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, 2017, 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 /* VBR: update version if attr changed are important for recovery */
640 /* update on-disk version of changed object */
641 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
642 rc = mdt_version_get_check_save(info, mo, 0);
644 GOTO(out_unlock, rc);
647 /* Ensure constant striping during chown(). See LU-2789. */
648 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
649 mutex_lock(&mo->mot_lov_mutex);
651 /* all attrs are packed into mti_attr in unpack_setattr */
652 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
654 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
655 mutex_unlock(&mo->mot_lov_mutex);
658 GOTO(out_unlock, rc);
659 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
662 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo, rc);
663 mdt_object_unlock(info, mo, lh, rc);
668 * Check HSM flags and add HS_DIRTY flag if relevant.
670 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
671 * and is not RELEASED.
673 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
679 /* If the file was modified, add the dirty flag */
680 ma->ma_need = MA_HSM;
681 rc = mdt_attr_get_complex(info, mo, ma);
683 CERROR("file attribute read error for "DFID": %d.\n",
684 PFID(mdt_object_fid(mo)), rc);
688 /* If an up2date copy exists in the backend, add dirty flag */
689 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
690 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
691 ma->ma_hsm.mh_flags |= HS_DIRTY;
693 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
695 CERROR("file attribute change error for "DFID": %d\n",
696 PFID(mdt_object_fid(mo)), rc);
702 static int mdt_reint_setattr(struct mdt_thread_info *info,
703 struct mdt_lock_handle *lhc)
705 struct md_attr *ma = &info->mti_attr;
706 struct mdt_reint_record *rr = &info->mti_rr;
707 struct ptlrpc_request *req = mdt_info_req(info);
708 struct mdt_object *mo;
709 struct mdt_body *repbody;
713 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
714 (unsigned int)ma->ma_attr.la_valid);
716 if (info->mti_dlm_req)
717 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
719 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
720 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
722 GOTO(out, rc = PTR_ERR(mo));
724 if (!mdt_object_exists(mo))
725 GOTO(out_put, rc = -ENOENT);
727 if (mdt_object_remote(mo))
728 GOTO(out_put, rc = -EREMOTE);
730 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
731 /* Check write access for the O_TRUNC case */
732 if (mdt_write_read(mo) < 0)
733 GOTO(out_put, rc = -ETXTBSY);
735 /* LU-10286: compatibility check for FLR.
736 * Please check the comment in mdt_finish_open() for details */
737 if (!exp_connect_flr(info->mti_exp)) {
738 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
739 if (rc < 0 && rc != -ENODATA)
742 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
743 GOTO(out_put, rc = -EOPNOTSUPP);
747 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
748 if (ma->ma_valid & MA_LOV)
749 GOTO(out_put, rc = -EPROTO);
751 rc = mdt_attr_set(info, mo, ma);
754 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
755 (ma->ma_valid & MA_INODE)) {
756 struct lu_buf *buf = &info->mti_buf;
757 struct mdt_lock_handle *lh;
759 rc = mdt_remote_permission(info);
763 if (ma->ma_attr.la_valid != 0)
764 GOTO(out_put, rc = -EPROTO);
766 lh = &info->mti_lh[MDT_LH_PARENT];
767 mdt_lock_reg_init(lh, LCK_PW);
769 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
773 if (ma->ma_valid & MA_LOV) {
774 buf->lb_buf = ma->ma_lmm;
775 buf->lb_len = ma->ma_lmm_size;
777 buf->lb_buf = ma->ma_lmv;
778 buf->lb_len = ma->ma_lmv_size;
780 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
781 (ma->ma_valid & MA_LOV) ?
782 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
785 mdt_object_unlock(info, mo, lh, rc);
789 GOTO(out_put, rc = -EPROTO);
792 /* If file data is modified, add the dirty flag */
793 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
794 rc = mdt_add_dirty_flag(info, mo, ma);
796 ma->ma_need = MA_INODE;
798 rc = mdt_attr_get_complex(info, mo, ma);
802 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
806 mdt_object_put(info->mti_env, mo);
809 mdt_counter_incr(req, LPROC_MDT_SETATTR);
811 mdt_client_compatibility(info);
812 rc2 = mdt_fix_reply(info);
818 static int mdt_reint_create(struct mdt_thread_info *info,
819 struct mdt_lock_handle *lhc)
821 struct ptlrpc_request *req = mdt_info_req(info);
825 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
826 RETURN(err_serious(-ESTALE));
828 if (info->mti_dlm_req)
829 ldlm_request_cancel(mdt_info_req(info),
830 info->mti_dlm_req, 0, LATF_SKIP);
832 if (!lu_name_is_valid(&info->mti_rr.rr_name))
835 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
837 mdt_counter_incr(req, LPROC_MDT_MKDIR);
845 /* Special file should stay on the same node as parent. */
846 mdt_counter_incr(req, LPROC_MDT_MKNOD);
849 CERROR("%s: Unsupported mode %o\n",
850 mdt_obd_name(info->mti_mdt),
851 info->mti_attr.ma_attr.la_mode);
852 RETURN(err_serious(-EOPNOTSUPP));
855 rc = mdt_create(info);
860 * VBR: save parent version in reply and child version getting by its name.
