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/
31 * lustre/mdt/mdt_reint.c
33 * Lustre Metadata Target (mdt) reintegration routines
35 * Author: Peter Braam <braam@clusterfs.com>
36 * Author: Andreas Dilger <adilger@clusterfs.com>
37 * Author: Phil Schwan <phil@clusterfs.com>
38 * Author: Huang Hua <huanghua@clusterfs.com>
39 * Author: Yury Umanets <umka@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_MDS
44 #include <lprocfs_status.h>
45 #include "mdt_internal.h"
46 #include <lustre_lmv.h>
48 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
51 ma->ma_need = MA_INODE;
56 * Get version of object by fid.
58 * Return real version or ENOENT_VERSION if object doesn't exist
60 static void mdt_obj_version_get(struct mdt_thread_info *info,
61 struct mdt_object *o, __u64 *version)
65 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
66 !fid_is_obf(mdt_object_fid(o)))
67 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
69 *version = ENOENT_VERSION;
70 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
71 PFID(mdt_object_fid(o)), *version);
75 * Check version is correct.
77 * Should be called only during replay.
79 static int mdt_version_check(struct ptlrpc_request *req,
80 __u64 version, int idx)
82 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
85 if (!exp_connect_vbr(req->rq_export))
88 LASSERT(req_is_replay(req));
89 /** VBR: version is checked always because costs nothing */
90 LASSERT(idx < PTLRPC_NUM_VERSIONS);
91 /** Sanity check for malformed buffers */
92 if (pre_ver == NULL) {
93 CERROR("No versions in request buffer\n");
94 spin_lock(&req->rq_export->exp_lock);
95 req->rq_export->exp_vbr_failed = 1;
96 spin_unlock(&req->rq_export->exp_lock);
98 } else if (pre_ver[idx] != version) {
99 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
100 pre_ver[idx], version);
101 spin_lock(&req->rq_export->exp_lock);
102 req->rq_export->exp_vbr_failed = 1;
103 spin_unlock(&req->rq_export->exp_lock);
110 * Save pre-versions in reply.
112 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
117 if (!exp_connect_vbr(req->rq_export))
120 LASSERT(!req_is_replay(req));
121 LASSERT(req->rq_repmsg != NULL);
122 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
124 reply_ver[idx] = version;
128 * Save enoent version, it is needed when it is obvious that object doesn't
129 * exist, e.g. child during create.
131 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
133 /* save version of file name for replay, it must be ENOENT here */
134 if (!req_is_replay(mdt_info_req(info))) {
135 info->mti_ver[idx] = ENOENT_VERSION;
136 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
141 * Get version from disk and save in reply buffer.
143 * Versions are saved in reply only during normal operations not replays.
145 void mdt_version_get_save(struct mdt_thread_info *info,
146 struct mdt_object *mto, int idx)
148 /* don't save versions during replay */
149 if (!req_is_replay(mdt_info_req(info))) {
150 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
151 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
156 * Get version from disk and check it, no save in reply.
158 int mdt_version_get_check(struct mdt_thread_info *info,
159 struct mdt_object *mto, int idx)
161 /* only check versions during replay */
162 if (!req_is_replay(mdt_info_req(info)))
165 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
166 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
170 * Get version from disk and check if recovery or just save.
172 int mdt_version_get_check_save(struct mdt_thread_info *info,
173 struct mdt_object *mto, int idx)
177 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
178 if (req_is_replay(mdt_info_req(info)))
179 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
182 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
187 * Lookup with version checking.
189 * This checks version of 'name'. Many reint functions uses 'name' for child not
190 * FID, therefore we need to get object by name and check its version.
192 int mdt_lookup_version_check(struct mdt_thread_info *info,
193 struct mdt_object *p,
194 const struct lu_name *lname,
195 struct lu_fid *fid, int idx)
199 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
201 /* Check version only during replay */
202 if (!req_is_replay(mdt_info_req(info)))
205 info->mti_ver[idx] = ENOENT_VERSION;
207 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 int mdt_unlock_slaves(struct mdt_thread_info *mti,
221 struct mdt_object *obj,
222 struct ldlm_enqueue_info *einfo,
225 union ldlm_policy_data *policy = &mti->mti_policy;
226 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
227 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
230 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
231 LASSERT(slave_locks);
233 memset(policy, 0, sizeof(*policy));
234 policy->l_inodebits.bits = einfo->ei_inodebits;
235 mdt_lock_handle_init(lh);
236 mdt_lock_reg_init(lh, einfo->ei_mode);
237 for (i = 0; i < slave_locks->ha_count; i++) {
238 if (test_bit(i, (void *)slave_locks->ha_map))
239 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
241 lh->mlh_reg_lh = slave_locks->ha_handles[i];
242 mdt_object_unlock(mti, NULL, lh, decref);
243 slave_locks->ha_handles[i].cookie = 0ull;
246 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
251 struct mdt_object *obj)
253 struct lu_device *bottom_dev;
254 struct lu_object *bottom_obj;
257 if (!S_ISDIR(obj->mot_header.loh_attr))
260 /* getxattr from bottom obj to avoid reading in shard FIDs */
261 bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
262 bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
263 mdt_object_fid(obj), NULL);
264 if (IS_ERR(bottom_obj))
265 return PTR_ERR(bottom_obj);
267 rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
269 lu_object_put(mti->mti_env, bottom_obj);
271 return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
275 * Lock slave stripes if necessary, the lock handles of slave stripes
276 * will be stored in einfo->ei_cbdata.
278 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
279 enum ldlm_mode mode, __u64 ibits,
280 struct ldlm_enqueue_info *einfo)
282 union ldlm_policy_data *policy = &mti->mti_policy;
284 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
286 einfo->ei_type = LDLM_IBITS;
287 einfo->ei_mode = mode;
288 einfo->ei_cb_bl = mdt_remote_blocking_ast;
289 einfo->ei_cb_local_bl = mdt_blocking_ast;
290 einfo->ei_cb_cp = ldlm_completion_ast;
291 einfo->ei_enq_slave = 1;
292 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
293 einfo->ei_inodebits = ibits;
294 memset(policy, 0, sizeof(*policy));
295 policy->l_inodebits.bits = ibits;
297 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
301 int mdt_reint_striped_lock(struct mdt_thread_info *info,
302 struct mdt_object *o,
303 struct mdt_lock_handle *lh,
305 struct ldlm_enqueue_info *einfo,
310 LASSERT(!mdt_object_remote(o));
312 memset(einfo, 0, sizeof(*einfo));
314 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
318 rc = mdt_object_striped(info, o);
321 mdt_object_unlock(info, o, lh, rc);
325 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
327 mdt_object_unlock(info, o, lh, rc);
328 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
335 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
336 struct mdt_object *o,
337 struct mdt_lock_handle *lh,
338 struct ldlm_enqueue_info *einfo, int decref)
340 if (einfo->ei_cbdata)
341 mdt_unlock_slaves(info, o, einfo, decref);
342 mdt_object_unlock(info, o, lh, decref);
345 static int mdt_restripe(struct mdt_thread_info *info,
346 struct mdt_object *parent,
347 const struct lu_name *lname,
348 const struct lu_fid *tfid,
349 struct md_op_spec *spec,
352 struct mdt_device *mdt = info->mti_mdt;
353 struct lu_fid *fid = &info->mti_tmp_fid2;
354 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
355 struct lmv_mds_md_v1 *lmv;
356 struct mdt_object *child;
357 struct mdt_lock_handle *lhp;
358 struct mdt_lock_handle *lhc;
359 struct mdt_body *repbody;
363 if (!mdt->mdt_enable_dir_restripe)
366 rc = mdt_version_get_check_save(info, parent, 0);
370 lhp = &info->mti_lh[MDT_LH_PARENT];
371 mdt_lock_pdo_init(lhp, LCK_PW, lname);
372 rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
377 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
379 GOTO(unlock_parent, rc);
381 if (ma->ma_valid & MA_LMV) {
382 /* don't allow restripe if parent dir layout is changing */
383 lmv = &ma->ma_lmv->lmv_md_v1;
384 if (!lmv_is_sane(lmv))
385 GOTO(unlock_parent, rc = -EBADF);
387 if (lmv_is_layout_changing(lmv))
388 GOTO(unlock_parent, rc = -EBUSY);
392 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
394 GOTO(unlock_parent, rc);
396 child = mdt_object_find(info->mti_env, mdt, fid);
398 GOTO(unlock_parent, rc = PTR_ERR(child));
400 if (!mdt_object_exists(child))
401 GOTO(out_child, rc = -ENOENT);
403 if (mdt_object_remote(child)) {
404 struct mdt_body *repbody;
406 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
408 GOTO(out_child, rc = -EPROTO);
410 repbody->mbo_fid1 = *fid;
411 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
412 GOTO(out_child, rc = -EREMOTE);
416 lhc = &info->mti_lh[MDT_LH_CHILD];
417 mdt_lock_reg_init(lhc, LCK_EX);
419 /* enqueue object remote LOOKUP lock */
420 if (mdt_object_remote(parent)) {
421 rc = mdt_remote_object_lock(info, parent, fid,
424 MDS_INODELOCK_LOOKUP, false);
429 rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
432 GOTO(unlock_child, rc);
434 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
435 rc = mdt_version_get_check_save(info, child, 1);
437 GOTO(unlock_child, rc);
439 spin_lock(&mdt->mdt_restriper.mdr_lock);
440 if (child->mot_restriping) {
442 spin_unlock(&mdt->mdt_restriper.mdr_lock);
443 GOTO(unlock_child, rc = -EBUSY);
445 child->mot_restriping = 1;
446 spin_unlock(&mdt->mdt_restriper.mdr_lock);
449 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
451 GOTO(restriping_clear, rc);
453 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
455 GOTO(restriping_clear, rc = -EPROTO);
457 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
461 child->mot_restriping = 0;
463 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
465 mdt_object_put(info->mti_env, child);
467 mdt_object_unlock(info, parent, lhp, rc);
473 * VBR: we save three versions in reply:
474 * 0 - parent. Check that parent version is the same during replay.
