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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_reint.c
34 * Lustre Metadata Target (mdt) reintegration routines
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Huang Hua <huanghua@clusterfs.com>
40 * Author: Yury Umanets <umka@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lprocfs_status.h>
46 #include "mdt_internal.h"
47 #include <lustre_lmv.h>
49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
52 ma->ma_need = MA_INODE;
57 * Get version of object by fid.
59 * Return real version or ENOENT_VERSION if object doesn't exist
61 static void mdt_obj_version_get(struct mdt_thread_info *info,
62 struct mdt_object *o, __u64 *version)
66 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
67 !fid_is_obf(mdt_object_fid(o)))
68 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
70 *version = ENOENT_VERSION;
71 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
72 PFID(mdt_object_fid(o)), *version);
76 * Check version is correct.
78 * Should be called only during replay.
80 static int mdt_version_check(struct ptlrpc_request *req,
81 __u64 version, int idx)
83 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
86 if (!exp_connect_vbr(req->rq_export))
89 LASSERT(req_is_replay(req));
90 /** VBR: version is checked always because costs nothing */
91 LASSERT(idx < PTLRPC_NUM_VERSIONS);
92 /** Sanity check for malformed buffers */
93 if (pre_ver == NULL) {
94 CERROR("No versions in request buffer\n");
95 spin_lock(&req->rq_export->exp_lock);
96 req->rq_export->exp_vbr_failed = 1;
97 spin_unlock(&req->rq_export->exp_lock);
99 } else if (pre_ver[idx] != version) {
100 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
101 pre_ver[idx], version);
102 spin_lock(&req->rq_export->exp_lock);
103 req->rq_export->exp_vbr_failed = 1;
104 spin_unlock(&req->rq_export->exp_lock);
111 * Save pre-versions in reply.
113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
118 if (!exp_connect_vbr(req->rq_export))
121 LASSERT(!req_is_replay(req));
122 LASSERT(req->rq_repmsg != NULL);
123 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
125 reply_ver[idx] = version;
129 * Save enoent version, it is needed when it is obvious that object doesn't
130 * exist, e.g. child during create.
132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
134 /* save version of file name for replay, it must be ENOENT here */
135 if (!req_is_replay(mdt_info_req(info))) {
136 info->mti_ver[idx] = ENOENT_VERSION;
137 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
142 * Get version from disk and save in reply buffer.
144 * Versions are saved in reply only during normal operations not replays.
146 void mdt_version_get_save(struct mdt_thread_info *info,
147 struct mdt_object *mto, int idx)
149 /* don't save versions during replay */
150 if (!req_is_replay(mdt_info_req(info))) {
151 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
152 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
157 * Get version from disk and check it, no save in reply.
159 int mdt_version_get_check(struct mdt_thread_info *info,
160 struct mdt_object *mto, int idx)
162 /* only check versions during replay */
163 if (!req_is_replay(mdt_info_req(info)))
166 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
167 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and check if recovery or just save.
173 int mdt_version_get_check_save(struct mdt_thread_info *info,
174 struct mdt_object *mto, int idx)
178 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
179 if (req_is_replay(mdt_info_req(info)))
180 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
183 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
188 * Lookup with version checking.
190 * This checks version of 'name'. Many reint functions uses 'name' for child not
191 * FID, therefore we need to get object by name and check its version.
193 static int mdt_lookup_version_check(struct mdt_thread_info *info,
194 struct mdt_object *p,
195 const struct lu_name *lname,
196 struct lu_fid *fid, int idx)
200 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
202 /* Check version only during replay */
203 if (!req_is_replay(mdt_info_req(info)))
206 info->mti_ver[idx] = ENOENT_VERSION;
208 struct mdt_object *child;
209 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
210 if (likely(!IS_ERR(child))) {
211 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
212 mdt_object_put(info->mti_env, child);
215 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
216 return vbrc ? vbrc : rc;
220 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
222 struct lu_ucred *uc = mdt_ucred(info);
223 struct mdt_device *mdt = info->mti_mdt;
225 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
226 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
227 mdt->mdt_enable_remote_dir_gid != -1)
235 * mdt_remote_permission: Check whether the remote operation is permitted,
237 * Only sysadmin can create remote directory / striped directory,
238 * migrate directory and set default stripedEA on directory, unless
240 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
242 * param[in] info: mdt_thread_info.
244 * retval = 0 remote operation is allowed.
245 * < 0 remote operation is denied.
247 static int mdt_remote_permission(struct mdt_thread_info *info)
249 struct md_op_spec *spec = &info->mti_spec;
250 struct lu_attr *attr = &info->mti_attr.ma_attr;
251 struct obd_export *exp = mdt_info_req(info)->rq_export;
254 if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
255 rc = mdt_remote_permission_check(info);
260 if (info->mti_rr.rr_opcode == REINT_CREATE &&
261 (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
262 spec->u.sp_ea.eadatalen != 0)) {
263 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
265 /* Only new clients can create remote dir( >= 2.4) and
266 * striped dir(>= 2.6), old client will return -ENOTSUPP */
267 if (!mdt_is_dne_client(exp))
270 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
271 !mdt_is_striped_client(exp))
274 rc = mdt_remote_permission_check(info);
279 if (info->mti_rr.rr_opcode == REINT_SETATTR) {
280 struct md_attr *ma = &info->mti_attr;
282 if ((ma->ma_valid & MA_LMV)) {
283 rc = mdt_remote_permission_check(info);
292 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
293 struct mdt_object *obj, __u64 ibits,
294 struct mdt_lock_handle *s0_lh,
295 struct mdt_object *s0_obj,
296 struct ldlm_enqueue_info *einfo,
299 union ldlm_policy_data *policy = &mti->mti_policy;
300 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
305 if (!S_ISDIR(obj->mot_header.loh_attr))
308 /* Unlock stripe 0 */
309 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
310 LASSERT(s0_obj != NULL);
311 mdt_object_unlock_put(mti, s0_obj, s0_lh, decref);
314 memset(policy, 0, sizeof(*policy));
315 policy->l_inodebits.bits = ibits;
317 if (slave_locks != NULL) {
318 LASSERT(s0_lh != NULL);
319 for (i = 1; i < slave_locks->count; i++) {
320 /* borrow s0_lh temporarily to do mdt unlock */
321 mdt_lock_reg_init(s0_lh, einfo->ei_mode);
322 s0_lh->mlh_rreg_lh = slave_locks->handles[i];
323 mdt_object_unlock(mti, NULL, s0_lh, decref);
324 slave_locks->handles[i].cookie = 0ull;
328 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
333 static int mdt_init_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
336 struct lu_buf *buf = &mti->mti_buf;
337 struct lmv_mds_md_v1 *lmv;
341 if (!S_ISDIR(obj->mot_header.loh_attr))
344 buf->lb_buf = mti->mti_xattr_buf;
345 buf->lb_len = sizeof(mti->mti_xattr_buf);
346 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
349 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
351 buf->lb_buf = mti->mti_big_lmm;
352 buf->lb_len = mti->mti_big_lmmsize;
356 if (rc == -ENODATA || rc == -ENOENT)
363 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
366 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
372 * Lock slave stripes if necessary, the lock handles of slave stripes
373 * will be stored in einfo->ei_cbdata.