861 * Version of child is getting and checking during its lookup. If
863 static int mdt_reint_unlink(struct mdt_thread_info *info,
864 struct mdt_lock_handle *lhc)
866 struct mdt_reint_record *rr = &info->mti_rr;
867 struct ptlrpc_request *req = mdt_info_req(info);
868 struct md_attr *ma = &info->mti_attr;
869 struct lu_fid *child_fid = &info->mti_tmp_fid1;
870 struct mdt_object *mp;
871 struct mdt_object *mc;
872 struct mdt_lock_handle *parent_lh;
873 struct mdt_lock_handle *child_lh;
874 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
875 struct lu_fid *s0_fid = &info->mti_tmp_fid2;
876 struct mdt_lock_handle *s0_lh = NULL;
877 struct mdt_object *s0_obj = NULL;
879 bool cos_incompat = false;
885 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
886 PNAME(&rr->rr_name));
888 if (info->mti_dlm_req)
889 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
891 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
892 RETURN(err_serious(-ENOENT));
894 if (!fid_is_md_operative(rr->rr_fid1))
897 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
901 if (mdt_object_remote(mp)) {
904 rc = mdt_version_get_check_save(info, mp, 0);
906 GOTO(put_parent, rc);
910 parent_lh = &info->mti_lh[MDT_LH_PARENT];
911 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
912 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
915 GOTO(put_parent, rc);
917 /* lookup child object along with version checking */
919 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
921 /* Name might not be able to find during resend of
922 * remote unlink, considering following case.
923 * dir_A is a remote directory, the name entry of
924 * dir_A is on MDT0, the directory is on MDT1,
926 * 1. client sends unlink req to MDT1.
927 * 2. MDT1 sends name delete update to MDT0.
928 * 3. name entry is being deleted in MDT0 synchronously.
929 * 4. MDT1 is restarted.
930 * 5. client resends unlink req to MDT1. So it can not
931 * find the name entry on MDT0 anymore.
932 * In this case, MDT1 only needs to destory the local
935 if (mdt_object_remote(mp) && rc == -ENOENT &&
936 !fid_is_zero(rr->rr_fid2) &&
937 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
939 *child_fid = *rr->rr_fid2;
941 GOTO(unlock_parent, rc);
945 if (!fid_is_md_operative(child_fid))
946 GOTO(unlock_parent, rc = -EPERM);
948 /* We will lock the child regardless it is local or remote. No harm. */
949 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
951 GOTO(unlock_parent, rc = PTR_ERR(mc));
953 if (!cos_incompat && mdt_init_slaves(info, mc, s0_fid) > 0) {
955 mdt_object_put(info->mti_env, mc);
956 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
960 child_lh = &info->mti_lh[MDT_LH_CHILD];
961 mdt_lock_reg_init(child_lh, LCK_EX);
962 if (info->mti_spec.sp_rm_entry) {
963 struct lu_ucred *uc = mdt_ucred(info);
965 if (!mdt_is_dne_client(req->rq_export))
966 /* Return -ENOTSUPP for old client */
967 GOTO(put_child, rc = -ENOTSUPP);
969 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
970 GOTO(put_child, rc = -EPERM);
972 ma->ma_need = MA_INODE;
974 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
975 NULL, &rr->rr_name, ma, no_name);
979 if (mdt_object_remote(mc)) {
980 struct mdt_body *repbody;
982 if (!fid_is_zero(rr->rr_fid2)) {
983 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
984 mdt_obd_name(info->mti_mdt),
985 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
986 GOTO(put_child, rc = -ENOENT);
988 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
989 mdt_obd_name(info->mti_mdt),
990 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
992 if (!mdt_is_dne_client(req->rq_export))
993 /* Return -ENOTSUPP for old client */
994 GOTO(put_child, rc = -ENOTSUPP);
996 /* Revoke the LOOKUP lock of the remote object granted by
997 * this MDT. Since the unlink will happen on another MDT,
998 * it will release the LOOKUP lock right away. Then What
999 * would happen if another client try to grab the LOOKUP
1000 * lock at the same time with unlink XXX */
1001 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1002 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1003 LASSERT(repbody != NULL);
1004 repbody->mbo_fid1 = *mdt_object_fid(mc);
1005 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1006 GOTO(unlock_child, rc = -EREMOTE);
1008 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1009 * this now because a running HSM restore on the child (unlink
1010 * victim) will hold the layout lock. See LU-4002. */
1011 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1012 if (mdt_object_remote(mp)) {
1013 /* Enqueue lookup lock from parent MDT */
1014 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1015 &child_lh->mlh_rreg_lh,
1016 child_lh->mlh_rreg_mode,
1017 MDS_INODELOCK_LOOKUP, false);
1019 GOTO(put_child, rc);
1021 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1024 rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits,
1027 GOTO(unlock_child, rc);
1030 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1031 * whether need MA_LOV and MA_COOKIE.
