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_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 &
599 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
600 __u64 lockpart = MDS_INODELOCK_UPDATE;
601 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
602 struct lu_fid *s0_fid = &info->mti_tmp_fid1;
603 struct mdt_lock_handle *s0_lh = NULL;
604 struct mdt_object *s0_obj = NULL;
605 bool cos_incompat = false;
609 rc = mdt_init_slaves(info, mo, s0_fid);
613 lh = &info->mti_lh[MDT_LH_PARENT];
614 mdt_lock_reg_init(lh, LCK_PW);
616 /* Even though the new MDT will grant PERM lock to the old
617 * client, but the old client will almost ignore that during
618 * So it needs to revoke both LOOKUP and PERM lock here, so
619 * both new and old client can cancel the dcache */
620 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
621 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
623 rc = mdt_reint_object_lock(info, mo, lh, lockpart, cos_incompat);
627 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
628 mdt_lock_reg_init(s0_lh, LCK_PW);
629 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_fid, s0_lh, &s0_obj,
632 GOTO(out_unlock, rc);
634 /* all attrs are packed into mti_attr in unpack_setattr */
635 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
636 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
638 /* This is only for set ctime when rename's source is on remote MDS. */
639 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
640 ma->ma_attr_flags |= MDS_VTX_BYPASS;
642 /* VBR: update version if attr changed are important for recovery */
644 /* update on-disk version of changed object */
645 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
646 rc = mdt_version_get_check_save(info, mo, 0);
648 GOTO(out_unlock, rc);
651 /* Ensure constant striping during chown(). See LU-2789. */
652 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
653 mutex_lock(&mo->mot_lov_mutex);
655 /* all attrs are packed into mti_attr in unpack_setattr */
656 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
658 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
659 mutex_unlock(&mo->mot_lov_mutex);
662 GOTO(out_unlock, rc);
663 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
666 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo, rc);
667 mdt_object_unlock(info, mo, lh, rc);
672 * Check HSM flags and add HS_DIRTY flag if relevant.
674 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
675 * and is not RELEASED.
677 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
683 /* If the file was modified, add the dirty flag */
684 ma->ma_need = MA_HSM;
685 rc = mdt_attr_get_complex(info, mo, ma);
687 CERROR("file attribute read error for "DFID": %d.\n",
688 PFID(mdt_object_fid(mo)), rc);
692 /* If an up2date copy exists in the backend, add dirty flag */
693 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
694 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
695 ma->ma_hsm.mh_flags |= HS_DIRTY;
697 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
699 CERROR("file attribute change error for "DFID": %d\n",
700 PFID(mdt_object_fid(mo)), rc);
706 static int mdt_reint_setattr(struct mdt_thread_info *info,
707 struct mdt_lock_handle *lhc)
709 struct md_attr *ma = &info->mti_attr;
710 struct mdt_reint_record *rr = &info->mti_rr;
711 struct ptlrpc_request *req = mdt_info_req(info);
712 struct mdt_object *mo;
713 struct mdt_body *repbody;
717 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
718 (unsigned int)ma->ma_attr.la_valid);
720 if (info->mti_dlm_req)
721 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
723 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
724 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
726 GOTO(out, rc = PTR_ERR(mo));
728 if (!mdt_object_exists(mo))
729 GOTO(out_put, rc = -ENOENT);
731 if (mdt_object_remote(mo))
732 GOTO(out_put, rc = -EREMOTE);
734 if ((ma->ma_attr.la_valid & LA_SIZE) ||
735 (rr->rr_flags & MRF_OPEN_TRUNC)) {
736 /* Check write access for the O_TRUNC case */
737 if (mdt_write_read(mo) < 0)
738 GOTO(out_put, rc = -ETXTBSY);
741 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
742 if (ma->ma_valid & MA_LOV)
743 GOTO(out_put, rc = -EPROTO);
745 rc = mdt_attr_set(info, mo, ma);
748 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
749 (ma->ma_valid & MA_INODE)) {
750 struct lu_buf *buf = &info->mti_buf;
751 struct mdt_lock_handle *lh;
753 rc = mdt_remote_permission(info);
757 if (ma->ma_attr.la_valid != 0)
758 GOTO(out_put, rc = -EPROTO);
760 lh = &info->mti_lh[MDT_LH_PARENT];
761 mdt_lock_reg_init(lh, LCK_PW);
763 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
767 if (ma->ma_valid & MA_LOV) {
768 buf->lb_buf = ma->ma_lmm;
769 buf->lb_len = ma->ma_lmm_size;
771 buf->lb_buf = ma->ma_lmv;
772 buf->lb_len = ma->ma_lmv_size;
774 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
775 (ma->ma_valid & MA_LOV) ?
