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,
294 struct ldlm_enqueue_info *einfo,
297 union ldlm_policy_data *policy = &mti->mti_policy;
298 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
299 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
302 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
303 LASSERT(slave_locks);
305 memset(policy, 0, sizeof(*policy));
306 policy->l_inodebits.bits = einfo->ei_inodebits;
307 mdt_lock_handle_init(lh);
308 mdt_lock_reg_init(lh, einfo->ei_mode);
309 for (i = 0; i < slave_locks->ha_count; i++) {
310 if (test_bit(i, (void *)slave_locks->ha_map))
311 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
313 lh->mlh_reg_lh = slave_locks->ha_handles[i];
314 mdt_object_unlock(mti, NULL, lh, decref);
315 slave_locks->ha_handles[i].cookie = 0ull;
318 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
322 static inline int mdt_object_striped(struct mdt_thread_info *mti,
323 struct mdt_object *obj)
327 if (!S_ISDIR(obj->mot_header.loh_attr))
330 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
333 return rc == -ENODATA ? 0 : rc;
339 * Lock slave stripes if necessary, the lock handles of slave stripes
340 * will be stored in einfo->ei_cbdata.
342 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
343 enum ldlm_mode mode, __u64 ibits,
344 struct ldlm_enqueue_info *einfo)
346 union ldlm_policy_data *policy = &mti->mti_policy;
348 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
350 einfo->ei_type = LDLM_IBITS;
351 einfo->ei_mode = mode;
352 einfo->ei_cb_bl = mdt_remote_blocking_ast;
353 einfo->ei_cb_local_bl = mdt_blocking_ast;
354 einfo->ei_cb_cp = ldlm_completion_ast;
355 einfo->ei_enq_slave = 1;
356 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
357 einfo->ei_inodebits = ibits;
358 memset(policy, 0, sizeof(*policy));
359 policy->l_inodebits.bits = ibits;
361 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
365 static inline int mdt_reint_striped_lock(struct mdt_thread_info *info,
366 struct mdt_object *o,
367 struct mdt_lock_handle *lh,
369 struct ldlm_enqueue_info *einfo,
374 LASSERT(!mdt_object_remote(o));
376 memset(einfo, 0, sizeof(*einfo));
378 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
382 rc = mdt_object_striped(info, o);
385 mdt_object_unlock(info, o, lh, rc);
389 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
391 mdt_object_unlock(info, o, lh, rc);
392 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
400 mdt_reint_striped_unlock(struct mdt_thread_info *info, struct mdt_object *o,
401 struct mdt_lock_handle *lh,
402 struct ldlm_enqueue_info *einfo, int decref)
404 if (einfo->ei_cbdata)
405 mdt_unlock_slaves(info, o, einfo, decref);
406 mdt_object_unlock(info, o, lh, decref);
410 * VBR: we save three versions in reply:
411 * 0 - parent. Check that parent version is the same during replay.
412 * 1 - name. Version of 'name' if file exists with the same name or
413 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
414 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
417 static int mdt_create(struct mdt_thread_info *info)
419 struct mdt_device *mdt = info->mti_mdt;
420 struct mdt_object *parent;
421 struct mdt_object *child;
422 struct mdt_lock_handle *lh;
423 struct mdt_body *repbody;
424 struct md_attr *ma = &info->mti_attr;
425 struct mdt_reint_record *rr = &info->mti_rr;
429 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
431 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
433 if (!fid_is_md_operative(rr->rr_fid1))
436 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
438 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
440 RETURN(PTR_ERR(parent));
442 if (!mdt_object_exists(parent))
443 GOTO(put_parent, rc = -ENOENT);
445 lh = &info->mti_lh[MDT_LH_PARENT];
446 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
447 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
449 GOTO(put_parent, rc);
451 if (!mdt_object_remote(parent)) {
452 rc = mdt_version_get_check_save(info, parent, 0);
454 GOTO(unlock_parent, rc);
458 * Check child name version during replay.
459 * During create replay a file may exist with same name.
461 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
462 &info->mti_tmp_fid1, 1);
464 GOTO(unlock_parent, rc = -EEXIST);
466 /* -ENOENT is expected here */
468 GOTO(unlock_parent, rc);
470 /* save version of file name for replay, it must be ENOENT here */
471 mdt_enoent_version_save(info, 1);
473 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
474 if (unlikely(IS_ERR(child)))
475 GOTO(unlock_parent, rc = PTR_ERR(child));
477 rc = mdt_remote_permission(info);
481 ma->ma_need = MA_INODE;
484 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
485 OBD_FAIL_MDS_REINT_CREATE_WRITE);
487 /* Version of child will be updated on disk. */
488 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
489 rc = mdt_version_get_check_save(info, child, 2);
493 /* Let lower layer know current lock mode. */
494 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
497 * Do not perform lookup sanity check. We know that name does
500 info->mti_spec.sp_cr_lookup = 0;
501 info->mti_spec.sp_feat = &dt_directory_features;
503 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
504 mdt_object_child(child), &info->mti_spec, ma);
506 rc = mdt_attr_get_complex(info, child, ma);
512 * On DNE, we need to eliminate dependey between 'mkdir a' and
513 * 'mkdir a/b' if b is a striped directory, to achieve this, two
514 * things are done below:
515 * 1. save child and slaves lock.
