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_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
275 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
277 lh = &info->mti_lh[MDT_LH_PARENT];
278 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
280 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
282 RETURN(PTR_ERR(parent));
284 if (!mdt_object_exists(parent))
285 GOTO(put_parent, rc = -ENOENT);
287 lh = &info->mti_lh[MDT_LH_PARENT];
288 if (mdt_object_remote(parent)) {
289 mdt_lock_reg_init(lh, LCK_EX);
290 rc = mdt_remote_object_lock(info, parent, &lh->mlh_rreg_lh,
292 MDS_INODELOCK_UPDATE);
294 GOTO(put_parent, rc);
297 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
298 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE,
301 GOTO(put_parent, rc);
303 rc = mdt_version_get_check_save(info, parent, 0);
305 GOTO(unlock_parent, rc);
309 * Check child name version during replay.
310 * During create replay a file may exist with same name.
312 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
313 &info->mti_tmp_fid1, 1);
315 GOTO(unlock_parent, rc = -EEXIST);
317 /* -ENOENT is expected here */
319 GOTO(unlock_parent, rc);
321 /* save version of file name for replay, it must be ENOENT here */
322 mdt_enoent_version_save(info, 1);
324 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
325 if (likely(!IS_ERR(child))) {
326 struct md_object *next = mdt_object_child(parent);
328 if (mdt_object_remote(child)) {
329 struct seq_server_site *ss;
330 struct lu_ucred *uc = mdt_ucred(info);
332 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
334 mdt->mdt_enable_remote_dir_gid &&
335 mdt->mdt_enable_remote_dir_gid != -1) {
336 CERROR("%s: Creating remote dir is only"
337 " permitted for administrator or"
338 " set mdt_enable_remote_dir_gid:"
340 mdt_obd_name(mdt), -EPERM);
341 GOTO(out_put_child, rc = -EPERM);
345 ss = mdt_seq_site(mdt);
346 if (ss->ss_node_id != 0 &&
347 mdt->mdt_enable_remote_dir == 0) {
348 CERROR("%s: remote dir is only permitted on"
350 " mdt.*.enable_remote_dir=1\n",
352 GOTO(out_put_child, rc = -EPERM);
354 if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
355 /* Return -EIO for old client */
356 GOTO(out_put_child, rc = -EIO);
360 ma->ma_need = MA_INODE;
362 /* capa for cross-ref will be stored here */
363 ma->ma_capa = req_capsule_server_get(info->mti_pill,
365 LASSERT(ma->ma_capa);
367 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
368 OBD_FAIL_MDS_REINT_CREATE_WRITE);
370 /* Version of child will be updated on disk. */
371 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
372 rc = mdt_version_get_check_save(info, child, 2);
374 GOTO(out_put_child, rc);
376 /* Let lower layer know current lock mode. */
377 info->mti_spec.sp_cr_mode =
378 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
381 * Do not perform lookup sanity check. We know that name does
384 info->mti_spec.sp_cr_lookup = 0;
385 info->mti_spec.sp_feat = &dt_directory_features;
387 rc = mdo_create(info->mti_env, next, &rr->rr_name,
388 mdt_object_child(child), &info->mti_spec, ma);
390 rc = mdt_attr_get_complex(info, child, ma);
393 /* Return fid & attr to client. */
394 if (ma->ma_valid & MA_INODE)
395 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
396 mdt_object_fid(child));
399 mdt_object_put(info->mti_env, child);
403 mdt_create_pack_capa(info, rc, child, repbody);
405 mdt_object_unlock(info, parent, lh, rc);
407 mdt_object_put(info->mti_env, parent);
411 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
412 struct mdt_object *obj, __u64 ibits,
413 struct ldlm_enqueue_info *einfo)
415 ldlm_policy_data_t *policy = &mti->mti_policy;
419 if (!S_ISDIR(obj->mot_header.loh_attr))
422 memset(policy, 0, sizeof(*policy));
423 policy->l_inodebits.bits = ibits;
425 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
431 * Lock slave stripes if necessary, the lock handles of slave stripes
432 * will be stored in einfo->ei_cbdata.