776 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
779 mdt_object_unlock(info, mo, lh, rc);
783 GOTO(out_put, rc = -EPROTO);
786 /* If file data is modified, add the dirty flag */
787 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
788 rc = mdt_add_dirty_flag(info, mo, ma);
790 ma->ma_need = MA_INODE;
792 rc = mdt_attr_get_complex(info, mo, ma);
796 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
800 mdt_object_put(info->mti_env, mo);
803 mdt_counter_incr(req, LPROC_MDT_SETATTR);
805 mdt_client_compatibility(info);
806 rc2 = mdt_fix_reply(info);
812 static int mdt_reint_create(struct mdt_thread_info *info,
813 struct mdt_lock_handle *lhc)
815 struct ptlrpc_request *req = mdt_info_req(info);
819 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
820 RETURN(err_serious(-ESTALE));
822 if (info->mti_dlm_req)
823 ldlm_request_cancel(mdt_info_req(info),
824 info->mti_dlm_req, 0, LATF_SKIP);
826 if (!lu_name_is_valid(&info->mti_rr.rr_name))
829 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
831 mdt_counter_incr(req, LPROC_MDT_MKDIR);
839 /* Special file should stay on the same node as parent. */
840 mdt_counter_incr(req, LPROC_MDT_MKNOD);
843 CERROR("%s: Unsupported mode %o\n",
844 mdt_obd_name(info->mti_mdt),
845 info->mti_attr.ma_attr.la_mode);
846 RETURN(err_serious(-EOPNOTSUPP));
849 rc = mdt_create(info);
854 * VBR: save parent version in reply and child version getting by its name.
855 * Version of child is getting and checking during its lookup. If
857 static int mdt_reint_unlink(struct mdt_thread_info *info,
858 struct mdt_lock_handle *lhc)
860 struct mdt_reint_record *rr = &info->mti_rr;
861 struct ptlrpc_request *req = mdt_info_req(info);
862 struct md_attr *ma = &info->mti_attr;
863 struct lu_fid *child_fid = &info->mti_tmp_fid1;
864 struct mdt_object *mp;
865 struct mdt_object *mc;
866 struct mdt_lock_handle *parent_lh;
867 struct mdt_lock_handle *child_lh;
868 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
869 struct lu_fid *s0_fid = &info->mti_tmp_fid2;
870 struct mdt_lock_handle *s0_lh = NULL;
871 struct mdt_object *s0_obj = NULL;
873 bool cos_incompat = false;
879 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
880 PNAME(&rr->rr_name));
882 if (info->mti_dlm_req)
883 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
885 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
886 RETURN(err_serious(-ENOENT));
888 if (!fid_is_md_operative(rr->rr_fid1))
891 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
895 if (mdt_object_remote(mp)) {
898 rc = mdt_version_get_check_save(info, mp, 0);
900 GOTO(put_parent, rc);
904 parent_lh = &info->mti_lh[MDT_LH_PARENT];
905 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
906 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
909 GOTO(put_parent, rc);
911 /* lookup child object along with version checking */
913 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
915 /* Name might not be able to find during resend of
916 * remote unlink, considering following case.
917 * dir_A is a remote directory, the name entry of
918 * dir_A is on MDT0, the directory is on MDT1,
920 * 1. client sends unlink req to MDT1.
921 * 2. MDT1 sends name delete update to MDT0.
922 * 3. name entry is being deleted in MDT0 synchronously.
923 * 4. MDT1 is restarted.
924 * 5. client resends unlink req to MDT1. So it can not
925 * find the name entry on MDT0 anymore.
926 * In this case, MDT1 only needs to destory the local
929 if (mdt_object_remote(mp) && rc == -ENOENT &&
930 !fid_is_zero(rr->rr_fid2) &&
931 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
933 *child_fid = *rr->rr_fid2;
935 GOTO(unlock_parent, rc);
939 if (!fid_is_md_operative(child_fid))
940 GOTO(unlock_parent, rc = -EPERM);
942 /* We will lock the child regardless it is local or remote. No harm. */
943 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
945 GOTO(unlock_parent, rc = PTR_ERR(mc));
947 if (!cos_incompat && mdt_init_slaves(info, mc, s0_fid) > 0) {
949 mdt_object_put(info->mti_env, mc);
950 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
954 child_lh = &info->mti_lh[MDT_LH_CHILD];
955 mdt_lock_reg_init(child_lh, LCK_EX);
956 if (info->mti_spec.sp_rm_entry) {
957 struct lu_ucred *uc = mdt_ucred(info);
959 if (!mdt_is_dne_client(req->rq_export))
960 /* Return -ENOTSUPP for old client */
961 GOTO(put_child, rc = -ENOTSUPP);
963 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
964 GOTO(put_child, rc = -EPERM);
966 ma->ma_need = MA_INODE;
968 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
969 NULL, &rr->rr_name, ma, no_name);
973 if (mdt_object_remote(mc)) {
974 struct mdt_body *repbody;
976 if (!fid_is_zero(rr->rr_fid2)) {
977 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
978 mdt_obd_name(info->mti_mdt),
979 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
980 GOTO(put_child, rc = -ENOENT);
982 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
983 mdt_obd_name(info->mti_mdt),
984 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
986 if (!mdt_is_dne_client(req->rq_export))
987 /* Return -ENOTSUPP for old client */
988 GOTO(put_child, rc = -ENOTSUPP);
990 /* Revoke the LOOKUP lock of the remote object granted by
991 * this MDT. Since the unlink will happen on another MDT,
992 * it will release the LOOKUP lock right away. Then What
993 * would happen if another client try to grab the LOOKUP
994 * lock at the same time with unlink XXX */
995 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
996 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
997 LASSERT(repbody != NULL);
998 repbody->mbo_fid1 = *mdt_object_fid(mc);
999 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1000 GOTO(unlock_child, rc = -EREMOTE);
1002 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1003 * this now because a running HSM restore on the child (unlink
1004 * victim) will hold the layout lock. See LU-4002. */
1005 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1006 if (mdt_object_remote(mp)) {
1007 /* Enqueue lookup lock from parent MDT */
1008 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1009 &child_lh->mlh_rreg_lh,
1010 child_lh->mlh_rreg_mode,
1011 MDS_INODELOCK_LOOKUP, false);
1013 GOTO(put_child, rc);
1015 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1018 rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits,
1021 GOTO(unlock_child, rc);
1024 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1025 * whether need MA_LOV and MA_COOKIE.
