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)
69 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
70 !fid_is_obf(mdt_object_fid(o)))
71 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
73 *version = ENOENT_VERSION;
74 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
75 PFID(mdt_object_fid(o)), *version);
79 * Check version is correct.
81 * Should be called only during replay.
83 static int mdt_version_check(struct ptlrpc_request *req,
84 __u64 version, int idx)
86 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
89 if (!exp_connect_vbr(req->rq_export))
92 LASSERT(req_is_replay(req));
93 /** VBR: version is checked always because costs nothing */
94 LASSERT(idx < PTLRPC_NUM_VERSIONS);
95 /** Sanity check for malformed buffers */
96 if (pre_ver == NULL) {
97 CERROR("No versions in request buffer\n");
98 spin_lock(&req->rq_export->exp_lock);
99 req->rq_export->exp_vbr_failed = 1;
100 spin_unlock(&req->rq_export->exp_lock);
102 } else if (pre_ver[idx] != version) {
103 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
104 pre_ver[idx], version);
105 spin_lock(&req->rq_export->exp_lock);
106 req->rq_export->exp_vbr_failed = 1;
107 spin_unlock(&req->rq_export->exp_lock);
114 * Save pre-versions in reply.
116 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
121 if (!exp_connect_vbr(req->rq_export))
124 LASSERT(!req_is_replay(req));
125 LASSERT(req->rq_repmsg != NULL);
126 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
128 reply_ver[idx] = version;
132 * Save enoent version, it is needed when it is obvious that object doesn't
133 * exist, e.g. child during create.
135 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
137 /* save version of file name for replay, it must be ENOENT here */
138 if (!req_is_replay(mdt_info_req(info))) {
139 info->mti_ver[idx] = ENOENT_VERSION;
140 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
145 * Get version from disk and save in reply buffer.
147 * Versions are saved in reply only during normal operations not replays.
149 void mdt_version_get_save(struct mdt_thread_info *info,
150 struct mdt_object *mto, int idx)
152 /* don't save versions during replay */
153 if (!req_is_replay(mdt_info_req(info))) {
154 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
155 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
160 * Get version from disk and check it, no save in reply.
162 int mdt_version_get_check(struct mdt_thread_info *info,
163 struct mdt_object *mto, int idx)
165 /* only check versions during replay */
166 if (!req_is_replay(mdt_info_req(info)))
169 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
170 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
174 * Get version from disk and check if recovery or just save.
176 int mdt_version_get_check_save(struct mdt_thread_info *info,
177 struct mdt_object *mto, int idx)
181 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
182 if (req_is_replay(mdt_info_req(info)))
183 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
186 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
191 * Lookup with version checking.
193 * This checks version of 'name'. Many reint functions uses 'name' for child not
194 * FID, therefore we need to get object by name and check its version.
196 static int mdt_lookup_version_check(struct mdt_thread_info *info,
197 struct mdt_object *p,
198 const struct lu_name *lname,
199 struct lu_fid *fid, int idx)
203 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
205 /* Check version only during replay */
206 if (!req_is_replay(mdt_info_req(info)))
209 info->mti_ver[idx] = ENOENT_VERSION;
211 struct mdt_object *child;
212 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
213 if (likely(!IS_ERR(child))) {
214 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
215 mdt_object_put(info->mti_env, child);
218 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
219 return vbrc ? vbrc : rc;
223 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
225 struct lu_ucred *uc = mdt_ucred(info);
226 struct mdt_device *mdt = info->mti_mdt;
228 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
229 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
230 mdt->mdt_enable_remote_dir_gid != -1)
238 * mdt_remote_permission: Check whether the remote operation is permitted,
240 * Only sysadmin can create remote directory / striped directory,
241 * migrate directory and set default stripedEA on directory, unless
243 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
245 * param[in] info: mdt_thread_info.
247 * retval = 0 remote operation is allowed.
248 * < 0 remote operation is denied.
250 static int mdt_remote_permission(struct mdt_thread_info *info)
252 struct md_op_spec *spec = &info->mti_spec;
253 struct lu_attr *attr = &info->mti_attr.ma_attr;
254 struct obd_export *exp = mdt_info_req(info)->rq_export;
257 if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
258 rc = mdt_remote_permission_check(info);
263 if (info->mti_rr.rr_opcode == REINT_CREATE &&
264 (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
265 spec->u.sp_ea.eadatalen != 0)) {
266 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
268 /* Only new clients can create remote dir( >= 2.4) and
269 * striped dir(>= 2.6), old client will return -ENOTSUPP */
270 if (!mdt_is_dne_client(exp))
273 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
274 !mdt_is_striped_client(exp))
277 rc = mdt_remote_permission_check(info);
282 if (info->mti_rr.rr_opcode == REINT_SETATTR) {
283 struct md_attr *ma = &info->mti_attr;
285 if ((ma->ma_valid & MA_LMV)) {
286 rc = mdt_remote_permission_check(info);
295 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
296 struct mdt_object *obj, __u64 ibits,
297 struct mdt_lock_handle *s0_lh,
298 struct mdt_object *s0_obj,
299 struct ldlm_enqueue_info *einfo,
302 union ldlm_policy_data *policy = &mti->mti_policy;
303 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
308 if (!S_ISDIR(obj->mot_header.loh_attr))
311 /* Unlock stripe 0 */
312 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
313 LASSERT(s0_obj != NULL);
314 mdt_object_unlock_put(mti, s0_obj, s0_lh, decref);
317 memset(policy, 0, sizeof(*policy));
318 policy->l_inodebits.bits = ibits;
320 if (slave_locks != NULL) {
321 LASSERT(s0_lh != NULL);
322 for (i = 1; i < slave_locks->count; i++) {
323 /* borrow s0_lh temporarily to do mdt unlock */
324 mdt_lock_reg_init(s0_lh, einfo->ei_mode);
325 s0_lh->mlh_rreg_lh = slave_locks->handles[i];
326 mdt_object_unlock(mti, NULL, s0_lh, decref);
327 slave_locks->handles[i].cookie = 0ull;
331 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
336 static int mdt_init_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
339 struct lu_buf *buf = &mti->mti_buf;
340 struct lmv_mds_md_v1 *lmv;
344 if (!S_ISDIR(obj->mot_header.loh_attr))
347 buf->lb_buf = mti->mti_xattr_buf;
348 buf->lb_len = sizeof(mti->mti_xattr_buf);
349 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
352 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
354 buf->lb_buf = mti->mti_big_lmm;
355 buf->lb_len = mti->mti_big_lmmsize;
359 if (rc == -ENODATA || rc == -ENOENT)
366 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
369 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
375 * Lock slave stripes if necessary, the lock handles of slave stripes
376 * will be stored in einfo->ei_cbdata.