434 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
435 ldlm_mode_t mode, __u64 ibits,
436 struct ldlm_enqueue_info *einfo)
438 ldlm_policy_data_t *policy = &mti->mti_policy;
442 if (!S_ISDIR(obj->mot_header.loh_attr))
445 memset(einfo, 0, sizeof(*einfo));
446 einfo->ei_type = LDLM_IBITS;
447 einfo->ei_mode = mode;
448 einfo->ei_cb_bl = mdt_remote_blocking_ast;
449 einfo->ei_cb_cp = ldlm_completion_ast;
450 einfo->ei_enq_slave = 1;
451 memset(policy, 0, sizeof(*policy));
452 policy->l_inodebits.bits = ibits;
454 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
459 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
460 struct md_attr *ma, int flags)
462 struct mdt_lock_handle *lh;
463 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
464 __u64 lockpart = MDS_INODELOCK_UPDATE;
465 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
469 /* attr shouldn't be set on remote object */
470 LASSERT(!mdt_object_remote(mo));
472 lh = &info->mti_lh[MDT_LH_PARENT];
473 mdt_lock_reg_init(lh, LCK_PW);
475 /* Even though the new MDT will grant PERM lock to the old
476 * client, but the old client will almost ignore that during
477 * So it needs to revoke both LOOKUP and PERM lock here, so
478 * both new and old client can cancel the dcache */
479 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
480 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
482 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
486 rc = mdt_lock_slaves(info, mo, LCK_EX, lockpart, einfo);
488 GOTO(out_unlock, rc);
490 if (mdt_object_exists(mo) == 0)
491 GOTO(out_unlock, rc = -ENOENT);
493 /* all attrs are packed into mti_attr in unpack_setattr */
494 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
495 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
497 /* This is only for set ctime when rename's source is on remote MDS. */
498 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
499 ma->ma_attr_flags |= MDS_VTX_BYPASS;
501 /* VBR: update version if attr changed are important for recovery */
503 /* update on-disk version of changed object */
504 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
505 rc = mdt_version_get_check_save(info, mo, 0);
507 GOTO(out_unlock, rc);
510 /* Ensure constant striping during chown(). See LU-2789. */
511 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
512 mutex_lock(&mo->mot_lov_mutex);
514 /* all attrs are packed into mti_attr in unpack_setattr */
515 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
517 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
518 mutex_unlock(&mo->mot_lov_mutex);
521 GOTO(out_unlock, rc);
525 mdt_unlock_slaves(info, mo, lockpart, einfo);
526 mdt_object_unlock(info, mo, lh, rc);
531 * Check HSM flags and add HS_DIRTY flag if relevant.
533 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
534 * and is not RELEASED.