516 * 2. if the child is a striped directory, relock parent so to
517 * compare against with COS locks to ensure parent was
520 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
521 struct mdt_lock_handle *lhc;
522 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
525 rc = mdt_object_striped(info, child);
531 if (!mdt_object_remote(parent)) {
532 mdt_object_unlock(info, parent, lh, 1);
533 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
534 rc = mdt_reint_object_lock(info, parent, lh,
535 MDS_INODELOCK_UPDATE,
542 lhc = &info->mti_lh[MDT_LH_CHILD];
543 mdt_lock_handle_init(lhc);
544 mdt_lock_reg_init(lhc, LCK_PW);
545 rc = mdt_reint_striped_lock(info, child, lhc,
546 MDS_INODELOCK_UPDATE, einfo,
551 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
554 /* Return fid & attr to client. */
555 if (ma->ma_valid & MA_INODE)
556 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
557 mdt_object_fid(child));
559 mdt_object_put(info->mti_env, child);
561 mdt_object_unlock(info, parent, lh, rc);
563 mdt_object_put(info->mti_env, parent);
567 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
570 struct mdt_lock_handle *lh;
571 int do_vbr = ma->ma_attr.la_valid &
572 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
573 __u64 lockpart = MDS_INODELOCK_UPDATE;
574 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
579 rc = mdt_object_striped(info, mo);
585 lh = &info->mti_lh[MDT_LH_PARENT];
586 mdt_lock_reg_init(lh, LCK_PW);
588 /* Even though the new MDT will grant PERM lock to the old
589 * client, but the old client will almost ignore that during
590 * So it needs to revoke both LOOKUP and PERM lock here, so
591 * both new and old client can cancel the dcache */
592 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
593 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
595 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
600 /* all attrs are packed into mti_attr in unpack_setattr */
601 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
602 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
604 /* VBR: update version if attr changed are important for recovery */
606 /* update on-disk version of changed object */
607 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
608 rc = mdt_version_get_check_save(info, mo, 0);
610 GOTO(out_unlock, rc);
613 /* Ensure constant striping during chown(). See LU-2789. */
614 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
615 mutex_lock(&mo->mot_lov_mutex);
617 /* all attrs are packed into mti_attr in unpack_setattr */
618 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
620 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
621 mutex_unlock(&mo->mot_lov_mutex);
624 GOTO(out_unlock, rc);
625 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
628 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
633 * Check HSM flags and add HS_DIRTY flag if relevant.
635 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
636 * and is not RELEASED.
638 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
644 /* If the file was modified, add the dirty flag */
645 ma->ma_need = MA_HSM;
646 rc = mdt_attr_get_complex(info, mo, ma);
648 CERROR("file attribute read error for "DFID": %d.\n",
649 PFID(mdt_object_fid(mo)), rc);
653 /* If an up2date copy exists in the backend, add dirty flag */
654 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
655 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
656 ma->ma_hsm.mh_flags |= HS_DIRTY;
658 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
660 CERROR("file attribute change error for "DFID": %d\n",
661 PFID(mdt_object_fid(mo)), rc);
667 static int mdt_reint_setattr(struct mdt_thread_info *info,
668 struct mdt_lock_handle *lhc)
670 struct md_attr *ma = &info->mti_attr;
671 struct mdt_reint_record *rr = &info->mti_rr;
672 struct ptlrpc_request *req = mdt_info_req(info);
673 struct mdt_object *mo;
674 struct mdt_body *repbody;
678 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
679 (unsigned int)ma->ma_attr.la_valid);
681 if (info->mti_dlm_req)
682 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
684 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
685 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
687 GOTO(out, rc = PTR_ERR(mo));
689 if (!mdt_object_exists(mo))
690 GOTO(out_put, rc = -ENOENT);
692 if (mdt_object_remote(mo))
693 GOTO(out_put, rc = -EREMOTE);
695 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
696 /* Check write access for the O_TRUNC case */
697 if (mdt_write_read(mo) < 0)
698 GOTO(out_put, rc = -ETXTBSY);
700 /* LU-10286: compatibility check for FLR.
701 * Please check the comment in mdt_finish_open() for details */
702 if (!exp_connect_flr(info->mti_exp)) {
703 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
704 if (rc < 0 && rc != -ENODATA)
707 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
708 GOTO(out_put, rc = -EOPNOTSUPP);
711 /* For truncate, the file size sent from client
712 * is believable, but the blocks are incorrect,
713 * which makes the block size in LSOM attribute
714 * inconsisent with the real block size.
716 rc = mdt_lsom_update(info, mo, true);
721 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
722 if (ma->ma_valid & MA_LOV)
723 GOTO(out_put, rc = -EPROTO);
725 rc = mdt_attr_set(info, mo, ma);
728 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
729 (ma->ma_valid & MA_INODE)) {
730 struct lu_buf *buf = &info->mti_buf;
731 struct mdt_lock_handle *lh;
733 rc = mdt_remote_permission(info);
737 if (ma->ma_attr.la_valid != 0)
738 GOTO(out_put, rc = -EPROTO);
740 lh = &info->mti_lh[MDT_LH_PARENT];
741 mdt_lock_reg_init(lh, LCK_PW);
743 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
747 if (ma->ma_valid & MA_LOV) {
748 buf->lb_buf = ma->ma_lmm;
749 buf->lb_len = ma->ma_lmm_size;
751 buf->lb_buf = ma->ma_lmv;
752 buf->lb_len = ma->ma_lmv_size;
754 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
755 (ma->ma_valid & MA_LOV) ?
756 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
759 mdt_object_unlock(info, mo, lh, rc);
763 GOTO(out_put, rc = -EPROTO);
766 /* If file data is modified, add the dirty flag */
767 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
768 rc = mdt_add_dirty_flag(info, mo, ma);
770 ma->ma_need = MA_INODE;
772 rc = mdt_attr_get_complex(info, mo, ma);
776 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
780 mdt_object_put(info->mti_env, mo);
783 mdt_counter_incr(req, LPROC_MDT_SETATTR);
785 mdt_client_compatibility(info);
786 rc2 = mdt_fix_reply(info);
792 static int mdt_reint_create(struct mdt_thread_info *info,
793 struct mdt_lock_handle *lhc)
795 struct ptlrpc_request *req = mdt_info_req(info);
799 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
800 RETURN(err_serious(-ESTALE));
802 if (info->mti_dlm_req)
803 ldlm_request_cancel(mdt_info_req(info),
804 info->mti_dlm_req, 0, LATF_SKIP);
806 if (!lu_name_is_valid(&info->mti_rr.rr_name))
809 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
811 mdt_counter_incr(req, LPROC_MDT_MKDIR);
819 /* Special file should stay on the same node as parent. */
820 mdt_counter_incr(req, LPROC_MDT_MKNOD);
823 CERROR("%s: Unsupported mode %o\n",
824 mdt_obd_name(info->mti_mdt),
825 info->mti_attr.ma_attr.la_mode);
826 RETURN(err_serious(-EOPNOTSUPP));
829 rc = mdt_create(info);
834 * VBR: save parent version in reply and child version getting by its name.