378 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
379 enum ldlm_mode mode, __u64 ibits,
380 struct lu_fid *s0_fid,
381 struct mdt_lock_handle *s0_lh,
382 struct mdt_object **s0_objp,
383 struct ldlm_enqueue_info *einfo)
385 union ldlm_policy_data *policy = &mti->mti_policy;
389 memset(einfo, 0, sizeof(*einfo));
391 rc = mdt_init_slaves(mti, obj, s0_fid);
395 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
397 if (!lu_fid_eq(s0_fid, mdt_object_fid(obj))) {
398 /* Except migrating object, whose 0_stripe and master
399 * object are the same object, 0_stripe and master
400 * object are different, though they are in the same
401 * MDT, to avoid adding osd_object_lock here, so we
402 * will enqueue the stripe0 lock in MDT0 for now */
403 *s0_objp = mdt_object_find(mti->mti_env, mti->mti_mdt, s0_fid);
404 if (IS_ERR(*s0_objp))
405 RETURN(PTR_ERR(*s0_objp));
407 rc = mdt_reint_object_lock(mti, *s0_objp, s0_lh, ibits, true);
409 mdt_object_put(mti->mti_env, *s0_objp);
414 einfo->ei_type = LDLM_IBITS;
415 einfo->ei_mode = mode;
416 einfo->ei_cb_bl = mdt_remote_blocking_ast;
417 einfo->ei_cb_local_bl = mdt_blocking_ast;
418 einfo->ei_cb_cp = ldlm_completion_ast;
419 einfo->ei_enq_slave = 1;
420 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
421 memset(policy, 0, sizeof(*policy));
422 policy->l_inodebits.bits = ibits;
424 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
430 * VBR: we save three versions in reply:
431 * 0 - parent. Check that parent version is the same during replay.
432 * 1 - name. Version of 'name' if file exists with the same name or
433 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
434 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
437 static int mdt_md_create(struct mdt_thread_info *info)
439 struct mdt_device *mdt = info->mti_mdt;
440 struct mdt_object *parent;
441 struct mdt_object *child;
442 struct mdt_lock_handle *lh;
443 struct mdt_body *repbody;
444 struct md_attr *ma = &info->mti_attr;
445 struct mdt_reint_record *rr = &info->mti_rr;
449 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
451 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
453 if (!fid_is_md_operative(rr->rr_fid1))
456 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
458 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
460 RETURN(PTR_ERR(parent));
462 if (!mdt_object_exists(parent))
463 GOTO(put_parent, rc = -ENOENT);
465 lh = &info->mti_lh[MDT_LH_PARENT];
466 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
467 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
469 GOTO(put_parent, rc);
471 if (!mdt_object_remote(parent)) {
472 rc = mdt_version_get_check_save(info, parent, 0);
474 GOTO(unlock_parent, rc);
478 * Check child name version during replay.
479 * During create replay a file may exist with same name.
481 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
482 &info->mti_tmp_fid1, 1);
484 GOTO(unlock_parent, rc = -EEXIST);
486 /* -ENOENT is expected here */
488 GOTO(unlock_parent, rc);
490 /* save version of file name for replay, it must be ENOENT here */
491 mdt_enoent_version_save(info, 1);
493 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
494 if (unlikely(IS_ERR(child)))
495 GOTO(unlock_parent, rc = PTR_ERR(child));
497 rc = mdt_remote_permission(info);
501 ma->ma_need = MA_INODE;
504 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
505 OBD_FAIL_MDS_REINT_CREATE_WRITE);
507 /* Version of child will be updated on disk. */
508 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
509 rc = mdt_version_get_check_save(info, child, 2);
513 /* Let lower layer know current lock mode. */
514 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
517 * Do not perform lookup sanity check. We know that name does
520 info->mti_spec.sp_cr_lookup = 0;
521 info->mti_spec.sp_feat = &dt_directory_features;
523 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
524 mdt_object_child(child), &info->mti_spec, ma);
526 rc = mdt_attr_get_complex(info, child, ma);
532 * On DNE, we need to eliminate dependey between 'mkdir a' and
533 * 'mkdir a/b' if b is a striped directory, to achieve this, two
534 * things are done below:
535 * 1. save child and slaves lock.
536 * 2. if the child is a striped directory, relock parent so to
537 * compare against with COS locks to ensure parent was
540 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
541 struct mdt_lock_handle *lhc;
542 struct mdt_lock_handle *s0_lh;
543 struct mdt_object *s0_obj = NULL;
544 struct ldlm_enqueue_info *einfo;
545 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
546 bool cos_incompat = false;
548 rc = mdt_init_slaves(info, child, s0_fid);
551 if (!mdt_object_remote(parent)) {
552 mdt_object_unlock(info, parent, lh, 1);
553 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
554 rc = mdt_reint_object_lock(info, parent, lh,
555 MDS_INODELOCK_UPDATE,
562 einfo = &info->mti_einfo;
563 lhc = &info->mti_lh[MDT_LH_CHILD];
564 mdt_lock_handle_init(lhc);
565 mdt_lock_reg_init(lhc, LCK_PW);
566 rc = mdt_reint_object_lock(info, child, lhc,
567 MDS_INODELOCK_UPDATE,
571 mdt_object_unlock(info, child, lhc, rc);
573 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
574 mdt_lock_handle_init(s0_lh);
575 mdt_lock_reg_init(s0_lh, LCK_PW);
576 rc = mdt_lock_slaves(info, child, LCK_PW, MDS_INODELOCK_UPDATE,
577 s0_fid, s0_lh, &s0_obj, einfo);
578 mdt_unlock_slaves(info, child, MDS_INODELOCK_UPDATE, s0_lh,
580 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME) && rc == -EIO)
584 /* Return fid & attr to client. */
585 if (ma->ma_valid & MA_INODE)
586 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
587 mdt_object_fid(child));
589 mdt_object_put(info->mti_env, child);
591 mdt_object_unlock(info, parent, lh, rc);
593 mdt_object_put(info->mti_env, parent);
597 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
600 struct mdt_lock_handle *lh;
601 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
602 __u64 lockpart = MDS_INODELOCK_UPDATE;
603 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
604 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
605 struct mdt_lock_handle *s0_lh = NULL;
606 struct mdt_object *s0_obj = NULL;
607 bool cos_incompat = false;
611 rc = mdt_init_slaves(info, mo, s0_fid);
615 lh = &info->mti_lh[MDT_LH_PARENT];
616 mdt_lock_reg_init(lh, LCK_PW);
618 /* Even though the new MDT will grant PERM lock to the old
619 * client, but the old client will almost ignore that during
620 * So it needs to revoke both LOOKUP and PERM lock here, so
621 * both new and old client can cancel the dcache */
622 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
623 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
625 rc = mdt_reint_object_lock(info, mo, lh, lockpart, cos_incompat);
629 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
630 mdt_lock_reg_init(s0_lh, LCK_PW);
631 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_fid, s0_lh, &s0_obj,
634 GOTO(out_unlock, rc);
636 /* all attrs are packed into mti_attr in unpack_setattr */
637 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
638 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
640 /* This is only for set ctime when rename's source is on remote MDS. */
641 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
642 ma->ma_attr_flags |= MDS_VTX_BYPASS;
644 /* VBR: update version if attr changed are important for recovery */
646 /* update on-disk version of changed object */
647 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
648 rc = mdt_version_get_check_save(info, mo, 0);
650 GOTO(out_unlock, rc);
653 /* Ensure constant striping during chown(). See LU-2789. */
654 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
655 mutex_lock(&mo->mot_lov_mutex);
657 /* all attrs are packed into mti_attr in unpack_setattr */
658 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
660 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
661 mutex_unlock(&mo->mot_lov_mutex);
664 GOTO(out_unlock, rc);
668 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo, rc);
669 mdt_object_unlock(info, mo, lh, rc);
674 * Check HSM flags and add HS_DIRTY flag if relevant.