536 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
542 /* If the file was modified, add the dirty flag */
543 ma->ma_need = MA_HSM;
544 rc = mdt_attr_get_complex(info, mo, ma);
546 CERROR("file attribute read error for "DFID": %d.\n",
547 PFID(mdt_object_fid(mo)), rc);
551 /* If an up2date copy exists in the backend, add dirty flag */
552 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
553 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
554 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
556 ma->ma_hsm.mh_flags |= HS_DIRTY;
558 mdt_lock_reg_init(lh, LCK_PW);
559 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
564 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
566 CERROR("file attribute change error for "DFID": %d\n",
567 PFID(mdt_object_fid(mo)), rc);
568 mdt_object_unlock(info, mo, lh, rc);
574 static int mdt_reint_setattr(struct mdt_thread_info *info,
575 struct mdt_lock_handle *lhc)
577 struct md_attr *ma = &info->mti_attr;
578 struct mdt_reint_record *rr = &info->mti_rr;
579 struct ptlrpc_request *req = mdt_info_req(info);
580 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
581 struct mdt_file_data *mfd;
582 struct mdt_object *mo;
583 struct mdt_body *repbody;
587 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
588 (unsigned int)ma->ma_attr.la_valid);
590 if (info->mti_dlm_req)
591 ldlm_request_cancel(req, info->mti_dlm_req, 0);
593 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
594 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
596 GOTO(out, rc = PTR_ERR(mo));
598 /* start a log jounal handle if needed */
599 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
600 if ((ma->ma_attr.la_valid & LA_SIZE) ||
601 (rr->rr_flags & MRF_OPEN_TRUNC)) {
602 /* Check write access for the O_TRUNC case */
603 if (mdt_write_read(mo) < 0)
604 GOTO(out_put, rc = -ETXTBSY);
606 } else if (info->mti_ioepoch &&
607 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
608 /* Truncate case. IOEpoch is opened. */
609 rc = mdt_write_get(mo);
613 mfd = mdt_mfd_new(med);
616 GOTO(out_put, rc = -ENOMEM);
619 mdt_ioepoch_open(info, mo, 0);
620 repbody->ioepoch = mo->mot_ioepoch;
622 mdt_object_get(info->mti_env, mo);
623 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
624 mfd->mfd_object = mo;
625 mfd->mfd_xid = req->rq_xid;
627 spin_lock(&med->med_open_lock);
628 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
629 spin_unlock(&med->med_open_lock);
630 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
633 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
635 /* SOM Attribute update case. Find the proper mfd and update
636 * SOM attributes on the proper object. */
637 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
638 LASSERT(info->mti_ioepoch);
640 spin_lock(&med->med_open_lock);
641 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
644 spin_unlock(&med->med_open_lock);
645 CDEBUG(D_INODE, "no handle for file close: "
646 "fid = "DFID": cookie = "LPX64"\n",
647 PFID(info->mti_rr.rr_fid1),
648 info->mti_ioepoch->handle.cookie);
649 GOTO(out_put, rc = -ESTALE);
651 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
652 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
654 class_handle_unhash(&mfd->mfd_handle);
655 cfs_list_del_init(&mfd->mfd_list);
656 spin_unlock(&med->med_open_lock);
658 mdt_mfd_close(info, mfd);
659 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
660 LASSERT((ma->ma_valid & MA_LOV) == 0);
661 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
664 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
665 struct lu_buf *buf = &info->mti_buf;
666 LASSERT(ma->ma_attr.la_valid == 0);
667 buf->lb_buf = ma->ma_lmm;
668 buf->lb_len = ma->ma_lmm_size;
669 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
670 buf, XATTR_NAME_LOV, 0);
673 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
674 struct lu_buf *buf = &info->mti_buf;
676 LASSERT(ma->ma_attr.la_valid == 0);
677 buf->lb_buf = ma->ma_lmv;
678 buf->lb_len = ma->ma_lmv_size;
679 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
680 buf, XATTR_NAME_DEFAULT_LMV, 0);
686 /* If file data is modified, add the dirty flag */
687 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
688 rc = mdt_add_dirty_flag(info, mo, ma);
690 ma->ma_need = MA_INODE;
692 rc = mdt_attr_get_complex(info, mo, ma);
696 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
698 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
699 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
700 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
701 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
702 struct lustre_capa *capa;
704 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
706 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
707 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
710 repbody->valid |= OBD_MD_FLOSSCAPA;
715 mdt_object_put(info->mti_env, mo);
718 mdt_counter_incr(req, LPROC_MDT_SETATTR);
720 mdt_client_compatibility(info);
721 rc2 = mdt_fix_reply(info);
727 static int mdt_reint_create(struct mdt_thread_info *info,
728 struct mdt_lock_handle *lhc)
730 struct ptlrpc_request *req = mdt_info_req(info);
734 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
735 RETURN(err_serious(-ESTALE));
737 if (info->mti_dlm_req)
738 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
740 if (!lu_name_is_valid(&info->mti_rr.rr_name))
743 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
745 mdt_counter_incr(req, LPROC_MDT_MKDIR);
753 /* Special file should stay on the same node as parent. */
754 mdt_counter_incr(req, LPROC_MDT_MKNOD);
757 CERROR("%s: Unsupported mode %o\n",
758 mdt_obd_name(info->mti_mdt),
759 info->mti_attr.ma_attr.la_mode);
760 RETURN(err_serious(-EOPNOTSUPP));
763 rc = mdt_md_create(info);
768 * VBR: save parent version in reply and child version getting by its name.