375 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
376 enum ldlm_mode mode, __u64 ibits,
377 struct lu_fid *s0_fid,
378 struct mdt_lock_handle *s0_lh,
379 struct mdt_object **s0_objp,
380 struct ldlm_enqueue_info *einfo)
382 union ldlm_policy_data *policy = &mti->mti_policy;
386 memset(einfo, 0, sizeof(*einfo));
388 rc = mdt_init_slaves(mti, obj, s0_fid);
392 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
394 if (!lu_fid_eq(s0_fid, mdt_object_fid(obj))) {
395 /* Except migrating object, whose 0_stripe and master
396 * object are the same object, 0_stripe and master
397 * object are different, though they are in the same
398 * MDT, to avoid adding osd_object_lock here, so we
399 * will enqueue the stripe0 lock in MDT0 for now */
400 *s0_objp = mdt_object_find(mti->mti_env, mti->mti_mdt, s0_fid);
401 if (IS_ERR(*s0_objp))
402 RETURN(PTR_ERR(*s0_objp));
404 rc = mdt_reint_object_lock(mti, *s0_objp, s0_lh, ibits, true);
406 mdt_object_put(mti->mti_env, *s0_objp);
411 einfo->ei_type = LDLM_IBITS;
412 einfo->ei_mode = mode;
413 einfo->ei_cb_bl = mdt_remote_blocking_ast;
414 einfo->ei_cb_local_bl = mdt_blocking_ast;
415 einfo->ei_cb_cp = ldlm_completion_ast;
416 einfo->ei_enq_slave = 1;
417 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
418 memset(policy, 0, sizeof(*policy));
419 policy->l_inodebits.bits = ibits;
421 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
427 * VBR: we save three versions in reply:
428 * 0 - parent. Check that parent version is the same during replay.
429 * 1 - name. Version of 'name' if file exists with the same name or
430 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
431 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
434 static int mdt_md_create(struct mdt_thread_info *info)
436 struct mdt_device *mdt = info->mti_mdt;
437 struct mdt_object *parent;
438 struct mdt_object *child;
439 struct mdt_lock_handle *lh;
440 struct mdt_body *repbody;
441 struct md_attr *ma = &info->mti_attr;
442 struct mdt_reint_record *rr = &info->mti_rr;
446 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
448 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
450 if (!fid_is_md_operative(rr->rr_fid1))
453 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
455 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
457 RETURN(PTR_ERR(parent));
459 if (!mdt_object_exists(parent))
460 GOTO(put_parent, rc = -ENOENT);
462 lh = &info->mti_lh[MDT_LH_PARENT];
463 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
464 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
466 GOTO(put_parent, rc);
468 if (!mdt_object_remote(parent)) {
469 rc = mdt_version_get_check_save(info, parent, 0);
471 GOTO(unlock_parent, rc);
475 * Check child name version during replay.
476 * During create replay a file may exist with same name.
478 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
479 &info->mti_tmp_fid1, 1);
481 GOTO(unlock_parent, rc = -EEXIST);
483 /* -ENOENT is expected here */
485 GOTO(unlock_parent, rc);
487 /* save version of file name for replay, it must be ENOENT here */
488 mdt_enoent_version_save(info, 1);
490 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
491 if (unlikely(IS_ERR(child)))
492 GOTO(unlock_parent, rc = PTR_ERR(child));
494 rc = mdt_remote_permission(info);
498 ma->ma_need = MA_INODE;
501 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
502 OBD_FAIL_MDS_REINT_CREATE_WRITE);
504 /* Version of child will be updated on disk. */
505 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
506 rc = mdt_version_get_check_save(info, child, 2);
510 /* Let lower layer know current lock mode. */
511 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
514 * Do not perform lookup sanity check. We know that name does
517 info->mti_spec.sp_cr_lookup = 0;
518 info->mti_spec.sp_feat = &dt_directory_features;
520 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
521 mdt_object_child(child), &info->mti_spec, ma);
523 rc = mdt_attr_get_complex(info, child, ma);
529 * On DNE, we need to eliminate dependey between 'mkdir a' and
530 * 'mkdir a/b' if b is a striped directory, to achieve this, two
531 * things are done below:
532 * 1. save child and slaves lock.
533 * 2. if the child is a striped directory, relock parent so to
534 * compare against with COS locks to ensure parent was
537 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
538 struct mdt_lock_handle *lhc;
539 struct mdt_lock_handle *s0_lh;
540 struct mdt_object *s0_obj = NULL;
541 struct ldlm_enqueue_info *einfo;
542 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
543 bool cos_incompat = false;
545 rc = mdt_init_slaves(info, child, s0_fid);
548 if (!mdt_object_remote(parent)) {
549 mdt_object_unlock(info, parent, lh, 1);
550 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
551 rc = mdt_reint_object_lock(info, parent, lh,
552 MDS_INODELOCK_UPDATE,
559 einfo = &info->mti_einfo;
560 lhc = &info->mti_lh[MDT_LH_CHILD];
561 mdt_lock_handle_init(lhc);
562 mdt_lock_reg_init(lhc, LCK_PW);
563 rc = mdt_reint_object_lock(info, child, lhc,
564 MDS_INODELOCK_UPDATE,
568 mdt_object_unlock(info, child, lhc, rc);
570 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
571 mdt_lock_handle_init(s0_lh);
572 mdt_lock_reg_init(s0_lh, LCK_PW);
573 rc = mdt_lock_slaves(info, child, LCK_PW, MDS_INODELOCK_UPDATE,
574 s0_fid, s0_lh, &s0_obj, einfo);
575 mdt_unlock_slaves(info, child, MDS_INODELOCK_UPDATE, s0_lh,
577 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME) && rc == -EIO)
581 /* Return fid & attr to client. */
582 if (ma->ma_valid & MA_INODE)
583 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
584 mdt_object_fid(child));
586 mdt_object_put(info->mti_env, child);
588 mdt_object_unlock(info, parent, lh, rc);
590 mdt_object_put(info->mti_env, parent);
594 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
597 struct mdt_lock_handle *lh;
598 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
599 __u64 lockpart = MDS_INODELOCK_UPDATE;
600 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
601 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
602 struct mdt_lock_handle *s0_lh = NULL;
603 struct mdt_object *s0_obj = NULL;
604 bool cos_incompat = false;
608 rc = mdt_init_slaves(info, mo, s0_fid);
612 lh = &info->mti_lh[MDT_LH_PARENT];
613 mdt_lock_reg_init(lh, LCK_PW);
615 /* Even though the new MDT will grant PERM lock to the old
616 * client, but the old client will almost ignore that during
617 * So it needs to revoke both LOOKUP and PERM lock here, so
618 * both new and old client can cancel the dcache */
619 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
620 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
622 rc = mdt_reint_object_lock(info, mo, lh, lockpart, cos_incompat);
626 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
627 mdt_lock_reg_init(s0_lh, LCK_PW);
628 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_fid, s0_lh, &s0_obj,
631 GOTO(out_unlock, rc);
633 /* all attrs are packed into mti_attr in unpack_setattr */
634 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
635 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
637 /* This is only for set ctime when rename's source is on remote MDS. */
638 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
639 ma->ma_attr_flags |= MDS_VTX_BYPASS;
641 /* VBR: update version if attr changed are important for recovery */
643 /* update on-disk version of changed object */
644 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
645 rc = mdt_version_get_check_save(info, mo, 0);
647 GOTO(out_unlock, rc);
650 /* Ensure constant striping during chown(). See LU-2789. */
651 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
652 mutex_lock(&mo->mot_lov_mutex);
654 /* all attrs are packed into mti_attr in unpack_setattr */
655 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
657 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
658 mutex_unlock(&mo->mot_lov_mutex);
661 GOTO(out_unlock, rc);
665 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo, rc);
666 mdt_object_unlock(info, mo, lh, rc);
671 * Check HSM flags and add HS_DIRTY flag if relevant.