1027 ma->ma_need = MA_INODE;
1030 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1031 mdt_lock_reg_init(s0_lh, LCK_EX);
1032 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_fid,
1033 s0_lh, &s0_obj, einfo);
1035 GOTO(unlock_child, rc);
1037 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1038 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1039 /* save version when object is locked */
1040 mdt_version_get_save(info, mc, 1);
1042 mutex_lock(&mc->mot_lov_mutex);
1044 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1045 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1047 mutex_unlock(&mc->mot_lov_mutex);
1049 GOTO(unlock_child, rc);
1051 if (!lu_object_is_dying(&mc->mot_header)) {
1052 rc = mdt_attr_get_complex(info, mc, ma);
1056 mdt_dom_check_and_discard(info, mc);
1058 mdt_handle_last_unlink(info, mc, ma);
1061 if (ma->ma_valid & MA_INODE) {
1062 switch (ma->ma_attr.la_mode & S_IFMT) {
1064 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1072 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1075 LASSERTF(0, "bad file type %o unlinking\n",
1076 ma->ma_attr.la_mode);
1083 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo,
1085 mdt_object_unlock(info, mc, child_lh, rc);
1087 mdt_object_put(info->mti_env, mc);
1089 mdt_object_unlock(info, mp, parent_lh, rc);
1091 mdt_object_put(info->mti_env, mp);
1096 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1099 static int mdt_reint_link(struct mdt_thread_info *info,
1100 struct mdt_lock_handle *lhc)
1102 struct mdt_reint_record *rr = &info->mti_rr;
1103 struct ptlrpc_request *req = mdt_info_req(info);
1104 struct md_attr *ma = &info->mti_attr;
1105 struct mdt_object *ms;
1106 struct mdt_object *mp;
1107 struct mdt_lock_handle *lhs;
1108 struct mdt_lock_handle *lhp;
1113 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1114 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1116 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1117 RETURN(err_serious(-ENOENT));
1119 if (info->mti_dlm_req)
1120 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1122 /* Invalid case so return error immediately instead of
1124 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1127 if (!fid_is_md_operative(rr->rr_fid1) ||
1128 !fid_is_md_operative(rr->rr_fid2))
1131 /* step 1: find target parent dir */
1132 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1134 RETURN(PTR_ERR(mp));
1136 rc = mdt_version_get_check_save(info, mp, 0);
1138 GOTO(put_parent, rc);
1140 /* step 2: find source */
1141 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1143 GOTO(put_parent, rc = PTR_ERR(ms));
1145 if (!mdt_object_exists(ms)) {
1146 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1147 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1148 GOTO(put_source, rc = -ENOENT);
1151 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1153 lhp = &info->mti_lh[MDT_LH_PARENT];
1154 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1155 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1158 GOTO(put_source, rc);
1160 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1162 lhs = &info->mti_lh[MDT_LH_CHILD];
1163 mdt_lock_reg_init(lhs, LCK_EX);
1164 rc = mdt_reint_object_lock(info, ms, lhs,
1165 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1168 GOTO(unlock_parent, rc);
1170 /* step 3: link it */
1171 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1172 OBD_FAIL_MDS_REINT_LINK_WRITE);
1174 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1175 rc = mdt_version_get_check_save(info, ms, 1);
1177 GOTO(unlock_source, rc);
1179 /** check target version by name during replay */
1180 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1181 &info->mti_tmp_fid1, 2);
1182 if (rc != 0 && rc != -ENOENT)
1183 GOTO(unlock_source, rc);
1184 /* save version of file name for replay, it must be ENOENT here */
1185 if (!req_is_replay(mdt_info_req(info))) {
1186 if (rc != -ENOENT) {
1187 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1188 PNAME(&rr->rr_name));
1189 GOTO(unlock_source, rc = -EEXIST);
1191 info->mti_ver[2] = ENOENT_VERSION;
1192 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1195 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1196 mdt_object_child(ms), &rr->rr_name, ma);
1199 mdt_counter_incr(req, LPROC_MDT_LINK);
1203 mdt_object_unlock(info, ms, lhs, rc);
1205 mdt_object_unlock(info, mp, lhp, rc);
1207 mdt_object_put(info->mti_env, ms);
1209 mdt_object_put(info->mti_env, mp);
1213 * lock the part of the directory according to the hash of the name
1214 * (lh->mlh_pdo_hash) in parallel directory lock.
1216 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1217 struct mdt_lock_handle *lh,
1218 struct mdt_object *obj, __u64 ibits,
1221 struct ldlm_res_id *res = &info->mti_res_id;
1222 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1223 union ldlm_policy_data *policy = &info->mti_policy;
1224 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1228 * Finish res_id initializing by name hash marking part of
1229 * directory which is taking modification.
1231 LASSERT(lh->mlh_pdo_hash != 0);
1232 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1233 memset(policy, 0, sizeof(*policy));
1234 policy->l_inodebits.bits = ibits;
1236 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1237 dlmflags |= LDLM_FL_COS_INCOMPAT;
1239 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1240 * going to be sent to client. If it is - mdt_intent_policy() path will
1241 * fix it up and turn FL_LOCAL flag off.
1243 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1244 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1249 * Get BFL lock for rename or migrate process.
