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);
712 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
713 if (ma->ma_valid & MA_LOV)
714 GOTO(out_put, rc = -EPROTO);
716 rc = mdt_attr_set(info, mo, ma);
719 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
720 (ma->ma_valid & MA_INODE)) {
721 struct lu_buf *buf = &info->mti_buf;
722 struct mdt_lock_handle *lh;
724 rc = mdt_remote_permission(info);
728 if (ma->ma_attr.la_valid != 0)
729 GOTO(out_put, rc = -EPROTO);
731 lh = &info->mti_lh[MDT_LH_PARENT];
732 mdt_lock_reg_init(lh, LCK_PW);
734 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
738 if (ma->ma_valid & MA_LOV) {
739 buf->lb_buf = ma->ma_lmm;
740 buf->lb_len = ma->ma_lmm_size;
742 buf->lb_buf = ma->ma_lmv;
743 buf->lb_len = ma->ma_lmv_size;
745 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
746 (ma->ma_valid & MA_LOV) ?
747 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
750 mdt_object_unlock(info, mo, lh, rc);
754 GOTO(out_put, rc = -EPROTO);
757 /* If file data is modified, add the dirty flag */
758 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
759 rc = mdt_add_dirty_flag(info, mo, ma);
761 ma->ma_need = MA_INODE;
763 rc = mdt_attr_get_complex(info, mo, ma);
767 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
771 mdt_object_put(info->mti_env, mo);
774 mdt_counter_incr(req, LPROC_MDT_SETATTR);
776 mdt_client_compatibility(info);
777 rc2 = mdt_fix_reply(info);
783 static int mdt_reint_create(struct mdt_thread_info *info,
784 struct mdt_lock_handle *lhc)
786 struct ptlrpc_request *req = mdt_info_req(info);
790 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
791 RETURN(err_serious(-ESTALE));
793 if (info->mti_dlm_req)
794 ldlm_request_cancel(mdt_info_req(info),
795 info->mti_dlm_req, 0, LATF_SKIP);
797 if (!lu_name_is_valid(&info->mti_rr.rr_name))
800 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
802 mdt_counter_incr(req, LPROC_MDT_MKDIR);
810 /* Special file should stay on the same node as parent. */
811 mdt_counter_incr(req, LPROC_MDT_MKNOD);
814 CERROR("%s: Unsupported mode %o\n",
815 mdt_obd_name(info->mti_mdt),
816 info->mti_attr.ma_attr.la_mode);
817 RETURN(err_serious(-EOPNOTSUPP));
820 rc = mdt_create(info);
825 * VBR: save parent version in reply and child version getting by its name.
826 * Version of child is getting and checking during its lookup. If
828 static int mdt_reint_unlink(struct mdt_thread_info *info,
829 struct mdt_lock_handle *lhc)
831 struct mdt_reint_record *rr = &info->mti_rr;
832 struct ptlrpc_request *req = mdt_info_req(info);
833 struct md_attr *ma = &info->mti_attr;
834 struct lu_fid *child_fid = &info->mti_tmp_fid1;
835 struct mdt_object *mp;
836 struct mdt_object *mc;
837 struct mdt_lock_handle *parent_lh;
838 struct mdt_lock_handle *child_lh;
839 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
841 bool cos_incompat = false;
847 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
848 PNAME(&rr->rr_name));
850 if (info->mti_dlm_req)
851 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
853 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
854 RETURN(err_serious(-ENOENT));
856 if (!fid_is_md_operative(rr->rr_fid1))
859 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
863 if (mdt_object_remote(mp)) {
866 rc = mdt_version_get_check_save(info, mp, 0);
868 GOTO(put_parent, rc);
872 parent_lh = &info->mti_lh[MDT_LH_PARENT];
873 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
874 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
877 GOTO(put_parent, rc);
879 /* lookup child object along with version checking */
881 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
883 /* Name might not be able to find during resend of
884 * remote unlink, considering following case.
885 * dir_A is a remote directory, the name entry of
886 * dir_A is on MDT0, the directory is on MDT1,
888 * 1. client sends unlink req to MDT1.
889 * 2. MDT1 sends name delete update to MDT0.
890 * 3. name entry is being deleted in MDT0 synchronously.
891 * 4. MDT1 is restarted.
892 * 5. client resends unlink req to MDT1. So it can not
893 * find the name entry on MDT0 anymore.