475 * 1 - name. Version of 'name' if file exists with the same name or
476 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
477 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
480 static int mdt_create(struct mdt_thread_info *info)
482 struct mdt_device *mdt = info->mti_mdt;
483 struct mdt_object *parent;
484 struct mdt_object *child;
485 struct mdt_lock_handle *lh;
486 struct mdt_body *repbody;
487 struct md_attr *ma = &info->mti_attr;
488 struct mdt_reint_record *rr = &info->mti_rr;
489 struct md_op_spec *spec = &info->mti_spec;
490 bool restripe = false;
494 DEBUG_REQ(D_INODE, mdt_info_req(info),
495 "Create ("DNAME"->"DFID") in "DFID,
496 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
498 if (!fid_is_md_operative(rr->rr_fid1))
501 if (S_ISDIR(ma->ma_attr.la_mode) &&
502 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
503 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
504 struct lu_ucred *uc = mdt_ucred(info);
505 struct obd_export *exp = mdt_info_req(info)->rq_export;
507 /* Only new clients can create remote dir( >= 2.4) and
508 * striped dir(>= 2.6), old client will return -ENOTSUPP
510 if (!mdt_is_dne_client(exp))
513 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
514 if (!mdt_is_striped_client(exp))
517 if (!mdt->mdt_enable_striped_dir)
519 } else if (!mdt->mdt_enable_remote_dir) {
523 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
524 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
528 if (!md_capable(uc, CAP_SYS_ADMIN) &&
529 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
530 mdt->mdt_enable_remote_dir_gid != -1)
533 /* restripe if later found dir exists, MDS_OPEN_CREAT means
534 * this is create only, don't try restripe.
536 if (mdt->mdt_enable_dir_restripe &&
537 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
538 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
542 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
544 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
546 RETURN(PTR_ERR(parent));
548 if (!mdt_object_exists(parent))
549 GOTO(put_parent, rc = -ENOENT);
552 * LU-10235: check if name exists locklessly first to avoid massive
553 * lock recalls on existing directories.
555 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
556 &info->mti_tmp_fid1, 1);
559 GOTO(put_parent, rc = -EEXIST);
561 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
565 /* -ENOENT is expected here */
567 GOTO(put_parent, rc);
569 /* save version of file name for replay, it must be ENOENT here */
570 mdt_enoent_version_save(info, 1);
572 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
574 lh = &info->mti_lh[MDT_LH_PARENT];
575 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
576 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
578 GOTO(put_parent, rc);
580 if (!mdt_object_remote(parent)) {
581 rc = mdt_version_get_check_save(info, parent, 0);
583 GOTO(unlock_parent, rc);
586 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
587 if (unlikely(IS_ERR(child)))
588 GOTO(unlock_parent, rc = PTR_ERR(child));
590 ma->ma_need = MA_INODE;
593 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
594 OBD_FAIL_MDS_REINT_CREATE_WRITE);
596 /* Version of child will be updated on disk. */
597 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
598 rc = mdt_version_get_check_save(info, child, 2);
602 /* Let lower layer know current lock mode. */
603 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
606 * Do not perform lookup sanity check. We know that name does
609 info->mti_spec.sp_cr_lookup = 0;
610 info->mti_spec.sp_feat = &dt_directory_features;
612 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
613 mdt_object_child(child), &info->mti_spec, ma);
615 rc = mdt_attr_get_complex(info, child, ma);
621 * On DNE, we need to eliminate dependey between 'mkdir a' and
622 * 'mkdir a/b' if b is a striped directory, to achieve this, two
623 * things are done below:
624 * 1. save child and slaves lock.
625 * 2. if the child is a striped directory, relock parent so to
626 * compare against with COS locks to ensure parent was
629 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
630 struct mdt_lock_handle *lhc;
631 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
634 rc = mdt_object_striped(info, child);
640 if (!mdt_object_remote(parent)) {
641 mdt_object_unlock(info, parent, lh, 1);
642 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
643 rc = mdt_reint_object_lock(info, parent, lh,
644 MDS_INODELOCK_UPDATE,
651 lhc = &info->mti_lh[MDT_LH_CHILD];
652 mdt_lock_handle_init(lhc);
653 mdt_lock_reg_init(lhc, LCK_PW);
654 rc = mdt_reint_striped_lock(info, child, lhc,
655 MDS_INODELOCK_UPDATE, einfo,
660 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
663 /* Return fid & attr to client. */
664 if (ma->ma_valid & MA_INODE)
665 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
666 mdt_object_fid(child));
669 mdt_object_put(info->mti_env, child);
671 mdt_object_unlock(info, parent, lh, rc);
673 mdt_object_put(info->mti_env, parent);
677 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
680 struct mdt_lock_handle *lh;
681 int do_vbr = ma->ma_attr.la_valid &
682 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
683 __u64 lockpart = MDS_INODELOCK_UPDATE;
684 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
689 rc = mdt_object_striped(info, mo);
695 lh = &info->mti_lh[MDT_LH_PARENT];
696 mdt_lock_reg_init(lh, LCK_PW);
698 /* Even though the new MDT will grant PERM lock to the old
699 * client, but the old client will almost ignore that during
700 * So it needs to revoke both LOOKUP and PERM lock here, so
701 * both new and old client can cancel the dcache
703 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
704 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
706 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
711 /* all attrs are packed into mti_attr in unpack_setattr */
712 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
713 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
715 /* VBR: update version if attr changed are important for recovery */
717 /* update on-disk version of changed object */
718 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
719 rc = mdt_version_get_check_save(info, mo, 0);
721 GOTO(out_unlock, rc);
724 /* Ensure constant striping during chown(). See LU-2789. */
725 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
726 mutex_lock(&mo->mot_lov_mutex);
728 /* all attrs are packed into mti_attr in unpack_setattr */
729 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
731 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
732 mutex_unlock(&mo->mot_lov_mutex);
735 GOTO(out_unlock, rc);
736 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
739 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
744 * Check HSM flags and add HS_DIRTY flag if relevant.
746 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
747 * and is not RELEASED.
749 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
752 struct lu_ucred *uc = mdt_ucred(info);
757 /* If the file was modified, add the dirty flag */
758 ma->ma_need = MA_HSM;
759 rc = mdt_attr_get_complex(info, mo, ma);
761 CERROR("file attribute read error for "DFID": %d.\n",
762 PFID(mdt_object_fid(mo)), rc);
766 /* If an up2date copy exists in the backend, add dirty flag */
767 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
768 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
769 ma->ma_hsm.mh_flags |= HS_DIRTY;
771 /* Bump cap so that closes from non-owner writers can
772 * set the HSM state to dirty.
774 cap_saved = uc->uc_cap;
775 uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
776 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
777 uc->uc_cap = cap_saved;
779 CERROR("file attribute change error for "DFID": %d\n",
780 PFID(mdt_object_fid(mo)), rc);
786 static int mdt_reint_setattr(struct mdt_thread_info *info,
787 struct mdt_lock_handle *lhc)
789 struct mdt_device *mdt = info->mti_mdt;
790 struct md_attr *ma = &info->mti_attr;
791 struct mdt_reint_record *rr = &info->mti_rr;
792 struct ptlrpc_request *req = mdt_info_req(info);
793 struct mdt_object *mo;
794 struct mdt_body *repbody;
795 ktime_t kstart = ktime_get();
799 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
800 (unsigned int)ma->ma_attr.la_valid);
802 if (info->mti_dlm_req)
803 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
805 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
807 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
808 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
810 GOTO(out, rc = PTR_ERR(mo));
812 if (!mdt_object_exists(mo))
813 GOTO(out_put, rc = -ENOENT);
815 if (mdt_object_remote(mo))
816 GOTO(out_put, rc = -EREMOTE);
818 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
819 /* revoke lease lock if size is going to be changed */
820 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
821 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
822 atomic_read(&mo->mot_lease_count) > 0)) {
823 down_read(&mo->mot_open_sem);
825 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
826 lhc = &info->mti_lh[MDT_LH_LOCAL];
827 mdt_lock_reg_init(lhc, LCK_CW);
829 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
831 up_read(&mo->mot_open_sem);
835 /* revoke lease lock */
836 mdt_object_unlock(info, mo, lhc, 1);
838 up_read(&mo->mot_open_sem);
841 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
842 /* Check write access for the O_TRUNC case */
843 if (mdt_write_read(mo) < 0)
844 GOTO(out_put, rc = -ETXTBSY);
846 /* LU-10286: compatibility check for FLR.
