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, 2015, 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 <lprocfs_status.h>
50 #include "mdt_internal.h"
51 #include <lustre_lmv.h>
53 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
56 ma->ma_need = MA_INODE;
61 * Get version of object by fid.
63 * Return real version or ENOENT_VERSION if object doesn't exist
65 static void mdt_obj_version_get(struct mdt_thread_info *info,
66 struct mdt_object *o, __u64 *version)
70 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
71 !fid_is_obf(mdt_object_fid(o)))
72 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
74 *version = ENOENT_VERSION;
75 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
76 PFID(mdt_object_fid(o)), *version);
80 * Check version is correct.
82 * Should be called only during replay.
84 static int mdt_version_check(struct ptlrpc_request *req,
85 __u64 version, int idx)
87 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
90 if (!exp_connect_vbr(req->rq_export))
93 LASSERT(req_is_replay(req));
94 /** VBR: version is checked always because costs nothing */
95 LASSERT(idx < PTLRPC_NUM_VERSIONS);
96 /** Sanity check for malformed buffers */
97 if (pre_ver == NULL) {
98 CERROR("No versions in request buffer\n");
99 spin_lock(&req->rq_export->exp_lock);
100 req->rq_export->exp_vbr_failed = 1;
101 spin_unlock(&req->rq_export->exp_lock);
103 } else if (pre_ver[idx] != version) {
104 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
105 pre_ver[idx], version);
106 spin_lock(&req->rq_export->exp_lock);
107 req->rq_export->exp_vbr_failed = 1;
108 spin_unlock(&req->rq_export->exp_lock);
115 * Save pre-versions in reply.
117 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
122 if (!exp_connect_vbr(req->rq_export))
125 LASSERT(!req_is_replay(req));
126 LASSERT(req->rq_repmsg != NULL);
127 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
129 reply_ver[idx] = version;
133 * Save enoent version, it is needed when it is obvious that object doesn't
134 * exist, e.g. child during create.
136 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
138 /* save version of file name for replay, it must be ENOENT here */
139 if (!req_is_replay(mdt_info_req(info))) {
140 info->mti_ver[idx] = ENOENT_VERSION;
141 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
146 * Get version from disk and save in reply buffer.
148 * Versions are saved in reply only during normal operations not replays.
150 void mdt_version_get_save(struct mdt_thread_info *info,
151 struct mdt_object *mto, int idx)
153 /* don't save versions during replay */
154 if (!req_is_replay(mdt_info_req(info))) {
155 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
156 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
161 * Get version from disk and check it, no save in reply.
163 int mdt_version_get_check(struct mdt_thread_info *info,
164 struct mdt_object *mto, int idx)
166 /* only check versions during replay */
167 if (!req_is_replay(mdt_info_req(info)))
170 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
171 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
175 * Get version from disk and check if recovery or just save.
177 int mdt_version_get_check_save(struct mdt_thread_info *info,
178 struct mdt_object *mto, int idx)
182 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
183 if (req_is_replay(mdt_info_req(info)))
184 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
187 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
192 * Lookup with version checking.
194 * This checks version of 'name'. Many reint functions uses 'name' for child not
195 * FID, therefore we need to get object by name and check its version.
197 static int mdt_lookup_version_check(struct mdt_thread_info *info,
198 struct mdt_object *p,
199 const struct lu_name *lname,
200 struct lu_fid *fid, int idx)
204 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
206 /* Check version only during replay */
207 if (!req_is_replay(mdt_info_req(info)))
210 info->mti_ver[idx] = ENOENT_VERSION;
212 struct mdt_object *child;
213 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
214 if (likely(!IS_ERR(child))) {
215 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
216 mdt_object_put(info->mti_env, child);
219 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
220 return vbrc ? vbrc : rc;
224 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
226 struct lu_ucred *uc = mdt_ucred(info);
227 struct mdt_device *mdt = info->mti_mdt;
229 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
230 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
231 mdt->mdt_enable_remote_dir_gid != -1)
239 * mdt_remote_permission: Check whether the remote operation is permitted,
241 * Only sysadmin can create remote directory / striped directory,
242 * migrate directory and set default stripedEA on directory, unless
244 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
246 * param[in] info: mdt_thread_info.
248 * retval = 0 remote operation is allowed.
249 * < 0 remote operation is denied.
251 static int mdt_remote_permission(struct mdt_thread_info *info)
253 struct md_op_spec *spec = &info->mti_spec;
254 struct lu_attr *attr = &info->mti_attr.ma_attr;
255 struct obd_export *exp = mdt_info_req(info)->rq_export;
258 if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
259 rc = mdt_remote_permission_check(info);
264 if (info->mti_rr.rr_opcode == REINT_CREATE &&
265 (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
266 spec->u.sp_ea.eadatalen != 0)) {
267 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
269 /* Only new clients can create remote dir( >= 2.4) and
270 * striped dir(>= 2.6), old client will return -ENOTSUPP */
271 if (!mdt_is_dne_client(exp))
274 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
275 !mdt_is_striped_client(exp))
278 rc = mdt_remote_permission_check(info);
283 if (info->mti_rr.rr_opcode == REINT_SETATTR) {
284 struct md_attr *ma = &info->mti_attr;
286 if ((ma->ma_valid & MA_LMV)) {
287 rc = mdt_remote_permission_check(info);
296 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
297 struct mdt_object *obj, __u64 ibits,
298 struct mdt_lock_handle *s0_lh,
299 struct mdt_object *s0_obj,
300 struct ldlm_enqueue_info *einfo,
303 union ldlm_policy_data *policy = &mti->mti_policy;
304 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
309 if (!S_ISDIR(obj->mot_header.loh_attr))
312 /* Unlock stripe 0 */
313 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
314 LASSERT(s0_obj != NULL);
315 mdt_object_unlock_put(mti, s0_obj, s0_lh, decref);
318 memset(policy, 0, sizeof(*policy));
319 policy->l_inodebits.bits = ibits;
321 if (slave_locks != NULL) {
322 LASSERT(s0_lh != NULL);
323 for (i = 1; i < slave_locks->count; i++) {
324 /* borrow s0_lh temporarily to do mdt unlock */
325 mdt_lock_reg_init(s0_lh, einfo->ei_mode);
326 s0_lh->mlh_rreg_lh = slave_locks->handles[i];
327 mdt_object_unlock(mti, NULL, s0_lh, decref);
328 slave_locks->handles[i].cookie = 0ull;
332 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
337 static int mdt_init_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
340 struct lu_buf *buf = &mti->mti_buf;
341 struct lmv_mds_md_v1 *lmv;
345 if (!S_ISDIR(obj->mot_header.loh_attr))
348 buf->lb_buf = mti->mti_xattr_buf;
349 buf->lb_len = sizeof(mti->mti_xattr_buf);
350 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
353 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
355 buf->lb_buf = mti->mti_big_lmm;
356 buf->lb_len = mti->mti_big_lmmsize;
360 if (rc == -ENODATA || rc == -ENOENT)
367 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
370 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
376 * Lock slave stripes if necessary, the lock handles of slave stripes
377 * will be stored in einfo->ei_cbdata.