894 * In this case, MDT1 only needs to destory the local
897 if (mdt_object_remote(mp) && rc == -ENOENT &&
898 !fid_is_zero(rr->rr_fid2) &&
899 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
901 *child_fid = *rr->rr_fid2;
903 GOTO(unlock_parent, rc);
907 if (!fid_is_md_operative(child_fid))
908 GOTO(unlock_parent, rc = -EPERM);
910 /* We will lock the child regardless it is local or remote. No harm. */
911 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
913 GOTO(unlock_parent, rc = PTR_ERR(mc));
916 rc = mdt_object_striped(info, mc);
918 GOTO(unlock_parent, rc = PTR_ERR(mc));
922 mdt_object_put(info->mti_env, mc);
923 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
928 child_lh = &info->mti_lh[MDT_LH_CHILD];
929 mdt_lock_reg_init(child_lh, LCK_EX);
930 if (info->mti_spec.sp_rm_entry) {
931 struct lu_ucred *uc = mdt_ucred(info);
933 if (!mdt_is_dne_client(req->rq_export))
934 /* Return -ENOTSUPP for old client */
935 GOTO(put_child, rc = -ENOTSUPP);
937 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
938 GOTO(put_child, rc = -EPERM);
940 ma->ma_need = MA_INODE;
942 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
943 NULL, &rr->rr_name, ma, no_name);
947 if (mdt_object_remote(mc)) {
948 struct mdt_body *repbody;
950 if (!fid_is_zero(rr->rr_fid2)) {
951 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
952 mdt_obd_name(info->mti_mdt),
953 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
954 GOTO(put_child, rc = -ENOENT);
956 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
957 mdt_obd_name(info->mti_mdt),
958 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
960 if (!mdt_is_dne_client(req->rq_export))
961 /* Return -ENOTSUPP for old client */
962 GOTO(put_child, rc = -ENOTSUPP);
964 /* Revoke the LOOKUP lock of the remote object granted by
965 * this MDT. Since the unlink will happen on another MDT,
966 * it will release the LOOKUP lock right away. Then What
967 * would happen if another client try to grab the LOOKUP
968 * lock at the same time with unlink XXX */
969 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
970 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
971 LASSERT(repbody != NULL);
972 repbody->mbo_fid1 = *mdt_object_fid(mc);
973 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
974 GOTO(unlock_child, rc = -EREMOTE);
976 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
977 * this now because a running HSM restore on the child (unlink
978 * victim) will hold the layout lock. See LU-4002. */
979 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
980 if (mdt_object_remote(mp)) {
981 /* Enqueue lookup lock from parent MDT */
982 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
983 &child_lh->mlh_rreg_lh,
984 child_lh->mlh_rreg_mode,
985 MDS_INODELOCK_LOOKUP, false);
989 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
992 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
998 * Now we can only make sure we need MA_INODE, in mdd layer, will check
999 * whether need MA_LOV and MA_COOKIE.
1001 ma->ma_need = MA_INODE;
1004 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1005 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1006 /* save version when object is locked */
1007 mdt_version_get_save(info, mc, 1);
1009 mutex_lock(&mc->mot_lov_mutex);
1011 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1012 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1014 mutex_unlock(&mc->mot_lov_mutex);
1016 GOTO(unlock_child, rc);
1018 if (!lu_object_is_dying(&mc->mot_header)) {
1019 rc = mdt_attr_get_complex(info, mc, ma);
1023 mdt_dom_check_and_discard(info, mc);
1025 mdt_handle_last_unlink(info, mc, ma);
1028 if (ma->ma_valid & MA_INODE) {
1029 switch (ma->ma_attr.la_mode & S_IFMT) {
1031 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1039 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1042 LASSERTF(0, "bad file type %o unlinking\n",
1043 ma->ma_attr.la_mode);
1050 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1052 mdt_object_put(info->mti_env, mc);
1054 mdt_object_unlock(info, mp, parent_lh, rc);
1056 mdt_object_put(info->mti_env, mp);
1061 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1064 static int mdt_reint_link(struct mdt_thread_info *info,
1065 struct mdt_lock_handle *lhc)
1067 struct mdt_reint_record *rr = &info->mti_rr;
1068 struct ptlrpc_request *req = mdt_info_req(info);
1069 struct md_attr *ma = &info->mti_attr;
1070 struct mdt_object *ms;
1071 struct mdt_object *mp;
1072 struct mdt_lock_handle *lhs;
1073 struct mdt_lock_handle *lhp;
1078 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1079 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1081 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1082 RETURN(err_serious(-ENOENT));
1084 if (info->mti_dlm_req)
1085 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1087 /* Invalid case so return error immediately instead of
1089 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1092 if (!fid_is_md_operative(rr->rr_fid1) ||
1093 !fid_is_md_operative(rr->rr_fid2))
1096 /* step 1: find target parent dir */
1097 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1099 RETURN(PTR_ERR(mp));
1101 rc = mdt_version_get_check_save(info, mp, 0);
1103 GOTO(put_parent, rc);
1105 /* step 2: find source */
1106 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1108 GOTO(put_parent, rc = PTR_ERR(ms));
1110 if (!mdt_object_exists(ms)) {
1111 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1112 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1113 GOTO(put_source, rc = -ENOENT);
1116 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1118 lhp = &info->mti_lh[MDT_LH_PARENT];
1119 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1120 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1123 GOTO(put_source, rc);
1125 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1127 lhs = &info->mti_lh[MDT_LH_CHILD];
1128 mdt_lock_reg_init(lhs, LCK_EX);
1129 rc = mdt_reint_object_lock(info, ms, lhs,
1130 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1133 GOTO(unlock_parent, rc);
1135 /* step 3: link it */
1136 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1137 OBD_FAIL_MDS_REINT_LINK_WRITE);
1139 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1140 rc = mdt_version_get_check_save(info, ms, 1);
1142 GOTO(unlock_source, rc);
1144 /** check target version by name during replay */
1145 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1146 &info->mti_tmp_fid1, 2);
1147 if (rc != 0 && rc != -ENOENT)
1148 GOTO(unlock_source, rc);
1149 /* save version of file name for replay, it must be ENOENT here */
1150 if (!req_is_replay(mdt_info_req(info))) {
1151 if (rc != -ENOENT) {
1152 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1153 PNAME(&rr->rr_name));
1154 GOTO(unlock_source, rc = -EEXIST);
1156 info->mti_ver[2] = ENOENT_VERSION;
1157 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1160 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1161 mdt_object_child(ms), &rr->rr_name, ma);
1164 mdt_counter_incr(req, LPROC_MDT_LINK);
1168 mdt_object_unlock(info, ms, lhs, rc);
1170 mdt_object_unlock(info, mp, lhp, rc);
1172 mdt_object_put(info->mti_env, ms);
1174 mdt_object_put(info->mti_env, mp);
1178 * lock the part of the directory according to the hash of the name
1179 * (lh->mlh_pdo_hash) in parallel directory lock.
