INODE;
896 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
897 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
898 NULL, &rr->rr_name, ma, no_name);
901 /* Revoke the LOOKUP lock of the remote object granted by
902 * this MDT. Since the unlink will happen on another MDT,
903 * it will release the LOOKUP lock right away. Then What
904 * would happen if another client try to grab the LOOKUP
905 * lock at the same time with unlink XXX */
906 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
908 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
909 LASSERT(repbody != NULL);
910 repbody->fid1 = *mdt_object_fid(mc);
911 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
912 GOTO(unlock_child, rc = -EREMOTE);
913 } else if (info->mti_spec.sp_rm_entry) {
915 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
917 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
921 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
922 * this now because a running HSM restore on the child (unlink
923 * victim) will hold the layout lock. See LU-4002. */
924 rc = mdt_object_lock(info, mc, child_lh,
925 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
930 * Now we can only make sure we need MA_INODE, in mdd layer, will check
931 * whether need MA_LOV and MA_COOKIE.
933 ma->ma_need = MA_INODE;
936 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, einfo);
938 GOTO(unlock_child, rc);
940 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
941 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
942 /* save version when object is locked */
943 mdt_version_get_save(info, mc, 1);
945 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
947 mutex_lock(&mc->mot_lov_mutex);
949 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
950 mdt_object_child(mc), &rr->rr_name, ma, no_name);
952 mutex_unlock(&mc->mot_lov_mutex);
954 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
955 rc = mdt_attr_get_complex(info, mc, ma);
957 mdt_handle_last_unlink(info, mc, ma);
959 if (ma->ma_valid & MA_INODE) {
960 switch (ma->ma_attr.la_mode & S_IFMT) {
962 mdt_counter_incr(req, LPROC_MDT_RMDIR);
970 mdt_counter_incr(req, LPROC_MDT_UNLINK);
973 LASSERTF(0, "bad file type %o unlinking\n",
974 ma->ma_attr.la_mode);
981 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, einfo);
982 mdt_object_unlock(info, mc, child_lh, rc);
984 /* Since we do not need reply md striped dir info to client, so
985 * reset mti_big_lmm_used to avoid confusing mdt_fix_reply */
986 if (info->mti_big_lmm_used)
987 info->mti_big_lmm_used = 0;
989 mdt_object_put(info->mti_env, mc);
991 mdt_object_unlock(info, mp, parent_lh, rc);
993 mdt_object_put(info->mti_env, mp);
999 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1002 static int mdt_reint_link(struct mdt_thread_info *info,
1003 struct mdt_lock_handle *lhc)
1005 struct mdt_reint_record *rr = &info->mti_rr;
1006 struct ptlrpc_request *req = mdt_info_req(info);
1007 struct md_attr *ma = &info->mti_attr;
1008 struct mdt_object *ms;
1009 struct mdt_object *mp;
1010 struct mdt_lock_handle *lhs;
1011 struct mdt_lock_handle *lhp;
1015 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1016 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1018 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1019 RETURN(err_serious(-ENOENT));
1021 if (info->mti_dlm_req)
1022 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1024 /* Invalid case so return error immediately instead of
1026 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1029 if (!fid_is_md_operative(rr->rr_fid1) ||
1030 !fid_is_md_operative(rr->rr_fid2))
1033 /* step 1: find & lock the target parent dir */
1034 lhp = &info->mti_lh[MDT_LH_PARENT];
1035 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1036 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1037 MDS_INODELOCK_UPDATE);
1039 RETURN(PTR_ERR(mp));
1041 rc = mdt_version_get_check_save(info, mp, 0);
1043 GOTO(out_unlock_parent, rc);
1045 /* step 2: find & lock the source */
1046 lhs = &info->mti_lh[MDT_LH_CHILD];
1047 mdt_lock_reg_init(lhs, LCK_EX);