379 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
380 enum ldlm_mode mode, __u64 ibits,
381 struct lu_fid *s0_fid,
382 struct mdt_lock_handle *s0_lh,
383 struct mdt_object **s0_objp,
384 struct ldlm_enqueue_info *einfo)
386 union ldlm_policy_data *policy = &mti->mti_policy;
390 memset(einfo, 0, sizeof(*einfo));
392 rc = mdt_init_slaves(mti, obj, s0_fid);
396 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
398 if (!lu_fid_eq(s0_fid, mdt_object_fid(obj))) {
399 /* Except migrating object, whose 0_stripe and master
400 * object are the same object, 0_stripe and master
401 * object are different, though they are in the same
402 * MDT, to avoid adding osd_object_lock here, so we
403 * will enqueue the stripe0 lock in MDT0 for now */
404 *s0_objp = mdt_object_find(mti->mti_env, mti->mti_mdt, s0_fid);
405 if (IS_ERR(*s0_objp))
406 RETURN(PTR_ERR(*s0_objp));
408 rc = mdt_reint_object_lock(mti, *s0_objp, s0_lh, ibits, true);
410 mdt_object_put(mti->mti_env, *s0_objp);
415 einfo->ei_type = LDLM_IBITS;
416 einfo->ei_mode = mode;
417 einfo->ei_cb_bl = mdt_remote_blocking_ast;
418 einfo->ei_cb_local_bl = mdt_blocking_ast;
419 einfo->ei_cb_cp = ldlm_completion_ast;
420 einfo->ei_enq_slave = 1;
421 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
422 memset(policy, 0, sizeof(*policy));
423 policy->l_inodebits.bits = ibits;
425 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
431 * VBR: we save three versions in reply:
432 * 0 - parent. Check that parent version is the same during replay.
433 * 1 - name. Version of 'name' if file exists with the same name or
434 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
435 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
438 static int mdt_md_create(struct mdt_thread_info *info)
440 struct mdt_device *mdt = info->mti_mdt;
441 struct mdt_object *parent;
442 struct mdt_object *child;
443 struct mdt_lock_handle *lh;
444 struct mdt_body *repbody;
445 struct md_attr *ma = &info->mti_attr;
446 struct mdt_reint_record *rr = &info->mti_rr;
450 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
452 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
454 if (!fid_is_md_operative(rr->rr_fid1))
457 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
459 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
461 RETURN(PTR_ERR(parent));
463 if (!mdt_object_exists(parent))
464 GOTO(put_parent, rc = -ENOENT);
466 lh = &info->mti_lh[MDT_LH_PARENT];
467 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
468 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
470 GOTO(put_parent, rc);
472 if (!mdt_object_remote(parent)) {
473 rc = mdt_version_get_check_save(info, parent, 0);
475 GOTO(unlock_parent, rc);
479 * Check child name version during replay.
480 * During create replay a file may exist with same name.
482 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
483 &info->mti_tmp_fid1, 1);
485 GOTO(unlock_parent, rc = -EEXIST);
487 /* -ENOENT is expected here */
489 GOTO(unlock_parent, rc);
491 /* save version of file name for replay, it must be ENOENT here */
492 mdt_enoent_version_save(info, 1);
494 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
495 if (unlikely(IS_ERR(child)))
496 GOTO(unlock_parent, rc = PTR_ERR(child));
498 rc = mdt_remote_permission(info);
502 ma->ma_need = MA_INODE;
505 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
506 OBD_FAIL_MDS_REINT_CREATE_WRITE);
508 /* Version of child will be updated on disk. */
509 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
510 rc = mdt_version_get_check_save(info, child, 2);
514 /* Let lower layer know current lock mode. */
515 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
518 * Do not perform lookup sanity check. We know that name does
521 info->mti_spec.sp_cr_lookup = 0;
522 info->mti_spec.sp_feat = &dt_directory_features;
524 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
525 mdt_object_child(child), &info->mti_spec, ma);
527 rc = mdt_attr_get_complex(info, child, ma);
533 * On DNE, we need to eliminate dependey between 'mkdir a' and
534 * 'mkdir a/b' if b is a striped directory, to achieve this, two
535 * things are done below:
536 * 1. save child and slaves lock.
537 * 2. if the child is a striped directory, relock parent so to
538 * compare against with COS locks to ensure parent was
541 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
542 struct mdt_lock_handle *lhc;
543 struct mdt_lock_handle *s0_lh;
544 struct mdt_object *s0_obj = NULL;
545 struct ldlm_enqueue_info *einfo;
546 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
547 bool cos_incompat = false;
549 rc = mdt_init_slaves(info, child, s0_fid);
552 if (!mdt_object_remote(parent)) {
553 mdt_object_unlock(info, parent, lh, 1);
554 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
555 rc = mdt_reint_object_lock(info, parent, lh,
556 MDS_INODELOCK_UPDATE,
563 einfo = &info->mti_einfo;
564 lhc = &info->mti_lh[MDT_LH_CHILD];
565 mdt_lock_handle_init(lhc);
566 mdt_lock_reg_init(lhc, LCK_PW);
567 rc = mdt_reint_object_lock(info, child, lhc,
568 MDS_INODELOCK_UPDATE,
572 mdt_object_unlock(info, child, lhc, rc);
574 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
575 mdt_lock_handle_init(s0_lh);
576 mdt_lock_reg_init(s0_lh, LCK_PW);
577 rc = mdt_lock_slaves(info, child, LCK_PW, MDS_INODELOCK_UPDATE,
578 s0_fid, s0_lh, &s0_obj, einfo);
579 mdt_unlock_slaves(info, child, MDS_INODELOCK_UPDATE, s0_lh,
581 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME) && rc == -EIO)
585 /* Return fid & attr to client. */
586 if (ma->ma_valid & MA_INODE)
587 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
588 mdt_object_fid(child));
590 mdt_object_put(info->mti_env, child);
592 mdt_object_unlock(info, parent, lh, rc);
594 mdt_object_put(info->mti_env, parent);
598 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
601 struct mdt_lock_handle *lh;
602 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
603 __u64 lockpart = MDS_INODELOCK_UPDATE;
604 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
605 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
606 struct mdt_lock_handle *s0_lh = NULL;
607 struct mdt_object *s0_obj = NULL;
608 bool cos_incompat = false;
612 rc = mdt_init_slaves(info, mo, s0_fid);
616 lh = &info->mti_lh[MDT_LH_PARENT];
617 mdt_lock_reg_init(lh, LCK_PW);
619 /* Even though the new MDT will grant PERM lock to the old
620 * client, but the old client will almost ignore that during
621 * So it needs to revoke both LOOKUP and PERM lock here, so
622 * both new and old client can cancel the dcache */
623 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
624 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
626 rc = mdt_reint_object_lock(info, mo, lh, lockpart, cos_incompat);
630 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
631 mdt_lock_reg_init(s0_lh, LCK_PW);
632 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_fid, s0_lh, &s0_obj,
635 GOTO(out_unlock, rc);
637 /* all attrs are packed into mti_attr in unpack_setattr */
638 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
639 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
641 /* This is only for set ctime when rename's source is on remote MDS. */
642 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
643 ma->ma_attr_flags |= MDS_VTX_BYPASS;
645 /* VBR: update version if attr changed are important for recovery */
647 /* update on-disk version of changed object */
648 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
649 rc = mdt_version_get_check_save(info, mo, 0);
651 GOTO(out_unlock, rc);
654 /* Ensure constant striping during chown(). See LU-2789. */
655 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
656 mutex_lock(&mo->mot_lov_mutex);
658 /* all attrs are packed into mti_attr in unpack_setattr */
659 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
661 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
662 mutex_unlock(&mo->mot_lov_mutex);
665 GOTO(out_unlock, rc);
669 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo, rc);
670 mdt_object_unlock(info, mo, lh, rc);
675 * Check HSM flags and add HS_DIRTY flag if relevant.
