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 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;
210 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
211 if (likely(!IS_ERR(child))) {
212 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
213 mdt_object_put(info->mti_env, child);
216 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
217 return vbrc ? vbrc : rc;
221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
222 struct mdt_object *obj,
223 struct ldlm_enqueue_info *einfo,
226 union ldlm_policy_data *policy = &mti->mti_policy;
227 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
228 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
231 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
232 LASSERT(slave_locks);
234 memset(policy, 0, sizeof(*policy));
235 policy->l_inodebits.bits = einfo->ei_inodebits;
236 mdt_lock_handle_init(lh);
237 mdt_lock_reg_init(lh, einfo->ei_mode);
238 for (i = 0; i < slave_locks->ha_count; i++) {
239 if (test_bit(i, (void *)slave_locks->ha_map))
240 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
242 lh->mlh_reg_lh = slave_locks->ha_handles[i];
243 mdt_object_unlock(mti, NULL, lh, decref);
244 slave_locks->ha_handles[i].cookie = 0ull;
247 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
252 struct mdt_object *obj)
254 struct lu_device *bottom_dev;
255 struct lu_object *bottom_obj;
258 if (!S_ISDIR(obj->mot_header.loh_attr))
261 /* getxattr from bottom obj to avoid reading in shard FIDs */
262 bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
263 bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
264 mdt_object_fid(obj), NULL);
265 if (IS_ERR(bottom_obj))
266 return PTR_ERR(bottom_obj);
268 rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
270 lu_object_put(mti->mti_env, bottom_obj);
272 return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
276 * Lock slave stripes if necessary, the lock handles of slave stripes
277 * will be stored in einfo->ei_cbdata.
279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
280 enum ldlm_mode mode, __u64 ibits,
281 struct ldlm_enqueue_info *einfo)
283 union ldlm_policy_data *policy = &mti->mti_policy;
285 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
287 einfo->ei_type = LDLM_IBITS;
288 einfo->ei_mode = mode;
289 einfo->ei_cb_bl = mdt_remote_blocking_ast;
290 einfo->ei_cb_local_bl = mdt_blocking_ast;
291 einfo->ei_cb_cp = ldlm_completion_ast;
292 einfo->ei_enq_slave = 1;
293 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
294 einfo->ei_inodebits = ibits;
295 memset(policy, 0, sizeof(*policy));
296 policy->l_inodebits.bits = ibits;
298 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
303 struct mdt_object *o,
304 struct mdt_lock_handle *lh,
306 struct ldlm_enqueue_info *einfo,
311 LASSERT(!mdt_object_remote(o));
313 memset(einfo, 0, sizeof(*einfo));
315 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
319 rc = mdt_object_striped(info, o);
322 mdt_object_unlock(info, o, lh, rc);
326 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
328 mdt_object_unlock(info, o, lh, rc);
329 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
337 struct mdt_object *o,
338 struct mdt_lock_handle *lh,
339 struct ldlm_enqueue_info *einfo, int decref)
341 if (einfo->ei_cbdata)
342 mdt_unlock_slaves(info, o, einfo, decref);
343 mdt_object_unlock(info, o, lh, decref);
346 static int mdt_restripe(struct mdt_thread_info *info,
347 struct mdt_object *parent,
348 const struct lu_name *lname,
349 const struct lu_fid *tfid,
350 struct md_op_spec *spec,
353 struct mdt_device *mdt = info->mti_mdt;
354 struct lu_fid *fid = &info->mti_tmp_fid2;
355 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
356 struct lmv_mds_md_v1 *lmv;
357 struct mdt_object *child;
358 struct mdt_lock_handle *lhp;
359 struct mdt_lock_handle *lhc;
360 struct mdt_body *repbody;
364 if (!mdt->mdt_enable_dir_restripe)
367 rc = mdt_version_get_check_save(info, parent, 0);
371 lhp = &info->mti_lh[MDT_LH_PARENT];
372 mdt_lock_pdo_init(lhp, LCK_PW, lname);
373 rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
378 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