769 * Version of child is getting and checking during its lookup. If
771 static int mdt_reint_unlink(struct mdt_thread_info *info,
772 struct mdt_lock_handle *lhc)
774 struct mdt_reint_record *rr = &info->mti_rr;
775 struct ptlrpc_request *req = mdt_info_req(info);
776 struct md_attr *ma = &info->mti_attr;
777 struct lu_fid *child_fid = &info->mti_tmp_fid1;
778 struct mdt_object *mp;
779 struct mdt_object *mc;
780 struct mdt_lock_handle *parent_lh;
781 struct mdt_lock_handle *child_lh;
782 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
787 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
788 PNAME(&rr->rr_name));
790 if (info->mti_dlm_req)
791 ldlm_request_cancel(req, info->mti_dlm_req, 0);
793 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
794 RETURN(err_serious(-ENOENT));
796 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
800 * step 1: Found the parent.
802 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
808 parent_lh = &info->mti_lh[MDT_LH_PARENT];
810 if (mdt_object_remote(mp)) {
811 mdt_lock_reg_init(parent_lh, LCK_EX);
812 rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
813 parent_lh->mlh_rreg_mode,
814 MDS_INODELOCK_UPDATE);
816 GOTO(put_parent, rc);
819 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
820 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
823 GOTO(put_parent, rc);
825 rc = mdt_version_get_check_save(info, mp, 0);
827 GOTO(unlock_parent, rc);
830 /* step 2: find & lock the child */
831 /* lookup child object along with version checking */
833 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
835 /* Name might not be able to find during resend of
836 * remote unlink, considering following case.
837 * dir_A is a remote directory, the name entry of
838 * dir_A is on MDT0, the directory is on MDT1,
840 * 1. client sends unlink req to MDT1.
841 * 2. MDT1 sends name delete update to MDT0.
842 * 3. name entry is being deleted in MDT0 synchronously.
843 * 4. MDT1 is restarted.
844 * 5. client resends unlink req to MDT1. So it can not
845 * find the name entry on MDT0 anymore.
846 * In this case, MDT1 only needs to destory the local
849 if (mdt_object_remote(mp) && rc == -ENOENT &&
850 !fid_is_zero(rr->rr_fid2) &&
851 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
853 *child_fid = *rr->rr_fid2;
855 GOTO(unlock_parent, rc);
859 if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid))
860 GOTO(unlock_parent, rc = -EPERM);
862 /* We will lock the child regardless it is local or remote. No harm. */
863 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
865 GOTO(unlock_parent, rc = PTR_ERR(mc));
867 child_lh = &info->mti_lh[MDT_LH_CHILD];
868 mdt_lock_reg_init(child_lh, LCK_EX);
869 if (mdt_object_remote(mc)) {
870 struct mdt_body *repbody;
872 if (!fid_is_zero(rr->rr_fid2)) {
873 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
874 mdt_obd_name(info->mti_mdt),
875 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
876 GOTO(put_child, rc = -ENOENT);
878 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
879 mdt_obd_name(info->mti_mdt),
880 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
882 if (!mdt_is_dne_client(req->rq_export))
883 /* Return -EIO for old client */
884 GOTO(put_child, rc = -EIO);
886 if (info->mti_spec.sp_rm_entry) {
887 struct lu_ucred *uc = mdt_ucred(info);
889 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
890 CERROR("%s: unlink remote entry is only "
891 "permitted for administrator: rc = %d\n",
892 mdt_obd_name(info->mti_mdt),
894 GOTO(put_child, rc = -EPERM);
897 ma->ma_need = MA_INODE;
899 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
900 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
901 NULL, &rr->rr_name, ma, no_name);
904 /* Revoke the LOOKUP lock of the remote object granted by
905 * this MDT. Since the unlink will happen on another MDT,
906 * it will release the LOOKUP lock right away. Then What
907 * would happen if another client try to grab the LOOKUP
908 * lock at the same time with unlink XXX */
909 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
911 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
912 LASSERT(repbody != NULL);
913 repbody->fid1 = *mdt_object_fid(mc);
914 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
915 GOTO(unlock_child, rc = -EREMOTE);
916 } else if (info->mti_spec.sp_rm_entry) {
918 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
920 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
924 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
925 * this now because a running HSM restore on the child (unlink
926 * victim) will hold the layout lock. See LU-4002. */
927 rc = mdt_object_lock(info, mc, child_lh,
928 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
933 * Now we can only make sure we need MA_INODE, in mdd layer, will check
934 * whether need MA_LOV and MA_COOKIE.