673 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
674 * and is not RELEASED.
676 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
682 /* If the file was modified, add the dirty flag */
683 ma->ma_need = MA_HSM;
684 rc = mdt_attr_get_complex(info, mo, ma);
686 CERROR("file attribute read error for "DFID": %d.\n",
687 PFID(mdt_object_fid(mo)), rc);
691 /* If an up2date copy exists in the backend, add dirty flag */
692 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
693 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
694 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
696 ma->ma_hsm.mh_flags |= HS_DIRTY;
698 mdt_lock_reg_init(lh, LCK_PW);
699 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
703 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
705 CERROR("file attribute change error for "DFID": %d\n",
706 PFID(mdt_object_fid(mo)), rc);
707 mdt_object_unlock(info, mo, lh, rc);
713 static int mdt_reint_setattr(struct mdt_thread_info *info,
714 struct mdt_lock_handle *lhc)
716 struct md_attr *ma = &info->mti_attr;
717 struct mdt_reint_record *rr = &info->mti_rr;
718 struct ptlrpc_request *req = mdt_info_req(info);
719 struct mdt_object *mo;
720 struct mdt_body *repbody;
724 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
725 (unsigned int)ma->ma_attr.la_valid);
727 if (info->mti_dlm_req)
728 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
730 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
731 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
733 GOTO(out, rc = PTR_ERR(mo));
735 if (!mdt_object_exists(mo))
736 GOTO(out_put, rc = -ENOENT);
738 if (mdt_object_remote(mo))
739 GOTO(out_put, rc = -EREMOTE);
741 if ((ma->ma_attr.la_valid & LA_SIZE) ||
742 (rr->rr_flags & MRF_OPEN_TRUNC)) {
743 /* Check write access for the O_TRUNC case */
744 if (mdt_write_read(mo) < 0)
745 GOTO(out_put, rc = -ETXTBSY);
748 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
749 if (ma->ma_valid & MA_LOV)
750 GOTO(out_put, rc = -EPROTO);
752 rc = mdt_attr_set(info, mo, ma);
755 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
756 (ma->ma_valid & MA_INODE)) {
757 struct lu_buf *buf = &info->mti_buf;
758 struct mdt_lock_handle *lh;
760 rc = mdt_remote_permission(info);
764 if (ma->ma_attr.la_valid != 0)
765 GOTO(out_put, rc = -EPROTO);
767 lh = &info->mti_lh[MDT_LH_PARENT];
768 mdt_lock_reg_init(lh, LCK_PW);
770 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
774 if (ma->ma_valid & MA_LOV) {
775 buf->lb_buf = ma->ma_lmm;
776 buf->lb_len = ma->ma_lmm_size;
778 buf->lb_buf = ma->ma_lmv;
779 buf->lb_len = ma->ma_lmv_size;
781 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
782 (ma->ma_valid & MA_LOV) ?
783 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
786 mdt_object_unlock(info, mo, lh, rc);
790 GOTO(out_put, rc = -EPROTO);
793 /* If file data is modified, add the dirty flag */
794 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
795 rc = mdt_add_dirty_flag(info, mo, ma);
797 ma->ma_need = MA_INODE;
799 rc = mdt_attr_get_complex(info, mo, ma);
803 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
807 mdt_object_put(info->mti_env, mo);
810 mdt_counter_incr(req, LPROC_MDT_SETATTR);
812 mdt_client_compatibility(info);
813 rc2 = mdt_fix_reply(info);
819 static int mdt_reint_create(struct mdt_thread_info *info,
820 struct mdt_lock_handle *lhc)
822 struct ptlrpc_request *req = mdt_info_req(info);
826 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
827 RETURN(err_serious(-ESTALE));
829 if (info->mti_dlm_req)
830 ldlm_request_cancel(mdt_info_req(info),
831 info->mti_dlm_req, 0, LATF_SKIP);
833 if (!lu_name_is_valid(&info->mti_rr.rr_name))
836 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
838 mdt_counter_incr(req, LPROC_MDT_MKDIR);
846 /* Special file should stay on the same node as parent. */
847 mdt_counter_incr(req, LPROC_MDT_MKNOD);
850 CERROR("%s: Unsupported mode %o\n",
851 mdt_obd_name(info->mti_mdt),
852 info->mti_attr.ma_attr.la_mode);
853 RETURN(err_serious(-EOPNOTSUPP));
856 rc = mdt_md_create(info);
861 * VBR: save parent version in reply and child version getting by its name.
862 * Version of child is getting and checking during its lookup. If
864 static int mdt_reint_unlink(struct mdt_thread_info *info,
865 struct mdt_lock_handle *lhc)
867 struct mdt_reint_record *rr = &info->mti_rr;
868 struct ptlrpc_request *req = mdt_info_req(info);
869 struct md_attr *ma = &info->mti_attr;
870 struct lu_fid *child_fid = &info->mti_tmp_fid1;
871 struct mdt_object *mp;
872 struct mdt_object *mc;
873 struct mdt_lock_handle *parent_lh;
874 struct mdt_lock_handle *child_lh;
875 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
876 struct lu_fid *s0_fid = &info->mti_tmp_fid2;
877 struct mdt_lock_handle *s0_lh = NULL;
878 struct mdt_object *s0_obj = NULL;
880 bool cos_incompat = false;
885 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
886 PNAME(&rr->rr_name));
888 if (info->mti_dlm_req)
889 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
891 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
892 RETURN(err_serious(-ENOENT));
894 if (!fid_is_md_operative(rr->rr_fid1))
897 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
901 if (mdt_object_remote(mp)) {
904 rc = mdt_version_get_check_save(info, mp, 0);
906 GOTO(put_parent, rc);
910 parent_lh = &info->mti_lh[MDT_LH_PARENT];
911 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
912 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
915 GOTO(put_parent, rc);
917 /* lookup child object along with version checking */
919 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
921 /* Name might not be able to find during resend of
922 * remote unlink, considering following case.