676 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
677 * and is not RELEASED.
679 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
685 /* If the file was modified, add the dirty flag */
686 ma->ma_need = MA_HSM;
687 rc = mdt_attr_get_complex(info, mo, ma);
689 CERROR("file attribute read error for "DFID": %d.\n",
690 PFID(mdt_object_fid(mo)), rc);
694 /* If an up2date copy exists in the backend, add dirty flag */
695 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
696 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
697 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
699 ma->ma_hsm.mh_flags |= HS_DIRTY;
701 mdt_lock_reg_init(lh, LCK_PW);
702 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
706 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
708 CERROR("file attribute change error for "DFID": %d\n",
709 PFID(mdt_object_fid(mo)), rc);
710 mdt_object_unlock(info, mo, lh, rc);
716 static int mdt_reint_setattr(struct mdt_thread_info *info,
717 struct mdt_lock_handle *lhc)
719 struct md_attr *ma = &info->mti_attr;
720 struct mdt_reint_record *rr = &info->mti_rr;
721 struct ptlrpc_request *req = mdt_info_req(info);
722 struct mdt_object *mo;
723 struct mdt_body *repbody;
727 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
728 (unsigned int)ma->ma_attr.la_valid);
730 if (info->mti_dlm_req)
731 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
733 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
734 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
736 GOTO(out, rc = PTR_ERR(mo));
738 if (!mdt_object_exists(mo))
739 GOTO(out_put, rc = -ENOENT);
741 if (mdt_object_remote(mo))
742 GOTO(out_put, rc = -EREMOTE);
744 if ((ma->ma_attr.la_valid & LA_SIZE) ||
745 (rr->rr_flags & MRF_OPEN_TRUNC)) {
746 /* Check write access for the O_TRUNC case */
747 if (mdt_write_read(mo) < 0)
748 GOTO(out_put, rc = -ETXTBSY);
751 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
752 if (ma->ma_valid & MA_LOV)
753 GOTO(out_put, rc = -EPROTO);
755 rc = mdt_attr_set(info, mo, ma);
758 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
759 (ma->ma_valid & MA_INODE)) {
760 struct lu_buf *buf = &info->mti_buf;
761 struct mdt_lock_handle *lh;
763 rc = mdt_remote_permission(info);
767 if (ma->ma_attr.la_valid != 0)
768 GOTO(out_put, rc = -EPROTO);
770 lh = &info->mti_lh[MDT_LH_PARENT];
771 mdt_lock_reg_init(lh, LCK_PW);
773 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
777 if (ma->ma_valid & MA_LOV) {
778 buf->lb_buf = ma->ma_lmm;
779 buf->lb_len = ma->ma_lmm_size;
781 buf->lb_buf = ma->ma_lmv;
782 buf->lb_len = ma->ma_lmv_size;
784 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
785 (ma->ma_valid & MA_LOV) ?
786 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
789 mdt_object_unlock(info, mo, lh, rc);
793 GOTO(out_put, rc = -EPROTO);
796 /* If file data is modified, add the dirty flag */
797 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
798 rc = mdt_add_dirty_flag(info, mo, ma);
800 ma->ma_need = MA_INODE;
802 rc = mdt_attr_get_complex(info, mo, ma);
806 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
810 mdt_object_put(info->mti_env, mo);
813 mdt_counter_incr(req, LPROC_MDT_SETATTR);
815 mdt_client_compatibility(info);
816 rc2 = mdt_fix_reply(info);
822 static int mdt_reint_create(struct mdt_thread_info *info,
823 struct mdt_lock_handle *lhc)
825 struct ptlrpc_request *req = mdt_info_req(info);
829 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
830 RETURN(err_serious(-ESTALE));
832 if (info->mti_dlm_req)
833 ldlm_request_cancel(mdt_info_req(info),
834 info->mti_dlm_req, 0, LATF_SKIP);
836 if (!lu_name_is_valid(&info->mti_rr.rr_name))
839 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
841 mdt_counter_incr(req, LPROC_MDT_MKDIR);
849 /* Special file should stay on the same node as parent. */
850 mdt_counter_incr(req, LPROC_MDT_MKNOD);
853 CERROR("%s: Unsupported mode %o\n",
854 mdt_obd_name(info->mti_mdt),
855 info->mti_attr.ma_attr.la_mode);
856 RETURN(err_serious(-EOPNOTSUPP));
859 rc = mdt_md_create(info);
864 * VBR: save parent version in reply and child version getting by its name.
865 * Version of child is getting and checking during its lookup. If
867 static int mdt_reint_unlink(struct mdt_thread_info *info,
868 struct mdt_lock_handle *lhc)
870 struct mdt_reint_record *rr = &info->mti_rr;
871 struct ptlrpc_request *req = mdt_info_req(info);
872 struct md_attr *ma = &info->mti_attr;
873 struct lu_fid *child_fid = &info->mti_tmp_fid1;
874 struct mdt_object *mp;
875 struct mdt_object *mc;
876 struct mdt_lock_handle *parent_lh;
877 struct mdt_lock_handle *child_lh;
878 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
879 struct lu_fid *s0_fid = &info->mti_tmp_fid2;
880 struct mdt_lock_handle *s0_lh = NULL;
881 struct mdt_object *s0_obj = NULL;
883 bool cos_incompat = false;
888 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
889 PNAME(&rr->rr_name));
891 if (info->mti_dlm_req)
892 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
894 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
895 RETURN(err_serious(-ENOENT));
897 if (!fid_is_md_operative(rr->rr_fid1))
900 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
904 if (mdt_object_remote(mp)) {
907 rc = mdt_version_get_check_save(info, mp, 0);
909 GOTO(put_parent, rc);
913 parent_lh = &info->mti_lh[MDT_LH_PARENT];
914 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
915 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
918 GOTO(put_parent, rc);
920 /* lookup child object along with version checking */
922 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
924 /* Name might not be able to find during resend of
925 * remote unlink, considering following case.