847 * Please check the comment in mdt_finish_open() for details
849 if (!exp_connect_flr(info->mti_exp) ||
850 !exp_connect_overstriping(info->mti_exp)) {
851 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
852 if (rc < 0 && rc != -ENODATA)
855 if (!exp_connect_flr(info->mti_exp)) {
857 mdt_lmm_is_flr(info->mti_big_lmm))
858 GOTO(out_put, rc = -EOPNOTSUPP);
861 if (!exp_connect_overstriping(info->mti_exp)) {
863 mdt_lmm_is_overstriping(info->mti_big_lmm))
864 GOTO(out_put, rc = -EOPNOTSUPP);
868 /* For truncate, the file size sent from client
869 * is believable, but the blocks are incorrect,
870 * which makes the block size in LSOM attribute
871 * inconsisent with the real block size.
873 rc = mdt_lsom_update(info, mo, true);
878 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
879 if (ma->ma_valid & MA_LOV)
880 GOTO(out_put, rc = -EPROTO);
882 /* MDT supports FMD for regular files due to Data-on-MDT */
883 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
884 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
885 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
888 if (ma->ma_attr.la_valid & LA_MTIME) {
889 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
891 ma->ma_valid |= MA_PFID;
895 rc = mdt_attr_set(info, mo, ma);
898 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
899 (ma->ma_valid & MA_INODE)) {
900 struct lu_buf *buf = &info->mti_buf;
901 struct lu_ucred *uc = mdt_ucred(info);
902 struct mdt_lock_handle *lh;
904 __u64 lockpart = MDS_INODELOCK_XATTR;
906 /* reject if either remote or striped dir is disabled */
907 if (ma->ma_valid & MA_LMV) {
908 if (!mdt->mdt_enable_remote_dir ||
909 !mdt->mdt_enable_striped_dir)
910 GOTO(out_put, rc = -EPERM);
912 if (!md_capable(uc, CAP_SYS_ADMIN) &&
913 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
914 mdt->mdt_enable_remote_dir_gid != -1)
915 GOTO(out_put, rc = -EPERM);
918 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
919 GOTO(out_put, rc = -ENOTDIR);
921 if (ma->ma_attr.la_valid != 0)
922 GOTO(out_put, rc = -EPROTO);
924 if (ma->ma_valid & MA_LOV) {
925 buf->lb_buf = ma->ma_lmm;
926 buf->lb_len = ma->ma_lmm_size;
927 name = XATTR_NAME_LOV;
929 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
932 buf->lb_len = ma->ma_lmv_size;
933 name = XATTR_NAME_DEFAULT_LMV;
934 /* force client to update dir default layout */
935 lockpart |= MDS_INODELOCK_LOOKUP;
938 lh = &info->mti_lh[MDT_LH_PARENT];
939 mdt_lock_reg_init(lh, LCK_PW);
941 rc = mdt_object_lock(info, mo, lh, lockpart);
945 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
948 mdt_object_unlock(info, mo, lh, rc);
952 GOTO(out_put, rc = -EPROTO);
955 /* If file data is modified, add the dirty flag */
956 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
957 rc = mdt_add_dirty_flag(info, mo, ma);
959 ma->ma_need = MA_INODE;
961 rc = mdt_attr_get_complex(info, mo, ma);
965 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
969 mdt_object_put(info->mti_env, mo);
972 mdt_counter_incr(req, LPROC_MDT_SETATTR,
973 ktime_us_delta(ktime_get(), kstart));
975 mdt_client_compatibility(info);
976 rc2 = mdt_fix_reply(info);
982 static int mdt_reint_create(struct mdt_thread_info *info,
983 struct mdt_lock_handle *lhc)
985 struct ptlrpc_request *req = mdt_info_req(info);
986 ktime_t kstart = ktime_get();
990 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
991 RETURN(err_serious(-ESTALE));
993 if (info->mti_dlm_req)
994 ldlm_request_cancel(mdt_info_req(info),
995 info->mti_dlm_req, 0, LATF_SKIP);
997 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1000 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1010 CERROR("%s: Unsupported mode %o\n",
1011 mdt_obd_name(info->mti_mdt),
1012 info->mti_attr.ma_attr.la_mode);
1013 RETURN(err_serious(-EOPNOTSUPP));
1016 rc = mdt_create(info);
1018 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1019 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1020 ktime_us_delta(ktime_get(), kstart));
1022 /* Special file should stay on the same node as parent*/
1023 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1024 ktime_us_delta(ktime_get(), kstart));
1031 * VBR: save parent version in reply and child version getting by its name.
1032 * Version of child is getting and checking during its lookup. If
1034 static int mdt_reint_unlink(struct mdt_thread_info *info,
1035 struct mdt_lock_handle *lhc)
1037 struct mdt_reint_record *rr = &info->mti_rr;
1038 struct ptlrpc_request *req = mdt_info_req(info);
1039 struct md_attr *ma = &info->mti_attr;
1040 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1041 struct mdt_object *mp;
1042 struct mdt_object *mc;
1043 struct mdt_lock_handle *parent_lh;
1044 struct mdt_lock_handle *child_lh;
1045 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1047 bool cos_incompat = false;
1049 ktime_t kstart = ktime_get();
1053 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1054 PNAME(&rr->rr_name));
1056 if (info->mti_dlm_req)
1057 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1059 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1060 RETURN(err_serious(-ENOENT));
1062 if (!fid_is_md_operative(rr->rr_fid1))
1065 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1067 RETURN(PTR_ERR(mp));
1069 if (mdt_object_remote(mp)) {
1070 cos_incompat = true;
1072 rc = mdt_version_get_check_save(info, mp, 0);
1074 GOTO(put_parent, rc);
1077 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1078 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1080 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1081 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1082 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1085 GOTO(put_parent, rc);
1087 /* lookup child object along with version checking */
1088 fid_zero(child_fid);
1089 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1091 /* Name might not be able to find during resend of
1092 * remote unlink, considering following case.
1093 * dir_A is a remote directory, the name entry of
1094 * dir_A is on MDT0, the directory is on MDT1,
1096 * 1. client sends unlink req to MDT1.
1097 * 2. MDT1 sends name delete update to MDT0.
1098 * 3. name entry is being deleted in MDT0 synchronously.
1099 * 4. MDT1 is restarted.
1100 * 5. client resends unlink req to MDT1. So it can not
1101 * find the name entry on MDT0 anymore.
1102 * In this case, MDT1 only needs to destory the local
1105 if (mdt_object_remote(mp) && rc == -ENOENT &&
1106 !fid_is_zero(rr->rr_fid2) &&
1107 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1109 *child_fid = *rr->rr_fid2;
1111 GOTO(unlock_parent, rc);
1115 if (!fid_is_md_operative(child_fid))
1116 GOTO(unlock_parent, rc = -EPERM);
1118 /* We will lock the child regardless it is local or remote. No harm. */
1119 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1121 GOTO(unlock_parent, rc = PTR_ERR(mc));
1123 if (!cos_incompat) {
1124 rc = mdt_object_striped(info, mc);
1126 GOTO(put_child, rc);
1130 mdt_object_put(info->mti_env, mc);
1131 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1136 child_lh = &info->mti_lh[MDT_LH_CHILD];
1137 mdt_lock_reg_init(child_lh, LCK_EX);
1138 if (info->mti_spec.sp_rm_entry) {
1139 struct lu_ucred *uc = mdt_ucred(info);
1141 if (!mdt_is_dne_client(req->rq_export))
1142 /* Return -ENOTSUPP for old client */
1143 GOTO(put_child, rc = -ENOTSUPP);
1145 if (!md_capable(uc, CAP_SYS_ADMIN))
1146 GOTO(put_child, rc = -EPERM);
1148 ma->ma_need = MA_INODE;
1150 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1151 NULL, &rr->rr_name, ma, no_name);
1152 GOTO(put_child, rc);
1155 if (mdt_object_remote(mc)) {
1156 struct mdt_body *repbody;
1158 if (!fid_is_zero(rr->rr_fid2)) {
1159 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1160 mdt_obd_name(info->mti_mdt),
1161 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1162 GOTO(put_child, rc = -ENOENT);
1164 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1165 mdt_obd_name(info->mti_mdt),
1166 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1168 if (!mdt_is_dne_client(req->rq_export))
1169 /* Return -ENOTSUPP for old client */
1170 GOTO(put_child, rc = -ENOTSUPP);
1172 /* Revoke the LOOKUP lock of the remote object granted by
1173 * this MDT. Since the unlink will happen on another MDT,
1174 * it will release the LOOKUP lock right away. Then What
1175 * would happen if another client try to grab the LOOKUP
1176 * lock at the same time with unlink XXX
1178 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1179 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1180 LASSERT(repbody != NULL);
1181 repbody->mbo_fid1 = *mdt_object_fid(mc);
1182 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1183 GOTO(unlock_child, rc = -EREMOTE);
1185 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1186 * this now because a running HSM restore on the child (unlink
1187 * victim) will hold the layout lock. See LU-4002.