835 * Version of child is getting and checking during its lookup. If
837 static int mdt_reint_unlink(struct mdt_thread_info *info,
838 struct mdt_lock_handle *lhc)
840 struct mdt_reint_record *rr = &info->mti_rr;
841 struct ptlrpc_request *req = mdt_info_req(info);
842 struct md_attr *ma = &info->mti_attr;
843 struct lu_fid *child_fid = &info->mti_tmp_fid1;
844 struct mdt_object *mp;
845 struct mdt_object *mc;
846 struct mdt_lock_handle *parent_lh;
847 struct mdt_lock_handle *child_lh;
848 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
850 bool cos_incompat = false;
856 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
857 PNAME(&rr->rr_name));
859 if (info->mti_dlm_req)
860 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
862 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
863 RETURN(err_serious(-ENOENT));
865 if (!fid_is_md_operative(rr->rr_fid1))
868 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
872 if (mdt_object_remote(mp)) {
875 rc = mdt_version_get_check_save(info, mp, 0);
877 GOTO(put_parent, rc);
881 parent_lh = &info->mti_lh[MDT_LH_PARENT];
882 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
883 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
886 GOTO(put_parent, rc);
888 /* lookup child object along with version checking */
890 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
892 /* Name might not be able to find during resend of
893 * remote unlink, considering following case.
894 * dir_A is a remote directory, the name entry of
895 * dir_A is on MDT0, the directory is on MDT1,
897 * 1. client sends unlink req to MDT1.
898 * 2. MDT1 sends name delete update to MDT0.
899 * 3. name entry is being deleted in MDT0 synchronously.
900 * 4. MDT1 is restarted.
901 * 5. client resends unlink req to MDT1. So it can not
902 * find the name entry on MDT0 anymore.
903 * In this case, MDT1 only needs to destory the local
906 if (mdt_object_remote(mp) && rc == -ENOENT &&
907 !fid_is_zero(rr->rr_fid2) &&
908 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
910 *child_fid = *rr->rr_fid2;
912 GOTO(unlock_parent, rc);
916 if (!fid_is_md_operative(child_fid))
917 GOTO(unlock_parent, rc = -EPERM);
919 /* We will lock the child regardless it is local or remote. No harm. */
920 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
922 GOTO(unlock_parent, rc = PTR_ERR(mc));
925 rc = mdt_object_striped(info, mc);
927 GOTO(unlock_parent, rc = PTR_ERR(mc));
931 mdt_object_put(info->mti_env, mc);
932 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
937 child_lh = &info->mti_lh[MDT_LH_CHILD];
938 mdt_lock_reg_init(child_lh, LCK_EX);
939 if (info->mti_spec.sp_rm_entry) {
940 struct lu_ucred *uc = mdt_ucred(info);
942 if (!mdt_is_dne_client(req->rq_export))
943 /* Return -ENOTSUPP for old client */
944 GOTO(put_child, rc = -ENOTSUPP);
946 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
947 GOTO(put_child, rc = -EPERM);
949 ma->ma_need = MA_INODE;
951 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
952 NULL, &rr->rr_name, ma, no_name);
956 if (mdt_object_remote(mc)) {
957 struct mdt_body *repbody;
959 if (!fid_is_zero(rr->rr_fid2)) {
960 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
961 mdt_obd_name(info->mti_mdt),
962 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
963 GOTO(put_child, rc = -ENOENT);
965 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
966 mdt_obd_name(info->mti_mdt),
967 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
969 if (!mdt_is_dne_client(req->rq_export))
970 /* Return -ENOTSUPP for old client */
971 GOTO(put_child, rc = -ENOTSUPP);
973 /* Revoke the LOOKUP lock of the remote object granted by
974 * this MDT. Since the unlink will happen on another MDT,
975 * it will release the LOOKUP lock right away. Then What
976 * would happen if another client try to grab the LOOKUP
977 * lock at the same time with unlink XXX */
978 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
979 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
980 LASSERT(repbody != NULL);
981 repbody->mbo_fid1 = *mdt_object_fid(mc);
982 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
983 GOTO(unlock_child, rc = -EREMOTE);
985 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
986 * this now because a running HSM restore on the child (unlink
987 * victim) will hold the layout lock. See LU-4002. */
988 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
989 if (mdt_object_remote(mp)) {
990 /* Enqueue lookup lock from parent MDT */
991 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
992 &child_lh->mlh_rreg_lh,
993 child_lh->mlh_rreg_mode,
994 MDS_INODELOCK_LOOKUP, false);
998 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1001 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1004 GOTO(put_child, rc);
1007 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1008 * whether need MA_LOV and MA_COOKIE.
1010 ma->ma_need = MA_INODE;
1013 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1014 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1015 /* save version when object is locked */
1016 mdt_version_get_save(info, mc, 1);
1018 mutex_lock(&mc->mot_lov_mutex);
1020 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1021 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1023 mutex_unlock(&mc->mot_lov_mutex);
1025 GOTO(unlock_child, rc);
1027 if (!lu_object_is_dying(&mc->mot_header)) {
1028 rc = mdt_attr_get_complex(info, mc, ma);
1032 mdt_dom_check_and_discard(info, mc);
1034 mdt_handle_last_unlink(info, mc, ma);
1037 if (ma->ma_valid & MA_INODE) {
1038 switch (ma->ma_attr.la_mode & S_IFMT) {
1040 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1048 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1051 LASSERTF(0, "bad file type %o unlinking\n",
1052 ma->ma_attr.la_mode);
1059 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1061 mdt_object_put(info->mti_env, mc);
1063 mdt_object_unlock(info, mp, parent_lh, rc);
1065 mdt_object_put(info->mti_env, mp);
1070 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1073 static int mdt_reint_link(struct mdt_thread_info *info,
1074 struct mdt_lock_handle *lhc)
1076 struct mdt_reint_record *rr = &info->mti_rr;
1077 struct ptlrpc_request *req = mdt_info_req(info);
1078 struct md_attr *ma = &info->mti_attr;
1079 struct mdt_object *ms;
1080 struct mdt_object *mp;
1081 struct mdt_lock_handle *lhs;
1082 struct mdt_lock_handle *lhp;
1087 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1088 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1090 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1091 RETURN(err_serious(-ENOENT));
1093 if (info->mti_dlm_req)