923 * dir_A is a remote directory, the name entry of
924 * dir_A is on MDT0, the directory is on MDT1,
926 * 1. client sends unlink req to MDT1.
927 * 2. MDT1 sends name delete update to MDT0.
928 * 3. name entry is being deleted in MDT0 synchronously.
929 * 4. MDT1 is restarted.
930 * 5. client resends unlink req to MDT1. So it can not
931 * find the name entry on MDT0 anymore.
932 * In this case, MDT1 only needs to destory the local
935 if (mdt_object_remote(mp) && rc == -ENOENT &&
936 !fid_is_zero(rr->rr_fid2) &&
937 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
939 *child_fid = *rr->rr_fid2;
941 GOTO(unlock_parent, rc);
945 if (!fid_is_md_operative(child_fid))
946 GOTO(unlock_parent, rc = -EPERM);
948 /* We will lock the child regardless it is local or remote. No harm. */
949 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
951 GOTO(unlock_parent, rc = PTR_ERR(mc));
953 if (!cos_incompat && mdt_init_slaves(info, mc, s0_fid) > 0) {
955 mdt_object_put(info->mti_env, mc);
956 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
960 child_lh = &info->mti_lh[MDT_LH_CHILD];
961 mdt_lock_reg_init(child_lh, LCK_EX);
962 if (info->mti_spec.sp_rm_entry) {
963 struct lu_ucred *uc = mdt_ucred(info);
965 if (!mdt_is_dne_client(req->rq_export))
966 /* Return -ENOTSUPP for old client */
967 GOTO(put_child, rc = -ENOTSUPP);
969 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
970 GOTO(put_child, rc = -EPERM);
972 ma->ma_need = MA_INODE;
974 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
975 NULL, &rr->rr_name, ma, no_name);
979 if (mdt_object_remote(mc)) {
980 struct mdt_body *repbody;
982 if (!fid_is_zero(rr->rr_fid2)) {
983 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
984 mdt_obd_name(info->mti_mdt),
985 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
986 GOTO(put_child, rc = -ENOENT);
988 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
989 mdt_obd_name(info->mti_mdt),
990 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
992 if (!mdt_is_dne_client(req->rq_export))
993 /* Return -ENOTSUPP for old client */
994 GOTO(put_child, rc = -ENOTSUPP);
996 /* Revoke the LOOKUP lock of the remote object granted by
997 * this MDT. Since the unlink will happen on another MDT,
998 * it will release the LOOKUP lock right away. Then What
999 * would happen if another client try to grab the LOOKUP
1000 * lock at the same time with unlink XXX */
1001 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1002 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1003 LASSERT(repbody != NULL);
1004 repbody->mbo_fid1 = *mdt_object_fid(mc);
1005 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1006 GOTO(unlock_child, rc = -EREMOTE);
1008 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1009 * this now because a running HSM restore on the child (unlink
1010 * victim) will hold the layout lock. See LU-4002. */
1011 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1012 if (mdt_object_remote(mp)) {
1013 /* Enqueue lookup lock from parent MDT */
1014 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1015 &child_lh->mlh_rreg_lh,
1016 child_lh->mlh_rreg_mode,
1017 MDS_INODELOCK_LOOKUP, false, false);
1019 GOTO(put_child, rc);
1021 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1024 rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits,
1027 GOTO(unlock_child, rc);
1030 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1031 * whether need MA_LOV and MA_COOKIE.
1033 ma->ma_need = MA_INODE;
1036 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1037 mdt_lock_reg_init(s0_lh, LCK_EX);
1038 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_fid,
1039 s0_lh, &s0_obj, einfo);
1041 GOTO(unlock_child, rc);
1043 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1044 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1045 /* save version when object is locked */
1046 mdt_version_get_save(info, mc, 1);
1048 mutex_lock(&mc->mot_lov_mutex);
1050 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1051 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1053 mutex_unlock(&mc->mot_lov_mutex);
1055 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1056 rc = mdt_attr_get_complex(info, mc, ma);
1058 mdt_handle_last_unlink(info, mc, ma);
1060 if (ma->ma_valid & MA_INODE) {
1061 switch (ma->ma_attr.la_mode & S_IFMT) {
1063 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1071 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1074 LASSERTF(0, "bad file type %o unlinking\n",
1075 ma->ma_attr.la_mode);
1082 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo,
1084 mdt_object_unlock(info, mc, child_lh, rc);
1086 mdt_object_put(info->mti_env, mc);
1088 mdt_object_unlock(info, mp, parent_lh, rc);
1090 mdt_object_put(info->mti_env, mp);
1095 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1098 static int mdt_reint_link(struct mdt_thread_info *info,
1099 struct mdt_lock_handle *lhc)
1101 struct mdt_reint_record *rr = &info->mti_rr;
1102 struct ptlrpc_request *req = mdt_info_req(info);
1103 struct md_attr *ma = &info->mti_attr;
1104 struct mdt_object *ms;
1105 struct mdt_object *mp;
1106 struct mdt_lock_handle *lhs;
1107 struct mdt_lock_handle *lhp;
1112 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1113 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1115 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1116 RETURN(err_serious(-ENOENT));
1118 if (info->mti_dlm_req)
1119 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1121 /* Invalid case so return error immediately instead of
1123 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1126 if (!fid_is_md_operative(rr->rr_fid1) ||
1127 !fid_is_md_operative(rr->rr_fid2))
1130 /* step 1: find target parent dir */
1131 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1133 RETURN(PTR_ERR(mp));
1135 rc = mdt_version_get_check_save(info, mp, 0);
1137 GOTO(put_parent, rc);
1139 /* step 2: find source */
1140 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1142 GOTO(put_parent, rc = PTR_ERR(ms));
1144 if (!mdt_object_exists(ms)) {
1145 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1146 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1147 GOTO(put_source, rc = -ENOENT);
1150 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1152 lhp = &info->mti_lh[MDT_LH_PARENT];
1153 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1154 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1157 GOTO(put_source, rc);
1159 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1161 lhs = &info->mti_lh[MDT_LH_CHILD];
1162 mdt_lock_reg_init(lhs, LCK_EX);
1163 rc = mdt_reint_object_lock(info, ms, lhs,
1164 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1167 GOTO(unlock_parent, rc);
1169 /* step 3: link it */
1170 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1171 OBD_FAIL_MDS_REINT_LINK_WRITE);
1173 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1174 rc = mdt_version_get_check_save(info, ms, 1);
1176 GOTO(unlock_source, rc);
1178 /** check target version by name during replay */
1179 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1180 &info->mti_tmp_fid1, 2);
1181 if (rc != 0 && rc != -ENOENT)
1182 GOTO(unlock_source, rc);
1183 /* save version of file name for replay, it must be ENOENT here */
1184 if (!req_is_replay(mdt_info_req(info))) {
1185 if (rc != -ENOENT) {
1186 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1187 PNAME(&rr->rr_name));
1188 GOTO(unlock_source, rc = -EEXIST);
1190 info->mti_ver[2] = ENOENT_VERSION;
1191 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1194 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1195 mdt_object_child(ms), &rr->rr_name, ma);
1198 mdt_counter_incr(req, LPROC_MDT_LINK);
1202 mdt_object_unlock(info, ms, lhs, rc);
1204 mdt_object_unlock(info, mp, lhp, rc);
1206 mdt_object_put(info->mti_env, ms);
1208 mdt_object_put(info->mti_env, mp);
1212 * lock the part of the directory according to the hash of the name
1213 * (lh->mlh_pdo_hash) in parallel directory lock.