1181 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1182 struct mdt_lock_handle *lh,
1183 struct mdt_object *obj, __u64 ibits,
1186 struct ldlm_res_id *res = &info->mti_res_id;
1187 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1188 union ldlm_policy_data *policy = &info->mti_policy;
1189 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1193 * Finish res_id initializing by name hash marking part of
1194 * directory which is taking modification.
1196 LASSERT(lh->mlh_pdo_hash != 0);
1197 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1198 memset(policy, 0, sizeof(*policy));
1199 policy->l_inodebits.bits = ibits;
1201 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1202 dlmflags |= LDLM_FL_COS_INCOMPAT;
1204 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1205 * going to be sent to client. If it is - mdt_intent_policy() path will
1206 * fix it up and turn FL_LOCAL flag off.
1208 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1209 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1214 * Get BFL lock for rename or migrate process.
1216 static int mdt_rename_lock(struct mdt_thread_info *info,
1217 struct lustre_handle *lh)
1222 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1223 struct lu_fid *fid = &info->mti_tmp_fid1;
1224 struct mdt_object *obj;
1226 /* XXX, right now, it has to use object API to
1227 * enqueue lock cross MDT, so it will enqueue
1228 * rename lock(with LUSTRE_BFL_FID) by root object */
1230 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1232 RETURN(PTR_ERR(obj));
1234 rc = mdt_remote_object_lock(info, obj,
1235 &LUSTRE_BFL_FID, lh,
1237 MDS_INODELOCK_UPDATE, false);
1238 mdt_object_put(info->mti_env, obj);
1240 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1241 union ldlm_policy_data *policy = &info->mti_policy;
1242 struct ldlm_res_id *res_id = &info->mti_res_id;
1245 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1246 memset(policy, 0, sizeof *policy);
1247 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1248 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1249 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1250 LCK_EX, &flags, ldlm_blocking_ast,
1251 ldlm_completion_ast, NULL, NULL, 0,
1253 &info->mti_exp->exp_handle.h_cookie,
1260 static void mdt_rename_unlock(struct lustre_handle *lh)
1263 LASSERT(lustre_handle_is_used(lh));
1264 /* Cancel the single rename lock right away */
1265 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1270 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1271 * target. Source should not be ancestor of target dir. May be other rename
1272 * checks can be moved here later.