1094 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1096 /* Invalid case so return error immediately instead of
1098 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1101 if (!fid_is_md_operative(rr->rr_fid1) ||
1102 !fid_is_md_operative(rr->rr_fid2))
1105 /* step 1: find target parent dir */
1106 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1108 RETURN(PTR_ERR(mp));
1110 rc = mdt_version_get_check_save(info, mp, 0);
1112 GOTO(put_parent, rc);
1114 /* step 2: find source */
1115 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1117 GOTO(put_parent, rc = PTR_ERR(ms));
1119 if (!mdt_object_exists(ms)) {
1120 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1121 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1122 GOTO(put_source, rc = -ENOENT);
1125 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1127 lhp = &info->mti_lh[MDT_LH_PARENT];
1128 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1129 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1132 GOTO(put_source, rc);
1134 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1136 lhs = &info->mti_lh[MDT_LH_CHILD];
1137 mdt_lock_reg_init(lhs, LCK_EX);
1138 rc = mdt_reint_object_lock(info, ms, lhs,
1139 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1142 GOTO(unlock_parent, rc);
1144 /* step 3: link it */
1145 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1146 OBD_FAIL_MDS_REINT_LINK_WRITE);
1148 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1149 rc = mdt_version_get_check_save(info, ms, 1);
1151 GOTO(unlock_source, rc);
1153 /** check target version by name during replay */
1154 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1155 &info->mti_tmp_fid1, 2);
1156 if (rc != 0 && rc != -ENOENT)
1157 GOTO(unlock_source, rc);
1158 /* save version of file name for replay, it must be ENOENT here */
1159 if (!req_is_replay(mdt_info_req(info))) {
1160 if (rc != -ENOENT) {
1161 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1162 PNAME(&rr->rr_name));
1163 GOTO(unlock_source, rc = -EEXIST);
1165 info->mti_ver[2] = ENOENT_VERSION;
1166 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1169 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1170 mdt_object_child(ms), &rr->rr_name, ma);
1173 mdt_counter_incr(req, LPROC_MDT_LINK);
1177 mdt_object_unlock(info, ms, lhs, rc);
1179 mdt_object_unlock(info, mp, lhp, rc);
1181 mdt_object_put(info->mti_env, ms);
1183 mdt_object_put(info->mti_env, mp);
1187 * lock the part of the directory according to the hash of the name
1188 * (lh->mlh_pdo_hash) in parallel directory lock.
1190 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1191 struct mdt_lock_handle *lh,
1192 struct mdt_object *obj, __u64 ibits,
1195 struct ldlm_res_id *res = &info->mti_res_id;
1196 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1197 union ldlm_policy_data *policy = &info->mti_policy;
1198 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1202 * Finish res_id initializing by name hash marking part of
1203 * directory which is taking modification.
1205 LASSERT(lh->mlh_pdo_hash != 0);
1206 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1207 memset(policy, 0, sizeof(*policy));
1208 policy->l_inodebits.bits = ibits;
1210 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1211 dlmflags |= LDLM_FL_COS_INCOMPAT;
1213 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1214 * going to be sent to client. If it is - mdt_intent_policy() path will
1215 * fix it up and turn FL_LOCAL flag off.
1217 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1218 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1223 * Get BFL lock for rename or migrate process.
1225 static int mdt_rename_lock(struct mdt_thread_info *info,
1226 struct lustre_handle *lh)
1231 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1232 struct lu_fid *fid = &info->mti_tmp_fid1;
1233 struct mdt_object *obj;
1235 /* XXX, right now, it has to use object API to
1236 * enqueue lock cross MDT, so it will enqueue
1237 * rename lock(with LUSTRE_BFL_FID) by root object */
1239 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1241 RETURN(PTR_ERR(obj));
1243 rc = mdt_remote_object_lock(info, obj,
1244 &LUSTRE_BFL_FID, lh,
1246 MDS_INODELOCK_UPDATE, false);
1247 mdt_object_put(info->mti_env, obj);
1249 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1250 union ldlm_policy_data *policy = &info->mti_policy;
1251 struct ldlm_res_id *res_id = &info->mti_res_id;
1254 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1255 memset(policy, 0, sizeof *policy);
1256 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1257 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1258 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1259 LCK_EX, &flags, ldlm_blocking_ast,
1260 ldlm_completion_ast, NULL, NULL, 0,
1262 &info->mti_exp->exp_handle.h_cookie,
1269 static void mdt_rename_unlock(struct lustre_handle *lh)
1272 LASSERT(lustre_handle_is_used(lh));
1273 /* Cancel the single rename lock right away */
1274 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1278 /* Update object linkEA */
1279 struct mdt_lock_list {
1280 struct mdt_object *mll_obj;
1281 struct mdt_lock_handle mll_lh;
1282 struct list_head mll_list;
1285 static void mdt_unlock_list(struct mdt_thread_info *info,
1286 struct list_head *list, int rc)
1288 struct mdt_lock_list *mll;
1289 struct mdt_lock_list *mll2;
1291 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1292 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1293 list_del(&mll->mll_list);
1298 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1299 struct mdt_object *obj,
1300 struct mdt_object *pobj,
1301 struct list_head *lock_list)
1303 struct lu_buf *buf = &info->mti_big_buf;
1304 struct linkea_data ldata = { NULL };
1310 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1313 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1314 if (buf->lb_buf == NULL)
1318 rc = mdt_links_read(info, obj, &ldata);
1320 if (rc == -ENOENT || rc == -ENODATA)
1325 /* ignore the migrating parent(@pobj) */
1326 retry_count = ldata.ld_leh->leh_reccount - 1;
1329 LASSERT(ldata.ld_leh != NULL);
1330 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1331 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1332 struct mdt_device *mdt = info->mti_mdt;
1333 struct mdt_object *mdt_pobj;
1334 struct mdt_lock_list *mll;
1335 struct lu_name name;
1339 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1341 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1342 if (IS_ERR(mdt_pobj)) {
1343 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1344 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1348 if (!mdt_object_exists(mdt_pobj)) {