1189 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1190 if (mdt_object_remote(mp)) {
1191 /* Enqueue lookup lock from parent MDT */
1192 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1193 &child_lh->mlh_rreg_lh,
1194 child_lh->mlh_rreg_mode,
1195 MDS_INODELOCK_LOOKUP, false);
1197 GOTO(put_child, rc);
1199 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1202 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1205 GOTO(put_child, rc);
1208 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1209 * whether need MA_LOV and MA_COOKIE.
1211 ma->ma_need = MA_INODE;
1214 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1215 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1216 /* save version when object is locked */
1217 mdt_version_get_save(info, mc, 1);
1219 mutex_lock(&mc->mot_lov_mutex);
1221 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1222 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1224 mutex_unlock(&mc->mot_lov_mutex);
1226 GOTO(unlock_child, rc);
1228 if (!lu_object_is_dying(&mc->mot_header)) {
1229 rc = mdt_attr_get_complex(info, mc, ma);
1232 } else if (mdt_dom_check_for_discard(info, mc)) {
1233 mdt_dom_discard_data(info, mc);
1235 mdt_handle_last_unlink(info, mc, ma);
1238 if (ma->ma_valid & MA_INODE) {
1239 switch (ma->ma_attr.la_mode & S_IFMT) {
1241 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1242 ktime_us_delta(ktime_get(), kstart));
1250 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1251 ktime_us_delta(ktime_get(), kstart));
1254 LASSERTF(0, "bad file type %o unlinking\n",
1255 ma->ma_attr.la_mode);
1262 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1264 mdt_object_put(info->mti_env, mc);
1266 mdt_object_unlock(info, mp, parent_lh, rc);
1268 mdt_object_put(info->mti_env, mp);
1269 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1274 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1277 static int mdt_reint_link(struct mdt_thread_info *info,
1278 struct mdt_lock_handle *lhc)
1280 struct mdt_reint_record *rr = &info->mti_rr;
1281 struct ptlrpc_request *req = mdt_info_req(info);
1282 struct md_attr *ma = &info->mti_attr;
1283 struct mdt_object *ms;
1284 struct mdt_object *mp;
1285 struct mdt_lock_handle *lhs;
1286 struct mdt_lock_handle *lhp;
1287 ktime_t kstart = ktime_get();
1292 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1293 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1295 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1296 RETURN(err_serious(-ENOENT));
1298 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1299 req->rq_no_reply = 1;
1300 RETURN(err_serious(-ENOENT));
1303 if (info->mti_dlm_req)
1304 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1306 /* Invalid case so return error immediately instead of
1309 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1312 if (!fid_is_md_operative(rr->rr_fid1) ||
1313 !fid_is_md_operative(rr->rr_fid2))
1316 /* step 1: find target parent dir */
1317 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1319 RETURN(PTR_ERR(mp));
1321 rc = mdt_version_get_check_save(info, mp, 0);
1323 GOTO(put_parent, rc);
1325 /* step 2: find source */
1326 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1328 GOTO(put_parent, rc = PTR_ERR(ms));
1330 if (!mdt_object_exists(ms)) {
1331 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1332 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1333 GOTO(put_source, rc = -ENOENT);
1336 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1338 lhp = &info->mti_lh[MDT_LH_PARENT];
1339 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1340 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1343 GOTO(put_source, rc);
1345 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1347 lhs = &info->mti_lh[MDT_LH_CHILD];
1348 mdt_lock_reg_init(lhs, LCK_EX);
1349 rc = mdt_reint_object_lock(info, ms, lhs,
1350 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1353 GOTO(unlock_parent, rc);
1355 /* step 3: link it */
1356 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1357 OBD_FAIL_MDS_REINT_LINK_WRITE);
1359 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1360 rc = mdt_version_get_check_save(info, ms, 1);
1362 GOTO(unlock_source, rc);
1364 /** check target version by name during replay */
1365 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1366 &info->mti_tmp_fid1, 2);
1367 if (rc != 0 && rc != -ENOENT)
1368 GOTO(unlock_source, rc);
1369 /* save version of file name for replay, it must be ENOENT here */
1370 if (!req_is_replay(mdt_info_req(info))) {
1371 if (rc != -ENOENT) {
1372 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1373 PNAME(&rr->rr_name));
1374 GOTO(unlock_source, rc = -EEXIST);
1376 info->mti_ver[2] = ENOENT_VERSION;
1377 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1380 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1381 mdt_object_child(ms), &rr->rr_name, ma);
1384 mdt_counter_incr(req, LPROC_MDT_LINK,
1385 ktime_us_delta(ktime_get(), kstart));
1389 mdt_object_unlock(info, ms, lhs, rc);
1391 mdt_object_unlock(info, mp, lhp, rc);
1393 mdt_object_put(info->mti_env, ms);
1395 mdt_object_put(info->mti_env, mp);
1399 * lock the part of the directory according to the hash of the name
1400 * (lh->mlh_pdo_hash) in parallel directory lock.
1402 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1403 struct mdt_lock_handle *lh,
1404 struct mdt_object *obj, __u64 ibits,
1407 struct ldlm_res_id *res = &info->mti_res_id;
1408 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1409 union ldlm_policy_data *policy = &info->mti_policy;
1410 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1414 * Finish res_id initializing by name hash marking part of
1415 * directory which is taking modification.
1417 LASSERT(lh->mlh_pdo_hash != 0);
1418 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1419 memset(policy, 0, sizeof(*policy));
1420 policy->l_inodebits.bits = ibits;
1422 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1423 dlmflags |= LDLM_FL_COS_INCOMPAT;
1425 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1426 * going to be sent to client. If it is - mdt_intent_policy() path will
1427 * fix it up and turn FL_LOCAL flag off.
1429 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1430 policy, res, dlmflags,
1431 &info->mti_exp->exp_handle.h_cookie);
1436 * Get BFL lock for rename or migrate process.
1438 static int mdt_rename_lock(struct mdt_thread_info *info,
1439 struct lustre_handle *lh)
1444 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1445 struct lu_fid *fid = &info->mti_tmp_fid1;
1446 struct mdt_object *obj;
1448 /* XXX, right now, it has to use object API to
1449 * enqueue lock cross MDT, so it will enqueue
1450 * rename lock(with LUSTRE_BFL_FID) by root object
1453 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1455 RETURN(PTR_ERR(obj));
1457 rc = mdt_remote_object_lock(info, obj,
1458 &LUSTRE_BFL_FID, lh,
1460 MDS_INODELOCK_UPDATE, false);
1461 mdt_object_put(info->mti_env, obj);
1463 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1464 union ldlm_policy_data *policy = &info->mti_policy;
1465 struct ldlm_res_id *res_id = &info->mti_res_id;
1468 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1469 memset(policy, 0, sizeof(*policy));
1470 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1471 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1472 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1473 LDLM_IBITS, policy, LCK_EX, &flags,
1475 ldlm_completion_ast, NULL, NULL, 0,
1477 &info->mti_exp->exp_handle.h_cookie,
1484 static void mdt_rename_unlock(struct lustre_handle *lh)
1487 LASSERT(lustre_handle_is_used(lh));
1488 /* Cancel the single rename lock right away */
1489 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1493 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1494 const struct lu_fid *fid,
1497 struct mdt_object *dir;
1501 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1505 /* check early, the real version will be saved after locking */
1506 rc = mdt_version_get_check(info, dir, idx);
1510 if (!mdt_object_exists(dir))
1511 GOTO(out_put, rc = -ENOENT);
1513 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1514 GOTO(out_put, rc = -ENOTDIR);
1518 mdt_object_put(info->mti_env, dir);
1523 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1524 * herein we don't really check it, just do revoke.
1526 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1527 struct mdt_object *pobj,
1528 struct mdt_object *obj)
1530 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1533 mdt_lock_handle_init(lh);
1534 mdt_lock_reg_init(lh, LCK_EX);
1536 if (mdt_object_remote(pobj)) {
1537 /* don't bother to check if pobj and obj are on the same MDT. */
1538 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1539 &lh->mlh_rreg_lh, LCK_EX,
1540 MDS_INODELOCK_LOOKUP, false);
1541 } else if (mdt_object_remote(obj)) {
1542 struct ldlm_res_id *res = &info->mti_res_id;
1543 union ldlm_policy_data *policy = &info->mti_policy;
1544 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1545 LDLM_FL_COS_INCOMPAT;
1547 fid_build_reg_res_name(mdt_object_fid(obj), res);
1548 memset(policy, 0, sizeof(*policy));
1549 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1550 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1551 &lh->mlh_reg_lh, LCK_EX, policy, res,
1554 /* do nothing if both are local */
1562 * TODO, currently we don't save this lock because there is no place to
1563 * hold this lock handle, but to avoid race we need to save this lock.