1033 ma->ma_need = MA_INODE;
1036 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1037 mdt_lock_reg_init(s0_lh, LCK_EX);
1038 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_fid,
1039 s0_lh, &s0_obj, einfo);
1041 GOTO(unlock_child, rc);
1043 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1044 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1045 /* save version when object is locked */
1046 mdt_version_get_save(info, mc, 1);
1048 mutex_lock(&mc->mot_lov_mutex);
1050 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1051 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1053 mutex_unlock(&mc->mot_lov_mutex);
1055 GOTO(unlock_child, rc);
1057 if (!lu_object_is_dying(&mc->mot_header)) {
1058 rc = mdt_attr_get_complex(info, mc, ma);
1062 mdt_dom_check_and_discard(info, mc);
1064 mdt_handle_last_unlink(info, mc, ma);
1067 if (ma->ma_valid & MA_INODE) {
1068 switch (ma->ma_attr.la_mode & S_IFMT) {
1070 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1078 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1081 LASSERTF(0, "bad file type %o unlinking\n",
1082 ma->ma_attr.la_mode);
1089 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo,
1091 mdt_object_unlock(info, mc, child_lh, rc);
1093 mdt_object_put(info->mti_env, mc);
1095 mdt_object_unlock(info, mp, parent_lh, rc);
1097 mdt_object_put(info->mti_env, mp);
1102 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1105 static int mdt_reint_link(struct mdt_thread_info *info,
1106 struct mdt_lock_handle *lhc)
1108 struct mdt_reint_record *rr = &info->mti_rr;
1109 struct ptlrpc_request *req = mdt_info_req(info);
1110 struct md_attr *ma = &info->mti_attr;
1111 struct mdt_object *ms;
1112 struct mdt_object *mp;
1113 struct mdt_lock_handle *lhs;
1114 struct mdt_lock_handle *lhp;
1119 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1120 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1122 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1123 RETURN(err_serious(-ENOENT));
1125 if (info->mti_dlm_req)
1126 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1128 /* Invalid case so return error immediately instead of
1130 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1133 if (!fid_is_md_operative(rr->rr_fid1) ||
1134 !fid_is_md_operative(rr->rr_fid2))
1137 /* step 1: find target parent dir */
1138 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1140 RETURN(PTR_ERR(mp));
1142 rc = mdt_version_get_check_save(info, mp, 0);
1144 GOTO(put_parent, rc);
1146 /* step 2: find source */
1147 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1149 GOTO(put_parent, rc = PTR_ERR(ms));
1151 if (!mdt_object_exists(ms)) {
1152 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1153 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1154 GOTO(put_source, rc = -ENOENT);
1157 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1159 lhp = &info->mti_lh[MDT_LH_PARENT];
1160 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1161 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1164 GOTO(put_source, rc);
1166 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1168 lhs = &info->mti_lh[MDT_LH_CHILD];
1169 mdt_lock_reg_init(lhs, LCK_EX);
1170 rc = mdt_reint_object_lock(info, ms, lhs,
1171 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1174 GOTO(unlock_parent, rc);
1176 /* step 3: link it */
1177 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1178 OBD_FAIL_MDS_REINT_LINK_WRITE);
1180 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1181 rc = mdt_version_get_check_save(info, ms, 1);
1183 GOTO(unlock_source, rc);
1185 /** check target version by name during replay */
1186 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1187 &info->mti_tmp_fid1, 2);
1188 if (rc != 0 && rc != -ENOENT)
1189 GOTO(unlock_source, rc);
1190 /* save version of file name for replay, it must be ENOENT here */
1191 if (!req_is_replay(mdt_info_req(info))) {
1192 if (rc != -ENOENT) {
1193 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1194 PNAME(&rr->rr_name));
1195 GOTO(unlock_source, rc = -EEXIST);
1197 info->mti_ver[2] = ENOENT_VERSION;
1198 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1201 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1202 mdt_object_child(ms), &rr->rr_name, ma);
1205 mdt_counter_incr(req, LPROC_MDT_LINK);
1209 mdt_object_unlock(info, ms, lhs, rc);
1211 mdt_object_unlock(info, mp, lhp, rc);
1213 mdt_object_put(info->mti_env, ms);
1215 mdt_object_put(info->mti_env, mp);
1219 * lock the part of the directory according to the hash of the name
1220 * (lh->mlh_pdo_hash) in parallel directory lock.
1222 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1223 struct mdt_lock_handle *lh,
1224 struct mdt_object *obj, __u64 ibits,
1227 struct ldlm_res_id *res = &info->mti_res_id;
1228 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1229 union ldlm_policy_data *policy = &info->mti_policy;
1230 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1234 * Finish res_id initializing by name hash marking part of
1235 * directory which is taking modification.
1237 LASSERT(lh->mlh_pdo_hash != 0);
1238 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1239 memset(policy, 0, sizeof(*policy));
1240 policy->l_inodebits.bits = ibits;
1242 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1243 dlmflags |= LDLM_FL_COS_INCOMPAT;
1245 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1246 * going to be sent to client. If it is - mdt_intent_policy() path will
1247 * fix it up and turn FL_LOCAL flag off.
1249 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1250 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1255 * Get BFL lock for rename or migrate process.
1257 static int mdt_rename_lock(struct mdt_thread_info *info,
1258 struct lustre_handle *lh)
1263 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1264 struct lu_fid *fid = &info->mti_tmp_fid1;
1265 struct mdt_object *obj;
1267 /* XXX, right now, it has to use object API to
1268 * enqueue lock cross MDT, so it will enqueue
1269 * rename lock(with LUSTRE_BFL_FID) by root object */
1271 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1273 RETURN(PTR_ERR(obj));
1275 rc = mdt_remote_object_lock(info, obj,
1276 &LUSTRE_BFL_FID, lh,
1278 MDS_INODELOCK_UPDATE, false);
1279 mdt_object_put(info->mti_env, obj);
1281 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1282 union ldlm_policy_data *policy = &info->mti_policy;
1283 struct ldlm_res_id *res_id = &info->mti_res_id;
1286 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1287 memset(policy, 0, sizeof *policy);
1288 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1289 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1290 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1291 LCK_EX, &flags, ldlm_blocking_ast,
1292 ldlm_completion_ast, NULL, NULL, 0,
1294 &info->mti_exp->exp_handle.h_cookie,
1301 static void mdt_rename_unlock(struct lustre_handle *lh)
1304 LASSERT(lustre_handle_is_used(lh));
1305 /* Cancel the single rename lock right away */
1306 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1311 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1312 * target. Source should not be ancestor of target dir. May be other rename
1313 * checks can be moved here later.