1251 static int mdt_rename_lock(struct mdt_thread_info *info,
1252 struct lustre_handle *lh)
1257 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1258 struct lu_fid *fid = &info->mti_tmp_fid1;
1259 struct mdt_object *obj;
1261 /* XXX, right now, it has to use object API to
1262 * enqueue lock cross MDT, so it will enqueue
1263 * rename lock(with LUSTRE_BFL_FID) by root object */
1265 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1267 RETURN(PTR_ERR(obj));
1269 rc = mdt_remote_object_lock(info, obj,
1270 &LUSTRE_BFL_FID, lh,
1272 MDS_INODELOCK_UPDATE, false);
1273 mdt_object_put(info->mti_env, obj);
1275 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1276 union ldlm_policy_data *policy = &info->mti_policy;
1277 struct ldlm_res_id *res_id = &info->mti_res_id;
1280 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1281 memset(policy, 0, sizeof *policy);
1282 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1283 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1284 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1285 LCK_EX, &flags, ldlm_blocking_ast,
1286 ldlm_completion_ast, NULL, NULL, 0,
1288 &info->mti_exp->exp_handle.h_cookie,
1295 static void mdt_rename_unlock(struct lustre_handle *lh)
1298 LASSERT(lustre_handle_is_used(lh));
1299 /* Cancel the single rename lock right away */
1300 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1305 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1306 * target. Source should not be ancestor of target dir. May be other rename
1307 * checks can be moved here later.
1309 static int mdt_is_subdir(struct mdt_thread_info *info,
1310 struct mdt_object *dir,
1311 const struct lu_fid *fid)
1313 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1317 /* If the source and target are in the same directory, they can not
1318 * be parent/child relationship, so subdir check is not needed */
1319 if (lu_fid_eq(&dir_fid, fid))
1322 if (!mdt_object_exists(dir))
1325 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1328 CERROR("%s: failed subdir check in "DFID" for "DFID
1329 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1330 PFID(&dir_fid), PFID(fid), rc);
1331 /* Return EINVAL only if a parent is the @fid */
1335 /* check the found fid */
1336 if (lu_fid_eq(&dir_fid, fid))
1343 /* Update object linkEA */
1344 struct mdt_lock_list {
1345 struct mdt_object *mll_obj;
1346 struct mdt_lock_handle mll_lh;
1347 struct list_head mll_list;
1350 static void mdt_unlock_list(struct mdt_thread_info *info,
1351 struct list_head *list, int rc)
1353 struct mdt_lock_list *mll;
1354 struct mdt_lock_list *mll2;
1356 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1357 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1358 list_del(&mll->mll_list);
1363 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1364 struct mdt_object *obj,
1365 struct mdt_object *pobj,
1366 struct list_head *lock_list)
1368 struct lu_buf *buf = &info->mti_big_buf;
1369 struct linkea_data ldata = { NULL };
1375 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1378 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1379 if (buf->lb_buf == NULL)
1383 rc = mdt_links_read(info, obj, &ldata);
1385 if (rc == -ENOENT || rc == -ENODATA)
1390 /* ignore the migrating parent(@pobj) */
1391 retry_count = ldata.ld_leh->leh_reccount - 1;
1394 LASSERT(ldata.ld_leh != NULL);
1395 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1396 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1397 struct mdt_device *mdt = info->mti_mdt;
1398 struct mdt_object *mdt_pobj;
1399 struct mdt_lock_list *mll;
1400 struct lu_name name;
1404 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1406 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1407 if (IS_ERR(mdt_pobj)) {
1408 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1409 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1413 if (!mdt_object_exists(mdt_pobj)) {
1414 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1415 mdt_obd_name(mdt), PFID(&fid));
1416 mdt_object_put(info->mti_env, mdt_pobj);
1420 /* Check if the object already exists in the list */
1421 list_for_each_entry(mll, lock_list, mll_list) {
1422 if (mll->mll_obj == mdt_pobj) {
1423 mdt_object_put(info->mti_env, mdt_pobj);
1428 if (mdt_pobj == pobj) {
1429 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1430 mdt_obd_name(mdt), PFID(&fid));
1431 mdt_object_put(info->mti_env, mdt_pobj);
1437 mdt_object_put(info->mti_env, mdt_pobj);
1438 GOTO(out, rc = -ENOMEM);
1441 /* Since this needs to lock all of objects in linkea, to avoid
1442 * deadlocks, because it does not follow parent-child order as
1443 * other MDT operation, let's use try_lock here and if the lock
1444 * cannot be gotten because of conflicting locks, then drop all
1445 * current locks, send an AST to the client, and start again. */
1446 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1448 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits,
1449 MDS_INODELOCK_UPDATE, true);
1450 if (!(ibits & MDS_INODELOCK_UPDATE)) {
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 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1464 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1465 MDS_INODELOCK_UPDATE);
1467 mdt_object_put(info->mti_env, mdt_pobj);
1472 if (mdt_object_remote(mdt_pobj)) {
1473 struct ldlm_lock *lock;
1475 /* For remote object, Set lock to cb_atomic,
1476 * so lock can be released in blocking_ast()
1477 * immediately, then the next try_lock will
1478 * have better chance to succeds */
1480 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1481 LASSERT(lock != NULL);
1482 lock_res_and_lock(lock);
1483 ldlm_set_atomic_cb(lock);
1484 unlock_res_and_lock(lock);