1565 mdt_object_unlock(info, NULL, lh, 1);
1571 * operation may takes locks of linkea, or directory stripes, group them in
1574 struct mdt_sub_lock {
1575 struct mdt_object *msl_obj;
1576 struct mdt_lock_handle msl_lh;
1577 struct list_head msl_linkage;
1580 static void mdt_unlock_list(struct mdt_thread_info *info,
1581 struct list_head *list, int decref)
1583 struct mdt_sub_lock *msl;
1584 struct mdt_sub_lock *tmp;
1586 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1587 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1588 list_del(&msl->msl_linkage);
1593 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1594 struct mdt_object *obj,
1595 struct mdt_lock_handle *lh,
1596 struct ldlm_enqueue_info *einfo,
1597 struct list_head *slave_locks,
1600 if (mdt_object_remote(obj)) {
1601 mdt_unlock_list(info, slave_locks, decref);
1602 mdt_object_unlock(info, obj, lh, decref);
1604 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1609 * lock parents of links, and also check whether total locks don't exceed
1612 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1613 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1614 * \retval -ev negative errno upon error
1616 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1617 struct mdt_object *pobj,
1618 const struct md_attr *ma,
1619 struct mdt_object *obj,
1620 struct mdt_lock_handle *lhp,
1621 struct ldlm_enqueue_info *peinfo,
1622 struct list_head *parent_slave_locks,
1623 struct list_head *link_locks)
1625 struct mdt_device *mdt = info->mti_mdt;
1626 struct lu_buf *buf = &info->mti_big_buf;
1627 struct lu_name *lname = &info->mti_name;
1628 struct linkea_data ldata = { NULL };
1629 bool blocked = false;
1630 int local_lnkp_cnt = 0;
1634 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1637 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1638 if (buf->lb_buf == NULL)
1642 rc = mdt_links_read(info, obj, &ldata);
1644 if (rc == -ENOENT || rc == -ENODATA)
1649 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1650 linkea_next_entry(&ldata)) {
1651 struct mdt_object *lnkp;
1652 struct mdt_sub_lock *msl;
1656 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1659 /* check if it's also linked to parent */
1660 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1661 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1662 PFID(&fid), PNAME(lname));
1663 /* in case link is remote object, revoke LOOKUP lock */
1664 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1670 /* check if it's linked to a stripe of parent */
1671 if (ma->ma_valid & MA_LMV) {
1672 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1673 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1676 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1677 fid_le_to_cpu(stripe_fid,
1678 &lmv->lmv_stripe_fids[j]);
1679 if (lu_fid_eq(stripe_fid, &fid)) {
1680 CDEBUG(D_INFO, "skip stripe "DFID
1681 ", reovke "DNAME"\n",
1682 PFID(&fid), PNAME(lname));
1683 lnkp = mdt_object_find(info->mti_env,
1686 GOTO(out, rc = PTR_ERR(lnkp));
1692 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1694 mdt_object_put(info->mti_env, lnkp);
1699 /* Check if it's already locked */
1700 list_for_each_entry(msl, link_locks, msl_linkage) {
1701 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1703 DFID" was locked, revoke "DNAME"\n",
1704 PFID(&fid), PNAME(lname));
1705 lnkp = msl->msl_obj;
1711 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1715 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1716 PFID(&fid), PNAME(lname));
1718 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1720 CWARN("%s: cannot find obj "DFID": %ld\n",
1721 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1725 if (!mdt_object_exists(lnkp)) {
1726 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1727 PFID(&fid), PNAME(lname));
1728 mdt_object_put(info->mti_env, lnkp);
1732 if (!mdt_object_remote(lnkp))
1737 GOTO(out, rc = -ENOMEM);
1740 * we can't follow parent-child lock order like other MD
1741 * operations, use lock_try here to avoid deadlock, if the lock
1742 * cannot be taken, drop all locks taken, revoke the blocked
1743 * one, and continue processing the remaining entries, and in
1744 * the end of the loop restart from beginning.
1746 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1748 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1749 MDS_INODELOCK_UPDATE, true);
1750 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1752 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1753 PFID(&fid), PNAME(lname));
1755 mdt_unlock_list(info, link_locks, 1);
1756 /* also unlock parent locks to avoid deadlock */
1758 mdt_migrate_object_unlock(info, pobj, lhp,
1765 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1766 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1767 MDS_INODELOCK_UPDATE);
1769 mdt_object_put(info->mti_env, lnkp);
1774 if (mdt_object_remote(lnkp)) {
1775 struct ldlm_lock *lock;
1778 * for remote object, set lock cb_atomic,
1779 * so lock can be released in blocking_ast()
1780 * immediately, then the next lock_try will
1781 * have better chance of success.
1783 lock = ldlm_handle2lock(
1784 &msl->msl_lh.mlh_rreg_lh);
1785 LASSERT(lock != NULL);
1786 lock_res_and_lock(lock);
1787 ldlm_set_atomic_cb(lock);
1788 unlock_res_and_lock(lock);
1789 LDLM_LOCK_PUT(lock);
1792 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1797 INIT_LIST_HEAD(&msl->msl_linkage);
1798 msl->msl_obj = lnkp;
1799 list_add_tail(&msl->msl_linkage, link_locks);
1801 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1805 GOTO(out, rc = -EBUSY);
1810 mdt_unlock_list(info, link_locks, rc);
1811 } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1812 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1815 * parent may have 3 local objects: master object and 2 stripes
1816 * (if it's being migrated too); source may have 1 local objects
1817 * as regular file; target has 1 local object.
1818 * Note, source may have 2 local locks if it is directory but it
1819 * can't have hardlinks, so it is not considered here.
1826 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1827 struct mdt_object *obj,
1828 const struct md_attr *ma,
1829 struct list_head *slave_locks)
1831 struct mdt_device *mdt = info->mti_mdt;
1832 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1833 struct lu_fid *fid = &info->mti_tmp_fid1;
1834 struct mdt_object *slave;
1835 struct mdt_sub_lock *msl;
1840 LASSERT(mdt_object_remote(obj));
1841 LASSERT(ma->ma_valid & MA_LMV);
1844 if (!lmv_is_sane(lmv))
1847 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1848 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1850 if (!fid_is_sane(fid))
1853 slave = mdt_object_find(info->mti_env, mdt, fid);
1855 GOTO(out, rc = PTR_ERR(slave));
1859 mdt_object_put(info->mti_env, slave);
1860 GOTO(out, rc = -ENOMEM);
1863 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1864 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1865 MDS_INODELOCK_UPDATE, true);
1868 mdt_object_put(info->mti_env, slave);
1872 INIT_LIST_HEAD(&msl->msl_linkage);
1873 msl->msl_obj = slave;
1874 list_add_tail(&msl->msl_linkage, slave_locks);
1880 mdt_unlock_list(info, slave_locks, rc);
1884 /* lock parent and its stripes */
1885 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1886 struct mdt_object *obj,
1887 const struct md_attr *ma,
1888 struct mdt_lock_handle *lh,
1889 struct ldlm_enqueue_info *einfo,
1890 struct list_head *slave_locks)
1894 if (mdt_object_remote(obj)) {
1895 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1896 &lh->mlh_rreg_lh, LCK_PW,
1897 MDS_INODELOCK_UPDATE, false);
1902 * if obj is remote and striped, lock its stripes explicitly
1903 * because it's not striped in LOD layer on this MDT.
1905 if (ma->ma_valid & MA_LMV) {
1906 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1908 mdt_object_unlock(info, obj, lh, rc);
1911 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1919 * in migration, object may be remote, and we need take full lock of it and its
1920 * stripes if it's directory, besides, object may be a remote object on its
1921 * parent, revoke its LOOKUP lock on where its parent is located.
1923 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1924 struct mdt_object *pobj,
1925 struct mdt_object *obj,
1926 struct mdt_lock_handle *lh,
1927 struct ldlm_enqueue_info *einfo,
1928 struct list_head *slave_locks)
1932 if (mdt_object_remote(obj)) {
1933 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1937 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1938 &lh->mlh_rreg_lh, LCK_EX,
1939 MDS_INODELOCK_FULL, false);
1944 * if obj is remote and striped, lock its stripes explicitly
1945 * because it's not striped in LOD layer on this MDT.