1315 static int mdt_is_subdir(struct mdt_thread_info *info,
1316 struct mdt_object *dir,
1317 const struct lu_fid *fid)
1319 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1323 /* If the source and target are in the same directory, they can not
1324 * be parent/child relationship, so subdir check is not needed */
1325 if (lu_fid_eq(&dir_fid, fid))
1328 if (!mdt_object_exists(dir))
1331 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1334 CERROR("%s: failed subdir check in "DFID" for "DFID
1335 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1336 PFID(&dir_fid), PFID(fid), rc);
1337 /* Return EINVAL only if a parent is the @fid */
1341 /* check the found fid */
1342 if (lu_fid_eq(&dir_fid, fid))
1349 /* Update object linkEA */
1350 struct mdt_lock_list {
1351 struct mdt_object *mll_obj;
1352 struct mdt_lock_handle mll_lh;
1353 struct list_head mll_list;
1356 static void mdt_unlock_list(struct mdt_thread_info *info,
1357 struct list_head *list, int rc)
1359 struct mdt_lock_list *mll;
1360 struct mdt_lock_list *mll2;
1362 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1363 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1364 list_del(&mll->mll_list);
1369 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1370 struct mdt_object *obj,
1371 struct mdt_object *pobj,
1372 struct list_head *lock_list)
1374 struct lu_buf *buf = &info->mti_big_buf;
1375 struct linkea_data ldata = { NULL };
1381 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1384 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1385 if (buf->lb_buf == NULL)
1389 rc = mdt_links_read(info, obj, &ldata);
1391 if (rc == -ENOENT || rc == -ENODATA)
1396 /* ignore the migrating parent(@pobj) */
1397 retry_count = ldata.ld_leh->leh_reccount - 1;
1400 LASSERT(ldata.ld_leh != NULL);
1401 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1402 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1403 struct mdt_device *mdt = info->mti_mdt;
1404 struct mdt_object *mdt_pobj;
1405 struct mdt_lock_list *mll;
1406 struct lu_name name;
1410 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1412 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1413 if (IS_ERR(mdt_pobj)) {
1414 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1415 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1419 if (!mdt_object_exists(mdt_pobj)) {
1420 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1421 mdt_obd_name(mdt), PFID(&fid));
1422 mdt_object_put(info->mti_env, mdt_pobj);
1426 /* Check if the object already exists in the list */
1427 list_for_each_entry(mll, lock_list, mll_list) {
1428 if (mll->mll_obj == mdt_pobj) {
1429 mdt_object_put(info->mti_env, mdt_pobj);
1434 if (mdt_pobj == pobj) {
1435 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1436 mdt_obd_name(mdt), PFID(&fid));
1437 mdt_object_put(info->mti_env, mdt_pobj);
1443 mdt_object_put(info->mti_env, mdt_pobj);
1444 GOTO(out, rc = -ENOMEM);
1447 /* Since this needs to lock all of objects in linkea, to avoid
1448 * deadlocks, because it does not follow parent-child order as
1449 * other MDT operation, let's use try_lock here and if the lock
1450 * cannot be gotten because of conflicting locks, then drop all
1451 * current locks, send an AST to the client, and start again. */
1452 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1454 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits,
1455 MDS_INODELOCK_UPDATE, true);
1456 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1457 mdt_unlock_list(info, lock_list, 0);
1459 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1460 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1463 if (retry_count == 0) {
1464 mdt_object_put(info->mti_env, mdt_pobj);
1466 GOTO(out, rc = -EBUSY);
1469 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1470 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1471 MDS_INODELOCK_UPDATE);
1473 mdt_object_put(info->mti_env, mdt_pobj);
1478 if (mdt_object_remote(mdt_pobj)) {
1479 struct ldlm_lock *lock;
1481 /* For remote object, Set lock to cb_atomic,
1482 * so lock can be released in blocking_ast()
1483 * immediately, then the next try_lock will
1484 * have better chance to succeds */
1486 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1487 LASSERT(lock != NULL);
1488 lock_res_and_lock(lock);
1489 ldlm_set_atomic_cb(lock);
1490 unlock_res_and_lock(lock);
1491 LDLM_LOCK_PUT(lock);
1493 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1499 INIT_LIST_HEAD(&mll->mll_list);
1500 mll->mll_obj = mdt_pobj;
1501 list_add_tail(&mll->mll_list, lock_list);
1503 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1508 mdt_unlock_list(info, lock_list, rc);
1512 /* migrate files from one MDT to another MDT */
1513 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1514 struct mdt_lock_handle *lhc)
1516 struct mdt_reint_record *rr = &info->mti_rr;
1517 struct md_attr *ma = &info->mti_attr;
1518 struct mdt_object *msrcdir;
1519 struct mdt_object *mold;
1520 struct mdt_object *mnew = NULL;
1521 struct mdt_lock_handle *lh_dirp;
1522 struct mdt_lock_handle *lh_childp;
1523 struct mdt_lock_handle *lh_tgtp = NULL;
1524 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1525 struct list_head lock_list;
1527 struct ldlm_lock *lease = NULL;
1528 bool lock_open_sem = false;
1532 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1533 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1535 /* 1: lock the source dir. */