677 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
678 * and is not RELEASED.
680 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
686 /* If the file was modified, add the dirty flag */
687 ma->ma_need = MA_HSM;
688 rc = mdt_attr_get_complex(info, mo, ma);
690 CERROR("file attribute read error for "DFID": %d.\n",
691 PFID(mdt_object_fid(mo)), rc);
695 /* If an up2date copy exists in the backend, add dirty flag */
696 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
697 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
698 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
700 ma->ma_hsm.mh_flags |= HS_DIRTY;
702 mdt_lock_reg_init(lh, LCK_PW);
703 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
707 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
709 CERROR("file attribute change error for "DFID": %d\n",
710 PFID(mdt_object_fid(mo)), rc);
711 mdt_object_unlock(info, mo, lh, rc);
717 static int mdt_reint_setattr(struct mdt_thread_info *info,
718 struct mdt_lock_handle *lhc)
720 struct md_attr *ma = &info->mti_attr;
721 struct mdt_reint_record *rr = &info->mti_rr;
722 struct ptlrpc_request *req = mdt_info_req(info);
723 struct mdt_object *mo;
724 struct mdt_body *repbody;
728 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
729 (unsigned int)ma->ma_attr.la_valid);
731 if (info->mti_dlm_req)
732 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
734 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
735 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
737 GOTO(out, rc = PTR_ERR(mo));
739 if (!mdt_object_exists(mo))
740 GOTO(out_put, rc = -ENOENT);
742 if (mdt_object_remote(mo))
743 GOTO(out_put, rc = -EREMOTE);
745 if ((ma->ma_attr.la_valid & LA_SIZE) ||
746 (rr->rr_flags & MRF_OPEN_TRUNC)) {
747 /* Check write access for the O_TRUNC case */
748 if (mdt_write_read(mo) < 0)
749 GOTO(out_put, rc = -ETXTBSY);
752 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
753 if (ma->ma_valid & MA_LOV)
754 GOTO(out_put, rc = -EPROTO);
756 rc = mdt_attr_set(info, mo, ma);
759 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
760 (ma->ma_valid & MA_INODE)) {
761 struct lu_buf *buf = &info->mti_buf;
762 struct mdt_lock_handle *lh;
764 rc = mdt_remote_permission(info);
768 if (ma->ma_attr.la_valid != 0)
769 GOTO(out_put, rc = -EPROTO);
771 lh = &info->mti_lh[MDT_LH_PARENT];
772 mdt_lock_reg_init(lh, LCK_PW);
774 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
778 if (ma->ma_valid & MA_LOV) {
779 buf->lb_buf = ma->ma_lmm;
780 buf->lb_len = ma->ma_lmm_size;
782 buf->lb_buf = ma->ma_lmv;
783 buf->lb_len = ma->ma_lmv_size;
785 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
786 (ma->ma_valid & MA_LOV) ?
787 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
790 mdt_object_unlock(info, mo, lh, rc);
794 GOTO(out_put, rc = -EPROTO);
797 /* If file data is modified, add the dirty flag */
798 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
799 rc = mdt_add_dirty_flag(info, mo, ma);
801 ma->ma_need = MA_INODE;
803 rc = mdt_attr_get_complex(info, mo, ma);
807 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
811 mdt_object_put(info->mti_env, mo);
814 mdt_counter_incr(req, LPROC_MDT_SETATTR);
816 mdt_client_compatibility(info);
817 rc2 = mdt_fix_reply(info);
823 static int mdt_reint_create(struct mdt_thread_info *info,
824 struct mdt_lock_handle *lhc)
826 struct ptlrpc_request *req = mdt_info_req(info);
830 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
831 RETURN(err_serious(-ESTALE));
833 if (info->mti_dlm_req)
834 ldlm_request_cancel(mdt_info_req(info),
835 info->mti_dlm_req, 0, LATF_SKIP);
837 if (!lu_name_is_valid(&info->mti_rr.rr_name))
840 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
842 mdt_counter_incr(req, LPROC_MDT_MKDIR);
850 /* Special file should stay on the same node as parent. */
851 mdt_counter_incr(req, LPROC_MDT_MKNOD);
854 CERROR("%s: Unsupported mode %o\n",
855 mdt_obd_name(info->mti_mdt),
856 info->mti_attr.ma_attr.la_mode);
857 RETURN(err_serious(-EOPNOTSUPP));
860 rc = mdt_md_create(info);
865 * VBR: save parent version in reply and child version getting by its name.
866 * Version of child is getting and checking during its lookup. If
868 static int mdt_reint_unlink(struct mdt_thread_info *info,
869 struct mdt_lock_handle *lhc)
871 struct mdt_reint_record *rr = &info->mti_rr;
872 struct ptlrpc_request *req = mdt_info_req(info);
873 struct md_attr *ma = &info->mti_attr;
874 struct lu_fid *child_fid = &info->mti_tmp_fid1;
875 struct mdt_object *mp;
876 struct mdt_object *mc;
877 struct mdt_lock_handle *parent_lh;
878 struct mdt_lock_handle *child_lh;
879 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
880 struct lu_fid *s0_fid = &info->mti_tmp_fid2;
881 struct mdt_lock_handle *s0_lh = NULL;
882 struct mdt_object *s0_obj = NULL;
884 bool cos_incompat = false;
889 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
890 PNAME(&rr->rr_name));
892 if (info->mti_dlm_req)
893 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
895 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
896 RETURN(err_serious(-ENOENT));
898 if (!fid_is_md_operative(rr->rr_fid1))
901 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
905 if (mdt_object_remote(mp)) {
908 rc = mdt_version_get_check_save(info, mp, 0);
910 GOTO(put_parent, rc);
914 parent_lh = &info->mti_lh[MDT_LH_PARENT];
915 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
916 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
919 GOTO(put_parent, rc);
921 /* lookup child object along with version checking */
923 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
925 /* Name might not be able to find during resend of
926 * remote unlink, considering following case.