1049 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1051 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1053 if (!mdt_object_exists(ms)) {
1054 mdt_object_put(info->mti_env, ms);
1055 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1056 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1057 GOTO(out_unlock_parent, rc = -ENOENT);
1060 if (mdt_object_remote(ms)) {
1061 mdt_object_put(info->mti_env, ms);
1062 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1063 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1065 GOTO(out_unlock_parent, rc = -EXDEV);
1068 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1069 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1071 mdt_object_put(info->mti_env, ms);
1072 GOTO(out_unlock_parent, rc);
1075 /* step 3: link it */
1076 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1077 OBD_FAIL_MDS_REINT_LINK_WRITE);
1079 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1080 rc = mdt_version_get_check_save(info, ms, 1);
1082 GOTO(out_unlock_child, rc);
1084 /** check target version by name during replay */
1085 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1086 &info->mti_tmp_fid1, 2);
1087 if (rc != 0 && rc != -ENOENT)
1088 GOTO(out_unlock_child, rc);
1089 /* save version of file name for replay, it must be ENOENT here */
1090 if (!req_is_replay(mdt_info_req(info))) {
1091 if (rc != -ENOENT) {
1092 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1093 PNAME(&rr->rr_name));
1094 GOTO(out_unlock_child, rc = -EEXIST);
1096 info->mti_ver[2] = ENOENT_VERSION;
1097 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1100 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1101 mdt_object_child(ms), &rr->rr_name, ma);
1104 mdt_counter_incr(req, LPROC_MDT_LINK);
1108 mdt_object_unlock_put(info, ms, lhs, rc);
1110 mdt_object_unlock_put(info, mp, lhp, rc);
1114 * lock the part of the directory according to the hash of the name
1115 * (lh->mlh_pdo_hash) in parallel directory lock.
1117 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1118 struct mdt_lock_handle *lh,
1119 struct mdt_object *obj, __u64 ibits)
1121 struct ldlm_res_id *res = &info->mti_res_id;
1122 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1123 ldlm_policy_data_t *policy = &info->mti_policy;
1127 * Finish res_id initializing by name hash marking part of
1128 * directory which is taking modification.
1130 LASSERT(lh->mlh_pdo_hash != 0);
1131 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1132 memset(policy, 0, sizeof(*policy));
1133 policy->l_inodebits.bits = ibits;
1135 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1136 * going to be sent to client. If it is - mdt_intent_policy() path will
1137 * fix it up and turn FL_LOCAL flag off.
1139 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1140 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1141 &info->mti_exp->exp_handle.h_cookie);
1145 static int mdt_rename_lock(struct mdt_thread_info *info,
1146 struct lustre_handle *lh)
1148 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1149 ldlm_policy_data_t *policy = &info->mti_policy;
1150 struct ldlm_res_id *res_id = &info->mti_res_id;
1155 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1157 memset(policy, 0, sizeof *policy);
1158 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1159 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 53, 0)
1160 /* In phase I, we will not do cross-rename, so local BFL lock would
1163 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1165 * Current node is controller, that is mdt0, where we should
1168 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1169 LCK_EX, &flags, ldlm_blocking_ast,
1170 ldlm_completion_ast, NULL, NULL, 0,
1172 &info->mti_exp->exp_handle.h_cookie,
1175 #warning "Local rename lock is invalid for DNE phase II."
1180 static void mdt_rename_unlock(struct lustre_handle *lh)
1183 LASSERT(lustre_handle_is_used(lh));
1184 ldlm_lock_decref(lh, LCK_EX);
1189 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1190 * target. Source should not be ancestor of target dir. May be other rename
1191 * checks can be moved here later.