926 * dir_A is a remote directory, the name entry of
927 * dir_A is on MDT0, the directory is on MDT1,
929 * 1. client sends unlink req to MDT1.
930 * 2. MDT1 sends name delete update to MDT0.
931 * 3. name entry is being deleted in MDT0 synchronously.
932 * 4. MDT1 is restarted.
933 * 5. client resends unlink req to MDT1. So it can not
934 * find the name entry on MDT0 anymore.
935 * In this case, MDT1 only needs to destory the local
938 if (mdt_object_remote(mp) && rc == -ENOENT &&
939 !fid_is_zero(rr->rr_fid2) &&
940 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
942 *child_fid = *rr->rr_fid2;
944 GOTO(unlock_parent, rc);
948 if (!fid_is_md_operative(child_fid))
949 GOTO(unlock_parent, rc = -EPERM);
951 /* We will lock the child regardless it is local or remote. No harm. */
952 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
954 GOTO(unlock_parent, rc = PTR_ERR(mc));
956 if (!cos_incompat && mdt_init_slaves(info, mc, s0_fid) > 0) {
958 mdt_object_put(info->mti_env, mc);
959 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
963 child_lh = &info->mti_lh[MDT_LH_CHILD];
964 mdt_lock_reg_init(child_lh, LCK_EX);
965 if (info->mti_spec.sp_rm_entry) {
966 struct lu_ucred *uc = mdt_ucred(info);
968 if (!mdt_is_dne_client(req->rq_export))
969 /* Return -ENOTSUPP for old client */
970 GOTO(put_child, rc = -ENOTSUPP);
972 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
973 GOTO(put_child, rc = -EPERM);
975 ma->ma_need = MA_INODE;
977 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
978 NULL, &rr->rr_name, ma, no_name);
982 if (mdt_object_remote(mc)) {
983 struct mdt_body *repbody;
985 if (!fid_is_zero(rr->rr_fid2)) {
986 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
987 mdt_obd_name(info->mti_mdt),
988 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
989 GOTO(put_child, rc = -ENOENT);
991 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
992 mdt_obd_name(info->mti_mdt),
993 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
995 if (!mdt_is_dne_client(req->rq_export))
996 /* Return -ENOTSUPP for old client */
997 GOTO(put_child, rc = -ENOTSUPP);
999 /* Revoke the LOOKUP lock of the remote object granted by
1000 * this MDT. Since the unlink will happen on another MDT,
1001 * it will release the LOOKUP lock right away. Then What
1002 * would happen if another client try to grab the LOOKUP
1003 * lock at the same time with unlink XXX */
1004 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1005 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1006 LASSERT(repbody != NULL);
1007 repbody->mbo_fid1 = *mdt_object_fid(mc);
1008 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1009 GOTO(unlock_child, rc = -EREMOTE);
1011 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1012 * this now because a running HSM restore on the child (unlink
1013 * victim) will hold the layout lock. See LU-4002. */
1014 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1015 if (mdt_object_remote(mp)) {
1016 /* Enqueue lookup lock from parent MDT */
1017 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1018 &child_lh->mlh_rreg_lh,
1019 child_lh->mlh_rreg_mode,
1020 MDS_INODELOCK_LOOKUP, false, false);
1022 GOTO(put_child, rc);
1024 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1027 rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits,
1030 GOTO(unlock_child, rc);
1033 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1034 * whether need MA_LOV and MA_COOKIE.
1036 ma->ma_need = MA_INODE;
1039 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1040 mdt_lock_reg_init(s0_lh, LCK_EX);
1041 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_fid,
1042 s0_lh, &s0_obj, einfo);
1044 GOTO(unlock_child, rc);
1046 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1047 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1048 /* save version when object is locked */
1049 mdt_version_get_save(info, mc, 1);
1051 mutex_lock(&mc->mot_lov_mutex);
1053 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1054 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1056 mutex_unlock(&mc->mot_lov_mutex);
1058 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1059 rc = mdt_attr_get_complex(info, mc, ma);
1061 mdt_handle_last_unlink(info, mc, ma);
1063 if (ma->ma_valid & MA_INODE) {
1064 switch (ma->ma_attr.la_mode & S_IFMT) {
1066 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1074 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1077 LASSERTF(0, "bad file type %o unlinking\n",
1078 ma->ma_attr.la_mode);
1085 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo,
1087 mdt_object_unlock(info, mc, child_lh, rc);
1089 mdt_object_put(info->mti_env, mc);
1091 mdt_object_unlock(info, mp, parent_lh, rc);
1093 mdt_object_put(info->mti_env, mp);
1098 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1101 static int mdt_reint_link(struct mdt_thread_info *info,
1102 struct mdt_lock_handle *lhc)
1104 struct mdt_reint_record *rr = &info->mti_rr;
1105 struct ptlrpc_request *req = mdt_info_req(info);
1106 struct md_attr *ma = &info->mti_attr;
1107 struct mdt_object *ms;
1108 struct mdt_object *mp;
1109 struct mdt_lock_handle *lhs;
1110 struct mdt_lock_handle *lhp;
1115 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1116 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1118 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1119 RETURN(err_serious(-ENOENT));
1121 if (info->mti_dlm_req)
1122 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1124 /* Invalid case so return error immediately instead of
1126 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1129 if (!fid_is_md_operative(rr->rr_fid1) ||
1130 !fid_is_md_operative(rr->rr_fid2))
1133 /* step 1: find target parent dir */
1134 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1136 RETURN(PTR_ERR(mp));
1138 rc = mdt_version_get_check_save(info, mp, 0);
1140 GOTO(put_parent, rc);
1142 /* step 2: find source */
1143 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1145 GOTO(put_parent, rc = PTR_ERR(ms));
1147 if (!mdt_object_exists(ms)) {
1148 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1149 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1150 GOTO(put_source, rc = -ENOENT);
1153 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1155 lhp = &info->mti_lh[MDT_LH_PARENT];
1156 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1157 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1160 GOTO(put_source, rc);
1162 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1164 lhs = &info->mti_lh[MDT_LH_CHILD];
1165 mdt_lock_reg_init(lhs, LCK_EX);
1166 rc = mdt_reint_object_lock(info, ms, lhs,
1167 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1170 GOTO(unlock_parent, rc);
1172 /* step 3: link it */
1173 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1174 OBD_FAIL_MDS_REINT_LINK_WRITE);
1176 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1177 rc = mdt_version_get_check_save(info, ms, 1);
1179 GOTO(unlock_source, rc);
1181 /** check target version by name during replay */
1182 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1183 &info->mti_tmp_fid1, 2);
1184 if (rc != 0 && rc != -ENOENT)
1185 GOTO(unlock_source, rc);
1186 /* save version of file name for replay, it must be ENOENT here */
1187 if (!req_is_replay(mdt_info_req(info))) {
1188 if (rc != -ENOENT) {
1189 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1190 PNAME(&rr->rr_name));
1191 GOTO(unlock_source, rc = -EEXIST);
1193 info->mti_ver[2] = ENOENT_VERSION;
1194 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1197 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1198 mdt_object_child(ms), &rr->rr_name, ma);
1201 mdt_counter_incr(req, LPROC_MDT_LINK);
1205 mdt_object_unlock(info, ms, lhs, rc);
1207 mdt_object_unlock(info, mp, lhp, rc);
1209 mdt_object_put(info->mti_env, ms);
1211 mdt_object_put(info->mti_env, mp);
1215 * lock the part of the directory according to the hash of the name
1216 * (lh->mlh_pdo_hash) in parallel directory lock.