1536 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1537 if (IS_ERR(msrcdir)) {
1538 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1539 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1540 (int)PTR_ERR(msrcdir));
1541 RETURN(PTR_ERR(msrcdir));
1544 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1545 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1546 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1549 GOTO(out_put_parent, rc);
1551 if (!mdt_object_remote(msrcdir)) {
1552 rc = mdt_version_get_check_save(info, msrcdir, 0);
1554 GOTO(out_unlock_parent, rc);
1557 /* 2: sanity check and find the object to be migrated. */
1559 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1561 GOTO(out_unlock_parent, rc);
1563 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1564 GOTO(out_unlock_parent, rc = -EINVAL);
1566 if (!fid_is_md_operative(old_fid))
1567 GOTO(out_unlock_parent, rc = -EPERM);
1569 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1570 GOTO(out_unlock_parent, rc = -EPERM);
1572 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1574 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1576 if (mdt_object_remote(mold)) {
1577 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1578 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1579 GOTO(out_put_child, rc = -EREMOTE);
1582 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1583 !mdt_object_remote(msrcdir)) {
1584 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1585 " MDT, which should be migrated first:"
1586 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1587 PFID(mdt_object_fid(msrcdir)), -EPERM);
1588 GOTO(out_put_child, rc = -EPERM);
1591 rc = mdt_remote_permission(info);
1593 GOTO(out_put_child, rc);
1595 /* 3: iterate the linkea of the object and lock all of the objects */
1596 INIT_LIST_HEAD(&lock_list);
1597 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1599 GOTO(out_put_child, rc);
1601 if (info->mti_spec.sp_migrate_close) {
1602 struct close_data *data;
1603 struct mdt_body *repbody;
1604 bool lease_broken = false;
1606 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1608 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1610 GOTO(out_lease, rc = -EPROTO);
1612 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1614 GOTO(out_lease, rc = -EPROTO);
1616 lease = ldlm_handle2lock(&data->cd_handle);
1618 GOTO(out_lease, rc = -ESTALE);
1620 /* try to hold open_sem so that nobody else can open the file */
1621 if (!down_write_trylock(&mold->mot_open_sem)) {
1622 ldlm_lock_cancel(lease);
1623 GOTO(out_lease, rc = -EBUSY);
1626 lock_open_sem = true;
1627 /* Check if the lease open lease has already canceled */
1628 lock_res_and_lock(lease);
1629 lease_broken = ldlm_is_cancel(lease);
1630 unlock_res_and_lock(lease);
1632 LDLM_DEBUG(lease, DFID " lease broken? %d",
1633 PFID(mdt_object_fid(mold)), lease_broken);
1635 /* Cancel server side lease. Client side counterpart should
1636 * have been cancelled. It's okay to cancel it now as we've
1637 * held mot_open_sem. */
1638 ldlm_lock_cancel(lease);
1641 GOTO(out_lease, rc = -EAGAIN);
1643 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1644 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1645 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1647 GOTO(out_unlock_list, rc);
1650 /* 4: lock of the object migrated object */
1651 lh_childp = &info->mti_lh[MDT_LH_OLD];
1652 mdt_lock_reg_init(lh_childp, LCK_EX);
1653 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1654 MDS_INODELOCK_LAYOUT;
1655 if (mdt_object_remote(msrcdir)) {
1656 /* Enqueue lookup lock from the parent MDT */
1657 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1658 &lh_childp->mlh_rreg_lh,
1659 lh_childp->mlh_rreg_mode,
1660 MDS_INODELOCK_LOOKUP, false);
1662 GOTO(out_unlock_list, rc);
1664 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1667 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1669 GOTO(out_unlock_child, rc);
1671 /* Migration is incompatible with HSM. */
1672 ma->ma_need = MA_HSM;
1674 rc = mdt_attr_get_complex(info, mold, ma);
1676 GOTO(out_unlock_child, rc);
1678 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1681 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1682 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1683 GOTO(out_unlock_child, rc);
1686 ma->ma_need = MA_LMV;
1688 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1689 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1690 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1692 GOTO(out_unlock_child, rc);
1694 if ((ma->ma_valid & MA_LMV)) {
1695 struct lmv_mds_md_v1 *lmm1;
1697 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1698 lmm1 = &ma->ma_lmv->lmv_md_v1;
1699 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1700 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1701 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1702 PFID(mdt_object_fid(mold)), -EPERM);
1703 GOTO(out_unlock_child, rc = -EPERM);
1706 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1707 GOTO(out_unlock_child, rc = -EINVAL);
1709 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1710 &lmm1->lmv_stripe_fids[1]);
1712 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1714 if (!mdt_object_remote(mnew)) {
1716 "%s: "DFID" being migrated is on this MDT:"
1717 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1718 PFID(rr->rr_fid2), -EPERM);