380 GOTO(unlock_parent, rc);
382 if (ma->ma_valid & MA_LMV) {
383 /* don't allow restripe if parent dir layout is changing */
384 lmv = &ma->ma_lmv->lmv_md_v1;
385 if (!lmv_is_sane(lmv))
386 GOTO(unlock_parent, rc = -EBADF);
388 if (lmv_is_layout_changing(lmv))
389 GOTO(unlock_parent, rc = -EBUSY);
393 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
395 GOTO(unlock_parent, rc);
397 child = mdt_object_find(info->mti_env, mdt, fid);
399 GOTO(unlock_parent, rc = PTR_ERR(child));
401 if (!mdt_object_exists(child))
402 GOTO(out_child, rc = -ENOENT);
404 if (mdt_object_remote(child)) {
405 struct mdt_body *repbody;
407 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
409 GOTO(out_child, rc = -EPROTO);
411 repbody->mbo_fid1 = *fid;
412 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
413 GOTO(out_child, rc = -EREMOTE);
417 lhc = &info->mti_lh[MDT_LH_CHILD];
418 mdt_lock_reg_init(lhc, LCK_EX);
420 /* enqueue object remote LOOKUP lock */
421 if (mdt_object_remote(parent)) {
422 rc = mdt_remote_object_lock(info, parent, fid,
425 MDS_INODELOCK_LOOKUP, false);
430 rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
433 GOTO(unlock_child, rc);
435 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
436 rc = mdt_version_get_check_save(info, child, 1);
438 GOTO(unlock_child, rc);
440 spin_lock(&mdt->mdt_restriper.mdr_lock);
441 if (child->mot_restriping) {
443 spin_unlock(&mdt->mdt_restriper.mdr_lock);
444 GOTO(unlock_child, rc = -EBUSY);
446 child->mot_restriping = 1;
447 spin_unlock(&mdt->mdt_restriper.mdr_lock);
450 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
452 GOTO(restriping_clear, rc);
454 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
456 GOTO(restriping_clear, rc = -EPROTO);
458 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
462 child->mot_restriping = 0;
464 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
466 mdt_object_put(info->mti_env, child);
468 mdt_object_unlock(info, parent, lhp, rc);
474 * VBR: we save three versions in reply:
475 * 0 - parent. Check that parent version is the same during replay.
476 * 1 - name. Version of 'name' if file exists with the same name or
477 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
478 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
481 static int mdt_create(struct mdt_thread_info *info)
483 struct mdt_device *mdt = info->mti_mdt;
484 struct mdt_object *parent;
485 struct mdt_object *child;
486 struct mdt_lock_handle *lh;
487 struct mdt_body *repbody;
488 struct md_attr *ma = &info->mti_attr;
489 struct mdt_reint_record *rr = &info->mti_rr;
490 struct md_op_spec *spec = &info->mti_spec;
491 bool restripe = false;
495 DEBUG_REQ(D_INODE, mdt_info_req(info),
496 "Create ("DNAME"->"DFID") in "DFID,
497 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
499 if (!fid_is_md_operative(rr->rr_fid1))
502 if (S_ISDIR(ma->ma_attr.la_mode) &&
503 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
504 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
505 struct lu_ucred *uc = mdt_ucred(info);
506 struct obd_export *exp = mdt_info_req(info)->rq_export;
508 /* Only new clients can create remote dir( >= 2.4) and
509 * striped dir(>= 2.6), old client will return -ENOTSUPP
511 if (!mdt_is_dne_client(exp))
514 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
515 if (!mdt_is_striped_client(exp))
518 if (!mdt->mdt_enable_striped_dir)
520 } else if (!mdt->mdt_enable_remote_dir) {
524 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
525 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
529 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
530 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
531 mdt->mdt_enable_remote_dir_gid != -1)
534 /* restripe if later found dir exists */
535 if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
539 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
541 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
543 RETURN(PTR_ERR(parent));
545 if (!mdt_object_exists(parent))
546 GOTO(put_parent, rc = -ENOENT);
549 * LU-10235: check if name exists locklessly first to avoid massive
550 * lock recalls on existing directories.