1218 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1219 struct mdt_lock_handle *lh,
1220 struct mdt_object *obj, __u64 ibits,
1223 struct ldlm_res_id *res = &info->mti_res_id;
1224 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1225 union ldlm_policy_data *policy = &info->mti_policy;
1226 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1230 * Finish res_id initializing by name hash marking part of
1231 * directory which is taking modification.
1233 LASSERT(lh->mlh_pdo_hash != 0);
1234 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1235 memset(policy, 0, sizeof(*policy));
1236 policy->l_inodebits.bits = ibits;
1238 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1239 dlmflags |= LDLM_FL_COS_INCOMPAT;
1241 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1242 * going to be sent to client. If it is - mdt_intent_policy() path will
1243 * fix it up and turn FL_LOCAL flag off.
1245 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1246 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1251 * Get BFL lock for rename or migrate process.
1253 static int mdt_rename_lock(struct mdt_thread_info *info,
1254 struct lustre_handle *lh)
1259 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1260 struct lu_fid *fid = &info->mti_tmp_fid1;
1261 struct mdt_object *obj;
1263 /* XXX, right now, it has to use object API to
1264 * enqueue lock cross MDT, so it will enqueue
1265 * rename lock(with LUSTRE_BFL_FID) by root object */
1267 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1269 RETURN(PTR_ERR(obj));
1271 rc = mdt_remote_object_lock(info, obj,
1272 &LUSTRE_BFL_FID, lh,
1274 MDS_INODELOCK_UPDATE, false, false);
1275 mdt_object_put(info->mti_env, obj);
1277 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1278 union ldlm_policy_data *policy = &info->mti_policy;
1279 struct ldlm_res_id *res_id = &info->mti_res_id;
1282 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1283 memset(policy, 0, sizeof *policy);
1284 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1285 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1286 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1287 LCK_EX, &flags, ldlm_blocking_ast,
1288 ldlm_completion_ast, NULL, NULL, 0,
1290 &info->mti_exp->exp_handle.h_cookie,
1297 static void mdt_rename_unlock(struct lustre_handle *lh)
1300 LASSERT(lustre_handle_is_used(lh));
1301 /* Cancel the single rename lock right away */
1302 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1307 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1308 * target. Source should not be ancestor of target dir. May be other rename
1309 * checks can be moved here later.
1311 static int mdt_is_subdir(struct mdt_thread_info *info,
1312 struct mdt_object *dir,
1313 const struct lu_fid *fid)
1315 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1319 /* If the source and target are in the same directory, they can not
1320 * be parent/child relationship, so subdir check is not needed */
1321 if (lu_fid_eq(&dir_fid, fid))
1324 if (!mdt_object_exists(dir))
1327 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1330 CERROR("%s: failed subdir check in "DFID" for "DFID
1331 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1332 PFID(&dir_fid), PFID(fid), rc);
1333 /* Return EINVAL only if a parent is the @fid */
1337 /* check the found fid */
1338 if (lu_fid_eq(&dir_fid, fid))
1345 /* Update object linkEA */
1346 struct mdt_lock_list {
1347 struct mdt_object *mll_obj;
1348 struct mdt_lock_handle mll_lh;
1349 struct list_head mll_list;
1352 static void mdt_unlock_list(struct mdt_thread_info *info,
1353 struct list_head *list, int rc)
1355 struct mdt_lock_list *mll;
1356 struct mdt_lock_list *mll2;
1358 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1359 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1360 list_del(&mll->mll_list);
1365 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1366 struct mdt_object *obj,
1367 struct mdt_object *pobj,
1368 struct list_head *lock_list)
1370 struct lu_buf *buf = &info->mti_big_buf;
1371 struct linkea_data ldata = { NULL };
1377 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1380 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1381 if (buf->lb_buf == NULL)
1385 rc = mdt_links_read(info, obj, &ldata);
1387 if (rc == -ENOENT || rc == -ENODATA)
1392 /* ignore the migrating parent(@pobj) */
1393 retry_count = ldata.ld_leh->leh_reccount - 1;
1396 LASSERT(ldata.ld_leh != NULL);
1397 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1398 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1399 struct mdt_device *mdt = info->mti_mdt;
1400 struct mdt_object *mdt_pobj;
1401 struct mdt_lock_list *mll;
1402 struct lu_name name;
1405 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1407 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1408 if (IS_ERR(mdt_pobj)) {
1409 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1410 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1414 if (!mdt_object_exists(mdt_pobj)) {
1415 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1416 mdt_obd_name(mdt), PFID(&fid));
1417 mdt_object_put(info->mti_env, mdt_pobj);
1421 /* Check if the object already exists in the list */
1422 list_for_each_entry(mll, lock_list, mll_list) {
1423 if (mll->mll_obj == mdt_pobj) {
1424 mdt_object_put(info->mti_env, mdt_pobj);
1429 if (mdt_pobj == pobj) {
1430 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1431 mdt_obd_name(mdt), PFID(&fid));
1432 mdt_object_put(info->mti_env, mdt_pobj);
1438 mdt_object_put(info->mti_env, mdt_pobj);
1439 GOTO(out, rc = -ENOMEM);
1442 /* Since this needs to lock all of objects in linkea, to avoid
1443 * deadlocks, because it does not follow parent-child order as
1444 * other MDT operation, let's use try_lock here and if the lock
1445 * cannot be gotten because of conflicting locks, then drop all
1446 * current locks, send an AST to the client, and start again. */
1447 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1448 rc = mdt_reint_object_lock_try(info, mdt_pobj, &mll->mll_lh,
1449 MDS_INODELOCK_UPDATE, true);
1451 mdt_unlock_list(info, lock_list, rc);
1453 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1454 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1457 if (retry_count == 0) {
1458 mdt_object_put(info->mti_env, mdt_pobj);
1460 GOTO(out, rc = -EBUSY);
1463 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1464 MDS_INODELOCK_UPDATE);
1466 mdt_object_put(info->mti_env, mdt_pobj);
1471 if (mdt_object_remote(mdt_pobj)) {
1472 struct ldlm_lock *lock;
1474 /* For remote object, Set lock to cb_atomic,
1475 * so lock can be released in blocking_ast()
1476 * immediately, then the next try_lock will
1477 * have better chance to succeds */
1479 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1480 LASSERT(lock != NULL);
1481 lock_res_and_lock(lock);
1482 ldlm_set_atomic_cb(lock);