1719 GOTO(out_put_new, rc = -EPERM);
1722 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1723 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1724 rc = mdt_remote_object_lock(info, mnew,
1725 mdt_object_fid(mnew),
1726 &lh_tgtp->mlh_rreg_lh,
1727 lh_tgtp->mlh_rreg_mode,
1728 MDS_INODELOCK_UPDATE, false);
1731 GOTO(out_put_new, rc);
1734 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1737 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1738 if (!mdt_object_remote(mnew)) {
1739 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1740 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1741 PFID(rr->rr_fid2), -EXDEV);
1742 GOTO(out_put_new, rc = -EXDEV);
1747 mdt_reint_init_ma(info, ma);
1749 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1750 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1752 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1753 mdt_object_child(mold), &rr->rr_name,
1754 mdt_object_child(mnew), ma);
1756 GOTO(out_unlock_new, rc);
1759 if (lh_tgtp != NULL)
1760 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1763 mdt_object_put(info->mti_env, mnew);
1765 mdt_object_unlock(info, mold, lh_childp, rc);
1767 /* we don't really modify linkea objects, so we can safely decref these
1768 * locks, and this can avoid saving them as COS locks, which may prevent
1769 * subsequent migrate. */
1770 mdt_unlock_list(info, &lock_list, 1);
1771 if (lease != NULL) {
1772 ldlm_reprocess_all(lease->l_resource);
1773 LDLM_LOCK_PUT(lease);
1777 up_write(&mold->mot_open_sem);
1779 mdt_object_put(info->mti_env, mold);
1781 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1783 mdt_object_put(info->mti_env, msrcdir);
1788 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1789 const struct lu_fid *fid,
1792 struct mdt_object *dir;
1796 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1800 /* check early, the real version will be saved after locking */
1801 rc = mdt_version_get_check(info, dir, idx);
1807 mdt_object_put(info->mti_env, dir);
1811 static int mdt_object_lock_save(struct mdt_thread_info *info,
1812 struct mdt_object *dir,
1813 struct mdt_lock_handle *lh,
1814 int idx, bool cos_incompat)
1818 /* we lock the target dir if it is local */
1819 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1824 /* get and save correct version after locking */
1825 mdt_version_get_save(info, dir, idx);
1830 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1831 * 2 - srcdir child; 3 - tgtdir child.
1832 * Update on disk version of srcdir child.
1835 * For DNE phase I, only these renames are allowed
1836 * mv src_p/src_c tgt_p/tgt_c
1837 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1838 * 2. src_p and tgt_p are same directory, and tgt_c does not
1839 * exists. In this case, all of modification will happen
1840 * in the MDT where ithesource parent is, only one remote
1841 * update is needed, i.e. set c_time/m_time on the child.
1842 * And tgt_c will be still in the same MDT as the original
1845 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1846 struct mdt_lock_handle *lhc)
1848 struct mdt_reint_record *rr = &info->mti_rr;
1849 struct md_attr *ma = &info->mti_attr;
1850 struct ptlrpc_request *req = mdt_info_req(info);
1851 struct mdt_object *msrcdir = NULL;
1852 struct mdt_object *mtgtdir = NULL;
1853 struct mdt_object *mold;
1854 struct mdt_object *mnew = NULL;
1855 struct mdt_lock_handle *lh_srcdirp;
1856 struct mdt_lock_handle *lh_tgtdirp;
1857 struct mdt_lock_handle *lh_oldp = NULL;
1858 struct mdt_lock_handle *lh_newp = NULL;
1859 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1860 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1862 bool reverse = false;
1867 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1868 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1869 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1871 /* find both parents. */
1872 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1873 if (IS_ERR(msrcdir))
1874 RETURN(PTR_ERR(msrcdir));
1876 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1878 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1880 mdt_object_get(info->mti_env, mtgtdir);
1882 /* Check if the @msrcdir is not a child of the @mtgtdir,
1883 * otherwise a reverse locking must take place. */
1884 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1888 GOTO(out_put_srcdir, rc);
1890 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1891 if (IS_ERR(mtgtdir))
1892 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1895 /* source needs to be looked up after locking source parent, otherwise
1896 * this rename may race with unlink source, and cause rename hang, see
1897 * sanityn.sh 55b, so check parents first, if later we found source is
1898 * remote, relock parents. */
1899 cos_incompat = (mdt_object_remote(msrcdir) ||
1900 mdt_object_remote(mtgtdir));
1902 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1904 /* lock parents in the proper order. */
1905 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1906 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1909 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1910 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1913 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1916 GOTO(out_put_tgtdir, rc);
1918 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1920 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1923 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1924 GOTO(out_put_tgtdir, rc);