1193 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1195 struct mdt_reint_record *rr = &info->mti_rr;
1196 struct lu_fid dst_fid = *rr->rr_fid2;
1197 struct mdt_object *dst;
1201 /* If the source and target are in the same directory, they can not
1202 * be parent/child relationship, so subdir check is not needed */
1203 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1207 LASSERT(fid_is_sane(&dst_fid));
1208 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1210 rc = mdo_is_subdir(info->mti_env,
1211 mdt_object_child(dst), fid,
1213 mdt_object_put(info->mti_env, dst);
1214 if (rc != -EREMOTE && rc < 0) {
1215 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1217 /* check the found fid */
1218 if (lu_fid_eq(&dst_fid, fid))
1224 } while (rc == -EREMOTE);
1229 /* Update object linkEA */
1230 struct mdt_lock_list {
1231 struct mdt_object *mll_obj;
1232 struct mdt_lock_handle mll_lh;
1233 struct list_head mll_list;
1236 static void mdt_unlock_list(struct mdt_thread_info *info,
1237 struct list_head *list, int rc)
1239 struct mdt_lock_list *mll;
1240 struct mdt_lock_list *mll2;
1242 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1243 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1244 list_del(&mll->mll_list);
1249 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1250 struct mdt_object *obj,
1251 struct mdt_object *pobj,
1252 struct list_head *lock_list)
1254 struct lu_buf *buf = &info->mti_big_buf;
1255 struct linkea_data ldata = { 0 };
1260 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1263 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1264 if (buf->lb_buf == NULL)
1268 rc = mdt_links_read(info, obj, &ldata);
1270 if (rc == -ENOENT || rc == -ENODATA)
1275 LASSERT(ldata.ld_leh != NULL);
1276 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1277 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1278 struct mdt_device *mdt = info->mti_mdt;
1279 struct mdt_object *mdt_pobj;
1280 struct mdt_lock_list *mll;
1281 struct lu_name name;
1284 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1286 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1288 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1289 if (IS_ERR(mdt_pobj)) {
1290 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1291 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1295 if (!mdt_object_exists(mdt_pobj)) {
1296 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1297 mdt_obd_name(mdt), PFID(&fid));
1298 mdt_object_put(info->mti_env, mdt_pobj);
1302 if (mdt_pobj == pobj) {
1303 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1304 mdt_obd_name(mdt), PFID(&fid));
1305 mdt_object_put(info->mti_env, mdt_pobj);
1311 mdt_object_put(info->mti_env, mdt_pobj);
1312 GOTO(out, rc = -ENOMEM);
1315 if (mdt_object_remote(mdt_pobj)) {
1316 mdt_lock_reg_init(&mll->mll_lh, LCK_EX);
1317 rc = mdt_remote_object_lock(info, mdt_pobj,
1318 &mll->mll_lh.mlh_rreg_lh,
1319 mll->mll_lh.mlh_rreg_mode,
1320 MDS_INODELOCK_UPDATE);
1322 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1323 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1324 MDS_INODELOCK_UPDATE,
1328 CERROR("%s: cannot lock "DFID": rc =%d\n",
1329 mdt_obd_name(mdt), PFID(&fid), rc);
1330 mdt_object_put(info->mti_env, mdt_pobj);
1335 CFS_INIT_LIST_HEAD(&mll->mll_list);
1336 mll->mll_obj = mdt_pobj;
1337 list_add_tail(&mll->mll_list, lock_list);
1341 mdt_unlock_list(info, lock_list, rc);
1345 /* migrate files from one MDT to another MDT */
1346 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1347 struct mdt_lock_handle *lhc)
1349 struct mdt_reint_record *rr = &info->mti_rr;
1350 struct md_attr *ma = &info->mti_attr;
1351 struct mdt_object *msrcdir;
1352 struct mdt_object *mold;
1353 struct mdt_object *mnew = NULL;
1354 struct mdt_lock_handle *lh_dirp;
1355 struct mdt_lock_handle *lh_childp;
1356 struct mdt_lock_handle *lh_tgtp = NULL;
1357 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1358 struct list_head lock_list;
1362 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1363 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1364 /* 1: lock the source dir. */
1365 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1366 if (IS_ERR(msrcdir)) {
1367 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1368 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1369 (int)PTR_ERR(msrcdir));
1370 RETURN(PTR_ERR(msrcdir));
1373 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1374 if (mdt_object_remote(msrcdir)) {
1375 mdt_lock_reg_init(lh_dirp, LCK_EX);
1376 rc = mdt_remote_object_lock(info, msrcdir,
1377 &lh_dirp->mlh_rreg_lh,
1378 lh_dirp->mlh_rreg_mode,
1379 MDS_INODELOCK_UPDATE);
1381 GOTO(out_put_parent, rc);