927 * dir_A is a remote directory, the name entry of
928 * dir_A is on MDT0, the directory is on MDT1,
930 * 1. client sends unlink req to MDT1.
931 * 2. MDT1 sends name delete update to MDT0.
932 * 3. name entry is being deleted in MDT0 synchronously.
933 * 4. MDT1 is restarted.
934 * 5. client resends unlink req to MDT1. So it can not
935 * find the name entry on MDT0 anymore.
936 * In this case, MDT1 only needs to destory the local
939 if (mdt_object_remote(mp) && rc == -ENOENT &&
940 !fid_is_zero(rr->rr_fid2) &&
941 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
943 *child_fid = *rr->rr_fid2;
945 GOTO(unlock_parent, rc);
949 if (!fid_is_md_operative(child_fid))
950 GOTO(unlock_parent, rc = -EPERM);
952 /* We will lock the child regardless it is local or remote. No harm. */
953 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
955 GOTO(unlock_parent, rc = PTR_ERR(mc));
957 if (!cos_incompat && mdt_init_slaves(info, mc, s0_fid) > 0) {
959 mdt_object_put(info->mti_env, mc);
960 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
964 child_lh = &info->mti_lh[MDT_LH_CHILD];
965 mdt_lock_reg_init(child_lh, LCK_EX);
966 if (info->mti_spec.sp_rm_entry) {
967 struct lu_ucred *uc = mdt_ucred(info);
969 if (!mdt_is_dne_client(req->rq_export))
970 /* Return -ENOTSUPP for old client */
971 GOTO(put_child, rc = -ENOTSUPP);
973 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
974 GOTO(put_child, rc = -EPERM);
976 ma->ma_need = MA_INODE;
978 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
979 NULL, &rr->rr_name, ma, no_name);
983 if (mdt_object_remote(mc)) {
984 struct mdt_body *repbody;
986 if (!fid_is_zero(rr->rr_fid2)) {
987 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
988 mdt_obd_name(info->mti_mdt),
989 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
990 GOTO(put_child, rc = -ENOENT);
992 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
993 mdt_obd_name(info->mti_mdt),
994 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
996 if (!mdt_is_dne_client(req->rq_export))
997 /* Return -ENOTSUPP for old client */
998 GOTO(put_child, rc = -ENOTSUPP);
1000 /* Revoke the LOOKUP lock of the remote object granted by
1001 * this MDT. Since the unlink will happen on another MDT,
1002 * it will release the LOOKUP lock right away. Then What
1003 * would happen if another client try to grab the LOOKUP
1004 * lock at the same time with unlink XXX */
1005 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1006 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1007 LASSERT(repbody != NULL);
1008 repbody->mbo_fid1 = *mdt_object_fid(mc);
1009 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1010 GOTO(unlock_child, rc = -EREMOTE);
1012 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1013 * this now because a running HSM restore on the child (unlink
1014 * victim) will hold the layout lock. See LU-4002. */
1015 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1016 if (mdt_object_remote(mp)) {
1017 /* Enqueue lookup lock from parent MDT */
1018 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1019 &child_lh->mlh_rreg_lh,
1020 child_lh->mlh_rreg_mode,
1021 MDS_INODELOCK_LOOKUP, false, false);
1023 GOTO(put_child, rc);
1025 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1028 rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits,
1031 GOTO(unlock_child, rc);
1034 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1035 * whether need MA_LOV and MA_COOKIE.
1037 ma->ma_need = MA_INODE;
1040 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1041 mdt_lock_reg_init(s0_lh, LCK_EX);
1042 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_fid,
1043 s0_lh, &s0_obj, einfo);
1045 GOTO(unlock_child, rc);
1047 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1048 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1049 /* save version when object is locked */
1050 mdt_version_get_save(info, mc, 1);
1052 mutex_lock(&mc->mot_lov_mutex);
1054 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1055 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1057 mutex_unlock(&mc->mot_lov_mutex);
1059 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1060 rc = mdt_attr_get_complex(info, mc, ma);
1062 mdt_handle_last_unlink(info, mc, ma);
1064 if (ma->ma_valid & MA_INODE) {
1065 switch (ma->ma_attr.la_mode & S_IFMT) {
1067 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1075 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1078 LASSERTF(0, "bad file type %o unlinking\n",
1079 ma->ma_attr.la_mode);
1086 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo,
1088 mdt_object_unlock(info, mc, child_lh, rc);
1090 mdt_object_put(info->mti_env, mc);
1092 mdt_object_unlock(info, mp, parent_lh, rc);
1094 mdt_object_put(info->mti_env, mp);
1099 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1102 static int mdt_reint_link(struct mdt_thread_info *info,
1103 struct mdt_lock_handle *lhc)
1105 struct mdt_reint_record *rr = &info->mti_rr;
1106 struct ptlrpc_request *req = mdt_info_req(info);
1107 struct md_attr *ma = &info->mti_attr;
1108 struct mdt_object *ms;
1109 struct mdt_object *mp;
1110 struct mdt_lock_handle *lhs;
1111 struct mdt_lock_handle *lhp;
1116 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1117 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1119 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1120 RETURN(err_serious(-ENOENT));
1122 if (info->mti_dlm_req)
1123 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1125 /* Invalid case so return error immediately instead of
1127 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1130 if (!fid_is_md_operative(rr->rr_fid1) ||
1131 !fid_is_md_operative(rr->rr_fid2))
1134 /* step 1: find target parent dir */
1135 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1137 RETURN(PTR_ERR(mp));
1139 rc = mdt_version_get_check_save(info, mp, 0);
1141 GOTO(put_parent, rc);
1143 /* step 2: find source */
1144 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1146 GOTO(put_parent, rc = PTR_ERR(ms));
1148 if (!mdt_object_exists(ms)) {
1149 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1150 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1151 GOTO(put_source, rc = -ENOENT);
1154 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1156 lhp = &info->mti_lh[MDT_LH_PARENT];
1157 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1158 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1161 GOTO(put_source, rc);
1163 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1165 lhs = &info->mti_lh[MDT_LH_CHILD];
1166 mdt_lock_reg_init(lhs, LCK_EX);
1167 rc = mdt_reint_object_lock(info, ms, lhs,
1168 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1171 GOTO(unlock_parent, rc);
1173 /* step 3: link it */
1174 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1175 OBD_FAIL_MDS_REINT_LINK_WRITE);
1177 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1178 rc = mdt_version_get_check_save(info, ms, 1);
1180 GOTO(unlock_source, rc);
1182 /** check target version by name during replay */
1183 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1184 &info->mti_tmp_fid1, 2);
1185 if (rc != 0 && rc != -ENOENT)
1186 GOTO(unlock_source, rc);
1187 /* save version of file name for replay, it must be ENOENT here */
1188 if (!req_is_replay(mdt_info_req(info))) {
1189 if (rc != -ENOENT) {
1190 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1191 PNAME(&rr->rr_name));
1192 GOTO(unlock_source, rc = -EEXIST);
1194 info->mti_ver[2] = ENOENT_VERSION;
1195 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1198 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1199 mdt_object_child(ms), &rr->rr_name, ma);
1202 mdt_counter_incr(req, LPROC_MDT_LINK);
1206 mdt_object_unlock(info, ms, lhs, rc);
1208 mdt_object_unlock(info, mp, lhp, rc);
1210 mdt_object_put(info->mti_env, ms);
1212 mdt_object_put(info->mti_env, mp);
1216 * lock the part of the directory according to the hash of the name
1217 * (lh->mlh_pdo_hash) in parallel directory lock.