1215 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1216 struct mdt_lock_handle *lh,
1217 struct mdt_object *obj, __u64 ibits,
1220 struct ldlm_res_id *res = &info->mti_res_id;
1221 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1222 union ldlm_policy_data *policy = &info->mti_policy;
1223 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1227 * Finish res_id initializing by name hash marking part of
1228 * directory which is taking modification.
1230 LASSERT(lh->mlh_pdo_hash != 0);
1231 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1232 memset(policy, 0, sizeof(*policy));
1233 policy->l_inodebits.bits = ibits;
1235 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1236 dlmflags |= LDLM_FL_COS_INCOMPAT;
1238 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1239 * going to be sent to client. If it is - mdt_intent_policy() path will
1240 * fix it up and turn FL_LOCAL flag off.
1242 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1243 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1248 * Get BFL lock for rename or migrate process.
1250 static int mdt_rename_lock(struct mdt_thread_info *info,
1251 struct lustre_handle *lh)
1256 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1257 struct lu_fid *fid = &info->mti_tmp_fid1;
1258 struct mdt_object *obj;
1260 /* XXX, right now, it has to use object API to
1261 * enqueue lock cross MDT, so it will enqueue
1262 * rename lock(with LUSTRE_BFL_FID) by root object */
1264 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1266 RETURN(PTR_ERR(obj));
1268 rc = mdt_remote_object_lock(info, obj,
1269 &LUSTRE_BFL_FID, lh,
1271 MDS_INODELOCK_UPDATE, false, false);
1272 mdt_object_put(info->mti_env, obj);
1274 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1275 union ldlm_policy_data *policy = &info->mti_policy;
1276 struct ldlm_res_id *res_id = &info->mti_res_id;
1279 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1280 memset(policy, 0, sizeof *policy);
1281 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1282 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1283 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1284 LCK_EX, &flags, ldlm_blocking_ast,
1285 ldlm_completion_ast, NULL, NULL, 0,
1287 &info->mti_exp->exp_handle.h_cookie,
1294 static void mdt_rename_unlock(struct lustre_handle *lh)
1297 LASSERT(lustre_handle_is_used(lh));
1298 /* Cancel the single rename lock right away */
1299 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1304 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1305 * target. Source should not be ancestor of target dir. May be other rename
1306 * checks can be moved here later.
1308 static int mdt_is_subdir(struct mdt_thread_info *info,
1309 struct mdt_object *dir,
1310 const struct lu_fid *fid)
1312 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1316 /* If the source and target are in the same directory, they can not
1317 * be parent/child relationship, so subdir check is not needed */
1318 if (lu_fid_eq(&dir_fid, fid))
1321 if (!mdt_object_exists(dir))
1324 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1327 CERROR("%s: failed subdir check in "DFID" for "DFID
1328 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1329 PFID(&dir_fid), PFID(fid), rc);
1330 /* Return EINVAL only if a parent is the @fid */
1334 /* check the found fid */
1335 if (lu_fid_eq(&dir_fid, fid))
1342 /* Update object linkEA */
1343 struct mdt_lock_list {
1344 struct mdt_object *mll_obj;
1345 struct mdt_lock_handle mll_lh;
1346 struct list_head mll_list;
1349 static void mdt_unlock_list(struct mdt_thread_info *info,
1350 struct list_head *list, int rc)
1352 struct mdt_lock_list *mll;
1353 struct mdt_lock_list *mll2;
1355 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1356 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1357 list_del(&mll->mll_list);
1362 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1363 struct mdt_object *obj,
1364 struct mdt_object *pobj,
1365 struct list_head *lock_list)
1367 struct lu_buf *buf = &info->mti_big_buf;
1368 struct linkea_data ldata = { NULL };
1374 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1377 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1378 if (buf->lb_buf == NULL)
1382 rc = mdt_links_read(info, obj, &ldata);
1384 if (rc == -ENOENT || rc == -ENODATA)
1389 /* ignore the migrating parent(@pobj) */
1390 retry_count = ldata.ld_leh->leh_reccount - 1;
1393 LASSERT(ldata.ld_leh != NULL);
1394 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1395 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1396 struct mdt_device *mdt = info->mti_mdt;
1397 struct mdt_object *mdt_pobj;
1398 struct mdt_lock_list *mll;
1399 struct lu_name name;
1402 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1404 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1405 if (IS_ERR(mdt_pobj)) {
1406 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1407 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1411 if (!mdt_object_exists(mdt_pobj)) {
1412 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1413 mdt_obd_name(mdt), PFID(&fid));
1414 mdt_object_put(info->mti_env, mdt_pobj);
1418 /* Check if the object already exists in the list */
1419 list_for_each_entry(mll, lock_list, mll_list) {
1420 if (mll->mll_obj == mdt_pobj) {
1421 mdt_object_put(info->mti_env, mdt_pobj);
1426 if (mdt_pobj == pobj) {
1427 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1428 mdt_obd_name(mdt), PFID(&fid));
1429 mdt_object_put(info->mti_env, mdt_pobj);
1435 mdt_object_put(info->mti_env, mdt_pobj);
1436 GOTO(out, rc = -ENOMEM);
1439 /* Since this needs to lock all of objects in linkea, to avoid
1440 * deadlocks, because it does not follow parent-child order as
1441 * other MDT operation, let's use try_lock here and if the lock
1442 * cannot be gotten because of conflicting locks, then drop all
1443 * current locks, send an AST to the client, and start again. */
1444 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1445 rc = mdt_reint_object_lock_try(info, mdt_pobj, &mll->mll_lh,
1446 MDS_INODELOCK_UPDATE, true);
1448 mdt_unlock_list(info, lock_list, rc);
1450 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1451 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1454 if (retry_count == 0) {
1455 mdt_object_put(info->mti_env, mdt_pobj);
1457 GOTO(out, rc = -EBUSY);
1460 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1461 MDS_INODELOCK_UPDATE);
1463 mdt_object_put(info->mti_env, mdt_pobj);
1468 if (mdt_object_remote(mdt_pobj)) {
1469 struct ldlm_lock *lock;
1471 /* For remote object, Set lock to cb_atomic,
1472 * so lock can be released in blocking_ast()
1473 * immediately, then the next try_lock will
1474 * have better chance to succeds */
1476 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1477 LASSERT(lock != NULL);
1478 lock_res_and_lock(lock);
1479 ldlm_set_atomic_cb(lock);