1383 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1384 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1385 MDS_INODELOCK_UPDATE,
1388 GOTO(out_put_parent, rc);
1390 rc = mdt_version_get_check_save(info, msrcdir, 0);
1392 GOTO(out_unlock_parent, rc);
1395 /* 2: sanity check and find the object to be migrated. */
1397 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1399 GOTO(out_unlock_parent, rc);
1401 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1402 GOTO(out_unlock_parent, rc = -EINVAL);
1404 if (!fid_is_md_operative(old_fid))
1405 GOTO(out_unlock_parent, rc = -EPERM);
1407 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1409 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1411 if (mdt_object_remote(mold)) {
1412 CERROR("%s: source "DFID" is on the remote MDT\n",
1413 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1414 GOTO(out_put_child, rc = -EREMOTE);
1417 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1418 !mdt_object_remote(msrcdir)) {
1419 CERROR("%s: parent "DFID" is still on the same"
1420 " MDT, which should be migrated first:"
1421 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1422 PFID(mdt_object_fid(msrcdir)), -EPERM);
1423 GOTO(out_put_child, rc = -EPERM);
1426 /* 3: iterate the linkea of the object and lock all of the objects */
1427 CFS_INIT_LIST_HEAD(&lock_list);
1428 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1430 GOTO(out_put_child, rc);
1432 /* 4: lock of the object migrated object */
1433 lh_childp = &info->mti_lh[MDT_LH_OLD];
1434 mdt_lock_reg_init(lh_childp, LCK_EX);
1435 rc = mdt_object_lock(info, mold, lh_childp,
1436 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1437 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1439 GOTO(out_unlock_list, rc);
1441 ma->ma_need = MA_LMV;
1443 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1444 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1445 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1447 GOTO(out_unlock_list, rc);
1449 if ((ma->ma_valid & MA_LMV)) {
1450 struct lmv_mds_md_v1 *lmm1;
1452 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1453 lmm1 = &ma->ma_lmv->lmv_md_v1;
1454 if (lmm1->lmv_magic != LMV_MAGIC_MIGRATE) {
1455 CERROR("%s: can not migrate striped dir "DFID
1456 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1457 PFID(mdt_object_fid(mold)), -EPERM);
1458 GOTO(out_unlock_child, rc = -EPERM);
1461 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1462 GOTO(out_unlock_child, rc = -EINVAL);
1464 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1465 &lmm1->lmv_stripe_fids[1]);
1467 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1469 if (!mdt_object_remote(mnew)) {
1470 CERROR("%s: "DFID" being migrated is on this MDT:"
1471 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1472 PFID(rr->rr_fid2), -EPERM);
1473 GOTO(out_put_new, rc = -EPERM);
1476 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1477 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1478 rc = mdt_remote_object_lock(info, mnew,
1479 &lh_tgtp->mlh_rreg_lh,
1480 lh_tgtp->mlh_rreg_mode,
1481 MDS_INODELOCK_UPDATE);
1484 GOTO(out_put_new, rc);
1487 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1490 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1491 if (!mdt_object_remote(mnew)) {
1492 CERROR("%s: Migration "DFID" is on this MDT !\n",
1493 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid2));
1494 GOTO(out_put_new, rc = -EXDEV);
1499 mdt_reint_init_ma(info, ma);
1501 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1502 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1504 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1505 old_fid, &rr->rr_name, mdt_object_child(mnew), ma);
1507 GOTO(out_unlock_new, rc);
1509 if (lh_tgtp != NULL)
1510 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1513 mdt_object_put(info->mti_env, mnew);
1515 mdt_object_unlock(info, mold, lh_childp, rc);
1517 mdt_unlock_list(info, &lock_list, rc);
1519 mdt_object_put(info->mti_env, mold);
1521 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1523 mdt_object_put(info->mti_env, msrcdir);
1529 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1530 * 2 - src child; 3 - tgt child.
1531 * Update on disk version of src child.
1534 * For DNE phase I, only these renames are allowed
1535 * mv src_p/src_c tgt_p/tgt_c
1536 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1537 * 2. src_p and tgt_p are same directory, and tgt_c does not
1538 * exists. In this case, all of modification will happen
1539 * in the MDT where ithesource parent is, only one remote
1540 * update is needed, i.e. set c_time/m_time on the child.