552 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
553 &info->mti_tmp_fid1, 1);
556 GOTO(put_parent, rc = -EEXIST);
558 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
562 /* -ENOENT is expected here */
564 GOTO(put_parent, rc);
566 /* save version of file name for replay, it must be ENOENT here */
567 mdt_enoent_version_save(info, 1);
569 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
571 lh = &info->mti_lh[MDT_LH_PARENT];
572 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
573 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
575 GOTO(put_parent, rc);
577 if (!mdt_object_remote(parent)) {
578 rc = mdt_version_get_check_save(info, parent, 0);
580 GOTO(unlock_parent, rc);
583 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
584 if (unlikely(IS_ERR(child)))
585 GOTO(unlock_parent, rc = PTR_ERR(child));
587 ma->ma_need = MA_INODE;
590 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
591 OBD_FAIL_MDS_REINT_CREATE_WRITE);
593 /* Version of child will be updated on disk. */
594 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
595 rc = mdt_version_get_check_save(info, child, 2);
599 /* Let lower layer know current lock mode. */
600 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
603 * Do not perform lookup sanity check. We know that name does
606 info->mti_spec.sp_cr_lookup = 0;
607 info->mti_spec.sp_feat = &dt_directory_features;
609 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
610 mdt_object_child(child), &info->mti_spec, ma);
612 rc = mdt_attr_get_complex(info, child, ma);
618 * On DNE, we need to eliminate dependey between 'mkdir a' and
619 * 'mkdir a/b' if b is a striped directory, to achieve this, two
620 * things are done below:
621 * 1. save child and slaves lock.
622 * 2. if the child is a striped directory, relock parent so to
623 * compare against with COS locks to ensure parent was
626 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
627 struct mdt_lock_handle *lhc;
628 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
631 rc = mdt_object_striped(info, child);
637 if (!mdt_object_remote(parent)) {
638 mdt_object_unlock(info, parent, lh, 1);
639 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
640 rc = mdt_reint_object_lock(info, parent, lh,
641 MDS_INODELOCK_UPDATE,
648 lhc = &info->mti_lh[MDT_LH_CHILD];
649 mdt_lock_handle_init(lhc);
650 mdt_lock_reg_init(lhc, LCK_PW);
651 rc = mdt_reint_striped_lock(info, child, lhc,
652 MDS_INODELOCK_UPDATE, einfo,
657 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
660 /* Return fid & attr to client. */
661 if (ma->ma_valid & MA_INODE)
662 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
663 mdt_object_fid(child));
666 mdt_object_put(info->mti_env, child);
668 mdt_object_unlock(info, parent, lh, rc);
670 mdt_object_put(info->mti_env, parent);
674 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
677 struct mdt_lock_handle *lh;
678 int do_vbr = ma->ma_attr.la_valid &
679 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
680 __u64 lockpart = MDS_INODELOCK_UPDATE;
681 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
686 rc = mdt_object_striped(info, mo);
692 lh = &info->mti_lh[MDT_LH_PARENT];
693 mdt_lock_reg_init(lh, LCK_PW);
695 /* Even though the new MDT will grant PERM lock to the old
696 * client, but the old client will almost ignore that during
697 * So it needs to revoke both LOOKUP and PERM lock here, so
698 * both new and old client can cancel the dcache
700 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
701 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
703 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
708 /* all attrs are packed into mti_attr in unpack_setattr */
709 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
710 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
712 /* VBR: update version if attr changed are important for recovery */
714 /* update on-disk version of changed object */
715 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
716 rc = mdt_version_get_check_save(info, mo, 0);
718 GOTO(out_unlock, rc);
721 /* Ensure constant striping during chown(). See LU-2789. */
722 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
723 mutex_lock(&mo->mot_lov_mutex);
725 /* all attrs are packed into mti_attr in unpack_setattr */
726 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
728 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
729 mutex_unlock(&mo->mot_lov_mutex);
732 GOTO(out_unlock, rc);
733 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
736 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
741 * Check HSM flags and add HS_DIRTY flag if relevant.