1947 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1948 struct md_attr *ma = &info->mti_attr;
1950 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1952 mdt_object_unlock(info, obj, lh, rc);
1956 if (ma->ma_valid & MA_LMV) {
1957 rc = mdt_lock_remote_slaves(info, obj, ma,
1960 mdt_object_unlock(info, obj, lh, rc);
1964 if (mdt_object_remote(pobj)) {
1965 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1970 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1978 * lookup source by name, if parent is striped directory, we need to find the
1979 * corresponding stripe where source is located, and then lookup there.
1981 * besides, if parent is migrating too, and file is already in target stripe,
1982 * this should be a redo of 'lfs migrate' on client side.
1984 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1985 struct mdt_object *pobj,
1986 const struct md_attr *ma,
1987 const struct lu_name *lname,
1988 struct mdt_object **spobj,
1989 struct mdt_object **sobj)
1991 const struct lu_env *env = info->mti_env;
1992 struct lu_fid *fid = &info->mti_tmp_fid1;
1993 struct mdt_object *stripe;
1996 if (ma->ma_valid & MA_LMV) {
1997 /* if parent is striped, lookup on corresponding stripe */
1998 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2000 if (!lmv_is_sane(lmv))
2003 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2008 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2010 stripe = mdt_object_find(env, info->mti_mdt, fid);
2012 return PTR_ERR(stripe);
2015 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2017 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2019 * if parent layout is changeing, and lookup child
2020 * failed on source stripe, lookup again on target
2021 * stripe, if it exists, it means previous migration
2022 * was interrupted, and current file was migrated
2025 mdt_object_put(env, stripe);
2027 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2032 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2034 stripe = mdt_object_find(env, info->mti_mdt, fid);
2036 return PTR_ERR(stripe);
2039 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2040 fid, &info->mti_spec);
2041 mdt_object_put(env, stripe);
2042 return rc ?: -EALREADY;
2044 mdt_object_put(env, stripe);
2049 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2055 mdt_object_get(env, stripe);
2060 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2061 if (IS_ERR(*sobj)) {
2062 mdt_object_put(env, stripe);
2063 rc = PTR_ERR(*sobj);
2071 /* end lease and close file for regular file */
2072 static int mdd_migrate_close(struct mdt_thread_info *info,
2073 struct mdt_object *obj)
2075 struct close_data *data;
2076 struct mdt_body *repbody;
2077 struct ldlm_lock *lease;
2082 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2084 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2088 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2093 lease = ldlm_handle2lock(&data->cd_handle);
2097 /* check if the lease was already canceled */
2098 lock_res_and_lock(lease);
2099 rc = ldlm_is_cancel(lease);
2100 unlock_res_and_lock(lease);
2104 LDLM_DEBUG(lease, DFID" lease broken",
2105 PFID(mdt_object_fid(obj)));
2109 * cancel server side lease, client side counterpart should have been
2110 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2112 ldlm_lock_cancel(lease);
2113 ldlm_reprocess_all(lease->l_resource, lease);
2114 LDLM_LOCK_PUT(lease);
2117 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2118 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2119 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2125 * migrate file in below steps:
2126 * 1. lock parent and its stripes
2127 * 2. lookup source by name
2128 * 3. lock parents of source links if source is not directory
2129 * 4. reject if source is in HSM
2130 * 5. take source open_sem and close file if source is regular file
2131 * 6. lock source and its stripes if it's directory
2132 * 7. lock target so subsequent change to it can trigger COS
2134 * 9. unlock above locks
2135 * 10. sync device if source has links
2137 int mdt_reint_migrate(struct mdt_thread_info *info,
2138 struct mdt_lock_handle *unused)
2140 const struct lu_env *env = info->mti_env;
2141 struct mdt_device *mdt = info->mti_mdt;
2142 struct ptlrpc_request *req = mdt_info_req(info);
2143 struct mdt_reint_record *rr = &info->mti_rr;
2144 struct lu_ucred *uc = mdt_ucred(info);
2145 struct md_attr *ma = &info->mti_attr;
2146 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2147 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2148 struct mdt_object *pobj;
2149 struct mdt_object *spobj = NULL;
2150 struct mdt_object *sobj = NULL;
2151 struct mdt_object *tobj;
2152 struct lustre_handle rename_lh = { 0 };
2153 struct mdt_lock_handle *lhp;
2154 struct mdt_lock_handle *lhs;
2155 struct mdt_lock_handle *lht;
2156 LIST_HEAD(parent_slave_locks);
2157 LIST_HEAD(child_slave_locks);
2158 LIST_HEAD(link_locks);
2159 int lock_retries = 5;
2160 bool open_sem_locked = false;
2161 bool do_sync = false;
2165 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2166 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2168 if (info->mti_dlm_req)
2169 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2171 if (!fid_is_md_operative(rr->rr_fid1) ||
2172 !fid_is_md_operative(rr->rr_fid2))
2175 /* don't allow migrate . or .. */
2176 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2179 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2182 if (uc && !md_capable(uc, CAP_SYS_ADMIN) &&
2183 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2184 mdt->mdt_enable_remote_dir_gid != -1)
2188 * Note: do not enqueue rename lock for replay request, because
2189 * if other MDT holds rename lock, but being blocked to wait for
2190 * this MDT to finish its recovery, and the failover MDT can not
2191 * get rename lock, which will cause deadlock.
2193 * req is NULL if this is called by directory auto-split.
2195 if (req && !req_is_replay(req)) {
2196 rc = mdt_rename_lock(info, &rename_lh);
2198 CERROR("%s: can't lock FS for rename: rc = %d\n",
2199 mdt_obd_name(info->mti_mdt), rc);
2204 /* pobj is master object of parent */
2205 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2207 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2210 rc = mdt_version_get_check(info, pobj, 0);
2212 GOTO(put_parent, rc);
2215 if (!mdt_object_exists(pobj))
2216 GOTO(put_parent, rc = -ENOENT);
2218 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2219 GOTO(put_parent, rc = -ENOTDIR);
2221 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2223 GOTO(put_parent, rc);
2226 /* lock parent object */
2227 lhp = &info->mti_lh[MDT_LH_PARENT];
2228 mdt_lock_reg_init(lhp, LCK_PW);
2229 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2230 &parent_slave_locks);
2232 GOTO(put_parent, rc);
2235 * spobj is the corresponding stripe against name if pobj is striped
2236 * directory, which is the real parent, and no need to lock, because
2237 * we've taken full lock of pobj.
2239 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2241 GOTO(unlock_parent, rc);
2243 /* lock parents of source links, and revoke LOOKUP lock of links */
2244 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2245 &parent_slave_locks, &link_locks);
2246 if (rc == -EBUSY && lock_retries-- > 0) {
2247 mdt_object_put(env, sobj);
2248 mdt_object_put(env, spobj);
2253 GOTO(put_source, rc);
2256 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2257 * with one request, if total lock count exceeds this limit, we will
2258 * drop all locks after migration, and synchronous device in the end.