1274 static int mdt_is_subdir(struct mdt_thread_info *info,
1275 struct mdt_object *dir,
1276 const struct lu_fid *fid)
1278 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1282 /* If the source and target are in the same directory, they can not
1283 * be parent/child relationship, so subdir check is not needed */
1284 if (lu_fid_eq(&dir_fid, fid))
1287 if (!mdt_object_exists(dir))
1290 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1293 CERROR("%s: failed subdir check in "DFID" for "DFID
1294 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1295 PFID(&dir_fid), PFID(fid), rc);
1296 /* Return EINVAL only if a parent is the @fid */
1300 /* check the found fid */
1301 if (lu_fid_eq(&dir_fid, fid))
1308 /* Update object linkEA */
1309 struct mdt_lock_list {
1310 struct mdt_object *mll_obj;
1311 struct mdt_lock_handle mll_lh;
1312 struct list_head mll_list;
1315 static void mdt_unlock_list(struct mdt_thread_info *info,
1316 struct list_head *list, int rc)
1318 struct mdt_lock_list *mll;
1319 struct mdt_lock_list *mll2;
1321 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1322 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1323 list_del(&mll->mll_list);
1328 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1329 struct mdt_object *obj,
1330 struct mdt_object *pobj,
1331 struct list_head *lock_list)
1333 struct lu_buf *buf = &info->mti_big_buf;
1334 struct linkea_data ldata = { NULL };
1340 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1343 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1344 if (buf->lb_buf == NULL)
1348 rc = mdt_links_read(info, obj, &ldata);
1350 if (rc == -ENOENT || rc == -ENODATA)
1355 /* ignore the migrating parent(@pobj) */
1356 retry_count = ldata.ld_leh->leh_reccount - 1;
1359 LASSERT(ldata.ld_leh != NULL);
1360 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1361 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1362 struct mdt_device *mdt = info->mti_mdt;
1363 struct mdt_object *mdt_pobj;
1364 struct mdt_lock_list *mll;
1365 struct lu_name name;
1369 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1371 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1372 if (IS_ERR(mdt_pobj)) {
1373 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1374 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1378 if (!mdt_object_exists(mdt_pobj)) {
1379 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1380 mdt_obd_name(mdt), PFID(&fid));
1381 mdt_object_put(info->mti_env, mdt_pobj);
1385 /* Check if the object already exists in the list */
1386 list_for_each_entry(mll, lock_list, mll_list) {
1387 if (mll->mll_obj == mdt_pobj) {
1388 mdt_object_put(info->mti_env, mdt_pobj);
1393 if (mdt_pobj == pobj) {
1394 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1395 mdt_obd_name(mdt), PFID(&fid));
1396 mdt_object_put(info->mti_env, mdt_pobj);
1402 mdt_object_put(info->mti_env, mdt_pobj);
1403 GOTO(out, rc = -ENOMEM);
1406 /* Since this needs to lock all of objects in linkea, to avoid
1407 * deadlocks, because it does not follow parent-child order as
1408 * other MDT operation, let's use try_lock here and if the lock
1409 * cannot be gotten because of conflicting locks, then drop all
1410 * current locks, send an AST to the client, and start again. */
1411 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1413 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits,
1414 MDS_INODELOCK_UPDATE, true);
1415 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1416 mdt_unlock_list(info, lock_list, 0);
1418 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1419 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1422 if (retry_count == 0) {
1423 mdt_object_put(info->mti_env, mdt_pobj);
1425 GOTO(out, rc = -EBUSY);
1428 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1429 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1430 MDS_INODELOCK_UPDATE);
1432 mdt_object_put(info->mti_env, mdt_pobj);
1437 if (mdt_object_remote(mdt_pobj)) {
1438 struct ldlm_lock *lock;
1440 /* For remote object, Set lock to cb_atomic,
1441 * so lock can be released in blocking_ast()
1442 * immediately, then the next try_lock will
1443 * have better chance to succeds */
1445 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1446 LASSERT(lock != NULL);
1447 lock_res_and_lock(lock);
1448 ldlm_set_atomic_cb(lock);
1449 unlock_res_and_lock(lock);
1450 LDLM_LOCK_PUT(lock);
1452 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1458 INIT_LIST_HEAD(&mll->mll_list);
1459 mll->mll_obj = mdt_pobj;
1460 list_add_tail(&mll->mll_list, lock_list);
1462 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1467 mdt_unlock_list(info, lock_list, rc);
1471 /* migrate files from one MDT to another MDT */
1472 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1473 struct mdt_lock_handle *lhc)
1475 struct mdt_reint_record *rr = &info->mti_rr;
1476 struct md_attr *ma = &info->mti_attr;
1477 struct mdt_object *msrcdir;
1478 struct mdt_object *mold;
1479 struct mdt_object *mnew = NULL;
1480 struct mdt_lock_handle *lh_dirp;
1481 struct mdt_lock_handle *lh_childp;
1482 struct mdt_lock_handle *lh_tgtp = NULL;
1483 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1484 struct list_head lock_list;
1486 struct ldlm_lock *lease = NULL;
1487 bool lock_open_sem = false;
1491 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1492 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1494 /* 1: lock the source dir. */
1495 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1496 if (IS_ERR(msrcdir)) {
1497 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1498 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1499 (int)PTR_ERR(msrcdir));
1500 RETURN(PTR_ERR(msrcdir));
1503 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1504 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1505 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1508 GOTO(out_put_parent, rc);
1510 if (!mdt_object_remote(msrcdir)) {
1511 rc = mdt_version_get_check_save(info, msrcdir, 0);
1513 GOTO(out_unlock_parent, rc);
1516 /* 2: sanity check and find the object to be migrated. */
1518 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1520 GOTO(out_unlock_parent, rc);
1522 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1523 GOTO(out_unlock_parent, rc = -EINVAL);
1525 if (!fid_is_md_operative(old_fid))
1526 GOTO(out_unlock_parent, rc = -EPERM);
1528 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1529 GOTO(out_unlock_parent, rc = -EPERM);
1531 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1533 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1535 if (mdt_object_remote(mold)) {