936 ma->ma_need = MA_INODE;
939 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, einfo);
941 GOTO(unlock_child, rc);
943 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
944 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
945 /* save version when object is locked */
946 mdt_version_get_save(info, mc, 1);
948 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
950 mutex_lock(&mc->mot_lov_mutex);
952 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
953 mdt_object_child(mc), &rr->rr_name, ma, no_name);
955 mutex_unlock(&mc->mot_lov_mutex);
957 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
958 rc = mdt_attr_get_complex(info, mc, ma);
960 mdt_handle_last_unlink(info, mc, ma);
962 if (ma->ma_valid & MA_INODE) {
963 switch (ma->ma_attr.la_mode & S_IFMT) {
965 mdt_counter_incr(req, LPROC_MDT_RMDIR);
973 mdt_counter_incr(req, LPROC_MDT_UNLINK);
976 LASSERTF(0, "bad file type %o unlinking\n",
977 ma->ma_attr.la_mode);
984 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, einfo);
985 mdt_object_unlock(info, mc, child_lh, rc);
987 /* Since we do not need reply md striped dir info to client, so
988 * reset mti_big_lmm_used to avoid confusing mdt_fix_reply */
989 if (info->mti_big_lmm_used)
990 info->mti_big_lmm_used = 0;
992 mdt_object_put(info->mti_env, mc);
994 mdt_object_unlock(info, mp, parent_lh, rc);
996 mdt_object_put(info->mti_env, mp);
1002 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1005 static int mdt_reint_link(struct mdt_thread_info *info,
1006 struct mdt_lock_handle *lhc)
1008 struct mdt_reint_record *rr = &info->mti_rr;
1009 struct ptlrpc_request *req = mdt_info_req(info);
1010 struct md_attr *ma = &info->mti_attr;
1011 struct mdt_object *ms;
1012 struct mdt_object *mp;
1013 struct mdt_lock_handle *lhs;
1014 struct mdt_lock_handle *lhp;
1018 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1019 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1021 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1022 RETURN(err_serious(-ENOENT));
1024 if (info->mti_dlm_req)
1025 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1027 /* Invalid case so return error immediately instead of
1029 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1032 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1033 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1036 /* step 1: find & lock the target parent dir */
1037 lhp = &info->mti_lh[MDT_LH_PARENT];
1038 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1039 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1040 MDS_INODELOCK_UPDATE);
1042 RETURN(PTR_ERR(mp));
1044 rc = mdt_version_get_check_save(info, mp, 0);
1046 GOTO(out_unlock_parent, rc);
1048 /* step 2: find & lock the source */
1049 lhs = &info->mti_lh[MDT_LH_CHILD];
1050 mdt_lock_reg_init(lhs, LCK_EX);
1052 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1054 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1056 if (!mdt_object_exists(ms)) {
1057 mdt_object_put(info->mti_env, ms);
1058 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1059 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1060 GOTO(out_unlock_parent, rc = -ENOENT);
1063 if (mdt_object_remote(ms)) {
1064 mdt_object_put(info->mti_env, ms);
1065 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1066 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1068 GOTO(out_unlock_parent, rc = -EXDEV);
1071 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1072 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1074 mdt_object_put(info->mti_env, ms);
1075 GOTO(out_unlock_parent, rc);
1078 /* step 3: link it */
1079 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1080 OBD_FAIL_MDS_REINT_LINK_WRITE);
1082 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1083 rc = mdt_version_get_check_save(info, ms, 1);