1541 * And tgt_c will be still in the same MDT as the original
1544 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1545 struct mdt_lock_handle *lhc)
1547 struct mdt_reint_record *rr = &info->mti_rr;
1548 struct md_attr *ma = &info->mti_attr;
1549 struct ptlrpc_request *req = mdt_info_req(info);
1550 struct mdt_object *msrcdir;
1551 struct mdt_object *mtgtdir;
1552 struct mdt_object *mold;
1553 struct mdt_object *mnew = NULL;
1554 struct mdt_lock_handle *lh_srcdirp;
1555 struct mdt_lock_handle *lh_tgtdirp;
1556 struct mdt_lock_handle *lh_oldp = NULL;
1557 struct mdt_lock_handle *lh_newp = NULL;
1558 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1559 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1563 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1564 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1565 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1567 /* step 1: lock the source dir. */
1568 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1569 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1570 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1571 MDS_INODELOCK_UPDATE);
1572 if (IS_ERR(msrcdir))
1573 RETURN(PTR_ERR(msrcdir));
1575 rc = mdt_version_get_check_save(info, msrcdir, 0);
1577 GOTO(out_unlock_source, rc);
1579 /* step 2: find & lock the target dir. */
1580 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1581 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1582 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1583 mdt_object_get(info->mti_env, msrcdir);
1585 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1586 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1587 MDS_INODELOCK_UPDATE);
1589 GOTO(out_unlock_source, rc);
1590 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1593 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1595 if (IS_ERR(mtgtdir))
1596 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1598 /* check early, the real version will be saved after locking */
1599 rc = mdt_version_get_check(info, mtgtdir, 1);
1601 GOTO(out_put_target, rc);
1603 if (unlikely(mdt_object_remote(mtgtdir))) {
1604 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1605 "on different MDTs\n", PFID(rr->rr_fid1),
1607 GOTO(out_put_target, rc = -EXDEV);
1609 if (likely(mdt_object_exists(mtgtdir))) {
1610 /* we lock the target dir if it is local */
1611 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1612 MDS_INODELOCK_UPDATE,
1615 GOTO(out_put_target, rc);
1616 /* get and save correct version after locking */
1617 mdt_version_get_save(info, mtgtdir, 1);
1619 GOTO(out_put_target, rc = -ESTALE);
1624 /* step 3: find & lock the old object. */
1626 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1628 GOTO(out_unlock_target, rc);
1630 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1631 GOTO(out_unlock_target, rc = -EINVAL);
1633 if (!fid_is_md_operative(old_fid))
1634 GOTO(out_unlock_target, rc = -EPERM);
1636 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1638 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1640 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1641 /* save version after locking */
1642 mdt_version_get_save(info, mold, 2);
1643 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1645 /* step 4: find & lock the new object. */
1646 /* new target object may not exist now */
1647 /* lookup with version checking */
1649 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1652 /* the new_fid should have been filled at this moment */
1653 if (lu_fid_eq(old_fid, new_fid))
1654 GOTO(out_put_old, rc);
1656 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1657 lu_fid_eq(new_fid, rr->rr_fid2))
1658 GOTO(out_put_old, rc = -EINVAL);
1660 if (!fid_is_md_operative(new_fid))
1661 GOTO(out_put_old, rc = -EPERM);
1663 if (mdt_object_remote(mold)) {
1664 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1666 GOTO(out_put_old, rc = -EXDEV);
1669 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1671 GOTO(out_put_old, rc = PTR_ERR(mnew));
1673 if (mdt_object_remote(mnew)) {
1674 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1676 GOTO(out_put_new, rc = -EXDEV);
1679 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1680 mdt_lock_reg_init(lh_oldp, LCK_EX);