1483 unlock_res_and_lock(lock);
1484 LDLM_LOCK_PUT(lock);
1486 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1492 INIT_LIST_HEAD(&mll->mll_list);
1493 mll->mll_obj = mdt_pobj;
1494 list_add_tail(&mll->mll_list, lock_list);
1496 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1501 mdt_unlock_list(info, lock_list, rc);
1505 /* migrate files from one MDT to another MDT */
1506 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1507 struct mdt_lock_handle *lhc)
1509 struct mdt_reint_record *rr = &info->mti_rr;
1510 struct md_attr *ma = &info->mti_attr;
1511 struct mdt_object *msrcdir;
1512 struct mdt_object *mold;
1513 struct mdt_object *mnew = NULL;
1514 struct mdt_lock_handle *lh_dirp;
1515 struct mdt_lock_handle *lh_childp;
1516 struct mdt_lock_handle *lh_tgtp = NULL;
1517 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1518 struct list_head lock_list;
1520 struct ldlm_lock *lease = NULL;
1521 bool lock_open_sem = false;
1525 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1526 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1528 /* 1: lock the source dir. */
1529 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1530 if (IS_ERR(msrcdir)) {
1531 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1532 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1533 (int)PTR_ERR(msrcdir));
1534 RETURN(PTR_ERR(msrcdir));
1537 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1538 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1539 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1542 GOTO(out_put_parent, rc);
1544 if (!mdt_object_remote(msrcdir)) {
1545 rc = mdt_version_get_check_save(info, msrcdir, 0);
1547 GOTO(out_unlock_parent, rc);
1550 /* 2: sanity check and find the object to be migrated. */
1552 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1554 GOTO(out_unlock_parent, rc);
1556 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1557 GOTO(out_unlock_parent, rc = -EINVAL);
1559 if (!fid_is_md_operative(old_fid))
1560 GOTO(out_unlock_parent, rc = -EPERM);
1562 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1563 GOTO(out_unlock_parent, rc = -EPERM);
1565 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1567 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1569 if (mdt_object_remote(mold)) {
1570 CERROR("%s: source "DFID" is on the remote MDT\n",
1571 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1572 GOTO(out_put_child, rc = -EREMOTE);
1575 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1576 !mdt_object_remote(msrcdir)) {
1577 CERROR("%s: parent "DFID" is still on the same"
1578 " MDT, which should be migrated first:"
1579 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1580 PFID(mdt_object_fid(msrcdir)), -EPERM);
1581 GOTO(out_put_child, rc = -EPERM);
1584 rc = mdt_remote_permission(info);
1586 GOTO(out_put_child, rc);
1588 /* 3: iterate the linkea of the object and lock all of the objects */
1589 INIT_LIST_HEAD(&lock_list);
1590 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1592 GOTO(out_put_child, rc);
1594 if (info->mti_spec.sp_migrate_close) {
1595 struct close_data *data;
1596 struct mdt_body *repbody;
1597 bool lease_broken = false;
1599 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1601 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1603 GOTO(out_lease, rc = -EPROTO);
1605 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1607 GOTO(out_lease, rc = -EPROTO);
1609 lease = ldlm_handle2lock(&data->cd_handle);
1611 GOTO(out_lease, rc = -ESTALE);
1613 /* try to hold open_sem so that nobody else can open the file */
1614 if (!down_write_trylock(&mold->mot_open_sem)) {
1615 ldlm_lock_cancel(lease);
1616 GOTO(out_lease, rc = -EBUSY);
1619 lock_open_sem = true;
1620 /* Check if the lease open lease has already canceled */
1621 lock_res_and_lock(lease);
1622 lease_broken = ldlm_is_cancel(lease);
1623 unlock_res_and_lock(lease);
1625 LDLM_DEBUG(lease, DFID " lease broken? %d",
1626 PFID(mdt_object_fid(mold)), lease_broken);
1628 /* Cancel server side lease. Client side counterpart should
1629 * have been cancelled. It's okay to cancel it now as we've
1630 * held mot_open_sem. */
1631 ldlm_lock_cancel(lease);
1634 GOTO(out_lease, rc = -EAGAIN);
1636 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1637 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1638 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1640 GOTO(out_unlock_list, rc);
1643 /* 4: lock of the object migrated object */
1644 lh_childp = &info->mti_lh[MDT_LH_OLD];
1645 mdt_lock_reg_init(lh_childp, LCK_EX);
1646 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1647 MDS_INODELOCK_LAYOUT;
1648 if (mdt_object_remote(msrcdir)) {
1649 /* Enqueue lookup lock from the parent MDT */
1650 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1651 &lh_childp->mlh_rreg_lh,
1652 lh_childp->mlh_rreg_mode,
1653 MDS_INODELOCK_LOOKUP, false, false);
1655 GOTO(out_unlock_list, rc);
1657 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1660 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1662 GOTO(out_unlock_child, rc);
1664 /* Migration is incompatible with HSM. */
1665 ma->ma_need = MA_HSM;
1667 rc = mdt_attr_get_complex(info, mold, ma);
1669 GOTO(out_unlock_child, rc);
1671 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1673 CERROR("%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1674 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1675 GOTO(out_unlock_child, rc);
1678 ma->ma_need = MA_LMV;
1680 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1681 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1682 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1684 GOTO(out_unlock_child, rc);
1686 if ((ma->ma_valid & MA_LMV)) {
1687 struct lmv_mds_md_v1 *lmm1;
1689 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1690 lmm1 = &ma->ma_lmv->lmv_md_v1;
1691 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1692 CERROR("%s: can not migrate striped dir "DFID
1693 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1694 PFID(mdt_object_fid(mold)), -EPERM);
1695 GOTO(out_unlock_child, rc = -EPERM);
1698 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1699 GOTO(out_unlock_child, rc = -EINVAL);
1701 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1702 &lmm1->lmv_stripe_fids[1]);
1704 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1706 if (!mdt_object_remote(mnew)) {
1707 CERROR("%s: "DFID" being migrated is on this MDT:"
1708 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1709 PFID(rr->rr_fid2), -EPERM);
1710 GOTO(out_put_new, rc = -EPERM);
1713 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1714 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1715 rc = mdt_remote_object_lock(info, mnew,
1716 mdt_object_fid(mnew),
1717 &lh_tgtp->mlh_rreg_lh,
1718 lh_tgtp->mlh_rreg_mode,
1719 MDS_INODELOCK_UPDATE, false, false);
1722 GOTO(out_put_new, rc);