1927 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1930 GOTO(out_put_tgtdir, rc);
1932 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1934 if (mtgtdir != msrcdir) {
1935 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1937 } else if (lh_srcdirp->mlh_pdo_hash !=
1938 lh_tgtdirp->mlh_pdo_hash) {
1939 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1940 MDS_INODELOCK_UPDATE,
1942 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1945 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1946 GOTO(out_put_tgtdir, rc);
1950 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1951 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1953 /* find mold object. */
1955 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1957 GOTO(out_unlock_parents, rc);
1959 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1960 GOTO(out_unlock_parents, rc = -EINVAL);
1962 if (!fid_is_md_operative(old_fid))
1963 GOTO(out_unlock_parents, rc = -EPERM);
1965 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1967 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1969 /* Check if @mtgtdir is subdir of @mold, before locking child
1970 * to avoid reverse locking. */
1971 if (mtgtdir != msrcdir) {
1972 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1974 GOTO(out_put_old, rc);
1977 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1978 /* save version after locking */
1979 mdt_version_get_save(info, mold, 2);
1981 if (!cos_incompat && mdt_object_remote(mold)) {
1982 cos_incompat = true;
1983 mdt_object_put(info->mti_env, mold);
1984 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1985 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1989 /* find mnew object:
1990 * mnew target object may not exist now
1991 * lookup with version checking */
1993 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1996 /* the new_fid should have been filled at this moment */
1997 if (lu_fid_eq(old_fid, new_fid))
1998 GOTO(out_put_old, rc);
2000 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2001 lu_fid_eq(new_fid, rr->rr_fid2))
2002 GOTO(out_put_old, rc = -EINVAL);
2004 if (!fid_is_md_operative(new_fid))
2005 GOTO(out_put_old, rc = -EPERM);
2007 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2009 GOTO(out_put_old, rc = PTR_ERR(mnew));
2011 if (mdt_object_remote(mnew)) {
2012 struct mdt_body *repbody;
2014 /* Always send rename req to the target child MDT */
2015 repbody = req_capsule_server_get(info->mti_pill,
2017 LASSERT(repbody != NULL);
2018 repbody->mbo_fid1 = *new_fid;
2019 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2020 GOTO(out_put_new, rc = -EXDEV);
2022 /* Before locking the target dir, check we do not replace
2023 * a dir with a non-dir, otherwise it may deadlock with
2024 * link op which tries to create a link in this dir
2025 * back to this non-dir. */
2026 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2027 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2028 GOTO(out_put_new, rc = -EISDIR);
2030 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2031 mdt_lock_reg_init(lh_oldp, LCK_EX);
2032 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2033 if (mdt_object_remote(msrcdir)) {
2034 /* Enqueue lookup lock from the parent MDT */
2035 rc = mdt_remote_object_lock(info, msrcdir,
2036 mdt_object_fid(mold),
2037 &lh_oldp->mlh_rreg_lh,
2038 lh_oldp->mlh_rreg_mode,
2039 MDS_INODELOCK_LOOKUP,
2042 GOTO(out_put_new, rc);
2044 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2047 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2050 GOTO(out_unlock_old, rc);
2052 /* Check if @msrcdir is subdir of @mnew, before locking child
2053 * to avoid reverse locking. */
2054 if (mtgtdir != msrcdir) {
2055 rc = mdt_is_subdir(info, msrcdir, new_fid);
2057 GOTO(out_unlock_old, rc);
2060 /* We used to acquire MDS_INODELOCK_FULL here but we
2061 * can't do this now because a running HSM restore on
2062 * the rename onto victim will hold the layout
2063 * lock. See LU-4002. */
2065 lh_newp = &info->mti_lh[MDT_LH_NEW];
2066 mdt_lock_reg_init(lh_newp, LCK_EX);
2067 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2068 MDS_INODELOCK_LOOKUP |
2069 MDS_INODELOCK_UPDATE,
2072 GOTO(out_unlock_old, rc);
2074 /* get and save version after locking */
2075 mdt_version_get_save(info, mnew, 3);
2076 } else if (rc != -EREMOTE && rc != -ENOENT) {
2077 GOTO(out_put_old, rc);
2079 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2080 mdt_lock_reg_init(lh_oldp, LCK_EX);
2081 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2082 if (mdt_object_remote(msrcdir)) {
2083 /* Enqueue lookup lock from the parent MDT */
2084 rc = mdt_remote_object_lock(info, msrcdir,
2085 mdt_object_fid(mold),
2086 &lh_oldp->mlh_rreg_lh,
2087 lh_oldp->mlh_rreg_mode,
2088 MDS_INODELOCK_LOOKUP,
2091 GOTO(out_put_old, rc);
2093 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2096 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2099 GOTO(out_unlock_old, rc);
2101 mdt_enoent_version_save(info, 3);
2104 /* step 5: rename it */
2105 mdt_reint_init_ma(info, ma);
2107 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2108 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2111 mutex_lock(&mnew->mot_lov_mutex);
2113 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2114 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2115 mnew != NULL ? mdt_object_child(mnew) : NULL,