1480 unlock_res_and_lock(lock);
1481 LDLM_LOCK_PUT(lock);
1483 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1489 INIT_LIST_HEAD(&mll->mll_list);
1490 mll->mll_obj = mdt_pobj;
1491 list_add_tail(&mll->mll_list, lock_list);
1493 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1498 mdt_unlock_list(info, lock_list, rc);
1502 /* migrate files from one MDT to another MDT */
1503 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1504 struct mdt_lock_handle *lhc)
1506 struct mdt_reint_record *rr = &info->mti_rr;
1507 struct md_attr *ma = &info->mti_attr;
1508 struct mdt_object *msrcdir;
1509 struct mdt_object *mold;
1510 struct mdt_object *mnew = NULL;
1511 struct mdt_lock_handle *lh_dirp;
1512 struct mdt_lock_handle *lh_childp;
1513 struct mdt_lock_handle *lh_tgtp = NULL;
1514 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1515 struct list_head lock_list;
1517 struct ldlm_lock *lease = NULL;
1518 bool lock_open_sem = false;
1522 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1523 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1525 /* 1: lock the source dir. */
1526 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1527 if (IS_ERR(msrcdir)) {
1528 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1529 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1530 (int)PTR_ERR(msrcdir));
1531 RETURN(PTR_ERR(msrcdir));
1534 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1535 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1536 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1539 GOTO(out_put_parent, rc);
1541 if (!mdt_object_remote(msrcdir)) {
1542 rc = mdt_version_get_check_save(info, msrcdir, 0);
1544 GOTO(out_unlock_parent, rc);
1547 /* 2: sanity check and find the object to be migrated. */
1549 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1551 GOTO(out_unlock_parent, rc);
1553 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1554 GOTO(out_unlock_parent, rc = -EINVAL);
1556 if (!fid_is_md_operative(old_fid))
1557 GOTO(out_unlock_parent, rc = -EPERM);
1559 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1560 GOTO(out_unlock_parent, rc = -EPERM);
1562 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1564 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1566 if (mdt_object_remote(mold)) {
1567 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1568 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1569 GOTO(out_put_child, rc = -EREMOTE);
1572 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1573 !mdt_object_remote(msrcdir)) {
1574 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1575 " MDT, which should be migrated first:"
1576 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1577 PFID(mdt_object_fid(msrcdir)), -EPERM);
1578 GOTO(out_put_child, rc = -EPERM);
1581 rc = mdt_remote_permission(info);
1583 GOTO(out_put_child, rc);
1585 /* 3: iterate the linkea of the object and lock all of the objects */
1586 INIT_LIST_HEAD(&lock_list);
1587 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1589 GOTO(out_put_child, rc);
1591 if (info->mti_spec.sp_migrate_close) {
1592 struct close_data *data;
1593 struct mdt_body *repbody;
1594 bool lease_broken = false;
1596 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1598 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1600 GOTO(out_lease, rc = -EPROTO);
1602 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1604 GOTO(out_lease, rc = -EPROTO);
1606 lease = ldlm_handle2lock(&data->cd_handle);
1608 GOTO(out_lease, rc = -ESTALE);
1610 /* try to hold open_sem so that nobody else can open the file */
1611 if (!down_write_trylock(&mold->mot_open_sem)) {
1612 ldlm_lock_cancel(lease);
1613 GOTO(out_lease, rc = -EBUSY);
1616 lock_open_sem = true;
1617 /* Check if the lease open lease has already canceled */
1618 lock_res_and_lock(lease);
1619 lease_broken = ldlm_is_cancel(lease);
1620 unlock_res_and_lock(lease);
1622 LDLM_DEBUG(lease, DFID " lease broken? %d",
1623 PFID(mdt_object_fid(mold)), lease_broken);
1625 /* Cancel server side lease. Client side counterpart should
1626 * have been cancelled. It's okay to cancel it now as we've
1627 * held mot_open_sem. */
1628 ldlm_lock_cancel(lease);
1631 GOTO(out_lease, rc = -EAGAIN);
1633 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1634 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1635 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1637 GOTO(out_unlock_list, rc);
1640 /* 4: lock of the object migrated object */
1641 lh_childp = &info->mti_lh[MDT_LH_OLD];
1642 mdt_lock_reg_init(lh_childp, LCK_EX);
1643 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1644 MDS_INODELOCK_LAYOUT;
1645 if (mdt_object_remote(msrcdir)) {
1646 /* Enqueue lookup lock from the parent MDT */
1647 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1648 &lh_childp->mlh_rreg_lh,
1649 lh_childp->mlh_rreg_mode,
1650 MDS_INODELOCK_LOOKUP, false, false);
1652 GOTO(out_unlock_list, rc);
1654 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1657 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1659 GOTO(out_unlock_child, rc);
1661 /* Migration is incompatible with HSM. */
1662 ma->ma_need = MA_HSM;
1664 rc = mdt_attr_get_complex(info, mold, ma);
1666 GOTO(out_unlock_child, rc);
1668 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1671 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1672 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1673 GOTO(out_unlock_child, rc);
1676 ma->ma_need = MA_LMV;
1678 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1679 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1680 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1682 GOTO(out_unlock_child, rc);
1684 if ((ma->ma_valid & MA_LMV)) {
1685 struct lmv_mds_md_v1 *lmm1;
1687 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1688 lmm1 = &ma->ma_lmv->lmv_md_v1;
1689 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1690 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1691 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1692 PFID(mdt_object_fid(mold)), -EPERM);
1693 GOTO(out_unlock_child, rc = -EPERM);
1696 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1697 GOTO(out_unlock_child, rc = -EINVAL);
1699 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1700 &lmm1->lmv_stripe_fids[1]);
1702 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1704 if (!mdt_object_remote(mnew)) {
1706 "%s: "DFID" being migrated is on this MDT:"
1707 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1708 PFID(rr->rr_fid2), -EPERM);
1709 GOTO(out_put_new, rc = -EPERM);
1712 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1713 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1714 rc = mdt_remote_object_lock(info, mnew,
1715 mdt_object_fid(mnew),
1716 &lh_tgtp->mlh_rreg_lh,
1717 lh_tgtp->mlh_rreg_mode,
1718 MDS_INODELOCK_UPDATE, false, false);
1721 GOTO(out_put_new, rc);