1349 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1350 mdt_obd_name(mdt), PFID(&fid));
1351 mdt_object_put(info->mti_env, mdt_pobj);
1355 /* Check if the object already exists in the list */
1356 list_for_each_entry(mll, lock_list, mll_list) {
1357 if (mll->mll_obj == mdt_pobj) {
1358 mdt_object_put(info->mti_env, mdt_pobj);
1363 if (mdt_pobj == pobj) {
1364 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1365 mdt_obd_name(mdt), PFID(&fid));
1366 mdt_object_put(info->mti_env, mdt_pobj);
1372 mdt_object_put(info->mti_env, mdt_pobj);
1373 GOTO(out, rc = -ENOMEM);
1376 /* Since this needs to lock all of objects in linkea, to avoid
1377 * deadlocks, because it does not follow parent-child order as
1378 * other MDT operation, let's use try_lock here and if the lock
1379 * cannot be gotten because of conflicting locks, then drop all
1380 * current locks, send an AST to the client, and start again. */
1381 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1383 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits,
1384 MDS_INODELOCK_UPDATE, true);
1385 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1386 mdt_unlock_list(info, lock_list, 0);
1388 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1389 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1392 if (retry_count == 0) {
1393 mdt_object_put(info->mti_env, mdt_pobj);
1395 GOTO(out, rc = -EBUSY);
1398 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1399 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1400 MDS_INODELOCK_UPDATE);
1402 mdt_object_put(info->mti_env, mdt_pobj);
1407 if (mdt_object_remote(mdt_pobj)) {
1408 struct ldlm_lock *lock;
1410 /* For remote object, Set lock to cb_atomic,
1411 * so lock can be released in blocking_ast()
1412 * immediately, then the next try_lock will
1413 * have better chance to succeds */
1415 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1416 LASSERT(lock != NULL);
1417 lock_res_and_lock(lock);
1418 ldlm_set_atomic_cb(lock);
1419 unlock_res_and_lock(lock);
1420 LDLM_LOCK_PUT(lock);
1422 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1428 INIT_LIST_HEAD(&mll->mll_list);
1429 mll->mll_obj = mdt_pobj;
1430 list_add_tail(&mll->mll_list, lock_list);
1432 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1437 mdt_unlock_list(info, lock_list, rc);
1441 /* migrate files from one MDT to another MDT */
1442 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1443 struct mdt_lock_handle *lhc)
1445 struct mdt_reint_record *rr = &info->mti_rr;
1446 struct md_attr *ma = &info->mti_attr;
1447 struct mdt_object *msrcdir;
1448 struct mdt_object *mold;
1449 struct mdt_object *mnew = NULL;
1450 struct mdt_lock_handle *lh_dirp;
1451 struct mdt_lock_handle *lh_childp;
1452 struct mdt_lock_handle *lh_tgtp = NULL;
1453 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1454 struct list_head lock_list;
1456 struct ldlm_lock *lease = NULL;
1457 bool lock_open_sem = false;
1461 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1462 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1464 /* 1: lock the source dir. */
1465 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1466 if (IS_ERR(msrcdir)) {
1467 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1468 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1469 (int)PTR_ERR(msrcdir));
1470 RETURN(PTR_ERR(msrcdir));
1473 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1474 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1475 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1478 GOTO(out_put_parent, rc);
1480 if (!mdt_object_remote(msrcdir)) {
1481 rc = mdt_version_get_check_save(info, msrcdir, 0);
1483 GOTO(out_unlock_parent, rc);
1486 /* 2: sanity check and find the object to be migrated. */
1488 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1490 GOTO(out_unlock_parent, rc);
1492 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1493 GOTO(out_unlock_parent, rc = -EINVAL);
1495 if (!fid_is_md_operative(old_fid))
1496 GOTO(out_unlock_parent, rc = -EPERM);
1498 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1499 GOTO(out_unlock_parent, rc = -EPERM);
1501 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1503 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1505 if (mdt_object_remote(mold)) {
1506 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1507 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1508 GOTO(out_put_child, rc = -EREMOTE);
1511 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1512 !mdt_object_remote(msrcdir)) {
1513 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1514 " MDT, which should be migrated first:"
1515 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1516 PFID(mdt_object_fid(msrcdir)), -EPERM);
1517 GOTO(out_put_child, rc = -EPERM);
1520 rc = mdt_remote_permission(info);
1522 GOTO(out_put_child, rc);
1524 /* 3: iterate the linkea of the object and lock all of the objects */
1525 INIT_LIST_HEAD(&lock_list);
1526 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1528 GOTO(out_put_child, rc);
1530 if (info->mti_spec.sp_migrate_close) {
1531 struct close_data *data;
1532 struct mdt_body *repbody;
1533 bool lease_broken = false;
1535 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1537 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1539 GOTO(out_lease, rc = -EPROTO);
1541 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1543 GOTO(out_lease, rc = -EPROTO);
1545 lease = ldlm_handle2lock(&data->cd_handle);
1547 GOTO(out_lease, rc = -ESTALE);
1549 /* try to hold open_sem so that nobody else can open the file */
1550 if (!down_write_trylock(&mold->mot_open_sem)) {
1551 ldlm_lock_cancel(lease);
1552 GOTO(out_lease, rc = -EBUSY);
1555 lock_open_sem = true;
1556 /* Check if the lease open lease has already canceled */
1557 lock_res_and_lock(lease);
1558 lease_broken = ldlm_is_cancel(lease);
1559 unlock_res_and_lock(lease);
1561 LDLM_DEBUG(lease, DFID " lease broken? %d",
1562 PFID(mdt_object_fid(mold)), lease_broken);
1564 /* Cancel server side lease. Client side counterpart should
1565 * have been cancelled. It's okay to cancel it now as we've
1566 * held mot_open_sem. */
1567 ldlm_lock_cancel(lease);
1570 GOTO(out_lease, rc = -EAGAIN);
1572 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1573 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1574 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1576 GOTO(out_unlock_list, rc);
1579 /* 4: lock of the object migrated object */
1580 lh_childp = &info->mti_lh[MDT_LH_OLD];
1581 mdt_lock_reg_init(lh_childp, LCK_EX);
1582 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1583 MDS_INODELOCK_LAYOUT;
1584 if (mdt_object_remote(msrcdir)) {