2116 &rr->rr_tgt_name, ma);
2119 mutex_unlock(&mnew->mot_lov_mutex);
2121 /* handle last link of tgt object */
2123 mdt_counter_incr(req, LPROC_MDT_RENAME);
2125 mdt_handle_last_unlink(info, mnew, ma);
2126 mdt_dom_check_and_discard(info, mnew);
2129 mdt_rename_counter_tally(info, info->mti_mdt, req,
2135 mdt_object_unlock(info, mnew, lh_newp, rc);
2137 mdt_object_unlock(info, mold, lh_oldp, rc);
2140 mdt_object_put(info->mti_env, mnew);
2142 mdt_object_put(info->mti_env, mold);
2144 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2145 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2147 mdt_object_put(info->mti_env, mtgtdir);
2149 mdt_object_put(info->mti_env, msrcdir);
2153 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2154 struct mdt_lock_handle *lhc, bool rename)
2156 struct mdt_reint_record *rr = &info->mti_rr;
2157 struct ptlrpc_request *req = mdt_info_req(info);
2158 struct lustre_handle rename_lh = { 0 };
2162 if (info->mti_dlm_req)
2163 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2165 if (!fid_is_md_operative(rr->rr_fid1) ||
2166 !fid_is_md_operative(rr->rr_fid2))
2169 /* Note: do not enqueue rename lock for replay request, because
2170 * if other MDT holds rename lock, but being blocked to wait for
2171 * this MDT to finish its recovery, and the failover MDT can not
2172 * get rename lock, which will cause deadlock. */
2173 if (!req_is_replay(req)) {
2174 rc = mdt_rename_lock(info, &rename_lh);
2176 CERROR("%s: can't lock FS for rename: rc = %d\n",
2177 mdt_obd_name(info->mti_mdt), rc);
2183 rc = mdt_reint_rename_internal(info, lhc);
2185 rc = mdt_reint_migrate_internal(info, lhc);
2187 if (lustre_handle_is_used(&rename_lh))
2188 mdt_rename_unlock(&rename_lh);
2193 static int mdt_reint_rename(struct mdt_thread_info *info,
2194 struct mdt_lock_handle *lhc)
2196 return mdt_reint_rename_or_migrate(info, lhc, true);
2199 static int mdt_reint_migrate(struct mdt_thread_info *info,
2200 struct mdt_lock_handle *lhc)
2202 return mdt_reint_rename_or_migrate(info, lhc, false);
2205 static int mdt_reint_resync(struct mdt_thread_info *info,
2206 struct mdt_lock_handle *lhc)
2208 struct mdt_reint_record *rr = &info->mti_rr;
2209 struct ptlrpc_request *req = mdt_info_req(info);
2210 struct md_attr *ma = &info->mti_attr;
2211 struct mdt_object *mo;
2212 struct ldlm_lock *lease;
2213 struct mdt_body *repbody;
2214 struct md_layout_change layout = { 0 };
2219 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2221 if (info->mti_dlm_req)
2222 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2224 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2226 GOTO(out, rc = PTR_ERR(mo));
2228 if (!mdt_object_exists(mo))
2229 GOTO(out_obj, rc = -ENOENT);
2231 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2232 GOTO(out_obj, rc = -EINVAL);
2234 if (mdt_object_remote(mo))
2235 GOTO(out_obj, rc = -EREMOTE);
2237 lease = ldlm_handle2lock(rr->rr_handle);
2239 GOTO(out_obj, rc = -ESTALE);
2241 /* It's really necessary to grab open_sem and check if the lease lock
2242 * has been lost. There would exist a concurrent writer coming in and
2243 * generating some dirty data in memory cache, the writeback would fail
2244 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2245 if (!down_write_trylock(&mo->mot_open_sem))
2246 GOTO(out_put_lease, rc = -EBUSY);
2248 lock_res_and_lock(lease);
2249 lease_broken = ldlm_is_cancel(lease);
2250 unlock_res_and_lock(lease);
2252 GOTO(out_unlock, rc = -EBUSY);
2254 /* the file has yet opened by anyone else after we took the lease. */
2255 layout.mlc_opc = MD_LAYOUT_RESYNC;
2256 rc = mdt_layout_change(info, mo, &layout);
2258 GOTO(out_unlock, rc);
2260 ma->ma_need = MA_INODE;
2262 rc = mdt_attr_get_complex(info, mo, ma);
2264 GOTO(out_unlock, rc);
2266 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2267 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2271 up_write(&mo->mot_open_sem);
2273 LDLM_LOCK_PUT(lease);
2275 mdt_object_put(info->mti_env, mo);
2277 mdt_client_compatibility(info);
2278 rc2 = mdt_fix_reply(info);
2284 struct mdt_reinter {
2285 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2286 enum lprocfs_extra_opc mr_extra_opc;
2289 static const struct mdt_reinter mdt_reinters[] = {
2291 .mr_handler = &mdt_reint_setattr,
2292 .mr_extra_opc = MDS_REINT_SETATTR,
2295 .mr_handler = &mdt_reint_create,
2296 .mr_extra_opc = MDS_REINT_CREATE,
2299 .mr_handler = &mdt_reint_link,
2300 .mr_extra_opc = MDS_REINT_LINK,
2303 .mr_handler = &mdt_reint_unlink,
2304 .mr_extra_opc = MDS_REINT_UNLINK,
2307 .mr_handler = &mdt_reint_rename,
2308 .mr_extra_opc = MDS_REINT_RENAME,
2311 .mr_handler = &mdt_reint_open,
2312 .mr_extra_opc = MDS_REINT_OPEN,
2314 [REINT_SETXATTR] = {
2315 .mr_handler = &mdt_reint_setxattr,
2316 .mr_extra_opc = MDS_REINT_SETXATTR,
2319 .mr_handler = &mdt_reint_unlink,
2320 .mr_extra_opc = MDS_REINT_UNLINK,
2323 .mr_handler = &mdt_reint_migrate,
2324 .mr_extra_opc = MDS_REINT_RENAME,
2327 .mr_handler = &mdt_reint_resync,
2328 .mr_extra_opc = MDS_REINT_RESYNC,
2332 int mdt_reint_rec(struct mdt_thread_info *info,
2333 struct mdt_lock_handle *lhc)
2335 const struct mdt_reinter *mr;
2339 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2342 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2343 if (mr->mr_handler == NULL)
2346 rc = (*mr->mr_handler)(info, lhc);
2348 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2349 PTLRPC_LAST_CNTR + mr->mr_extra_opc);