1085 GOTO(out_unlock_child, rc);
1087 /** check target version by name during replay */
1088 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1089 &info->mti_tmp_fid1, 2);
1090 if (rc != 0 && rc != -ENOENT)
1091 GOTO(out_unlock_child, rc);
1092 /* save version of file name for replay, it must be ENOENT here */
1093 if (!req_is_replay(mdt_info_req(info))) {
1094 if (rc != -ENOENT) {
1095 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1096 PNAME(&rr->rr_name));
1097 GOTO(out_unlock_child, rc = -EEXIST);
1099 info->mti_ver[2] = ENOENT_VERSION;
1100 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1103 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1104 mdt_object_child(ms), &rr->rr_name, ma);
1107 mdt_counter_incr(req, LPROC_MDT_LINK);
1111 mdt_object_unlock_put(info, ms, lhs, rc);
1113 mdt_object_unlock_put(info, mp, lhp, rc);
1117 * lock the part of the directory according to the hash of the name
1118 * (lh->mlh_pdo_hash) in parallel directory lock.
1120 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1121 struct mdt_lock_handle *lh,
1122 struct mdt_object *obj, __u64 ibits)
1124 struct ldlm_res_id *res = &info->mti_res_id;
1125 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1126 ldlm_policy_data_t *policy = &info->mti_policy;
1130 * Finish res_id initializing by name hash marking part of
1131 * directory which is taking modification.
1133 LASSERT(lh->mlh_pdo_hash != 0);
1134 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1135 memset(policy, 0, sizeof(*policy));
1136 policy->l_inodebits.bits = ibits;
1138 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1139 * going to be sent to client. If it is - mdt_intent_policy() path will
1140 * fix it up and turn FL_LOCAL flag off.
1142 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1143 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1144 &info->mti_exp->exp_handle.h_cookie);
1148 static int mdt_rename_lock(struct mdt_thread_info *info,
1149 struct lustre_handle *lh)
1151 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1152 ldlm_policy_data_t *policy = &info->mti_policy;
1153 struct ldlm_res_id *res_id = &info->mti_res_id;
1158 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1160 memset(policy, 0, sizeof *policy);
1161 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1162 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 53, 0)
1163 /* In phase I, we will not do cross-rename, so local BFL lock would
1166 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1168 * Current node is controller, that is mdt0, where we should
1171 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1172 LCK_EX, &flags, ldlm_blocking_ast,
1173 ldlm_completion_ast, NULL, NULL, 0,
1175 &info->mti_exp->exp_handle.h_cookie,
1178 #warning "Local rename lock is invalid for DNE phase II."
1183 static void mdt_rename_unlock(struct lustre_handle *lh)
1186 LASSERT(lustre_handle_is_used(lh));
1187 ldlm_lock_decref(lh, LCK_EX);
1192 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1193 * target. Source should not be ancestor of target dir. May be other rename
1194 * checks can be moved here later.
1196 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1198 struct mdt_reint_record *rr = &info->mti_rr;
1199 struct lu_fid dst_fid = *rr->rr_fid2;
1200 struct mdt_object *dst;
1205 LASSERT(fid_is_sane(&dst_fid));
1206 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1208 rc = mdo_is_subdir(info->mti_env,
1209 mdt_object_child(dst), fid,
1211 mdt_object_put(info->mti_env, dst);
1212 if (rc != -EREMOTE && rc < 0) {
1213 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1215 /* check the found fid */
1216 if (lu_fid_eq(&dst_fid, fid))
1222 } while (rc == -EREMOTE);
1228 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1229 * 2 - src child; 3 - tgt child.