1536 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1537 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1538 GOTO(out_put_child, rc = -EREMOTE);
1541 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1542 !mdt_object_remote(msrcdir)) {
1543 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1544 " MDT, which should be migrated first:"
1545 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1546 PFID(mdt_object_fid(msrcdir)), -EPERM);
1547 GOTO(out_put_child, rc = -EPERM);
1550 rc = mdt_remote_permission(info);
1552 GOTO(out_put_child, rc);
1554 /* 3: iterate the linkea of the object and lock all of the objects */
1555 INIT_LIST_HEAD(&lock_list);
1556 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1558 GOTO(out_put_child, rc);
1560 if (info->mti_spec.sp_migrate_close) {
1561 struct close_data *data;
1562 struct mdt_body *repbody;
1563 bool lease_broken = false;
1565 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1567 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1569 GOTO(out_lease, rc = -EPROTO);
1571 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1573 GOTO(out_lease, rc = -EPROTO);
1575 lease = ldlm_handle2lock(&data->cd_handle);
1577 GOTO(out_lease, rc = -ESTALE);
1579 /* try to hold open_sem so that nobody else can open the file */
1580 if (!down_write_trylock(&mold->mot_open_sem)) {
1581 ldlm_lock_cancel(lease);
1582 GOTO(out_lease, rc = -EBUSY);
1585 lock_open_sem = true;
1586 /* Check if the lease open lease has already canceled */
1587 lock_res_and_lock(lease);
1588 lease_broken = ldlm_is_cancel(lease);
1589 unlock_res_and_lock(lease);
1591 LDLM_DEBUG(lease, DFID " lease broken? %d",
1592 PFID(mdt_object_fid(mold)), lease_broken);
1594 /* Cancel server side lease. Client side counterpart should
1595 * have been cancelled. It's okay to cancel it now as we've
1596 * held mot_open_sem. */
1597 ldlm_lock_cancel(lease);
1600 GOTO(out_lease, rc = -EAGAIN);
1602 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1603 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1604 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1606 GOTO(out_unlock_list, rc);
1609 /* 4: lock of the object migrated object */
1610 lh_childp = &info->mti_lh[MDT_LH_OLD];
1611 mdt_lock_reg_init(lh_childp, LCK_EX);
1612 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1613 MDS_INODELOCK_LAYOUT;
1614 if (mdt_object_remote(msrcdir)) {
1615 /* Enqueue lookup lock from the parent MDT */
1616 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1617 &lh_childp->mlh_rreg_lh,
1618 lh_childp->mlh_rreg_mode,
1619 MDS_INODELOCK_LOOKUP, false);
1621 GOTO(out_unlock_list, rc);
1623 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1626 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1628 GOTO(out_unlock_child, rc);
1630 /* Migration is incompatible with HSM. */
1631 ma->ma_need = MA_HSM;
1633 rc = mdt_attr_get_complex(info, mold, ma);
1635 GOTO(out_unlock_child, rc);
1637 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1640 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1641 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1642 GOTO(out_unlock_child, rc);
1645 ma->ma_need = MA_LMV;
1647 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1648 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1649 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1651 GOTO(out_unlock_child, rc);
1653 if ((ma->ma_valid & MA_LMV)) {
1654 struct lmv_mds_md_v1 *lmm1;
1656 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1657 lmm1 = &ma->ma_lmv->lmv_md_v1;
1658 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1659 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1660 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1661 PFID(mdt_object_fid(mold)), -EPERM);
1662 GOTO(out_unlock_child, rc = -EPERM);
1665 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1666 GOTO(out_unlock_child, rc = -EINVAL);
1668 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1669 &lmm1->lmv_stripe_fids[1]);
1671 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1673 if (!mdt_object_remote(mnew)) {
1675 "%s: "DFID" being migrated is on this MDT:"
1676 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1677 PFID(rr->rr_fid2), -EPERM);
1678 GOTO(out_put_new, rc = -EPERM);
1681 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1682 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1683 rc = mdt_remote_object_lock(info, mnew,
1684 mdt_object_fid(mnew),
1685 &lh_tgtp->mlh_rreg_lh,
1686 lh_tgtp->mlh_rreg_mode,
1687 MDS_INODELOCK_UPDATE, false);
1690 GOTO(out_put_new, rc);
1693 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1696 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1697 if (!mdt_object_remote(mnew)) {
1698 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1699 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1700 PFID(rr->rr_fid2), -EXDEV);
1701 GOTO(out_put_new, rc = -EXDEV);
1706 mdt_reint_init_ma(info, ma);
1708 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1709 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1711 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1712 mdt_object_child(mold), &rr->rr_name,
1713 mdt_object_child(mnew), ma);
1715 GOTO(out_unlock_new, rc);
1718 if (lh_tgtp != NULL)
1719 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1722 mdt_object_put(info->mti_env, mnew);
1724 mdt_object_unlock(info, mold, lh_childp, rc);
1726 /* we don't really modify linkea objects, so we can safely decref these
1727 * locks, and this can avoid saving them as COS locks, which may prevent
1728 * subsequent migrate. */
1729 mdt_unlock_list(info, &lock_list, 1);
1730 if (lease != NULL) {
1731 ldlm_reprocess_all(lease->l_resource);
1732 LDLM_LOCK_PUT(lease);
1736 up_write(&mold->mot_open_sem);
1738 mdt_object_put(info->mti_env, mold);
1740 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1742 mdt_object_put(info->mti_env, msrcdir);
1747 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1748 const struct lu_fid *fid,
1751 struct mdt_object *dir;
1755 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1759 /* check early, the real version will be saved after locking */
1760 rc = mdt_version_get_check(info, dir, idx);
1766 mdt_object_put(info->mti_env, dir);
1770 static int mdt_object_lock_save(struct mdt_thread_info *info,
1771 struct mdt_object *dir,
1772 struct mdt_lock_handle *lh,
1773 int idx, bool cos_incompat)
1777 /* we lock the target dir if it is local */
1778 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1783 /* get and save correct version after locking */
1784 mdt_version_get_save(info, dir, idx);
1789 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1790 * 2 - srcdir child; 3 - tgtdir child.