743 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
744 * and is not RELEASED.
746 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
749 struct lu_ucred *uc = mdt_ucred(info);
754 /* If the file was modified, add the dirty flag */
755 ma->ma_need = MA_HSM;
756 rc = mdt_attr_get_complex(info, mo, ma);
758 CERROR("file attribute read error for "DFID": %d.\n",
759 PFID(mdt_object_fid(mo)), rc);
763 /* If an up2date copy exists in the backend, add dirty flag */
764 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
765 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
766 ma->ma_hsm.mh_flags |= HS_DIRTY;
768 /* Bump cap so that closes from non-owner writers can
769 * set the HSM state to dirty.
771 cap_saved = uc->uc_cap;
772 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
773 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
774 uc->uc_cap = cap_saved;
776 CERROR("file attribute change error for "DFID": %d\n",
777 PFID(mdt_object_fid(mo)), rc);
783 static int mdt_reint_setattr(struct mdt_thread_info *info,
784 struct mdt_lock_handle *lhc)
786 struct mdt_device *mdt = info->mti_mdt;
787 struct md_attr *ma = &info->mti_attr;
788 struct mdt_reint_record *rr = &info->mti_rr;
789 struct ptlrpc_request *req = mdt_info_req(info);
790 struct mdt_object *mo;
791 struct mdt_body *repbody;
795 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
796 (unsigned int)ma->ma_attr.la_valid);
798 if (info->mti_dlm_req)
799 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
801 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
803 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
804 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
806 GOTO(out, rc = PTR_ERR(mo));
808 if (!mdt_object_exists(mo))
809 GOTO(out_put, rc = -ENOENT);
811 if (mdt_object_remote(mo))
812 GOTO(out_put, rc = -EREMOTE);
814 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
815 /* revoke lease lock if size is going to be changed */
816 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
817 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
818 atomic_read(&mo->mot_lease_count) > 0)) {
819 down_read(&mo->mot_open_sem);
821 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
822 lhc = &info->mti_lh[MDT_LH_LOCAL];
823 mdt_lock_reg_init(lhc, LCK_CW);
825 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
827 up_read(&mo->mot_open_sem);
831 /* revoke lease lock */
832 mdt_object_unlock(info, mo, lhc, 1);
834 up_read(&mo->mot_open_sem);
837 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
838 /* Check write access for the O_TRUNC case */
839 if (mdt_write_read(mo) < 0)
840 GOTO(out_put, rc = -ETXTBSY);
842 /* LU-10286: compatibility check for FLR.
843 * Please check the comment in mdt_finish_open() for details
845 if (!exp_connect_flr(info->mti_exp) ||
846 !exp_connect_overstriping(info->mti_exp)) {
847 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
848 if (rc < 0 && rc != -ENODATA)
851 if (!exp_connect_flr(info->mti_exp)) {
853 mdt_lmm_is_flr(info->mti_big_lmm))
854 GOTO(out_put, rc = -EOPNOTSUPP);
857 if (!exp_connect_overstriping(info->mti_exp)) {
859 mdt_lmm_is_overstriping(info->mti_big_lmm))
860 GOTO(out_put, rc = -EOPNOTSUPP);
864 /* For truncate, the file size sent from client
865 * is believable, but the blocks are incorrect,
866 * which makes the block size in LSOM attribute
867 * inconsisent with the real block size.