1585 /* Enqueue lookup lock from the parent MDT */
1586 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1587 &lh_childp->mlh_rreg_lh,
1588 lh_childp->mlh_rreg_mode,
1589 MDS_INODELOCK_LOOKUP, false);
1591 GOTO(out_unlock_list, rc);
1593 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1596 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1598 GOTO(out_unlock_child, rc);
1600 /* Migration is incompatible with HSM. */
1601 ma->ma_need = MA_HSM;
1603 rc = mdt_attr_get_complex(info, mold, ma);
1605 GOTO(out_unlock_child, rc);
1607 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1610 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1611 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1612 GOTO(out_unlock_child, rc);
1615 ma->ma_need = MA_LMV;
1617 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1618 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1619 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1621 GOTO(out_unlock_child, rc);
1623 if ((ma->ma_valid & MA_LMV)) {
1624 struct lmv_mds_md_v1 *lmm1;
1626 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1627 lmm1 = &ma->ma_lmv->lmv_md_v1;
1628 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1629 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1630 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1631 PFID(mdt_object_fid(mold)), -EPERM);
1632 GOTO(out_unlock_child, rc = -EPERM);
1635 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1636 GOTO(out_unlock_child, rc = -EINVAL);
1638 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1639 &lmm1->lmv_stripe_fids[1]);
1641 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1643 if (!mdt_object_remote(mnew)) {
1645 "%s: "DFID" being migrated is on this MDT:"
1646 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1647 PFID(rr->rr_fid2), -EPERM);
1648 GOTO(out_put_new, rc = -EPERM);
1651 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1652 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1653 rc = mdt_remote_object_lock(info, mnew,
1654 mdt_object_fid(mnew),
1655 &lh_tgtp->mlh_rreg_lh,
1656 lh_tgtp->mlh_rreg_mode,
1657 MDS_INODELOCK_UPDATE, false);
1660 GOTO(out_put_new, rc);
1663 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1666 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1667 if (!mdt_object_remote(mnew)) {
1668 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1669 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1670 PFID(rr->rr_fid2), -EXDEV);
1671 GOTO(out_put_new, rc = -EXDEV);
1676 mdt_reint_init_ma(info, ma);
1678 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1679 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1681 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1682 mdt_object_child(mold), &rr->rr_name,
1683 mdt_object_child(mnew), ma);
1685 GOTO(out_unlock_new, rc);
1688 if (lh_tgtp != NULL)
1689 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1692 mdt_object_put(info->mti_env, mnew);
1694 mdt_object_unlock(info, mold, lh_childp, rc);
1696 /* we don't really modify linkea objects, so we can safely decref these
1697 * locks, and this can avoid saving them as COS locks, which may prevent
1698 * subsequent migrate. */
1699 mdt_unlock_list(info, &lock_list, 1);
1700 if (lease != NULL) {
1701 ldlm_reprocess_all(lease->l_resource);
1702 LDLM_LOCK_PUT(lease);
1706 up_write(&mold->mot_open_sem);
1708 mdt_object_put(info->mti_env, mold);
1710 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1712 mdt_object_put(info->mti_env, msrcdir);
1717 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1718 const struct lu_fid *fid,
1721 struct mdt_object *dir;
1725 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1729 /* check early, the real version will be saved after locking */
1730 rc = mdt_version_get_check(info, dir, idx);
1736 mdt_object_put(info->mti_env, dir);
1740 static int mdt_object_lock_save(struct mdt_thread_info *info,
1741 struct mdt_object *dir,
1742 struct mdt_lock_handle *lh,
1743 int idx, bool cos_incompat)
1747 /* we lock the target dir if it is local */
1748 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1753 /* get and save correct version after locking */
1754 mdt_version_get_save(info, dir, idx);
1759 * determine lock order of sobj and tobj
1761 * there are two situations we need to lock tobj before sobj:
1762 * 1. sobj is child of tobj
1763 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
1764 * larger than that of tobj
1766 * \retval 1 lock tobj before sobj
1767 * \retval 0 lock sobj before tobj
1768 * \retval -ev negative errno upon error
1770 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
1771 struct mdt_object *sobj,
1772 struct mdt_object *tobj)
1774 struct md_attr *ma = &info->mti_attr;
1775 struct lu_fid *spfid = &info->mti_tmp_fid1;
1776 struct lu_fid *tpfid = &info->mti_tmp_fid2;
1777 struct lmv_mds_md_v1 *lmv;
1782 /* sobj and tobj are the same */
1786 if (fid_is_root(mdt_object_fid(sobj)))
1789 if (fid_is_root(mdt_object_fid(tobj)))
1792 /* check whether sobj is child of tobj */
1793 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
1794 mdt_object_fid(tobj));
1801 /* check whether sobj and tobj are children of the same parent */
1802 rc = mdt_attr_get_pfid(info, sobj, spfid);
1806 rc = mdt_attr_get_pfid(info, tobj, tpfid);
1810 if (!lu_fid_eq(spfid, tpfid))
1813 /* check whether sobj and tobj are sibling stripes */
1814 ma->ma_need = MA_LMV;
1816 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1817 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1818 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
1822 if (!(ma->ma_valid & MA_LMV))
1825 lmv = &ma->ma_lmv->lmv_md_v1;
1826 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
1828 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
1831 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
1835 if (!(ma->ma_valid & MA_LMV))
1838 lmv = &ma->ma_lmv->lmv_md_v1;
1839 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
1841 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
1843 /* check stripe index of sobj and tobj */
1844 if (sindex == tindex)
1847 return sindex < tindex ? 0 : 1;
1851 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1852 * 2 - srcdir child; 3 - tgtdir child.
1853 * Update on disk version of srcdir child.
1856 * For DNE phase I, only these renames are allowed
1857 * mv src_p/src_c tgt_p/tgt_c
1858 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1859 * 2. src_p and tgt_p are same directory, and tgt_c does not
1860 * exists. In this case, all of modification will happen
1861 * in the MDT where ithesource parent is, only one remote
1862 * update is needed, i.e. set c_time/m_time on the child.