1791 * Update on disk version of srcdir child.
1794 * For DNE phase I, only these renames are allowed
1795 * mv src_p/src_c tgt_p/tgt_c
1796 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1797 * 2. src_p and tgt_p are same directory, and tgt_c does not
1798 * exists. In this case, all of modification will happen
1799 * in the MDT where ithesource parent is, only one remote
1800 * update is needed, i.e. set c_time/m_time on the child.
1801 * And tgt_c will be still in the same MDT as the original
1804 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1805 struct mdt_lock_handle *lhc)
1807 struct mdt_reint_record *rr = &info->mti_rr;
1808 struct md_attr *ma = &info->mti_attr;
1809 struct ptlrpc_request *req = mdt_info_req(info);
1810 struct mdt_object *msrcdir = NULL;
1811 struct mdt_object *mtgtdir = NULL;
1812 struct mdt_object *mold;
1813 struct mdt_object *mnew = NULL;
1814 struct mdt_lock_handle *lh_srcdirp;
1815 struct mdt_lock_handle *lh_tgtdirp;
1816 struct mdt_lock_handle *lh_oldp = NULL;
1817 struct mdt_lock_handle *lh_newp = NULL;
1818 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1819 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1821 bool reverse = false;
1826 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1827 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1828 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1830 /* find both parents. */
1831 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1832 if (IS_ERR(msrcdir))
1833 RETURN(PTR_ERR(msrcdir));
1835 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1837 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1839 mdt_object_get(info->mti_env, mtgtdir);
1841 /* Check if the @msrcdir is not a child of the @mtgtdir,
1842 * otherwise a reverse locking must take place. */
1843 rc = mdt_is_subdir(info, msrcdir, rr->rr_fid2);
1847 GOTO(out_put_srcdir, rc);
1849 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1850 if (IS_ERR(mtgtdir))
1851 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1854 /* source needs to be looked up after locking source parent, otherwise
1855 * this rename may race with unlink source, and cause rename hang, see
1856 * sanityn.sh 55b, so check parents first, if later we found source is
1857 * remote, relock parents. */
1858 cos_incompat = (mdt_object_remote(msrcdir) ||
1859 mdt_object_remote(mtgtdir));
1861 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1863 /* lock parents in the proper order. */
1864 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1865 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1868 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1869 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1872 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1875 GOTO(out_put_tgtdir, rc);
1877 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1879 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1882 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1883 GOTO(out_put_tgtdir, rc);
1886 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1889 GOTO(out_put_tgtdir, rc);
1891 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1893 if (mtgtdir != msrcdir) {
1894 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1896 } else if (lh_srcdirp->mlh_pdo_hash !=
1897 lh_tgtdirp->mlh_pdo_hash) {
1898 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1899 MDS_INODELOCK_UPDATE,
1901 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1904 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
1905 GOTO(out_put_tgtdir, rc);
1909 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1910 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1912 /* find mold object. */
1914 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1916 GOTO(out_unlock_parents, rc);
1918 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1919 GOTO(out_unlock_parents, rc = -EINVAL);
1921 if (!fid_is_md_operative(old_fid))
1922 GOTO(out_unlock_parents, rc = -EPERM);
1924 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1926 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1928 /* Check if @mtgtdir is subdir of @mold, before locking child
1929 * to avoid reverse locking. */
1930 if (mtgtdir != msrcdir) {
1931 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1933 GOTO(out_put_old, rc);
1936 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1937 /* save version after locking */
1938 mdt_version_get_save(info, mold, 2);
1940 if (!cos_incompat && mdt_object_remote(mold)) {
1941 cos_incompat = true;
1942 mdt_object_put(info->mti_env, mold);
1943 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