1485 LDLM_LOCK_PUT(lock);
1487 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1493 INIT_LIST_HEAD(&mll->mll_list);
1494 mll->mll_obj = mdt_pobj;
1495 list_add_tail(&mll->mll_list, lock_list);
1497 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1502 mdt_unlock_list(info, lock_list, rc);
1506 /* migrate files from one MDT to another MDT */
1507 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1508 struct mdt_lock_handle *lhc)
1510 struct mdt_reint_record *rr = &info->mti_rr;
1511 struct md_attr *ma = &info->mti_attr;
1512 struct mdt_object *msrcdir;
1513 struct mdt_object *mold;
1514 struct mdt_object *mnew = NULL;
1515 struct mdt_lock_handle *lh_dirp;
1516 struct mdt_lock_handle *lh_childp;
1517 struct mdt_lock_handle *lh_tgtp = NULL;
1518 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1519 struct list_head lock_list;
1521 struct ldlm_lock *lease = NULL;
1522 bool lock_open_sem = false;
1526 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1527 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1529 /* 1: lock the source dir. */
1530 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1531 if (IS_ERR(msrcdir)) {
1532 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1533 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1534 (int)PTR_ERR(msrcdir));
1535 RETURN(PTR_ERR(msrcdir));
1538 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1539 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1540 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1543 GOTO(out_put_parent, rc);
1545 if (!mdt_object_remote(msrcdir)) {
1546 rc = mdt_version_get_check_save(info, msrcdir, 0);
1548 GOTO(out_unlock_parent, rc);
1551 /* 2: sanity check and find the object to be migrated. */
1553 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1555 GOTO(out_unlock_parent, rc);
1557 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1558 GOTO(out_unlock_parent, rc = -EINVAL);
1560 if (!fid_is_md_operative(old_fid))
1561 GOTO(out_unlock_parent, rc = -EPERM);
1563 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1564 GOTO(out_unlock_parent, rc = -EPERM);
1566 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1568 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1570 if (mdt_object_remote(mold)) {
1571 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1572 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1573 GOTO(out_put_child, rc = -EREMOTE);
1576 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1577 !mdt_object_remote(msrcdir)) {
1578 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1579 " MDT, which should be migrated first:"
1580 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1581 PFID(mdt_object_fid(msrcdir)), -EPERM);
1582 GOTO(out_put_child, rc = -EPERM);
1585 rc = mdt_remote_permission(info);
1587 GOTO(out_put_child, rc);
1589 /* 3: iterate the linkea of the object and lock all of the objects */
1590 INIT_LIST_HEAD(&lock_list);
1591 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1593 GOTO(out_put_child, rc);
1595 if (info->mti_spec.sp_migrate_close) {
1596 struct close_data *data;
1597 struct mdt_body *repbody;
1598 bool lease_broken = false;
1600 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1602 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1604 GOTO(out_lease, rc = -EPROTO);
1606 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1608 GOTO(out_lease, rc = -EPROTO);
1610 lease = ldlm_handle2lock(&data->cd_handle);
1612 GOTO(out_lease, rc = -ESTALE);
1614 /* try to hold open_sem so that nobody else can open the file */
1615 if (!down_write_trylock(&mold->mot_open_sem)) {
1616 ldlm_lock_cancel(lease);
1617 GOTO(out_lease, rc = -EBUSY);
1620 lock_open_sem = true;
1621 /* Check if the lease open lease has already canceled */
1622 lock_res_and_lock(lease);
1623 lease_broken = ldlm_is_cancel(lease);
1624 unlock_res_and_lock(lease);
1626 LDLM_DEBUG(lease, DFID " lease broken? %d",
1627 PFID(mdt_object_fid(mold)), lease_broken);
1629 /* Cancel server side lease. Client side counterpart should
1630 * have been cancelled. It's okay to cancel it now as we've
1631 * held mot_open_sem. */
1632 ldlm_lock_cancel(lease);
1635 GOTO(out_lease, rc = -EAGAIN);
1637 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1638 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1639 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1641 GOTO(out_unlock_list, rc);
1644 /* 4: lock of the object migrated object */
1645 lh_childp = &info->mti_lh[MDT_LH_OLD];
1646 mdt_lock_reg_init(lh_childp, LCK_EX);
1647 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1648 MDS_INODELOCK_LAYOUT;
1649 if (mdt_object_remote(msrcdir)) {
1650 /* Enqueue lookup lock from the parent MDT */
1651 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1652 &lh_childp->mlh_rreg_lh,
1653 lh_childp->mlh_rreg_mode,
1654 MDS_INODELOCK_LOOKUP, false);
1656 GOTO(out_unlock_list, rc);
1658 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1661 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1663 GOTO(out_unlock_child, rc);
1665 /* Migration is incompatible with HSM. */
1666 ma->ma_need = MA_HSM;
1668 rc = mdt_attr_get_complex(info, mold, ma);
1670 GOTO(out_unlock_child, rc);
1672 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1675 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1676 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1677 GOTO(out_unlock_child, rc);
1680 ma->ma_need = MA_LMV;
1682 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1683 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1684 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1686 GOTO(out_unlock_child, rc);
1688 if ((ma->ma_valid & MA_LMV)) {