2262 /* TODO: DoM migration is not supported, migrate dirent only */
2263 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2264 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2266 GOTO(unlock_links, rc);
2268 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2269 info->mti_spec.sp_migrate_nsonly = 1;
2272 /* if migration HSM is allowed */
2273 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2274 ma->ma_need = MA_HSM;
2276 rc = mdt_attr_get_complex(info, sobj, ma);
2278 GOTO(unlock_links, rc);
2280 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2281 GOTO(unlock_links, rc = -EOPNOTSUPP);
2284 /* end lease and close file for regular file */
2285 if (info->mti_spec.sp_migrate_close) {
2286 /* try to hold open_sem so that nobody else can open the file */
2287 if (!down_write_trylock(&sobj->mot_open_sem)) {
2289 mdd_migrate_close(info, sobj);
2290 GOTO(unlock_links, rc = -EBUSY);
2292 open_sem_locked = true;
2293 rc = mdd_migrate_close(info, sobj);
2295 GOTO(unlock_open_sem, rc);
2300 lhs = &info->mti_lh[MDT_LH_OLD];
2301 mdt_lock_reg_init(lhs, LCK_EX);
2302 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2303 &child_slave_locks);
2305 GOTO(unlock_open_sem, rc);
2308 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2310 GOTO(unlock_source, rc = PTR_ERR(tobj));
2312 lht = &info->mti_lh[MDT_LH_NEW];
2313 mdt_lock_reg_init(lht, LCK_EX);
2314 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2316 GOTO(put_target, rc);
2318 /* Don't do lookup sanity check. We know name doesn't exist. */
2319 info->mti_spec.sp_cr_lookup = 0;
2320 info->mti_spec.sp_feat = &dt_directory_features;
2322 rc = mdo_migrate(env, mdt_object_child(pobj),
2323 mdt_object_child(sobj), &rr->rr_name,
2324 mdt_object_child(tobj),
2325 &info->mti_spec, ma);
2327 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2328 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2331 mdt_object_unlock(info, tobj, lht, rc);
2333 mdt_object_put(env, tobj);
2335 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2336 &child_slave_locks, rc);
2338 if (open_sem_locked)
2339 up_write(&sobj->mot_open_sem);
2341 /* if we've got too many locks to save into RPC,
2342 * then just commit before the locks are released
2345 mdt_device_sync(env, mdt);
2346 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2348 mdt_object_put(env, sobj);
2349 mdt_object_put(env, spobj);
2351 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2352 &parent_slave_locks, rc);
2354 mdt_object_put(env, pobj);
2356 if (lustre_handle_is_used(&rename_lh))
2357 mdt_rename_unlock(&rename_lh);
2362 static int mdt_object_lock_save(struct mdt_thread_info *info,
2363 struct mdt_object *dir,
2364 struct mdt_lock_handle *lh,
2365 int idx, bool cos_incompat)
2369 /* we lock the target dir if it is local */
2370 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2375 /* get and save correct version after locking */
2376 mdt_version_get_save(info, dir, idx);
2381 * determine lock order of sobj and tobj
2383 * there are two situations we need to lock tobj before sobj:
2384 * 1. sobj is child of tobj
2385 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2386 * larger than that of tobj
2388 * \retval 1 lock tobj before sobj
2389 * \retval 0 lock sobj before tobj
2390 * \retval -ev negative errno upon error
2392 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2393 struct mdt_object *sobj,
2394 struct mdt_object *tobj)
2396 struct md_attr *ma = &info->mti_attr;
2397 struct lu_fid *spfid = &info->mti_tmp_fid1;
2398 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2399 struct lmv_mds_md_v1 *lmv;
2404 /* sobj and tobj are the same */
2408 if (fid_is_root(mdt_object_fid(sobj)))
2411 if (fid_is_root(mdt_object_fid(tobj)))
2414 /* check whether sobj is child of tobj */
2415 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2416 mdt_object_fid(tobj));
2423 /* check whether sobj and tobj are children of the same parent */
2424 rc = mdt_attr_get_pfid(info, sobj, spfid);
2428 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2432 if (!lu_fid_eq(spfid, tpfid))
2435 /* check whether sobj and tobj are sibling stripes */
2436 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2440 if (!(ma->ma_valid & MA_LMV))
2443 lmv = &ma->ma_lmv->lmv_md_v1;
2444 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2446 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2449 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2453 if (!(ma->ma_valid & MA_LMV))
2456 lmv = &ma->ma_lmv->lmv_md_v1;
2457 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2459 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2461 /* check stripe index of sobj and tobj */
2462 if (sindex == tindex)
2465 return sindex < tindex ? 0 : 1;
2469 * lock rename source object.
2471 * Both source and source parent may be remote, and source may be a remote
2472 * object on source parent, to avoid overriding lock handle, store remote
2473 * LOOKUP lock separately in @lhr.
2475 * \retval 0 on success
2476 * \retval -ev negative errno upon error
2478 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2479 struct mdt_object *parent,
2480 struct mdt_object *child,
2481 struct mdt_lock_handle *lhc,
2482 struct mdt_lock_handle *lhr,
2488 rc = mdt_is_remote_object(info, parent, child);
2493 /* enqueue remote LOOKUP lock from the parent MDT */
2494 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2496 if (mdt_object_remote(parent)) {
2497 rc = mdt_remote_object_lock(info, parent,
2498 mdt_object_fid(child),
2505 LASSERT(mdt_object_remote(child));
2506 rc = mdt_object_local_lock(info, child, lhr,
2507 &rmt_ibits, 0, true);
2512 ibits &= ~MDS_INODELOCK_LOOKUP;
2515 if (mdt_object_remote(child)) {
2516 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2523 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2528 mdt_object_unlock(info, child, lhr, rc);
2534 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2535 * 2 - srcdir child; 3 - tgtdir child.
2536 * Update on disk version of srcdir child.
2538 static int mdt_reint_rename(struct mdt_thread_info *info,
2539 struct mdt_lock_handle *unused)
2541 struct mdt_device *mdt = info->mti_mdt;
2542 struct mdt_reint_record *rr = &info->mti_rr;
2543 struct md_attr *ma = &info->mti_attr;
2544 struct ptlrpc_request *req = mdt_info_req(info);
2545 struct mdt_object *msrcdir = NULL;
2546 struct mdt_object *mtgtdir = NULL;
2547 struct mdt_object *mold;
2548 struct mdt_object *mnew = NULL;
2549 struct lustre_handle rename_lh = { 0 };
2550 struct mdt_lock_handle *lh_srcdirp;
2551 struct mdt_lock_handle *lh_tgtdirp;
2552 struct mdt_lock_handle *lh_oldp = NULL;
2553 struct mdt_lock_handle *lh_rmt = NULL;
2554 struct mdt_lock_handle *lh_newp = NULL;
2555 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2556 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2558 bool reverse = false, discard = false;
2560 ktime_t kstart = ktime_get();
2564 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2565 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2566 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2568 if (info->mti_dlm_req)
2569 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2571 if (!fid_is_md_operative(rr->rr_fid1) ||
2572 !fid_is_md_operative(rr->rr_fid2))
2575 /* find both parents. */
2576 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2577 if (IS_ERR(msrcdir))
2578 RETURN(PTR_ERR(msrcdir));
2580 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2582 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2584 mdt_object_get(info->mti_env, mtgtdir);
2586 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2587 if (IS_ERR(mtgtdir))
2588 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2592 * Note: do not enqueue rename lock for replay request, because
2593 * if other MDT holds rename lock, but being blocked to wait for
2594 * this MDT to finish its recovery, and the failover MDT can not
2595 * get rename lock, which will cause deadlock.
2597 if (!req_is_replay(req)) {
2599 * Normally rename RPC is handled on the MDT with the target
2600 * directory (if target exists, it's on the MDT with the
2601 * target), if the source directory is remote, it's a hint that
2602 * source is remote too (this may not be true, but it won't
2603 * cause any issue), return -EXDEV early to avoid taking
2606 if (!mdt->mdt_enable_remote_rename &&
2607 mdt_object_remote(msrcdir))
2608 GOTO(out_put_tgtdir, rc = -EXDEV);
2610 /* This might be further relaxed in the future for regular file
2611 * renames in different source and target parents. Start with
2612 * only same-directory renames for simplicity and because this
2613 * is by far the most the common use case.
2615 if (msrcdir != mtgtdir) {
2616 rc = mdt_rename_lock(info, &rename_lh);
2618 CERROR("%s: cannot lock for rename: rc = %d\n",
2619 mdt_obd_name(mdt), rc);
2620 GOTO(out_put_tgtdir, rc);
2623 CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2624 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2625 PNAME(&rr->rr_name));
2629 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2631 GOTO(out_unlock_rename, rc);
2635 /* source needs to be looked up after locking source parent, otherwise
2636 * this rename may race with unlink source, and cause rename hang, see
2637 * sanityn.sh 55b, so check parents first, if later we found source is
2638 * remote, relock parents.
2640 cos_incompat = (mdt_object_remote(msrcdir) ||
2641 mdt_object_remote(mtgtdir));
2643 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2645 /* lock parents in the proper order. */
2646 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2647 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2649 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2650 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2652 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2653 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2656 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2659 GOTO(out_unlock_rename, rc);
2661 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2663 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2666 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2667 GOTO(out_unlock_rename, rc);
2670 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2673 GOTO(out_unlock_rename, rc);
2675 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2677 if (mtgtdir != msrcdir) {
2678 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2680 } else if (!mdt_object_remote(mtgtdir) &&
2681 lh_srcdirp->mlh_pdo_hash !=
2682 lh_tgtdirp->mlh_pdo_hash) {
2683 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2684 MDS_INODELOCK_UPDATE,
2686 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2689 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2690 GOTO(out_unlock_rename, rc);
2694 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2695 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2697 /* find mold object. */
2699 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2701 GOTO(out_unlock_parents, rc);
2703 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2704 GOTO(out_unlock_parents, rc = -EINVAL);
2706 if (!fid_is_md_operative(old_fid))
2707 GOTO(out_unlock_parents, rc = -EPERM);
2709 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2711 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2713 if (!mdt_object_exists(mold)) {
2714 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2716 "object does not exist");
2717 GOTO(out_put_old, rc = -ENOENT);
2720 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2721 GOTO(out_put_old, rc = -EXDEV);
2723 /* Check if @mtgtdir is subdir of @mold, before locking child
2724 * to avoid reverse locking.