1944 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
1948 /* find mnew object:
1949 * mnew target object may not exist now
1950 * lookup with version checking */
1952 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1955 /* the new_fid should have been filled at this moment */
1956 if (lu_fid_eq(old_fid, new_fid))
1957 GOTO(out_put_old, rc);
1959 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1960 lu_fid_eq(new_fid, rr->rr_fid2))
1961 GOTO(out_put_old, rc = -EINVAL);
1963 if (!fid_is_md_operative(new_fid))
1964 GOTO(out_put_old, rc = -EPERM);
1966 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1968 GOTO(out_put_old, rc = PTR_ERR(mnew));
1970 if (mdt_object_remote(mnew)) {
1971 struct mdt_body *repbody;
1973 /* Always send rename req to the target child MDT */
1974 repbody = req_capsule_server_get(info->mti_pill,
1976 LASSERT(repbody != NULL);
1977 repbody->mbo_fid1 = *new_fid;
1978 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1979 GOTO(out_put_new, rc = -EXDEV);
1981 /* Before locking the target dir, check we do not replace
1982 * a dir with a non-dir, otherwise it may deadlock with
1983 * link op which tries to create a link in this dir
1984 * back to this non-dir. */
1985 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
1986 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
1987 GOTO(out_put_new, rc = -EISDIR);
1989 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1990 mdt_lock_reg_init(lh_oldp, LCK_EX);
1991 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
1992 if (mdt_object_remote(msrcdir)) {
1993 /* Enqueue lookup lock from the parent MDT */
1994 rc = mdt_remote_object_lock(info, msrcdir,
1995 mdt_object_fid(mold),
1996 &lh_oldp->mlh_rreg_lh,
1997 lh_oldp->mlh_rreg_mode,
1998 MDS_INODELOCK_LOOKUP,
2001 GOTO(out_put_new, rc);
2003 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2006 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2009 GOTO(out_unlock_old, rc);
2011 /* Check if @msrcdir is subdir of @mnew, before locking child
2012 * to avoid reverse locking. */
2013 if (mtgtdir != msrcdir) {
2014 rc = mdt_is_subdir(info, msrcdir, new_fid);
2016 GOTO(out_unlock_old, rc);
2019 /* We used to acquire MDS_INODELOCK_FULL here but we
2020 * can't do this now because a running HSM restore on
2021 * the rename onto victim will hold the layout
2022 * lock. See LU-4002. */
2024 lh_newp = &info->mti_lh[MDT_LH_NEW];
2025 mdt_lock_reg_init(lh_newp, LCK_EX);
2026 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2027 MDS_INODELOCK_LOOKUP |
2028 MDS_INODELOCK_UPDATE,
2031 GOTO(out_unlock_old, rc);
2033 /* get and save version after locking */
2034 mdt_version_get_save(info, mnew, 3);
2035 } else if (rc != -EREMOTE && rc != -ENOENT) {
2036 GOTO(out_put_old, rc);
2038 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2039 mdt_lock_reg_init(lh_oldp, LCK_EX);
2040 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2041 if (mdt_object_remote(msrcdir)) {
2042 /* Enqueue lookup lock from the parent MDT */
2043 rc = mdt_remote_object_lock(info, msrcdir,
2044 mdt_object_fid(mold),
2045 &lh_oldp->mlh_rreg_lh,
2046 lh_oldp->mlh_rreg_mode,
2047 MDS_INODELOCK_LOOKUP,
2050 GOTO(out_put_old, rc);
2052 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2055 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2058 GOTO(out_unlock_old, rc);
2060 mdt_enoent_version_save(info, 3);
2063 /* step 5: rename it */
2064 mdt_reint_init_ma(info, ma);
2066 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2067 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2070 mutex_lock(&mnew->mot_lov_mutex);
2072 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2073 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2074 mnew != NULL ? mdt_object_child(mnew) : NULL,
2075 &rr->rr_tgt_name, ma);
2078 mutex_unlock(&mnew->mot_lov_mutex);
2080 /* handle last link of tgt object */
2082 mdt_counter_incr(req, LPROC_MDT_RENAME);
2084 mdt_handle_last_unlink(info, mnew, ma);
2085 mdt_dom_check_and_discard(info, mnew);
2088 mdt_rename_counter_tally(info, info->mti_mdt, req,
2094 mdt_object_unlock(info, mnew, lh_newp, rc);
2096 mdt_object_unlock(info, mold, lh_oldp, rc);
2099 mdt_object_put(info->mti_env, mnew);
2101 mdt_object_put(info->mti_env, mold);
2103 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2104 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2106 mdt_object_put(info->mti_env, mtgtdir);
2108 mdt_object_put(info->mti_env, msrcdir);
2112 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2113 struct mdt_lock_handle *lhc, bool rename)
2115 struct mdt_reint_record *rr = &info->mti_rr;