869 rc = mdt_lsom_update(info, mo, true);
874 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
875 if (ma->ma_valid & MA_LOV)
876 GOTO(out_put, rc = -EPROTO);
878 /* MDT supports FMD for regular files due to Data-on-MDT */
879 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
880 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
881 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
884 rc = mdt_attr_set(info, mo, ma);
887 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
888 (ma->ma_valid & MA_INODE)) {
889 struct lu_buf *buf = &info->mti_buf;
890 struct lu_ucred *uc = mdt_ucred(info);
891 struct mdt_lock_handle *lh;
893 __u64 lockpart = MDS_INODELOCK_XATTR;
895 /* reject if either remote or striped dir is disabled */
896 if (ma->ma_valid & MA_LMV) {
897 if (!mdt->mdt_enable_remote_dir ||
898 !mdt->mdt_enable_striped_dir)
899 GOTO(out_put, rc = -EPERM);
901 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
902 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
903 mdt->mdt_enable_remote_dir_gid != -1)
904 GOTO(out_put, rc = -EPERM);
907 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
908 GOTO(out_put, rc = -ENOTDIR);
910 if (ma->ma_attr.la_valid != 0)
911 GOTO(out_put, rc = -EPROTO);
913 if (ma->ma_valid & MA_LOV) {
914 buf->lb_buf = ma->ma_lmm;
915 buf->lb_len = ma->ma_lmm_size;
916 name = XATTR_NAME_LOV;
918 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
921 buf->lb_len = ma->ma_lmv_size;
922 name = XATTR_NAME_DEFAULT_LMV;
923 /* force client to update dir default layout */
924 lockpart |= MDS_INODELOCK_LOOKUP;
927 lh = &info->mti_lh[MDT_LH_PARENT];
928 mdt_lock_reg_init(lh, LCK_PW);
930 rc = mdt_object_lock(info, mo, lh, lockpart);
934 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
937 mdt_object_unlock(info, mo, lh, rc);
941 GOTO(out_put, rc = -EPROTO);
944 /* If file data is modified, add the dirty flag */
945 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
946 rc = mdt_add_dirty_flag(info, mo, ma);
948 ma->ma_need = MA_INODE;
950 rc = mdt_attr_get_complex(info, mo, ma);
954 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
958 mdt_object_put(info->mti_env, mo);
961 mdt_counter_incr(req, LPROC_MDT_SETATTR);
963 mdt_client_compatibility(info);
964 rc2 = mdt_fix_reply(info);
970 static int mdt_reint_create(struct mdt_thread_info *info,
971 struct mdt_lock_handle *lhc)
973 struct ptlrpc_request *req = mdt_info_req(info);
977 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
978 RETURN(err_serious(-ESTALE));
980 if (info->mti_dlm_req)
981 ldlm_request_cancel(mdt_info_req(info),
982 info->mti_dlm_req, 0, LATF_SKIP);
984 if (!lu_name_is_valid(&info->mti_rr.rr_name))
987 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
989 mdt_counter_incr(req, LPROC_MDT_MKDIR);
997 /* Special file should stay on the same node as parent. */
998 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1001 CERROR("%s: Unsupported mode %o\n",
1002 mdt_obd_name(info->mti_mdt),
1003 info->mti_attr.ma_attr.la_mode);
1004 RETURN(err_serious(-EOPNOTSUPP));
1007 rc = mdt_create(info);
1012 * VBR: save parent version in reply and child version getting by its name.
1013 * Version of child is getting and checking during its lookup. If
1015 static int mdt_reint_unlink(struct mdt_thread_info *info,
1016 struct mdt_lock_handle *lhc)
1018 struct mdt_reint_record *rr = &info->mti_rr;
1019 struct ptlrpc_request *req = mdt_info_req(info);
1020 struct md_attr *ma = &info->mti_attr;
1021 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1022 struct mdt_object *mp;
1023 struct mdt_object *mc;
1024 struct mdt_lock_handle *parent_lh;
1025 struct mdt_lock_handle *child_lh;
1026 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1028 bool cos_incompat = false;
1033 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1034 PNAME(&rr->rr_name));
1036 if (info->mti_dlm_req)
1037 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1039 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1040 RETURN(err_serious(-ENOENT));
1042 if (!fid_is_md_operative(rr->rr_fid1))
1045 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1047 RETURN(PTR_ERR(mp));
1049 if (mdt_object_remote(mp)) {
1050 cos_incompat = true;
1052 rc = mdt_version_get_check_save(info, mp, 0);
1054 GOTO(put_parent, rc);
1058 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1059 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1060 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1063 GOTO(put_parent, rc);
1065 /* lookup child object along with version checking */
1066 fid_zero(child_fid);
1067 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1069 /* Name might not be able to find during resend of
1070 * remote unlink, considering following case.