1219 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1220 struct mdt_lock_handle *lh,
1221 struct mdt_object *obj, __u64 ibits,
1224 struct ldlm_res_id *res = &info->mti_res_id;
1225 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1226 union ldlm_policy_data *policy = &info->mti_policy;
1227 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1231 * Finish res_id initializing by name hash marking part of
1232 * directory which is taking modification.
1234 LASSERT(lh->mlh_pdo_hash != 0);
1235 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1236 memset(policy, 0, sizeof(*policy));
1237 policy->l_inodebits.bits = ibits;
1239 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1240 dlmflags |= LDLM_FL_COS_INCOMPAT;
1242 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1243 * going to be sent to client. If it is - mdt_intent_policy() path will
1244 * fix it up and turn FL_LOCAL flag off.
1246 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1247 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1252 * Get BFL lock for rename or migrate process.
1254 static int mdt_rename_lock(struct mdt_thread_info *info,
1255 struct lustre_handle *lh)
1260 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1261 struct lu_fid *fid = &info->mti_tmp_fid1;
1262 struct mdt_object *obj;
1264 /* XXX, right now, it has to use object API to
1265 * enqueue lock cross MDT, so it will enqueue
1266 * rename lock(with LUSTRE_BFL_FID) by root object */
1268 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1270 RETURN(PTR_ERR(obj));
1272 rc = mdt_remote_object_lock(info, obj,
1273 &LUSTRE_BFL_FID, lh,
1275 MDS_INODELOCK_UPDATE, false, false);
1276 mdt_object_put(info->mti_env, obj);
1278 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1279 union ldlm_policy_data *policy = &info->mti_policy;
1280 struct ldlm_res_id *res_id = &info->mti_res_id;
1283 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1284 memset(policy, 0, sizeof *policy);
1285 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1286 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1287 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1288 LCK_EX, &flags, ldlm_blocking_ast,
1289 ldlm_completion_ast, NULL, NULL, 0,
1291 &info->mti_exp->exp_handle.h_cookie,
1298 static void mdt_rename_unlock(struct lustre_handle *lh)
1301 LASSERT(lustre_handle_is_used(lh));
1302 /* Cancel the single rename lock right away */
1303 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1308 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1309 * target. Source should not be ancestor of target dir. May be other rename
1310 * checks can be moved here later.
1312 static int mdt_is_subdir(struct mdt_thread_info *info,
1313 struct mdt_object *dir,
1314 const struct lu_fid *fid)
1316 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1320 /* If the source and target are in the same directory, they can not
1321 * be parent/child relationship, so subdir check is not needed */
1322 if (lu_fid_eq(&dir_fid, fid))
1325 if (!mdt_object_exists(dir))
1328 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1331 CERROR("%s: failed subdir check in "DFID" for "DFID
1332 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1333 PFID(&dir_fid), PFID(fid), rc);
1334 /* Return EINVAL only if a parent is the @fid */
1338 /* check the found fid */
1339 if (lu_fid_eq(&dir_fid, fid))
1346 /* Update object linkEA */
1347 struct mdt_lock_list {
1348 struct mdt_object *mll_obj;
1349 struct mdt_lock_handle mll_lh;
1350 struct list_head mll_list;
1353 static void mdt_unlock_list(struct mdt_thread_info *info,
1354 struct list_head *list, int rc)
1356 struct mdt_lock_list *mll;
1357 struct mdt_lock_list *mll2;
1359 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1360 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1361 list_del(&mll->mll_list);
1366 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1367 struct mdt_object *obj,
1368 struct mdt_object *pobj,
1369 struct list_head *lock_list)
1371 struct lu_buf *buf = &info->mti_big_buf;
1372 struct linkea_data ldata = { NULL };
1378 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1381 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1382 if (buf->lb_buf == NULL)
1386 rc = mdt_links_read(info, obj, &ldata);
1388 if (rc == -ENOENT || rc == -ENODATA)
1393 /* ignore the migrating parent(@pobj) */
1394 retry_count = ldata.ld_leh->leh_reccount - 1;
1397 LASSERT(ldata.ld_leh != NULL);
1398 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1399 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1400 struct mdt_device *mdt = info->mti_mdt;
1401 struct mdt_object *mdt_pobj;
1402 struct mdt_lock_list *mll;
1403 struct lu_name name;
1406 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1408 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1409 if (IS_ERR(mdt_pobj)) {
1410 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1411 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1415 if (!mdt_object_exists(mdt_pobj)) {
1416 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1417 mdt_obd_name(mdt), PFID(&fid));
1418 mdt_object_put(info->mti_env, mdt_pobj);
1422 /* Check if the object already exists in the list */
1423 list_for_each_entry(mll, lock_list, mll_list) {
1424 if (mll->mll_obj == mdt_pobj) {
1425 mdt_object_put(info->mti_env, mdt_pobj);
1430 if (mdt_pobj == pobj) {
1431 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1432 mdt_obd_name(mdt), PFID(&fid));
1433 mdt_object_put(info->mti_env, mdt_pobj);
1439 mdt_object_put(info->mti_env, mdt_pobj);
1440 GOTO(out, rc = -ENOMEM);
1443 /* Since this needs to lock all of objects in linkea, to avoid
1444 * deadlocks, because it does not follow parent-child order as
1445 * other MDT operation, let's use try_lock here and if the lock
1446 * cannot be gotten because of conflicting locks, then drop all
1447 * current locks, send an AST to the client, and start again. */
1448 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1449 rc = mdt_reint_object_lock_try(info, mdt_pobj, &mll->mll_lh,
1450 MDS_INODELOCK_UPDATE, true);
1452 mdt_unlock_list(info, lock_list, rc);
1454 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1455 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1458 if (retry_count == 0) {
1459 mdt_object_put(info->mti_env, mdt_pobj);
1461 GOTO(out, rc = -EBUSY);
1464 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1465 MDS_INODELOCK_UPDATE);
1467 mdt_object_put(info->mti_env, mdt_pobj);
1472 if (mdt_object_remote(mdt_pobj)) {
1473 struct ldlm_lock *lock;
1475 /* For remote object, Set lock to cb_atomic,
1476 * so lock can be released in blocking_ast()
1477 * immediately, then the next try_lock will
1478 * have better chance to succeds */
1480 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1481 LASSERT(lock != NULL);
1482 lock_res_and_lock(lock);
1483 ldlm_set_atomic_cb(lock);