2726 if (mtgtdir != msrcdir) {
2727 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2732 GOTO(out_put_old, rc);
2736 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2737 /* save version after locking */
2738 mdt_version_get_save(info, mold, 2);
2740 if (!cos_incompat && mdt_object_remote(mold)) {
2741 cos_incompat = true;
2742 mdt_object_put(info->mti_env, mold);
2743 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2744 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2748 /* find mnew object:
2749 * mnew target object may not exist now
2750 * lookup with version checking
2753 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2756 /* the new_fid should have been filled at this moment */
2757 if (lu_fid_eq(old_fid, new_fid))
2758 GOTO(out_put_old, rc);
2760 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2761 lu_fid_eq(new_fid, rr->rr_fid2))
2762 GOTO(out_put_old, rc = -EINVAL);
2764 if (!fid_is_md_operative(new_fid))
2765 GOTO(out_put_old, rc = -EPERM);
2767 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2769 GOTO(out_put_old, rc = PTR_ERR(mnew));
2771 if (!mdt_object_exists(mnew)) {
2772 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2774 "object does not exist");
2775 GOTO(out_put_new, rc = -ENOENT);
2778 if (mdt_object_remote(mnew)) {
2779 struct mdt_body *repbody;
2781 /* Always send rename req to the target child MDT */
2782 repbody = req_capsule_server_get(info->mti_pill,
2784 LASSERT(repbody != NULL);
2785 repbody->mbo_fid1 = *new_fid;
2786 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2787 GOTO(out_put_new, rc = -EXDEV);
2789 /* Before locking the target dir, check we do not replace
2790 * a dir with a non-dir, otherwise it may deadlock with
2791 * link op which tries to create a link in this dir
2792 * back to this non-dir.
2794 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2795 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2796 GOTO(out_put_new, rc = -EISDIR);
2798 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2799 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2800 mdt_lock_reg_init(lh_oldp, LCK_EX);
2801 mdt_lock_reg_init(lh_rmt, LCK_EX);
2802 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2803 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2804 lh_rmt, lock_ibits, cos_incompat);
2806 GOTO(out_put_new, rc);
2808 /* Check if @msrcdir is subdir of @mnew, before locking child
2809 * to avoid reverse locking.
2811 if (mtgtdir != msrcdir) {
2812 rc = mdo_is_subdir(info->mti_env,
2813 mdt_object_child(msrcdir), new_fid);
2817 GOTO(out_unlock_old, rc);
2821 /* We used to acquire MDS_INODELOCK_FULL here but we
2822 * can't do this now because a running HSM restore on
2823 * the rename onto victim will hold the layout
2824 * lock. See LU-4002.
2827 lh_newp = &info->mti_lh[MDT_LH_NEW];
2828 mdt_lock_reg_init(lh_newp, LCK_EX);
2829 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2830 if (mdt_object_remote(mtgtdir)) {
2831 rc = mdt_remote_object_lock(info, mtgtdir,
2832 mdt_object_fid(mnew),
2833 &lh_newp->mlh_rreg_lh,
2834 lh_newp->mlh_rreg_mode,
2835 MDS_INODELOCK_LOOKUP,
2838 GOTO(out_unlock_old, rc);
2840 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2842 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2845 GOTO(out_unlock_new, rc);
2847 /* get and save version after locking */
2848 mdt_version_get_save(info, mnew, 3);
2849 } else if (rc != -ENOENT) {
2850 GOTO(out_put_old, rc);
2852 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2853 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2854 mdt_lock_reg_init(lh_oldp, LCK_EX);
2855 mdt_lock_reg_init(lh_rmt, LCK_EX);
2856 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2857 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2858 lh_rmt, lock_ibits, cos_incompat);
2860 GOTO(out_put_old, rc);
2862 mdt_enoent_version_save(info, 3);
2865 /* step 5: rename it */
2866 mdt_reint_init_ma(info, ma);
2868 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2869 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2872 mutex_lock(&mnew->mot_lov_mutex);
2874 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2875 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2876 mnew != NULL ? mdt_object_child(mnew) : NULL,
2877 &rr->rr_tgt_name, ma);
2880 mutex_unlock(&mnew->mot_lov_mutex);
2882 /* handle last link of tgt object */
2884 mdt_counter_incr(req, LPROC_MDT_RENAME,
2885 ktime_us_delta(ktime_get(), kstart));
2887 mdt_handle_last_unlink(info, mnew, ma);
2888 discard = mdt_dom_check_for_discard(info, mnew);
2890 mdt_rename_counter_tally(info, info->mti_mdt, req,
2892 ktime_us_delta(ktime_get(), kstart));
2898 mdt_object_unlock(info, mnew, lh_newp, rc);
2900 mdt_object_unlock(info, NULL, lh_rmt, rc);
2901 mdt_object_unlock(info, mold, lh_oldp, rc);
2903 if (mnew && !discard)
2904 mdt_object_put(info->mti_env, mnew);
2906 mdt_object_put(info->mti_env, mold);
2908 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2909 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2911 if (lustre_handle_is_used(&rename_lh))
2912 mdt_rename_unlock(&rename_lh);
2914 mdt_object_put(info->mti_env, mtgtdir);
2916 mdt_object_put(info->mti_env, msrcdir);
2918 /* The DoM discard can be done right in the place above where it is
2919 * assigned, meanwhile it is done here after rename unlock due to
2920 * compatibility with old clients, for them the discard blocks
2921 * the main thread until completion. Check LU-11359 for details.
2924 mdt_dom_discard_data(info, mnew);
2925 mdt_object_put(info->mti_env, mnew);
2930 static int mdt_reint_resync(struct mdt_thread_info *info,
2931 struct mdt_lock_handle *lhc)
2933 struct mdt_reint_record *rr = &info->mti_rr;
2934 struct ptlrpc_request *req = mdt_info_req(info);
2935 struct md_attr *ma = &info->mti_attr;
2936 struct mdt_object *mo;
2937 struct ldlm_lock *lease;
2938 struct mdt_body *repbody;
2939 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2944 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2946 if (info->mti_dlm_req)
2947 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2949 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2951 GOTO(out, rc = PTR_ERR(mo));
2953 if (!mdt_object_exists(mo))
2954 GOTO(out_obj, rc = -ENOENT);
2956 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2957 GOTO(out_obj, rc = -EINVAL);
2959 if (mdt_object_remote(mo))
2960 GOTO(out_obj, rc = -EREMOTE);
2962 lease = ldlm_handle2lock(rr->rr_lease_handle);
2964 GOTO(out_obj, rc = -ESTALE);
2966 /* It's really necessary to grab open_sem and check if the lease lock
2967 * has been lost. There would exist a concurrent writer coming in and
2968 * generating some dirty data in memory cache, the writeback would fail
2969 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2971 if (!down_write_trylock(&mo->mot_open_sem))
2972 GOTO(out_put_lease, rc = -EBUSY);
2974 lock_res_and_lock(lease);
2975 lease_broken = ldlm_is_cancel(lease);
2976 unlock_res_and_lock(lease);
2978 GOTO(out_unlock, rc = -EBUSY);
2980 /* the file has yet opened by anyone else after we took the lease. */
2981 layout.mlc_opc = MD_LAYOUT_RESYNC;
2982 lhc = &info->mti_lh[MDT_LH_LOCAL];
2983 rc = mdt_layout_change(info, mo, lhc, &layout);
2985 GOTO(out_unlock, rc);
2987 mdt_object_unlock(info, mo, lhc, 0);
2989 ma->ma_need = MA_INODE;
2991 rc = mdt_attr_get_complex(info, mo, ma);
2993 GOTO(out_unlock, rc);
2995 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2996 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3000 up_write(&mo->mot_open_sem);
3002 LDLM_LOCK_PUT(lease);
3004 mdt_object_put(info->mti_env, mo);
3006 mdt_client_compatibility(info);
3007 rc2 = mdt_fix_reply(info);
3013 struct mdt_reinter {
3014 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3015 enum lprocfs_extra_opc mr_extra_opc;
3018 static const struct mdt_reinter mdt_reinters[] = {
3020 .mr_handler = &mdt_reint_setattr,
3021 .mr_extra_opc = MDS_REINT_SETATTR,
3024 .mr_handler = &mdt_reint_create,
3025 .mr_extra_opc = MDS_REINT_CREATE,
3028 .mr_handler = &mdt_reint_link,
3029 .mr_extra_opc = MDS_REINT_LINK,
3032 .mr_handler = &mdt_reint_unlink,
3033 .mr_extra_opc = MDS_REINT_UNLINK,
3036 .mr_handler = &mdt_reint_rename,
3037 .mr_extra_opc = MDS_REINT_RENAME,
3040 .mr_handler = &mdt_reint_open,
3041 .mr_extra_opc = MDS_REINT_OPEN,
3043 [REINT_SETXATTR] = {
3044 .mr_handler = &mdt_reint_setxattr,
3045 .mr_extra_opc = MDS_REINT_SETXATTR,
3048 .mr_handler = &mdt_reint_unlink,
3049 .mr_extra_opc = MDS_REINT_UNLINK,
3052 .mr_handler = &mdt_reint_migrate,
3053 .mr_extra_opc = MDS_REINT_RENAME,
3056 .mr_handler = &mdt_reint_resync,
3057 .mr_extra_opc = MDS_REINT_RESYNC,
3061 int mdt_reint_rec(struct mdt_thread_info *info,
3062 struct mdt_lock_handle *lhc)
3064 const struct mdt_reinter *mr;
3068 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3071 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3072 if (mr->mr_handler == NULL)
3075 rc = (*mr->mr_handler)(info, lhc);
3077 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3078 PTLRPC_LAST_CNTR + mr->mr_extra_opc);