1863 * And tgt_c will be still in the same MDT as the original
1866 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1867 struct mdt_lock_handle *lhc)
1869 struct mdt_reint_record *rr = &info->mti_rr;
1870 struct md_attr *ma = &info->mti_attr;
1871 struct ptlrpc_request *req = mdt_info_req(info);
1872 struct mdt_object *msrcdir = NULL;
1873 struct mdt_object *mtgtdir = NULL;
1874 struct mdt_object *mold;
1875 struct mdt_object *mnew = NULL;
1876 struct mdt_lock_handle *lh_srcdirp;
1877 struct mdt_lock_handle *lh_tgtdirp;
1878 struct mdt_lock_handle *lh_oldp = NULL;
1879 struct mdt_lock_handle *lh_newp = NULL;
1880 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1881 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1883 bool reverse = false;
1888 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1889 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1890 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1892 /* find both parents. */
1893 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1894 if (IS_ERR(msrcdir))
1895 RETURN(PTR_ERR(msrcdir));
1897 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1899 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1901 mdt_object_get(info->mti_env, mtgtdir);
1903 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1904 if (IS_ERR(mtgtdir))
1905 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1908 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
1910 GOTO(out_put_tgtdir, rc);
1914 /* source needs to be looked up after locking source parent, otherwise
1915 * this rename may race with unlink source, and cause rename hang, see
1916 * sanityn.sh 55b, so check parents first, if later we found source is
1917 * remote, relock parents. */
1918 cos_incompat = (mdt_object_remote(msrcdir) ||
1919 mdt_object_remote(mtgtdir));
1921 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1923 /* lock parents in the proper order. */
1924 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1925 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1928 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1929 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1932 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1935 GOTO(out_put_tgtdir, rc);
1937 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1939 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1942 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1943 GOTO(out_put_tgtdir, rc);
1946 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1949 GOTO(out_put_tgtdir, rc);
1951 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1953 if (mtgtdir != msrcdir) {
1954 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1956 } else if (lh_srcdirp->mlh_pdo_hash !=
1957 lh_tgtdirp->mlh_pdo_hash) {
1958 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1959 MDS_INODELOCK_UPDATE,
1961 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1964 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1965 GOTO(out_put_tgtdir, rc);
1969 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1970 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1972 /* find mold object. */
1974 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1976 GOTO(out_unlock_parents, rc);
1978 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1979 GOTO(out_unlock_parents, rc = -EINVAL);
1981 if (!fid_is_md_operative(old_fid))
1982 GOTO(out_unlock_parents, rc = -EPERM);
1984 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1986 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1988 /* Check if @mtgtdir is subdir of @mold, before locking child
1989 * to avoid reverse locking. */
1990 if (mtgtdir != msrcdir) {
1991 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
1996 GOTO(out_put_old, rc);
2000 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2001 /* save version after locking */
2002 mdt_version_get_save(info, mold, 2);
2004 if (!cos_incompat && mdt_object_remote(mold)) {
2005 cos_incompat = true;
2006 mdt_object_put(info->mti_env, mold);
2007 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2008 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2012 /* find mnew object:
2013 * mnew target object may not exist now
2014 * lookup with version checking */
2016 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2019 /* the new_fid should have been filled at this moment */
2020 if (lu_fid_eq(old_fid, new_fid))
2021 GOTO(out_put_old, rc);
2023 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2024 lu_fid_eq(new_fid, rr->rr_fid2))
2025 GOTO(out_put_old, rc = -EINVAL);
2027 if (!fid_is_md_operative(new_fid))
2028 GOTO(out_put_old, rc = -EPERM);
2030 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2032 GOTO(out_put_old, rc = PTR_ERR(mnew));
2034 if (mdt_object_remote(mnew)) {
2035 struct mdt_body *repbody;
2037 /* Always send rename req to the target child MDT */
2038 repbody = req_capsule_server_get(info->mti_pill,
2040 LASSERT(repbody != NULL);
2041 repbody->mbo_fid1 = *new_fid;
2042 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2043 GOTO(out_put_new, rc = -EXDEV);
2045 /* Before locking the target dir, check we do not replace
2046 * a dir with a non-dir, otherwise it may deadlock with
2047 * link op which tries to create a link in this dir
2048 * back to this non-dir. */
2049 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2050 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2051 GOTO(out_put_new, rc = -EISDIR);
2053 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2054 mdt_lock_reg_init(lh_oldp, LCK_EX);
2055 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2056 if (mdt_object_remote(msrcdir)) {
2057 /* Enqueue lookup lock from the parent MDT */
2058 rc = mdt_remote_object_lock(info, msrcdir,
2059 mdt_object_fid(mold),
2060 &lh_oldp->mlh_rreg_lh,
2061 lh_oldp->mlh_rreg_mode,
2062 MDS_INODELOCK_LOOKUP,
2065 GOTO(out_put_new, rc);
2067 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2070 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2073 GOTO(out_unlock_old, rc);
2075 /* Check if @msrcdir is subdir of @mnew, before locking child
2076 * to avoid reverse locking. */
2077 if (mtgtdir != msrcdir) {
2078 rc = mdo_is_subdir(info->mti_env,
2079 mdt_object_child(msrcdir), new_fid);
2083 GOTO(out_unlock_old, rc);
2087 /* We used to acquire MDS_INODELOCK_FULL here but we
2088 * can't do this now because a running HSM restore on
2089 * the rename onto victim will hold the layout
2090 * lock. See LU-4002. */
2092 lh_newp = &info->mti_lh[MDT_LH_NEW];
2093 mdt_lock_reg_init(lh_newp, LCK_EX);
2094 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2095 MDS_INODELOCK_LOOKUP |
2096 MDS_INODELOCK_UPDATE,