1484 unlock_res_and_lock(lock);
1485 LDLM_LOCK_PUT(lock);
1487 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1493 INIT_LIST_HEAD(&mll->mll_list);
1494 mll->mll_obj = mdt_pobj;
1495 list_add_tail(&mll->mll_list, lock_list);
1497 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1502 mdt_unlock_list(info, lock_list, rc);
1506 /* migrate files from one MDT to another MDT */
1507 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1508 struct mdt_lock_handle *lhc)
1510 struct mdt_reint_record *rr = &info->mti_rr;
1511 struct md_attr *ma = &info->mti_attr;
1512 struct mdt_object *msrcdir;
1513 struct mdt_object *mold;
1514 struct mdt_object *mnew = NULL;
1515 struct mdt_lock_handle *lh_dirp;
1516 struct mdt_lock_handle *lh_childp;
1517 struct mdt_lock_handle *lh_tgtp = NULL;
1518 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1519 struct list_head lock_list;
1521 struct ldlm_lock *lease = NULL;
1522 bool lock_open_sem = false;
1526 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1527 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1529 /* 1: lock the source dir. */
1530 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1531 if (IS_ERR(msrcdir)) {
1532 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1533 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1534 (int)PTR_ERR(msrcdir));
1535 RETURN(PTR_ERR(msrcdir));
1538 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1539 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1540 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1543 GOTO(out_put_parent, rc);
1545 if (!mdt_object_remote(msrcdir)) {
1546 rc = mdt_version_get_check_save(info, msrcdir, 0);
1548 GOTO(out_unlock_parent, rc);
1551 /* 2: sanity check and find the object to be migrated. */
1553 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1555 GOTO(out_unlock_parent, rc);
1557 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1558 GOTO(out_unlock_parent, rc = -EINVAL);
1560 if (!fid_is_md_operative(old_fid))
1561 GOTO(out_unlock_parent, rc = -EPERM);
1563 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1564 GOTO(out_unlock_parent, rc = -EPERM);
1566 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1568 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1570 if (mdt_object_remote(mold)) {
1571 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1572 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1573 GOTO(out_put_child, rc = -EREMOTE);
1576 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1577 !mdt_object_remote(msrcdir)) {
1578 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1579 " MDT, which should be migrated first:"
1580 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1581 PFID(mdt_object_fid(msrcdir)), -EPERM);
1582 GOTO(out_put_child, rc = -EPERM);
1585 rc = mdt_remote_permission(info);
1587 GOTO(out_put_child, rc);
1589 /* 3: iterate the linkea of the object and lock all of the objects */
1590 INIT_LIST_HEAD(&lock_list);
1591 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1593 GOTO(out_put_child, rc);
1595 if (info->mti_spec.sp_migrate_close) {
1596 struct close_data *data;
1597 struct mdt_body *repbody;
1598 bool lease_broken = false;
1600 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1602 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1604 GOTO(out_lease, rc = -EPROTO);
1606 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1608 GOTO(out_lease, rc = -EPROTO);
1610 lease = ldlm_handle2lock(&data->cd_handle);
1612 GOTO(out_lease, rc = -ESTALE);
1614 /* try to hold open_sem so that nobody else can open the file */
1615 if (!down_write_trylock(&mold->mot_open_sem)) {
1616 ldlm_lock_cancel(lease);
1617 GOTO(out_lease, rc = -EBUSY);
1620 lock_open_sem = true;
1621 /* Check if the lease open lease has already canceled */
1622 lock_res_and_lock(lease);
1623 lease_broken = ldlm_is_cancel(lease);
1624 unlock_res_and_lock(lease);
1626 LDLM_DEBUG(lease, DFID " lease broken? %d",
1627 PFID(mdt_object_fid(mold)), lease_broken);
1629 /* Cancel server side lease. Client side counterpart should
1630 * have been cancelled. It's okay to cancel it now as we've
1631 * held mot_open_sem. */
1632 ldlm_lock_cancel(lease);
1635 GOTO(out_lease, rc = -EAGAIN);
1637 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1638 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1639 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1641 GOTO(out_unlock_list, rc);
1644 /* 4: lock of the object migrated object */
1645 lh_childp = &info->mti_lh[MDT_LH_OLD];
1646 mdt_lock_reg_init(lh_childp, LCK_EX);
1647 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1648 MDS_INODELOCK_LAYOUT;
1649 if (mdt_object_remote(msrcdir)) {
1650 /* Enqueue lookup lock from the parent MDT */
1651 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1652 &lh_childp->mlh_rreg_lh,
1653 lh_childp->mlh_rreg_mode,
1654 MDS_INODELOCK_LOOKUP, false, false);
1656 GOTO(out_unlock_list, rc);
1658 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1661 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1663 GOTO(out_unlock_child, rc);
1665 /* Migration is incompatible with HSM. */
1666 ma->ma_need = MA_HSM;
1668 rc = mdt_attr_get_complex(info, mold, ma);
1670 GOTO(out_unlock_child, rc);
1672 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1675 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1676 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1677 GOTO(out_unlock_child, rc);
1680 ma->ma_need = MA_LMV;
1682 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1683 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1684 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1686 GOTO(out_unlock_child, rc);
1688 if ((ma->ma_valid & MA_LMV)) {
1689 struct lmv_mds_md_v1 *lmm1;
1691 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1692 lmm1 = &ma->ma_lmv->lmv_md_v1;
1693 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1694 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1695 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1696 PFID(mdt_object_fid(mold)), -EPERM);
1697 GOTO(out_unlock_child, rc = -EPERM);
1700 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1701 GOTO(out_unlock_child, rc = -EINVAL);
1703 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1704 &lmm1->lmv_stripe_fids[1]);
1706 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1708 if (!mdt_object_remote(mnew)) {
1710 "%s: "DFID" being migrated is on this MDT:"
1711 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1712 PFID(rr->rr_fid2), -EPERM);
1713 GOTO(out_put_new, rc = -EPERM);
1716 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1717 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1718 rc = mdt_remote_object_lock(info, mnew,
1719 mdt_object_fid(mnew),
1720 &lh_tgtp->mlh_rreg_lh,
1721 lh_tgtp->mlh_rreg_mode,
1722 MDS_INODELOCK_UPDATE, false, false);
1725 GOTO(out_put_new, rc);