1230 * Update on disk version of src child.
1233 * For DNE phase I, only these renames are allowed
1234 * mv src_p/src_c tgt_p/tgt_c
1235 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1236 * 2. src_p and tgt_p are same directory, and tgt_c does not
1237 * exists. In this case, all of modification will happen
1238 * in the MDT where ithesource parent is, only one remote
1239 * update is needed, i.e. set c_time/m_time on the child.
1240 * And tgt_c will be still in the same MDT as the original
1243 static int mdt_reint_rename(struct mdt_thread_info *info,
1244 struct mdt_lock_handle *lhc)
1246 struct mdt_reint_record *rr = &info->mti_rr;
1247 struct md_attr *ma = &info->mti_attr;
1248 struct ptlrpc_request *req = mdt_info_req(info);
1249 struct mdt_object *msrcdir;
1250 struct mdt_object *mtgtdir;
1251 struct mdt_object *mold;
1252 struct mdt_object *mnew = NULL;
1253 struct mdt_lock_handle *lh_srcdirp;
1254 struct mdt_lock_handle *lh_tgtdirp;
1255 struct mdt_lock_handle *lh_oldp;
1256 struct mdt_lock_handle *lh_newp;
1257 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1258 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1259 struct lustre_handle rename_lh = { 0 };
1263 if (info->mti_dlm_req)
1264 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1266 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1267 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1268 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1270 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1271 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1274 rc = mdt_rename_lock(info, &rename_lh);
1276 CERROR("Can't lock FS for rename, rc %d\n", rc);
1280 lh_newp = &info->mti_lh[MDT_LH_NEW];
1282 /* step 1: lock the source dir. */
1283 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1284 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1285 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1286 MDS_INODELOCK_UPDATE);
1287 if (IS_ERR(msrcdir))
1288 GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
1290 rc = mdt_version_get_check_save(info, msrcdir, 0);
1292 GOTO(out_unlock_source, rc);
1294 /* step 2: find & lock the target dir. */
1295 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1296 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1297 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1298 mdt_object_get(info->mti_env, msrcdir);
1300 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1301 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1302 MDS_INODELOCK_UPDATE);
1304 GOTO(out_unlock_source, rc);
1305 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1308 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1310 if (IS_ERR(mtgtdir))
1311 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1313 /* check early, the real version will be saved after locking */
1314 rc = mdt_version_get_check(info, mtgtdir, 1);
1316 GOTO(out_put_target, rc);
1318 if (unlikely(mdt_object_remote(mtgtdir))) {
1319 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1320 "on different MDTs\n", PFID(rr->rr_fid1),
1322 GOTO(out_put_target, rc = -EXDEV);
1324 if (likely(mdt_object_exists(mtgtdir))) {
1325 /* we lock the target dir if it is local */
1326 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1327 MDS_INODELOCK_UPDATE,
1330 GOTO(out_put_target, rc);
1331 /* get and save correct version after locking */
1332 mdt_version_get_save(info, mtgtdir, 1);
1334 GOTO(out_put_target, rc = -ESTALE);
1339 /* step 3: find & lock the old object. */
1341 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1343 GOTO(out_unlock_target, rc);
1345 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1346 GOTO(out_unlock_target, rc = -EINVAL);
1348 if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid))
1349 GOTO(out_unlock_target, rc = -EPERM);
1351 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1353 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1355 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1356 mdt_lock_reg_init(lh_oldp, LCK_EX);