1724 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1727 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1728 if (!mdt_object_remote(mnew)) {
1729 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1730 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1731 PFID(rr->rr_fid2), -EXDEV);
1732 GOTO(out_put_new, rc = -EXDEV);
1737 mdt_reint_init_ma(info, ma);
1739 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1740 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1742 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1743 mdt_object_child(mold), &rr->rr_name,
1744 mdt_object_child(mnew), ma);
1746 GOTO(out_unlock_new, rc);
1749 if (lh_tgtp != NULL)
1750 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1753 mdt_object_put(info->mti_env, mnew);
1755 mdt_object_unlock(info, mold, lh_childp, rc);
1757 /* we don't really modify linkea objects, so we can safely decref these
1758 * locks, and this can avoid saving them as COS locks, which may prevent
1759 * subsequent migrate. */
1760 mdt_unlock_list(info, &lock_list, 1);
1761 if (lease != NULL) {
1762 ldlm_reprocess_all(lease->l_resource);
1763 LDLM_LOCK_PUT(lease);
1767 up_write(&mold->mot_open_sem);
1769 mdt_object_put(info->mti_env, mold);
1771 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1773 mdt_object_put(info->mti_env, msrcdir);
1778 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1779 const struct lu_fid *fid,
1782 struct mdt_object *dir;
1786 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1790 /* check early, the real version will be saved after locking */
1791 rc = mdt_version_get_check(info, dir, idx);
1797 mdt_object_put(info->mti_env, dir);
1801 static int mdt_object_lock_save(struct mdt_thread_info *info,
1802 struct mdt_object *dir,
1803 struct mdt_lock_handle *lh,
1804 int idx, bool cos_incompat)
1808 /* we lock the target dir if it is local */
1809 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1814 /* get and save correct version after locking */
1815 mdt_version_get_save(info, dir, idx);
1820 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1821 * 2 - srcdir child; 3 - tgtdir child.
1822 * Update on disk version of srcdir child.
1825 * For DNE phase I, only these renames are allowed
1826 * mv src_p/src_c tgt_p/tgt_c
1827 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1828 * 2. src_p and tgt_p are same directory, and tgt_c does not
1829 * exists. In this case, all of modification will happen
1830 * in the MDT where ithesource parent is, only one remote
1831 * update is needed, i.e. set c_time/m_time on the child.
1832 * And tgt_c will be still in the same MDT as the original
1835 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1836 struct mdt_lock_handle *lhc)
1838 struct mdt_reint_record *rr = &info->mti_rr;
1839 struct md_attr *ma = &info->mti_attr;
1840 struct ptlrpc_request *req = mdt_info_req(info);
1841 struct mdt_object *msrcdir = NULL;
1842 struct mdt_object *mtgtdir = NULL;
1843 struct mdt_object *mold;
1844 struct mdt_object *mnew = NULL;
1845 struct mdt_lock_handle *lh_srcdirp;
1846 struct mdt_lock_handle *lh_tgtdirp;
1847 struct mdt_lock_handle *lh_oldp = NULL;
1848 struct mdt_lock_handle *lh_newp = NULL;
1849 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1850 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1852 bool reverse = false;
1857 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1858 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1859 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1861 /* find both parents. */
1862 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1863 if (IS_ERR(msrcdir))
1864 RETURN(PTR_ERR(msrcdir));
1866 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1868 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1870 mdt_object_get(info->mti_env, mtgtdir);
1872 /* Check if the @msrcdir is not a child of the @mtgtdir,
1873 * otherwise a reverse locking must take place. */
1874 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1878 GOTO(out_put_srcdir, rc);
1880 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1881 if (IS_ERR(mtgtdir))
1882 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1885 /* source needs to be looked up after locking source parent, otherwise
1886 * this rename may race with unlink source, and cause rename hang, see
1887 * sanityn.sh 55b, so check parents first, if later we found source is
1888 * remote, relock parents. */
1889 cos_incompat = (mdt_object_remote(msrcdir) ||
1890 mdt_object_remote(mtgtdir));
1892 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1894 /* lock parents in the proper order. */
1895 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1896 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1899 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1900 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1903 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1906 GOTO(out_put_tgtdir, rc);
1908 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1910 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1913 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1914 GOTO(out_put_tgtdir, rc);
1917 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1920 GOTO(out_put_tgtdir, rc);
1922 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1924 if (mtgtdir != msrcdir) {
1925 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1927 } else if (lh_srcdirp->mlh_pdo_hash !=
1928 lh_tgtdirp->mlh_pdo_hash) {
1929 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1930 MDS_INODELOCK_UPDATE,
1932 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1935 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1936 GOTO(out_put_tgtdir, rc);
1940 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1941 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1943 /* find mold object. */
1945 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1947 GOTO(out_unlock_parents, rc);
1949 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1950 GOTO(out_unlock_parents, rc = -EINVAL);
1952 if (!fid_is_md_operative(old_fid))
1953 GOTO(out_unlock_parents, rc = -EPERM);
1955 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1957 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1959 /* Check if @mtgtdir is subdir of @mold, before locking child
1960 * to avoid reverse locking. */
1961 if (mtgtdir != msrcdir) {
1962 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1964 GOTO(out_put_old, rc);
1967 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1968 /* save version after locking */
1969 mdt_version_get_save(info, mold, 2);
1971 if (!cos_incompat && mdt_object_remote(mold)) {
1972 cos_incompat = true;
1973 mdt_object_put(info->mti_env, mold);
1974 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1975 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1979 /* find mnew object:
1980 * mnew target object may not exist now
1981 * lookup with version checking */