1689 struct lmv_mds_md_v1 *lmm1;
1691 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1692 lmm1 = &ma->ma_lmv->lmv_md_v1;
1693 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1694 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1695 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1696 PFID(mdt_object_fid(mold)), -EPERM);
1697 GOTO(out_unlock_child, rc = -EPERM);
1700 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1701 GOTO(out_unlock_child, rc = -EINVAL);
1703 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1704 &lmm1->lmv_stripe_fids[1]);
1706 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1708 if (!mdt_object_remote(mnew)) {
1710 "%s: "DFID" being migrated is on this MDT:"
1711 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1712 PFID(rr->rr_fid2), -EPERM);
1713 GOTO(out_put_new, rc = -EPERM);
1716 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1717 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1718 rc = mdt_remote_object_lock(info, mnew,
1719 mdt_object_fid(mnew),
1720 &lh_tgtp->mlh_rreg_lh,
1721 lh_tgtp->mlh_rreg_mode,
1722 MDS_INODELOCK_UPDATE, false);
1725 GOTO(out_put_new, rc);
1728 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1731 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1732 if (!mdt_object_remote(mnew)) {
1733 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1734 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1735 PFID(rr->rr_fid2), -EXDEV);
1736 GOTO(out_put_new, rc = -EXDEV);
1741 mdt_reint_init_ma(info, ma);
1743 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1744 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1746 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1747 mdt_object_child(mold), &rr->rr_name,
1748 mdt_object_child(mnew), ma);
1750 GOTO(out_unlock_new, rc);
1753 if (lh_tgtp != NULL)
1754 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1757 mdt_object_put(info->mti_env, mnew);
1759 mdt_object_unlock(info, mold, lh_childp, rc);
1761 /* we don't really modify linkea objects, so we can safely decref these
1762 * locks, and this can avoid saving them as COS locks, which may prevent
1763 * subsequent migrate. */
1764 mdt_unlock_list(info, &lock_list, 1);
1765 if (lease != NULL) {
1766 ldlm_reprocess_all(lease->l_resource);
1767 LDLM_LOCK_PUT(lease);
1771 up_write(&mold->mot_open_sem);
1773 mdt_object_put(info->mti_env, mold);
1775 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1777 mdt_object_put(info->mti_env, msrcdir);
1782 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1783 const struct lu_fid *fid,
1786 struct mdt_object *dir;
1790 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1794 /* check early, the real version will be saved after locking */
1795 rc = mdt_version_get_check(info, dir, idx);
1801 mdt_object_put(info->mti_env, dir);
1805 static int mdt_object_lock_save(struct mdt_thread_info *info,
1806 struct mdt_object *dir,
1807 struct mdt_lock_handle *lh,
1808 int idx, bool cos_incompat)
1812 /* we lock the target dir if it is local */
1813 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1818 /* get and save correct version after locking */
1819 mdt_version_get_save(info, dir, idx);
1824 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1825 * 2 - srcdir child; 3 - tgtdir child.
1826 * Update on disk version of srcdir child.
1829 * For DNE phase I, only these renames are allowed
1830 * mv src_p/src_c tgt_p/tgt_c
1831 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1832 * 2. src_p and tgt_p are same directory, and tgt_c does not
1833 * exists. In this case, all of modification will happen
1834 * in the MDT where ithesource parent is, only one remote
1835 * update is needed, i.e. set c_time/m_time on the child.
1836 * And tgt_c will be still in the same MDT as the original
1839 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1840 struct mdt_lock_handle *lhc)
1842 struct mdt_reint_record *rr = &info->mti_rr;
1843 struct md_attr *ma = &info->mti_attr;
1844 struct ptlrpc_request *req = mdt_info_req(info);
1845 struct mdt_object *msrcdir = NULL;
1846 struct mdt_object *mtgtdir = NULL;
1847 struct mdt_object *mold;
1848 struct mdt_object *mnew = NULL;
1849 struct mdt_lock_handle *lh_srcdirp;
1850 struct mdt_lock_handle *lh_tgtdirp;
1851 struct mdt_lock_handle *lh_oldp = NULL;
1852 struct mdt_lock_handle *lh_newp = NULL;
1853 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1854 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1856 bool reverse = false;
1861 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1862 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1863 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1865 /* find both parents. */
1866 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1867 if (IS_ERR(msrcdir))
1868 RETURN(PTR_ERR(msrcdir));
1870 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1872 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1874 mdt_object_get(info->mti_env, mtgtdir);
1876 /* Check if the @msrcdir is not a child of the @mtgtdir,
1877 * otherwise a reverse locking must take place. */
1878 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1882 GOTO(out_put_srcdir, rc);
1884 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1885 if (IS_ERR(mtgtdir))
1886 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1889 /* source needs to be looked up after locking source parent, otherwise
1890 * this rename may race with unlink source, and cause rename hang, see
1891 * sanityn.sh 55b, so check parents first, if later we found source is
1892 * remote, relock parents. */
1893 cos_incompat = (mdt_object_remote(msrcdir) ||
1894 mdt_object_remote(mtgtdir));