1728 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1731 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1732 if (!mdt_object_remote(mnew)) {
1733 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1734 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1735 PFID(rr->rr_fid2), -EXDEV);
1736 GOTO(out_put_new, rc = -EXDEV);
1741 mdt_reint_init_ma(info, ma);
1743 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1744 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1746 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1747 mdt_object_child(mold), &rr->rr_name,
1748 mdt_object_child(mnew), ma);
1750 GOTO(out_unlock_new, rc);
1753 if (lh_tgtp != NULL)
1754 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1757 mdt_object_put(info->mti_env, mnew);
1759 mdt_object_unlock(info, mold, lh_childp, rc);
1761 /* we don't really modify linkea objects, so we can safely decref these
1762 * locks, and this can avoid saving them as COS locks, which may prevent
1763 * subsequent migrate. */
1764 mdt_unlock_list(info, &lock_list, 1);
1765 if (lease != NULL) {
1766 ldlm_reprocess_all(lease->l_resource);
1767 LDLM_LOCK_PUT(lease);
1771 up_write(&mold->mot_open_sem);
1773 mdt_object_put(info->mti_env, mold);
1775 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1777 mdt_object_put(info->mti_env, msrcdir);
1782 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1783 const struct lu_fid *fid,
1786 struct mdt_object *dir;
1790 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1794 /* check early, the real version will be saved after locking */
1795 rc = mdt_version_get_check(info, dir, idx);
1801 mdt_object_put(info->mti_env, dir);
1805 static int mdt_object_lock_save(struct mdt_thread_info *info,
1806 struct mdt_object *dir,
1807 struct mdt_lock_handle *lh,
1808 int idx, bool cos_incompat)
1812 /* we lock the target dir if it is local */
1813 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1818 /* get and save correct version after locking */
1819 mdt_version_get_save(info, dir, idx);
1824 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1825 * 2 - srcdir child; 3 - tgtdir child.
1826 * Update on disk version of srcdir child.
1829 * For DNE phase I, only these renames are allowed
1830 * mv src_p/src_c tgt_p/tgt_c
1831 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1832 * 2. src_p and tgt_p are same directory, and tgt_c does not
1833 * exists. In this case, all of modification will happen
1834 * in the MDT where ithesource parent is, only one remote
1835 * update is needed, i.e. set c_time/m_time on the child.
1836 * And tgt_c will be still in the same MDT as the original
1839 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1840 struct mdt_lock_handle *lhc)
1842 struct mdt_reint_record *rr = &info->mti_rr;
1843 struct md_attr *ma = &info->mti_attr;
1844 struct ptlrpc_request *req = mdt_info_req(info);
1845 struct mdt_object *msrcdir = NULL;
1846 struct mdt_object *mtgtdir = NULL;
1847 struct mdt_object *mold;
1848 struct mdt_object *mnew = NULL;
1849 struct mdt_lock_handle *lh_srcdirp;
1850 struct mdt_lock_handle *lh_tgtdirp;
1851 struct mdt_lock_handle *lh_oldp = NULL;
1852 struct mdt_lock_handle *lh_newp = NULL;
1853 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1854 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1856 bool reverse = false;
1861 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1862 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1863 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1865 /* find both parents. */
1866 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1867 if (IS_ERR(msrcdir))
1868 RETURN(PTR_ERR(msrcdir));
1870 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1872 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1874 mdt_object_get(info->mti_env, mtgtdir);
1876 /* Check if the @msrcdir is not a child of the @mtgtdir,
1877 * otherwise a reverse locking must take place. */
1878 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1882 GOTO(out_put_srcdir, rc);
1884 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1885 if (IS_ERR(mtgtdir))
1886 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1889 /* source needs to be looked up after locking source parent, otherwise
1890 * this rename may race with unlink source, and cause rename hang, see
1891 * sanityn.sh 55b, so check parents first, if later we found source is
1892 * remote, relock parents. */
1893 cos_incompat = (mdt_object_remote(msrcdir) ||
1894 mdt_object_remote(mtgtdir));
1896 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1898 /* lock parents in the proper order. */
1899 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1900 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1903 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1904 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1907 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1910 GOTO(out_put_tgtdir, rc);
1912 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1914 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1917 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1918 GOTO(out_put_tgtdir, rc);
1921 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1924 GOTO(out_put_tgtdir, rc);
1926 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1928 if (mtgtdir != msrcdir) {
1929 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1931 } else if (lh_srcdirp->mlh_pdo_hash !=
1932 lh_tgtdirp->mlh_pdo_hash) {
1933 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1934 MDS_INODELOCK_UPDATE,
1936 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1939 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1940 GOTO(out_put_tgtdir, rc);
1944 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1945 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1947 /* find mold object. */
1949 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1951 GOTO(out_unlock_parents, rc);
1953 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1954 GOTO(out_unlock_parents, rc = -EINVAL);
1956 if (!fid_is_md_operative(old_fid))
1957 GOTO(out_unlock_parents, rc = -EPERM);
1959 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1961 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1963 /* Check if @mtgtdir is subdir of @mold, before locking child
1964 * to avoid reverse locking. */
1965 if (mtgtdir != msrcdir) {
1966 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1968 GOTO(out_put_old, rc);
1971 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1972 /* save version after locking */
1973 mdt_version_get_save(info, mold, 2);
1975 if (!cos_incompat && mdt_object_remote(mold)) {
1976 cos_incompat = true;
1977 mdt_object_put(info->mti_env, mold);
1978 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1979 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1983 /* find mnew object:
1984 * mnew target object may not exist now
1985 * lookup with version checking */