1357 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1358 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1360 mdt_object_put(info->mti_env, mold);
1361 GOTO(out_unlock_target, rc);
1364 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1365 /* save version after locking */
1366 mdt_version_get_save(info, mold, 2);
1367 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1369 /* step 4: find & lock the new object. */
1370 /* new target object may not exist now */
1371 /* lookup with version checking */
1373 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1376 /* the new_fid should have been filled at this moment */
1377 if (lu_fid_eq(old_fid, new_fid))
1378 GOTO(out_unlock_old, rc);
1380 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1381 lu_fid_eq(new_fid, rr->rr_fid2))
1382 GOTO(out_unlock_old, rc = -EINVAL);
1384 if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid))
1385 GOTO(out_unlock_old, rc = -EPERM);
1387 if (mdt_object_remote(mold)) {
1388 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1390 GOTO(out_unlock_old, rc = -EXDEV);
1393 mdt_lock_reg_init(lh_newp, LCK_EX);
1394 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1396 GOTO(out_unlock_old, rc = PTR_ERR(mnew));
1398 if (mdt_object_remote(mnew)) {
1399 mdt_object_put(info->mti_env, mnew);
1400 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1402 GOTO(out_unlock_old, rc = -EXDEV);
1405 /* We used to acquire MDS_INODELOCK_FULL here but we
1406 * can't do this now because a running HSM restore on
1407 * the rename onto victim will hold the layout
1408 * lock. See LU-4002. */
1409 rc = mdt_object_lock(info, mnew, lh_newp,
1410 MDS_INODELOCK_LOOKUP |
1411 MDS_INODELOCK_UPDATE,
1414 mdt_object_put(info->mti_env, mnew);
1415 GOTO(out_unlock_old, rc);
1417 /* get and save version after locking */
1418 mdt_version_get_save(info, mnew, 3);
1419 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1420 } else if (rc != -EREMOTE && rc != -ENOENT) {
1421 GOTO(out_unlock_old, rc);
1423 /* If mnew does not exist and mold are remote directory,
1424 * it only allows rename if they are under same directory */
1425 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1426 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1428 GOTO(out_unlock_old, rc = -EXDEV);
1430 mdt_enoent_version_save(info, 3);
1433 /* step 5: rename it */
1434 mdt_reint_init_ma(info, ma);
1436 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1437 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1439 /* Check if @dst is subdir of @src. */
1440 rc = mdt_rename_sanity(info, old_fid);
1442 GOTO(out_unlock_new, rc);
1445 mutex_lock(&mnew->mot_lov_mutex);
1447 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1448 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1449 mnew != NULL ? mdt_object_child(mnew) : NULL,
1450 &rr->rr_tgt_name, ma);
1453 mutex_unlock(&mnew->mot_lov_mutex);
1455 /* handle last link of tgt object */
1457 mdt_counter_incr(req, LPROC_MDT_RENAME);
1459 mdt_handle_last_unlink(info, mnew, ma);
1461 mdt_rename_counter_tally(info, info->mti_mdt, req,
1468 mdt_object_unlock_put(info, mnew, lh_newp, rc);
1470 mdt_object_unlock_put(info, mold, lh_oldp, rc);
1472 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1474 mdt_object_put(info->mti_env, mtgtdir);
1476 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1478 if (lustre_handle_is_used(&rename_lh))
1479 mdt_rename_unlock(&rename_lh);
1483 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1484 struct mdt_lock_handle *lhc);
1486 static mdt_reinter reinters[REINT_MAX] = {
1487 [REINT_SETATTR] = mdt_reint_setattr,
1488 [REINT_CREATE] = mdt_reint_create,
1489 [REINT_LINK] = mdt_reint_link,
1490 [REINT_UNLINK] = mdt_reint_unlink,
1491 [REINT_RENAME] = mdt_reint_rename,
1492 [REINT_OPEN] = mdt_reint_open,
1493 [REINT_SETXATTR] = mdt_reint_setxattr,
1494 [REINT_RMENTRY] = mdt_reint_unlink
1497 int mdt_reint_rec(struct mdt_thread_info *info,
1498 struct mdt_lock_handle *lhc)
1503 rc = reinters[info->mti_rr.rr_opcode](info, lhc);