1896 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1898 /* lock parents in the proper order. */
1899 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1900 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1903 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1904 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1907 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1910 GOTO(out_put_tgtdir, rc);
1912 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1914 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1917 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1918 GOTO(out_put_tgtdir, rc);
1921 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1924 GOTO(out_put_tgtdir, rc);
1926 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1928 if (mtgtdir != msrcdir) {
1929 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1931 } else if (lh_srcdirp->mlh_pdo_hash !=
1932 lh_tgtdirp->mlh_pdo_hash) {
1933 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1934 MDS_INODELOCK_UPDATE,
1936 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1939 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1940 GOTO(out_put_tgtdir, rc);
1944 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1945 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1947 /* find mold object. */
1949 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1951 GOTO(out_unlock_parents, rc);
1953 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1954 GOTO(out_unlock_parents, rc = -EINVAL);
1956 if (!fid_is_md_operative(old_fid))
1957 GOTO(out_unlock_parents, rc = -EPERM);
1959 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1961 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1963 /* Check if @mtgtdir is subdir of @mold, before locking child
1964 * to avoid reverse locking. */
1965 if (mtgtdir != msrcdir) {
1966 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1968 GOTO(out_put_old, rc);
1971 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1972 /* save version after locking */
1973 mdt_version_get_save(info, mold, 2);
1975 if (!cos_incompat && mdt_object_remote(mold)) {
1976 cos_incompat = true;
1977 mdt_object_put(info->mti_env, mold);
1978 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1979 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1983 /* find mnew object:
1984 * mnew target object may not exist now
1985 * lookup with version checking */
1987 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1990 /* the new_fid should have been filled at this moment */
1991 if (lu_fid_eq(old_fid, new_fid))
1992 GOTO(out_put_old, rc);
1994 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1995 lu_fid_eq(new_fid, rr->rr_fid2))
1996 GOTO(out_put_old, rc = -EINVAL);
1998 if (!fid_is_md_operative(new_fid))
1999 GOTO(out_put_old, rc = -EPERM);
2001 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2003 GOTO(out_put_old, rc = PTR_ERR(mnew));
2005 if (mdt_object_remote(mnew)) {
2006 struct mdt_body *repbody;
2008 /* Always send rename req to the target child MDT */
2009 repbody = req_capsule_server_get(info->mti_pill,
2011 LASSERT(repbody != NULL);
2012 repbody->mbo_fid1 = *new_fid;
2013 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2014 GOTO(out_put_new, rc = -EXDEV);
2016 /* Before locking the target dir, check we do not replace
2017 * a dir with a non-dir, otherwise it may deadlock with
2018 * link op which tries to create a link in this dir
2019 * back to this non-dir. */
2020 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2021 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2022 GOTO(out_put_new, rc = -EISDIR);
2024 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2025 mdt_lock_reg_init(lh_oldp, LCK_EX);
2026 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2027 if (mdt_object_remote(msrcdir)) {
2028 /* Enqueue lookup lock from the parent MDT */
2029 rc = mdt_remote_object_lock(info, msrcdir,
2030 mdt_object_fid(mold),
2031 &lh_oldp->mlh_rreg_lh,
2032 lh_oldp->mlh_rreg_mode,
2033 MDS_INODELOCK_LOOKUP,
2036 GOTO(out_put_new, rc);
2038 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2041 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2044 GOTO(out_unlock_old, rc);
2046 /* Check if @msrcdir is subdir of @mnew, before locking child
2047 * to avoid reverse locking. */
2048 if (mtgtdir != msrcdir) {
2049 rc = mdt_is_subdir(info, msrcdir, new_fid);
2051 GOTO(out_unlock_old, rc);
2054 /* We used to acquire MDS_INODELOCK_FULL here but we
2055 * can't do this now because a running HSM restore on
2056 * the rename onto victim will hold the layout
2057 * lock. See LU-4002. */
2059 lh_newp = &info->mti_lh[MDT_LH_NEW];
2060 mdt_lock_reg_init(lh_newp, LCK_EX);
2061 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2062 MDS_INODELOCK_LOOKUP |
2063 MDS_INODELOCK_UPDATE,
2066 GOTO(out_unlock_old, rc);
2068 /* get and save version after locking */
2069 mdt_version_get_save(info, mnew, 3);
2070 } else if (rc != -EREMOTE && rc != -ENOENT) {
2071 GOTO(out_put_old, rc);
2073 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2074 mdt_lock_reg_init(lh_oldp, LCK_EX);
2075 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2076 if (mdt_object_remote(msrcdir)) {
2077 /* Enqueue lookup lock from the parent MDT */
2078 rc = mdt_remote_object_lock(info, msrcdir,
2079 mdt_object_fid(mold),
2080 &lh_oldp->mlh_rreg_lh,
2081 lh_oldp->mlh_rreg_mode,
2082 MDS_INODELOCK_LOOKUP,
2085 GOTO(out_put_old, rc);
2087 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2090 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2093 GOTO(out_unlock_old, rc);
2095 mdt_enoent_version_save(info, 3);
2098 /* step 5: rename it */
2099 mdt_reint_init_ma(info, ma);