1725 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1728 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1729 if (!mdt_object_remote(mnew)) {
1730 CERROR("%s: Migration "DFID" is on this MDT:"
1731 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1732 PFID(rr->rr_fid2), -EXDEV);
1733 GOTO(out_put_new, rc = -EXDEV);
1738 mdt_reint_init_ma(info, ma);
1740 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1741 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1743 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1744 mdt_object_child(mold), &rr->rr_name,
1745 mdt_object_child(mnew), ma);
1747 GOTO(out_unlock_new, rc);
1750 if (lh_tgtp != NULL)
1751 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1754 mdt_object_put(info->mti_env, mnew);
1756 mdt_object_unlock(info, mold, lh_childp, rc);
1758 /* we don't really modify linkea objects, so we can safely decref these
1759 * locks, and this can avoid saving them as COS locks, which may prevent
1760 * subsequent migrate. */
1761 mdt_unlock_list(info, &lock_list, 1);
1762 if (lease != NULL) {
1763 ldlm_reprocess_all(lease->l_resource);
1764 LDLM_LOCK_PUT(lease);
1768 up_write(&mold->mot_open_sem);
1770 mdt_object_put(info->mti_env, mold);
1772 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1774 mdt_object_put(info->mti_env, msrcdir);
1779 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1780 const struct lu_fid *fid,
1783 struct mdt_object *dir;
1787 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1791 /* check early, the real version will be saved after locking */
1792 rc = mdt_version_get_check(info, dir, idx);
1798 mdt_object_put(info->mti_env, dir);
1802 static int mdt_object_lock_save(struct mdt_thread_info *info,
1803 struct mdt_object *dir,
1804 struct mdt_lock_handle *lh,
1805 int idx, bool cos_incompat)
1809 /* we lock the target dir if it is local */
1810 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1815 /* get and save correct version after locking */
1816 mdt_version_get_save(info, dir, idx);
1821 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1822 * 2 - srcdir child; 3 - tgtdir child.
1823 * Update on disk version of srcdir child.
1826 * For DNE phase I, only these renames are allowed
1827 * mv src_p/src_c tgt_p/tgt_c
1828 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1829 * 2. src_p and tgt_p are same directory, and tgt_c does not
1830 * exists. In this case, all of modification will happen
1831 * in the MDT where ithesource parent is, only one remote
1832 * update is needed, i.e. set c_time/m_time on the child.
1833 * And tgt_c will be still in the same MDT as the original
1836 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1837 struct mdt_lock_handle *lhc)
1839 struct mdt_reint_record *rr = &info->mti_rr;
1840 struct md_attr *ma = &info->mti_attr;
1841 struct ptlrpc_request *req = mdt_info_req(info);
1842 struct mdt_object *msrcdir = NULL;
1843 struct mdt_object *mtgtdir = NULL;
1844 struct mdt_object *mold;
1845 struct mdt_object *mnew = NULL;
1846 struct mdt_lock_handle *lh_srcdirp;
1847 struct mdt_lock_handle *lh_tgtdirp;
1848 struct mdt_lock_handle *lh_oldp = NULL;
1849 struct mdt_lock_handle *lh_newp = NULL;
1850 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1851 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1853 bool reverse = false;
1858 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1859 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1860 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1862 /* find both parents. */
1863 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1864 if (IS_ERR(msrcdir))
1865 RETURN(PTR_ERR(msrcdir));
1867 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1869 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1871 mdt_object_get(info->mti_env, mtgtdir);
1873 /* Check if the @msrcdir is not a child of the @mtgtdir,
1874 * otherwise a reverse locking must take place. */
1875 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1879 GOTO(out_put_srcdir, rc);
1881 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1882 if (IS_ERR(mtgtdir))
1883 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1886 /* source needs to be looked up after locking source parent, otherwise
1887 * this rename may race with unlink source, and cause rename hang, see
1888 * sanityn.sh 55b, so check parents first, if later we found source is
1889 * remote, relock parents. */
1890 cos_incompat = (mdt_object_remote(msrcdir) ||
1891 mdt_object_remote(mtgtdir));
1893 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1895 /* lock parents in the proper order. */
1896 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1897 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1900 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1901 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1904 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1907 GOTO(out_put_tgtdir, rc);
1909 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1911 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1914 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1915 GOTO(out_put_tgtdir, rc);
1918 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1921 GOTO(out_put_tgtdir, rc);
1923 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1925 if (mtgtdir != msrcdir) {
1926 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1928 } else if (lh_srcdirp->mlh_pdo_hash !=
1929 lh_tgtdirp->mlh_pdo_hash) {
1930 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1931 MDS_INODELOCK_UPDATE,
1933 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1936 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1937 GOTO(out_put_tgtdir, rc);
1941 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1942 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1944 /* find mold object. */
1946 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1948 GOTO(out_unlock_parents, rc);
1950 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1951 GOTO(out_unlock_parents, rc = -EINVAL);
1953 if (!fid_is_md_operative(old_fid))
1954 GOTO(out_unlock_parents, rc = -EPERM);
1956 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1958 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1960 /* Check if @mtgtdir is subdir of @mold, before locking child
1961 * to avoid reverse locking. */
1962 if (mtgtdir != msrcdir) {
1963 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1965 GOTO(out_put_old, rc);
1968 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1969 /* save version after locking */
1970 mdt_version_get_save(info, mold, 2);
1972 if (!cos_incompat && mdt_object_remote(mold)) {
1973 cos_incompat = true;
1974 mdt_object_put(info->mti_env, mold);
1975 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1976 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1980 /* find mnew object:
1981 * mnew target object may not exist now
1982 * lookup with version checking */