1681 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1682 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1684 GOTO(out_put_new, rc);
1686 /* We used to acquire MDS_INODELOCK_FULL here but we
1687 * can't do this now because a running HSM restore on
1688 * the rename onto victim will hold the layout
1689 * lock. See LU-4002. */
1691 lh_newp = &info->mti_lh[MDT_LH_NEW];
1692 mdt_lock_reg_init(lh_newp, LCK_EX);
1693 rc = mdt_object_lock(info, mnew, lh_newp,
1694 MDS_INODELOCK_LOOKUP |
1695 MDS_INODELOCK_UPDATE,
1698 GOTO(out_unlock_old, rc);
1700 /* get and save version after locking */
1701 mdt_version_get_save(info, mnew, 3);
1702 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1703 } else if (rc != -EREMOTE && rc != -ENOENT) {
1704 GOTO(out_put_old, rc);
1706 /* If mnew does not exist and mold are remote directory,
1707 * it only allows rename if they are under same directory */
1708 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1709 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1711 GOTO(out_put_old, rc = -EXDEV);
1714 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1715 mdt_lock_reg_init(lh_oldp, LCK_EX);
1716 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1717 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1719 GOTO(out_put_old, rc);
1721 mdt_enoent_version_save(info, 3);
1724 /* step 5: rename it */
1725 mdt_reint_init_ma(info, ma);
1727 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1728 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1730 /* Check if @dst is subdir of @src. */
1731 rc = mdt_rename_sanity(info, old_fid);
1733 GOTO(out_unlock_new, rc);
1736 mutex_lock(&mnew->mot_lov_mutex);
1738 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1739 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1740 mnew != NULL ? mdt_object_child(mnew) : NULL,
1741 &rr->rr_tgt_name, ma);
1744 mutex_unlock(&mnew->mot_lov_mutex);
1746 /* handle last link of tgt object */
1748 mdt_counter_incr(req, LPROC_MDT_RENAME);
1750 mdt_handle_last_unlink(info, mnew, ma);
1752 mdt_rename_counter_tally(info, info->mti_mdt, req,
1759 mdt_object_unlock(info, mnew, lh_newp, rc);
1761 mdt_object_unlock(info, mold, lh_oldp, rc);
1764 mdt_object_put(info->mti_env, mnew);
1766 mdt_object_put(info->mti_env, mold);
1768 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1770 mdt_object_put(info->mti_env, mtgtdir);
1772 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1776 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1777 struct mdt_lock_handle *lhc,
1780 struct mdt_reint_record *rr = &info->mti_rr;
1781 struct ptlrpc_request *req = mdt_info_req(info);
1782 struct lustre_handle rename_lh = { 0 };
1786 if (info->mti_dlm_req)
1787 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1789 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1790 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1793 rc = mdt_rename_lock(info, &rename_lh);
1795 CERROR("%s: can't lock FS for rename: rc = %d\n",
1796 mdt_obd_name(info->mti_mdt), rc);
1801 rc = mdt_reint_rename_internal(info, lhc);
1803 rc = mdt_reint_migrate_internal(info, lhc);
1805 if (lustre_handle_is_used(&rename_lh))
1806 mdt_rename_unlock(&rename_lh);
1811 static int mdt_reint_rename(struct mdt_thread_info *info,
1812 struct mdt_lock_handle *lhc)
1814 return mdt_reint_rename_or_migrate(info, lhc, true);
1817 static int mdt_reint_migrate(struct mdt_thread_info *info,
1818 struct mdt_lock_handle *lhc)
1820 return mdt_reint_rename_or_migrate(info, lhc, false);
1823 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1824 struct mdt_lock_handle *lhc);
1826 static mdt_reinter reinters[REINT_MAX] = {
1827 [REINT_SETATTR] = mdt_reint_setattr,
1828 [REINT_CREATE] = mdt_reint_create,
1829 [REINT_LINK] = mdt_reint_link,
1830 [REINT_UNLINK] = mdt_reint_unlink,
1831 [REINT_RENAME] = mdt_reint_rename,
1832 [REINT_OPEN] = mdt_reint_open,
1833 [REINT_SETXATTR] = mdt_reint_setxattr,
1834 [REINT_RMENTRY] = mdt_reint_unlink,
1835 [REINT_MIGRATE] = mdt_reint_migrate,
1838 int mdt_reint_rec(struct mdt_thread_info *info,
1839 struct mdt_lock_handle *lhc)
1844 rc = reinters[info->mti_rr.rr_opcode](info, lhc);