2101 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2102 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2105 mutex_lock(&mnew->mot_lov_mutex);
2107 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2108 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2109 mnew != NULL ? mdt_object_child(mnew) : NULL,
2110 &rr->rr_tgt_name, ma);
2113 mutex_unlock(&mnew->mot_lov_mutex);
2115 /* handle last link of tgt object */
2117 mdt_counter_incr(req, LPROC_MDT_RENAME);
2119 mdt_handle_last_unlink(info, mnew, ma);
2120 mdt_dom_check_and_discard(info, mnew);
2123 mdt_rename_counter_tally(info, info->mti_mdt, req,
2129 mdt_object_unlock(info, mnew, lh_newp, rc);
2131 mdt_object_unlock(info, mold, lh_oldp, rc);
2134 mdt_object_put(info->mti_env, mnew);
2136 mdt_object_put(info->mti_env, mold);
2138 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2139 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2141 mdt_object_put(info->mti_env, mtgtdir);
2143 mdt_object_put(info->mti_env, msrcdir);
2147 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2148 struct mdt_lock_handle *lhc, bool rename)
2150 struct mdt_reint_record *rr = &info->mti_rr;
2151 struct ptlrpc_request *req = mdt_info_req(info);
2152 struct lustre_handle rename_lh = { 0 };
2156 if (info->mti_dlm_req)
2157 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2159 if (!fid_is_md_operative(rr->rr_fid1) ||
2160 !fid_is_md_operative(rr->rr_fid2))
2163 /* Note: do not enqueue rename lock for replay request, because
2164 * if other MDT holds rename lock, but being blocked to wait for
2165 * this MDT to finish its recovery, and the failover MDT can not
2166 * get rename lock, which will cause deadlock. */
2167 if (!req_is_replay(req)) {
2168 rc = mdt_rename_lock(info, &rename_lh);
2170 CERROR("%s: can't lock FS for rename: rc = %d\n",
2171 mdt_obd_name(info->mti_mdt), rc);
2177 rc = mdt_reint_rename_internal(info, lhc);
2179 rc = mdt_reint_migrate_internal(info, lhc);
2181 if (lustre_handle_is_used(&rename_lh))
2182 mdt_rename_unlock(&rename_lh);
2187 static int mdt_reint_rename(struct mdt_thread_info *info,
2188 struct mdt_lock_handle *lhc)
2190 return mdt_reint_rename_or_migrate(info, lhc, true);
2193 static int mdt_reint_migrate(struct mdt_thread_info *info,
2194 struct mdt_lock_handle *lhc)
2196 return mdt_reint_rename_or_migrate(info, lhc, false);
2199 static int mdt_reint_resync(struct mdt_thread_info *info,
2200 struct mdt_lock_handle *lhc)
2202 struct mdt_reint_record *rr = &info->mti_rr;
2203 struct ptlrpc_request *req = mdt_info_req(info);
2204 struct md_attr *ma = &info->mti_attr;
2205 struct mdt_object *mo;
2206 struct ldlm_lock *lease;
2207 struct mdt_body *repbody;
2208 struct md_layout_change layout = { 0 };
2213 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2215 if (info->mti_dlm_req)
2216 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2218 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2220 GOTO(out, rc = PTR_ERR(mo));
2222 if (!mdt_object_exists(mo))
2223 GOTO(out_obj, rc = -ENOENT);
2225 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2226 GOTO(out_obj, rc = -EINVAL);
2228 if (mdt_object_remote(mo))
2229 GOTO(out_obj, rc = -EREMOTE);
2231 lease = ldlm_handle2lock(rr->rr_handle);
2233 GOTO(out_obj, rc = -ESTALE);
2235 /* It's really necessary to grab open_sem and check if the lease lock
2236 * has been lost. There would exist a concurrent writer coming in and
2237 * generating some dirty data in memory cache, the writeback would fail
2238 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2239 if (!down_write_trylock(&mo->mot_open_sem))
2240 GOTO(out_put_lease, rc = -EBUSY);
2242 lock_res_and_lock(lease);
2243 lease_broken = ldlm_is_cancel(lease);
2244 unlock_res_and_lock(lease);
2246 GOTO(out_unlock, rc = -EBUSY);
2248 /* the file has yet opened by anyone else after we took the lease. */
2249 layout.mlc_opc = MD_LAYOUT_RESYNC;
2250 rc = mdt_layout_change(info, mo, &layout);
2252 GOTO(out_unlock, rc = -EBUSY);
2254 ma->ma_need = MA_INODE;
2256 rc = mdt_attr_get_complex(info, mo, ma);
2258 GOTO(out_unlock, rc);
2260 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2261 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2265 up_write(&mo->mot_open_sem);
2267 LDLM_LOCK_PUT(lease);
2269 mdt_object_put(info->mti_env, mo);
2271 mdt_client_compatibility(info);
2272 rc2 = mdt_fix_reply(info);
2278 struct mdt_reinter {
2279 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2280 enum lprocfs_extra_opc mr_extra_opc;
2283 static const struct mdt_reinter mdt_reinters[] = {
2285 .mr_handler = &mdt_reint_setattr,
2286 .mr_extra_opc = MDS_REINT_SETATTR,
2289 .mr_handler = &mdt_reint_create,
2290 .mr_extra_opc = MDS_REINT_CREATE,
2293 .mr_handler = &mdt_reint_link,
2294 .mr_extra_opc = MDS_REINT_LINK,
2297 .mr_handler = &mdt_reint_unlink,
2298 .mr_extra_opc = MDS_REINT_UNLINK,
2301 .mr_handler = &mdt_reint_rename,
2302 .mr_extra_opc = MDS_REINT_RENAME,
2305 .mr_handler = &mdt_reint_open,
2306 .mr_extra_opc = MDS_REINT_OPEN,
2308 [REINT_SETXATTR] = {
2309 .mr_handler = &mdt_reint_setxattr,
2310 .mr_extra_opc = MDS_REINT_SETXATTR,
2313 .mr_handler = &mdt_reint_unlink,
2314 .mr_extra_opc = MDS_REINT_UNLINK,
2317 .mr_handler = &mdt_reint_migrate,
2318 .mr_extra_opc = MDS_REINT_RENAME,
2321 .mr_handler = &mdt_reint_resync,
2322 .mr_extra_opc = MDS_REINT_RESYNC,
2326 int mdt_reint_rec(struct mdt_thread_info *info,
2327 struct mdt_lock_handle *lhc)
2329 const struct mdt_reinter *mr;
2333 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2336 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2337 if (mr->mr_handler == NULL)
2340 rc = (*mr->mr_handler)(info, lhc);
2342 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2343 PTLRPC_LAST_CNTR + mr->mr_extra_opc);