1984 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1987 /* the new_fid should have been filled at this moment */
1988 if (lu_fid_eq(old_fid, new_fid))
1989 GOTO(out_put_old, rc);
1991 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1992 lu_fid_eq(new_fid, rr->rr_fid2))
1993 GOTO(out_put_old, rc = -EINVAL);
1995 if (!fid_is_md_operative(new_fid))
1996 GOTO(out_put_old, rc = -EPERM);
1998 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2000 GOTO(out_put_old, rc = PTR_ERR(mnew));
2002 if (mdt_object_remote(mnew)) {
2003 struct mdt_body *repbody;
2005 /* Always send rename req to the target child MDT */
2006 repbody = req_capsule_server_get(info->mti_pill,
2008 LASSERT(repbody != NULL);
2009 repbody->mbo_fid1 = *new_fid;
2010 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2011 GOTO(out_put_new, rc = -EXDEV);
2013 /* Before locking the target dir, check we do not replace
2014 * a dir with a non-dir, otherwise it may deadlock with
2015 * link op which tries to create a link in this dir
2016 * back to this non-dir. */
2017 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2018 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2019 GOTO(out_put_new, rc = -EISDIR);
2021 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2022 mdt_lock_reg_init(lh_oldp, LCK_EX);
2023 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2024 if (mdt_object_remote(msrcdir)) {
2025 /* Enqueue lookup lock from the parent MDT */
2026 rc = mdt_remote_object_lock(info, msrcdir,
2027 mdt_object_fid(mold),
2028 &lh_oldp->mlh_rreg_lh,
2029 lh_oldp->mlh_rreg_mode,
2030 MDS_INODELOCK_LOOKUP,
2033 GOTO(out_put_new, rc);
2035 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2038 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2041 GOTO(out_unlock_old, rc);
2043 /* Check if @msrcdir is subdir of @mnew, before locking child
2044 * to avoid reverse locking. */
2045 if (mtgtdir != msrcdir) {
2046 rc = mdt_is_subdir(info, msrcdir, new_fid);
2048 GOTO(out_unlock_old, rc);
2051 /* We used to acquire MDS_INODELOCK_FULL here but we
2052 * can't do this now because a running HSM restore on
2053 * the rename onto victim will hold the layout
2054 * lock. See LU-4002. */
2056 lh_newp = &info->mti_lh[MDT_LH_NEW];
2057 mdt_lock_reg_init(lh_newp, LCK_EX);
2058 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2059 MDS_INODELOCK_LOOKUP |
2060 MDS_INODELOCK_UPDATE,
2063 GOTO(out_unlock_old, rc);
2065 /* get and save version after locking */
2066 mdt_version_get_save(info, mnew, 3);
2067 } else if (rc != -EREMOTE && rc != -ENOENT) {
2068 GOTO(out_put_old, rc);
2070 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2071 mdt_lock_reg_init(lh_oldp, LCK_EX);
2072 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2073 if (mdt_object_remote(msrcdir)) {
2074 /* Enqueue lookup lock from the parent MDT */
2075 rc = mdt_remote_object_lock(info, msrcdir,
2076 mdt_object_fid(mold),
2077 &lh_oldp->mlh_rreg_lh,
2078 lh_oldp->mlh_rreg_mode,
2079 MDS_INODELOCK_LOOKUP,
2082 GOTO(out_put_old, rc);
2084 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2087 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2090 GOTO(out_unlock_old, rc);
2092 mdt_enoent_version_save(info, 3);
2095 /* step 5: rename it */
2096 mdt_reint_init_ma(info, ma);
2098 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2099 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2102 mutex_lock(&mnew->mot_lov_mutex);
2104 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2105 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2106 mnew != NULL ? mdt_object_child(mnew) : NULL,
2107 &rr->rr_tgt_name, ma);
2110 mutex_unlock(&mnew->mot_lov_mutex);
2112 /* handle last link of tgt object */
2114 mdt_counter_incr(req, LPROC_MDT_RENAME);
2116 mdt_handle_last_unlink(info, mnew, ma);
2118 mdt_rename_counter_tally(info, info->mti_mdt, req,
2124 mdt_object_unlock(info, mnew, lh_newp, rc);
2126 mdt_object_unlock(info, mold, lh_oldp, rc);
2129 mdt_object_put(info->mti_env, mnew);
2131 mdt_object_put(info->mti_env, mold);
2133 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2134 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2136 mdt_object_put(info->mti_env, mtgtdir);
2138 mdt_object_put(info->mti_env, msrcdir);
2142 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2143 struct mdt_lock_handle *lhc, bool rename)
2145 struct mdt_reint_record *rr = &info->mti_rr;
2146 struct ptlrpc_request *req = mdt_info_req(info);
2147 struct lustre_handle rename_lh = { 0 };
2151 if (info->mti_dlm_req)
2152 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2154 if (!fid_is_md_operative(rr->rr_fid1) ||
2155 !fid_is_md_operative(rr->rr_fid2))
2158 /* Note: do not enqueue rename lock for replay request, because
2159 * if other MDT holds rename lock, but being blocked to wait for
2160 * this MDT to finish its recovery, and the failover MDT can not
2161 * get rename lock, which will cause deadlock. */
2162 if (!req_is_replay(req)) {
2163 rc = mdt_rename_lock(info, &rename_lh);
2165 CERROR("%s: can't lock FS for rename: rc = %d\n",
2166 mdt_obd_name(info->mti_mdt), rc);
2172 rc = mdt_reint_rename_internal(info, lhc);
2174 rc = mdt_reint_migrate_internal(info, lhc);
2176 if (lustre_handle_is_used(&rename_lh))
2177 mdt_rename_unlock(&rename_lh);
2182 static int mdt_reint_rename(struct mdt_thread_info *info,
2183 struct mdt_lock_handle *lhc)
2185 return mdt_reint_rename_or_migrate(info, lhc, true);
2188 static int mdt_reint_migrate(struct mdt_thread_info *info,
2189 struct mdt_lock_handle *lhc)
2191 return mdt_reint_rename_or_migrate(info, lhc, false);
2194 struct mdt_reinter {
2195 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2196 enum lprocfs_extra_opc mr_extra_opc;
2199 static const struct mdt_reinter mdt_reinters[] = {
2201 .mr_handler = &mdt_reint_setattr,
2202 .mr_extra_opc = MDS_REINT_SETATTR,
2205 .mr_handler = &mdt_reint_create,
2206 .mr_extra_opc = MDS_REINT_CREATE,
2209 .mr_handler = &mdt_reint_link,
2210 .mr_extra_opc = MDS_REINT_LINK,
2213 .mr_handler = &mdt_reint_unlink,
2214 .mr_extra_opc = MDS_REINT_UNLINK,
2217 .mr_handler = &mdt_reint_rename,
2218 .mr_extra_opc = MDS_REINT_RENAME,
2221 .mr_handler = &mdt_reint_open,
2222 .mr_extra_opc = MDS_REINT_OPEN,
2224 [REINT_SETXATTR] = {
2225 .mr_handler = &mdt_reint_setxattr,
2226 .mr_extra_opc = MDS_REINT_SETXATTR,
2229 .mr_handler = &mdt_reint_unlink,
2230 .mr_extra_opc = MDS_REINT_UNLINK,
2233 .mr_handler = &mdt_reint_migrate,
2234 .mr_extra_opc = MDS_REINT_RENAME,
2238 int mdt_reint_rec(struct mdt_thread_info *info,
2239 struct mdt_lock_handle *lhc)
2241 const struct mdt_reinter *mr;
2245 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2248 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2249 if (mr->mr_handler == NULL)
2252 rc = (*mr->mr_handler)(info, lhc);
2254 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2255 PTLRPC_LAST_CNTR + mr->mr_extra_opc);