1983 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1986 /* the new_fid should have been filled at this moment */
1987 if (lu_fid_eq(old_fid, new_fid))
1988 GOTO(out_put_old, rc);
1990 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1991 lu_fid_eq(new_fid, rr->rr_fid2))
1992 GOTO(out_put_old, rc = -EINVAL);
1994 if (!fid_is_md_operative(new_fid))
1995 GOTO(out_put_old, rc = -EPERM);
1997 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1999 GOTO(out_put_old, rc = PTR_ERR(mnew));
2001 if (mdt_object_remote(mnew)) {
2002 struct mdt_body *repbody;
2004 /* Always send rename req to the target child MDT */
2005 repbody = req_capsule_server_get(info->mti_pill,
2007 LASSERT(repbody != NULL);
2008 repbody->mbo_fid1 = *new_fid;
2009 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2010 GOTO(out_put_new, rc = -EXDEV);
2012 /* Before locking the target dir, check we do not replace
2013 * a dir with a non-dir, otherwise it may deadlock with
2014 * link op which tries to create a link in this dir
2015 * back to this non-dir. */
2016 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2017 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2018 GOTO(out_put_new, rc = -EISDIR);
2020 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2021 mdt_lock_reg_init(lh_oldp, LCK_EX);
2022 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2023 if (mdt_object_remote(msrcdir)) {
2024 /* Enqueue lookup lock from the parent MDT */
2025 rc = mdt_remote_object_lock(info, msrcdir,
2026 mdt_object_fid(mold),
2027 &lh_oldp->mlh_rreg_lh,
2028 lh_oldp->mlh_rreg_mode,
2029 MDS_INODELOCK_LOOKUP,
2032 GOTO(out_put_new, rc);
2034 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2037 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2040 GOTO(out_unlock_old, rc);
2042 /* Check if @msrcdir is subdir of @mnew, before locking child
2043 * to avoid reverse locking. */
2044 if (mtgtdir != msrcdir) {
2045 rc = mdt_is_subdir(info, msrcdir, new_fid);
2047 GOTO(out_unlock_old, rc);
2050 /* We used to acquire MDS_INODELOCK_FULL here but we
2051 * can't do this now because a running HSM restore on
2052 * the rename onto victim will hold the layout
2053 * lock. See LU-4002. */
2055 lh_newp = &info->mti_lh[MDT_LH_NEW];
2056 mdt_lock_reg_init(lh_newp, LCK_EX);
2057 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2058 MDS_INODELOCK_LOOKUP |
2059 MDS_INODELOCK_UPDATE,
2062 GOTO(out_unlock_old, rc);
2064 /* get and save version after locking */
2065 mdt_version_get_save(info, mnew, 3);
2066 } else if (rc != -EREMOTE && rc != -ENOENT) {
2067 GOTO(out_put_old, rc);
2069 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2070 mdt_lock_reg_init(lh_oldp, LCK_EX);
2071 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2072 if (mdt_object_remote(msrcdir)) {
2073 /* Enqueue lookup lock from the parent MDT */
2074 rc = mdt_remote_object_lock(info, msrcdir,
2075 mdt_object_fid(mold),
2076 &lh_oldp->mlh_rreg_lh,
2077 lh_oldp->mlh_rreg_mode,
2078 MDS_INODELOCK_LOOKUP,
2081 GOTO(out_put_old, rc);
2083 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2086 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2089 GOTO(out_unlock_old, rc);
2091 mdt_enoent_version_save(info, 3);
2094 /* step 5: rename it */
2095 mdt_reint_init_ma(info, ma);
2097 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2098 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2101 mutex_lock(&mnew->mot_lov_mutex);
2103 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2104 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2105 mnew != NULL ? mdt_object_child(mnew) : NULL,
2106 &rr->rr_tgt_name, ma);
2109 mutex_unlock(&mnew->mot_lov_mutex);
2111 /* handle last link of tgt object */
2113 mdt_counter_incr(req, LPROC_MDT_RENAME);
2115 mdt_handle_last_unlink(info, mnew, ma);
2117 mdt_rename_counter_tally(info, info->mti_mdt, req,
2123 mdt_object_unlock(info, mnew, lh_newp, rc);
2125 mdt_object_unlock(info, mold, lh_oldp, rc);
2128 mdt_object_put(info->mti_env, mnew);
2130 mdt_object_put(info->mti_env, mold);
2132 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2133 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2135 mdt_object_put(info->mti_env, mtgtdir);
2137 mdt_object_put(info->mti_env, msrcdir);
2141 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2142 struct mdt_lock_handle *lhc, bool rename)
2144 struct mdt_reint_record *rr = &info->mti_rr;
2145 struct ptlrpc_request *req = mdt_info_req(info);
2146 struct lustre_handle rename_lh = { 0 };
2150 if (info->mti_dlm_req)
2151 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2153 if (!fid_is_md_operative(rr->rr_fid1) ||
2154 !fid_is_md_operative(rr->rr_fid2))
2157 /* Note: do not enqueue rename lock for replay request, because
2158 * if other MDT holds rename lock, but being blocked to wait for
2159 * this MDT to finish its recovery, and the failover MDT can not
2160 * get rename lock, which will cause deadlock. */
2161 if (!req_is_replay(req)) {
2162 rc = mdt_rename_lock(info, &rename_lh);
2164 CERROR("%s: can't lock FS for rename: rc = %d\n",
2165 mdt_obd_name(info->mti_mdt), rc);
2171 rc = mdt_reint_rename_internal(info, lhc);
2173 rc = mdt_reint_migrate_internal(info, lhc);
2175 if (lustre_handle_is_used(&rename_lh))
2176 mdt_rename_unlock(&rename_lh);
2181 static int mdt_reint_rename(struct mdt_thread_info *info,
2182 struct mdt_lock_handle *lhc)
2184 return mdt_reint_rename_or_migrate(info, lhc, true);
2187 static int mdt_reint_migrate(struct mdt_thread_info *info,
2188 struct mdt_lock_handle *lhc)
2190 return mdt_reint_rename_or_migrate(info, lhc, false);
2193 struct mdt_reinter {
2194 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2195 enum lprocfs_extra_opc mr_extra_opc;
2198 static const struct mdt_reinter mdt_reinters[] = {
2200 .mr_handler = &mdt_reint_setattr,
2201 .mr_extra_opc = MDS_REINT_SETATTR,
2204 .mr_handler = &mdt_reint_create,
2205 .mr_extra_opc = MDS_REINT_CREATE,
2208 .mr_handler = &mdt_reint_link,
2209 .mr_extra_opc = MDS_REINT_LINK,
2212 .mr_handler = &mdt_reint_unlink,
2213 .mr_extra_opc = MDS_REINT_UNLINK,
2216 .mr_handler = &mdt_reint_rename,
2217 .mr_extra_opc = MDS_REINT_RENAME,
2220 .mr_handler = &mdt_reint_open,
2221 .mr_extra_opc = MDS_REINT_OPEN,
2223 [REINT_SETXATTR] = {
2224 .mr_handler = &mdt_reint_setxattr,
2225 .mr_extra_opc = MDS_REINT_SETXATTR,
2228 .mr_handler = &mdt_reint_unlink,
2229 .mr_extra_opc = MDS_REINT_UNLINK,
2232 .mr_handler = &mdt_reint_migrate,
2233 .mr_extra_opc = MDS_REINT_RENAME,
2237 int mdt_reint_rec(struct mdt_thread_info *info,
2238 struct mdt_lock_handle *lhc)
2240 const struct mdt_reinter *mr;
2244 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2247 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2248 if (mr->mr_handler == NULL)
2251 rc = (*mr->mr_handler)(info, lhc);
2253 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2254 PTLRPC_LAST_CNTR + mr->mr_extra_opc);