1987 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1990 /* the new_fid should have been filled at this moment */
1991 if (lu_fid_eq(old_fid, new_fid))
1992 GOTO(out_put_old, rc);
1994 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1995 lu_fid_eq(new_fid, rr->rr_fid2))
1996 GOTO(out_put_old, rc = -EINVAL);
1998 if (!fid_is_md_operative(new_fid))
1999 GOTO(out_put_old, rc = -EPERM);
2001 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2003 GOTO(out_put_old, rc = PTR_ERR(mnew));
2005 if (mdt_object_remote(mnew)) {
2006 struct mdt_body *repbody;
2008 /* Always send rename req to the target child MDT */
2009 repbody = req_capsule_server_get(info->mti_pill,
2011 LASSERT(repbody != NULL);
2012 repbody->mbo_fid1 = *new_fid;
2013 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2014 GOTO(out_put_new, rc = -EXDEV);
2016 /* Before locking the target dir, check we do not replace
2017 * a dir with a non-dir, otherwise it may deadlock with
2018 * link op which tries to create a link in this dir
2019 * back to this non-dir. */
2020 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2021 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2022 GOTO(out_put_new, rc = -EISDIR);
2024 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2025 mdt_lock_reg_init(lh_oldp, LCK_EX);
2026 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2027 if (mdt_object_remote(msrcdir)) {
2028 /* Enqueue lookup lock from the parent MDT */
2029 rc = mdt_remote_object_lock(info, msrcdir,
2030 mdt_object_fid(mold),
2031 &lh_oldp->mlh_rreg_lh,
2032 lh_oldp->mlh_rreg_mode,
2033 MDS_INODELOCK_LOOKUP,
2036 GOTO(out_put_new, rc);
2038 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2041 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2044 GOTO(out_unlock_old, rc);
2046 /* Check if @msrcdir is subdir of @mnew, before locking child
2047 * to avoid reverse locking. */
2048 if (mtgtdir != msrcdir) {
2049 rc = mdt_is_subdir(info, msrcdir, new_fid);
2051 GOTO(out_unlock_old, rc);
2054 /* We used to acquire MDS_INODELOCK_FULL here but we
2055 * can't do this now because a running HSM restore on
2056 * the rename onto victim will hold the layout
2057 * lock. See LU-4002. */
2059 lh_newp = &info->mti_lh[MDT_LH_NEW];
2060 mdt_lock_reg_init(lh_newp, LCK_EX);
2061 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2062 MDS_INODELOCK_LOOKUP |
2063 MDS_INODELOCK_UPDATE,
2066 GOTO(out_unlock_old, rc);
2068 /* get and save version after locking */
2069 mdt_version_get_save(info, mnew, 3);
2070 } else if (rc != -EREMOTE && rc != -ENOENT) {
2071 GOTO(out_put_old, rc);
2073 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2074 mdt_lock_reg_init(lh_oldp, LCK_EX);
2075 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2076 if (mdt_object_remote(msrcdir)) {
2077 /* Enqueue lookup lock from the parent MDT */
2078 rc = mdt_remote_object_lock(info, msrcdir,
2079 mdt_object_fid(mold),
2080 &lh_oldp->mlh_rreg_lh,
2081 lh_oldp->mlh_rreg_mode,
2082 MDS_INODELOCK_LOOKUP,
2085 GOTO(out_put_old, rc);
2087 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2090 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2093 GOTO(out_unlock_old, rc);
2095 mdt_enoent_version_save(info, 3);
2098 /* step 5: rename it */
2099 mdt_reint_init_ma(info, ma);
2101 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2102 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2105 mutex_lock(&mnew->mot_lov_mutex);
2107 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2108 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2109 mnew != NULL ? mdt_object_child(mnew) : NULL,
2110 &rr->rr_tgt_name, ma);
2113 mutex_unlock(&mnew->mot_lov_mutex);
2115 /* handle last link of tgt object */
2117 mdt_counter_incr(req, LPROC_MDT_RENAME);
2119 mdt_handle_last_unlink(info, mnew, ma);
2121 mdt_rename_counter_tally(info, info->mti_mdt, req,
2127 mdt_object_unlock(info, mnew, lh_newp, rc);
2129 mdt_object_unlock(info, mold, lh_oldp, rc);
2132 mdt_object_put(info->mti_env, mnew);
2134 mdt_object_put(info->mti_env, mold);
2136 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2137 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2139 mdt_object_put(info->mti_env, mtgtdir);
2141 mdt_object_put(info->mti_env, msrcdir);
2145 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2146 struct mdt_lock_handle *lhc, bool rename)
2148 struct mdt_reint_record *rr = &info->mti_rr;
2149 struct ptlrpc_request *req = mdt_info_req(info);
2150 struct lustre_handle rename_lh = { 0 };
2154 if (info->mti_dlm_req)
2155 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2157 if (!fid_is_md_operative(rr->rr_fid1) ||
2158 !fid_is_md_operative(rr->rr_fid2))
2161 /* Note: do not enqueue rename lock for replay request, because
2162 * if other MDT holds rename lock, but being blocked to wait for
2163 * this MDT to finish its recovery, and the failover MDT can not
2164 * get rename lock, which will cause deadlock. */
2165 if (!req_is_replay(req)) {
2166 rc = mdt_rename_lock(info, &rename_lh);
2168 CERROR("%s: can't lock FS for rename: rc = %d\n",
2169 mdt_obd_name(info->mti_mdt), rc);
2175 rc = mdt_reint_rename_internal(info, lhc);
2177 rc = mdt_reint_migrate_internal(info, lhc);
2179 if (lustre_handle_is_used(&rename_lh))
2180 mdt_rename_unlock(&rename_lh);
2185 static int mdt_reint_rename(struct mdt_thread_info *info,
2186 struct mdt_lock_handle *lhc)
2188 return mdt_reint_rename_or_migrate(info, lhc, true);
2191 static int mdt_reint_migrate(struct mdt_thread_info *info,
2192 struct mdt_lock_handle *lhc)
2194 return mdt_reint_rename_or_migrate(info, lhc, false);
2197 struct mdt_reinter {
2198 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2199 enum lprocfs_extra_opc mr_extra_opc;
2202 static const struct mdt_reinter mdt_reinters[] = {
2204 .mr_handler = &mdt_reint_setattr,
2205 .mr_extra_opc = MDS_REINT_SETATTR,
2208 .mr_handler = &mdt_reint_create,
2209 .mr_extra_opc = MDS_REINT_CREATE,
2212 .mr_handler = &mdt_reint_link,
2213 .mr_extra_opc = MDS_REINT_LINK,
2216 .mr_handler = &mdt_reint_unlink,
2217 .mr_extra_opc = MDS_REINT_UNLINK,
2220 .mr_handler = &mdt_reint_rename,
2221 .mr_extra_opc = MDS_REINT_RENAME,
2224 .mr_handler = &mdt_reint_open,
2225 .mr_extra_opc = MDS_REINT_OPEN,
2227 [REINT_SETXATTR] = {
2228 .mr_handler = &mdt_reint_setxattr,
2229 .mr_extra_opc = MDS_REINT_SETXATTR,
2232 .mr_handler = &mdt_reint_unlink,
2233 .mr_extra_opc = MDS_REINT_UNLINK,
2236 .mr_handler = &mdt_reint_migrate,
2237 .mr_extra_opc = MDS_REINT_RENAME,
2241 int mdt_reint_rec(struct mdt_thread_info *info,
2242 struct mdt_lock_handle *lhc)
2244 const struct mdt_reinter *mr;
2248 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2251 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2252 if (mr->mr_handler == NULL)
2255 rc = (*mr->mr_handler)(info, lhc);
2257 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2258 PTLRPC_LAST_CNTR + mr->mr_extra_opc);