2116 struct ptlrpc_request *req = mdt_info_req(info);
2117 struct lustre_handle rename_lh = { 0 };
2121 if (info->mti_dlm_req)
2122 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2124 if (!fid_is_md_operative(rr->rr_fid1) ||
2125 !fid_is_md_operative(rr->rr_fid2))
2128 /* Note: do not enqueue rename lock for replay request, because
2129 * if other MDT holds rename lock, but being blocked to wait for
2130 * this MDT to finish its recovery, and the failover MDT can not
2131 * get rename lock, which will cause deadlock. */
2132 if (!req_is_replay(req)) {
2133 rc = mdt_rename_lock(info, &rename_lh);
2135 CERROR("%s: can't lock FS for rename: rc = %d\n",
2136 mdt_obd_name(info->mti_mdt), rc);
2142 rc = mdt_reint_rename_internal(info, lhc);
2144 rc = mdt_reint_migrate_internal(info, lhc);
2146 if (lustre_handle_is_used(&rename_lh))
2147 mdt_rename_unlock(&rename_lh);
2152 static int mdt_reint_rename(struct mdt_thread_info *info,
2153 struct mdt_lock_handle *lhc)
2155 return mdt_reint_rename_or_migrate(info, lhc, true);
2158 static int mdt_reint_migrate(struct mdt_thread_info *info,
2159 struct mdt_lock_handle *lhc)
2161 return mdt_reint_rename_or_migrate(info, lhc, false);
2164 static int mdt_reint_resync(struct mdt_thread_info *info,
2165 struct mdt_lock_handle *lhc)
2167 struct mdt_reint_record *rr = &info->mti_rr;
2168 struct ptlrpc_request *req = mdt_info_req(info);
2169 struct md_attr *ma = &info->mti_attr;
2170 struct mdt_object *mo;
2171 struct ldlm_lock *lease;
2172 struct mdt_body *repbody;
2173 struct md_layout_change layout = { 0 };
2178 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2180 if (info->mti_dlm_req)
2181 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2183 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2185 GOTO(out, rc = PTR_ERR(mo));
2187 if (!mdt_object_exists(mo))
2188 GOTO(out_obj, rc = -ENOENT);
2190 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2191 GOTO(out_obj, rc = -EINVAL);
2193 if (mdt_object_remote(mo))
2194 GOTO(out_obj, rc = -EREMOTE);
2196 lease = ldlm_handle2lock(rr->rr_handle);
2198 GOTO(out_obj, rc = -ESTALE);
2200 /* It's really necessary to grab open_sem and check if the lease lock
2201 * has been lost. There would exist a concurrent writer coming in and
2202 * generating some dirty data in memory cache, the writeback would fail
2203 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2204 if (!down_write_trylock(&mo->mot_open_sem))
2205 GOTO(out_put_lease, rc = -EBUSY);
2207 lock_res_and_lock(lease);
2208 lease_broken = ldlm_is_cancel(lease);
2209 unlock_res_and_lock(lease);
2211 GOTO(out_unlock, rc = -EBUSY);
2213 /* the file has yet opened by anyone else after we took the lease. */
2214 layout.mlc_opc = MD_LAYOUT_RESYNC;
2215 rc = mdt_layout_change(info, mo, &layout);
2217 GOTO(out_unlock, rc);
2219 ma->ma_need = MA_INODE;
2221 rc = mdt_attr_get_complex(info, mo, ma);
2223 GOTO(out_unlock, rc);
2225 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2226 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2230 up_write(&mo->mot_open_sem);
2232 LDLM_LOCK_PUT(lease);
2234 mdt_object_put(info->mti_env, mo);
2236 mdt_client_compatibility(info);
2237 rc2 = mdt_fix_reply(info);
2243 struct mdt_reinter {
2244 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2245 enum lprocfs_extra_opc mr_extra_opc;
2248 static const struct mdt_reinter mdt_reinters[] = {
2250 .mr_handler = &mdt_reint_setattr,
2251 .mr_extra_opc = MDS_REINT_SETATTR,
2254 .mr_handler = &mdt_reint_create,
2255 .mr_extra_opc = MDS_REINT_CREATE,
2258 .mr_handler = &mdt_reint_link,
2259 .mr_extra_opc = MDS_REINT_LINK,
2262 .mr_handler = &mdt_reint_unlink,
2263 .mr_extra_opc = MDS_REINT_UNLINK,
2266 .mr_handler = &mdt_reint_rename,
2267 .mr_extra_opc = MDS_REINT_RENAME,
2270 .mr_handler = &mdt_reint_open,
2271 .mr_extra_opc = MDS_REINT_OPEN,
2273 [REINT_SETXATTR] = {
2274 .mr_handler = &mdt_reint_setxattr,
2275 .mr_extra_opc = MDS_REINT_SETXATTR,
2278 .mr_handler = &mdt_reint_unlink,
2279 .mr_extra_opc = MDS_REINT_UNLINK,
2282 .mr_handler = &mdt_reint_migrate,
2283 .mr_extra_opc = MDS_REINT_RENAME,
2286 .mr_handler = &mdt_reint_resync,
2287 .mr_extra_opc = MDS_REINT_RESYNC,
2291 int mdt_reint_rec(struct mdt_thread_info *info,
2292 struct mdt_lock_handle *lhc)
2294 const struct mdt_reinter *mr;
2298 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2301 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2302 if (mr->mr_handler == NULL)
2305 rc = (*mr->mr_handler)(info, lhc);
2307 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2308 PTLRPC_LAST_CNTR + mr->mr_extra_opc);