2099 GOTO(out_unlock_old, rc);
2101 /* get and save version after locking */
2102 mdt_version_get_save(info, mnew, 3);
2103 } else if (rc != -EREMOTE && rc != -ENOENT) {
2104 GOTO(out_put_old, rc);
2106 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2107 mdt_lock_reg_init(lh_oldp, LCK_EX);
2108 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2109 if (mdt_object_remote(msrcdir)) {
2110 /* Enqueue lookup lock from the parent MDT */
2111 rc = mdt_remote_object_lock(info, msrcdir,
2112 mdt_object_fid(mold),
2113 &lh_oldp->mlh_rreg_lh,
2114 lh_oldp->mlh_rreg_mode,
2115 MDS_INODELOCK_LOOKUP,
2118 GOTO(out_put_old, rc);
2120 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2123 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2126 GOTO(out_unlock_old, rc);
2128 mdt_enoent_version_save(info, 3);
2131 /* step 5: rename it */
2132 mdt_reint_init_ma(info, ma);
2134 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2135 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2138 mutex_lock(&mnew->mot_lov_mutex);
2140 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2141 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2142 mnew != NULL ? mdt_object_child(mnew) : NULL,
2143 &rr->rr_tgt_name, ma);
2146 mutex_unlock(&mnew->mot_lov_mutex);
2148 /* handle last link of tgt object */
2150 mdt_counter_incr(req, LPROC_MDT_RENAME);
2152 mdt_handle_last_unlink(info, mnew, ma);
2153 mdt_dom_check_and_discard(info, mnew);
2156 mdt_rename_counter_tally(info, info->mti_mdt, req,
2162 mdt_object_unlock(info, mnew, lh_newp, rc);
2164 mdt_object_unlock(info, mold, lh_oldp, rc);
2167 mdt_object_put(info->mti_env, mnew);
2169 mdt_object_put(info->mti_env, mold);
2171 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2172 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2174 mdt_object_put(info->mti_env, mtgtdir);
2176 mdt_object_put(info->mti_env, msrcdir);
2180 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2181 struct mdt_lock_handle *lhc, bool rename)
2183 struct mdt_reint_record *rr = &info->mti_rr;
2184 struct ptlrpc_request *req = mdt_info_req(info);
2185 struct lustre_handle rename_lh = { 0 };
2189 if (info->mti_dlm_req)
2190 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2192 if (!fid_is_md_operative(rr->rr_fid1) ||
2193 !fid_is_md_operative(rr->rr_fid2))
2196 /* Note: do not enqueue rename lock for replay request, because
2197 * if other MDT holds rename lock, but being blocked to wait for
2198 * this MDT to finish its recovery, and the failover MDT can not
2199 * get rename lock, which will cause deadlock. */
2200 if (!req_is_replay(req)) {
2201 rc = mdt_rename_lock(info, &rename_lh);
2203 CERROR("%s: can't lock FS for rename: rc = %d\n",
2204 mdt_obd_name(info->mti_mdt), rc);
2210 rc = mdt_reint_rename_internal(info, lhc);
2212 rc = mdt_reint_migrate_internal(info, lhc);
2214 if (lustre_handle_is_used(&rename_lh))
2215 mdt_rename_unlock(&rename_lh);
2220 static int mdt_reint_rename(struct mdt_thread_info *info,
2221 struct mdt_lock_handle *lhc)
2223 return mdt_reint_rename_or_migrate(info, lhc, true);
2226 static int mdt_reint_migrate(struct mdt_thread_info *info,
2227 struct mdt_lock_handle *lhc)
2229 return mdt_reint_rename_or_migrate(info, lhc, false);
2232 static int mdt_reint_resync(struct mdt_thread_info *info,
2233 struct mdt_lock_handle *lhc)
2235 struct mdt_reint_record *rr = &info->mti_rr;
2236 struct ptlrpc_request *req = mdt_info_req(info);
2237 struct md_attr *ma = &info->mti_attr;
2238 struct mdt_object *mo;
2239 struct ldlm_lock *lease;
2240 struct mdt_body *repbody;
2241 struct md_layout_change layout = { 0 };
2246 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2248 if (info->mti_dlm_req)
2249 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2251 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2253 GOTO(out, rc = PTR_ERR(mo));
2255 if (!mdt_object_exists(mo))
2256 GOTO(out_obj, rc = -ENOENT);
2258 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2259 GOTO(out_obj, rc = -EINVAL);
2261 if (mdt_object_remote(mo))
2262 GOTO(out_obj, rc = -EREMOTE);
2264 lease = ldlm_handle2lock(rr->rr_handle);
2266 GOTO(out_obj, rc = -ESTALE);
2268 /* It's really necessary to grab open_sem and check if the lease lock
2269 * has been lost. There would exist a concurrent writer coming in and
2270 * generating some dirty data in memory cache, the writeback would fail
2271 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2272 if (!down_write_trylock(&mo->mot_open_sem))
2273 GOTO(out_put_lease, rc = -EBUSY);
2275 lock_res_and_lock(lease);
2276 lease_broken = ldlm_is_cancel(lease);
2277 unlock_res_and_lock(lease);
2279 GOTO(out_unlock, rc = -EBUSY);
2281 /* the file has yet opened by anyone else after we took the lease. */
2282 layout.mlc_opc = MD_LAYOUT_RESYNC;
2283 rc = mdt_layout_change(info, mo, &layout);
2285 GOTO(out_unlock, rc);
2287 ma->ma_need = MA_INODE;
2289 rc = mdt_attr_get_complex(info, mo, ma);
2291 GOTO(out_unlock, rc);
2293 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2294 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2298 up_write(&mo->mot_open_sem);
2300 LDLM_LOCK_PUT(lease);
2302 mdt_object_put(info->mti_env, mo);
2304 mdt_client_compatibility(info);
2305 rc2 = mdt_fix_reply(info);
2311 struct mdt_reinter {
2312 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2313 enum lprocfs_extra_opc mr_extra_opc;
2316 static const struct mdt_reinter mdt_reinters[] = {
2318 .mr_handler = &mdt_reint_setattr,
2319 .mr_extra_opc = MDS_REINT_SETATTR,
2322 .mr_handler = &mdt_reint_create,
2323 .mr_extra_opc = MDS_REINT_CREATE,
2326 .mr_handler = &mdt_reint_link,
2327 .mr_extra_opc = MDS_REINT_LINK,
2330 .mr_handler = &mdt_reint_unlink,
2331 .mr_extra_opc = MDS_REINT_UNLINK,
2334 .mr_handler = &mdt_reint_rename,
2335 .mr_extra_opc = MDS_REINT_RENAME,
2338 .mr_handler = &mdt_reint_open,
2339 .mr_extra_opc = MDS_REINT_OPEN,
2341 [REINT_SETXATTR] = {
2342 .mr_handler = &mdt_reint_setxattr,
2343 .mr_extra_opc = MDS_REINT_SETXATTR,
2346 .mr_handler = &mdt_reint_unlink,
2347 .mr_extra_opc = MDS_REINT_UNLINK,
2350 .mr_handler = &mdt_reint_migrate,
2351 .mr_extra_opc = MDS_REINT_RENAME,
2354 .mr_handler = &mdt_reint_resync,
2355 .mr_extra_opc = MDS_REINT_RESYNC,
2359 int mdt_reint_rec(struct mdt_thread_info *info,
2360 struct mdt_lock_handle *lhc)
2362 const struct mdt_reinter *mr;
2366 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2369 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2370 if (mr->mr_handler == NULL)
2373 rc = (*mr->mr_handler)(info, lhc);
2375 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2376 PTLRPC_LAST_CNTR + mr->mr_extra_opc);