1071 * dir_A is a remote directory, the name entry of
1072 * dir_A is on MDT0, the directory is on MDT1,
1074 * 1. client sends unlink req to MDT1.
1075 * 2. MDT1 sends name delete update to MDT0.
1076 * 3. name entry is being deleted in MDT0 synchronously.
1077 * 4. MDT1 is restarted.
1078 * 5. client resends unlink req to MDT1. So it can not
1079 * find the name entry on MDT0 anymore.
1080 * In this case, MDT1 only needs to destory the local
1083 if (mdt_object_remote(mp) && rc == -ENOENT &&
1084 !fid_is_zero(rr->rr_fid2) &&
1085 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1087 *child_fid = *rr->rr_fid2;
1089 GOTO(unlock_parent, rc);
1093 if (!fid_is_md_operative(child_fid))
1094 GOTO(unlock_parent, rc = -EPERM);
1096 /* We will lock the child regardless it is local or remote. No harm. */
1097 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1099 GOTO(unlock_parent, rc = PTR_ERR(mc));
1101 if (!cos_incompat) {
1102 rc = mdt_object_striped(info, mc);
1104 GOTO(put_child, rc);
1108 mdt_object_put(info->mti_env, mc);
1109 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1114 child_lh = &info->mti_lh[MDT_LH_CHILD];
1115 mdt_lock_reg_init(child_lh, LCK_EX);
1116 if (info->mti_spec.sp_rm_entry) {
1117 struct lu_ucred *uc = mdt_ucred(info);
1119 if (!mdt_is_dne_client(req->rq_export))
1120 /* Return -ENOTSUPP for old client */
1121 GOTO(put_child, rc = -ENOTSUPP);
1123 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
1124 GOTO(put_child, rc = -EPERM);
1126 ma->ma_need = MA_INODE;
1128 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1129 NULL, &rr->rr_name, ma, no_name);
1130 GOTO(put_child, rc);
1133 if (mdt_object_remote(mc)) {
1134 struct mdt_body *repbody;
1136 if (!fid_is_zero(rr->rr_fid2)) {
1137 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1138 mdt_obd_name(info->mti_mdt),
1139 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1140 GOTO(put_child, rc = -ENOENT);
1142 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1143 mdt_obd_name(info->mti_mdt),
1144 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1146 if (!mdt_is_dne_client(req->rq_export))
1147 /* Return -ENOTSUPP for old client */
1148 GOTO(put_child, rc = -ENOTSUPP);
1150 /* Revoke the LOOKUP lock of the remote object granted by
1151 * this MDT. Since the unlink will happen on another MDT,
1152 * it will release the LOOKUP lock right away. Then What
1153 * would happen if another client try to grab the LOOKUP
1154 * lock at the same time with unlink XXX
1156 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1157 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1158 LASSERT(repbody != NULL);
1159 repbody->mbo_fid1 = *mdt_object_fid(mc);
1160 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1161 GOTO(unlock_child, rc = -EREMOTE);
1163 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1164 * this now because a running HSM restore on the child (unlink
1165 * victim) will hold the layout lock. See LU-4002.
1167 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1168 if (mdt_object_remote(mp)) {
1169 /* Enqueue lookup lock from parent MDT */
1170 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1171 &child_lh->mlh_rreg_lh,
1172 child_lh->mlh_rreg_mode,
1173 MDS_INODELOCK_LOOKUP, false);
1175 GOTO(put_child, rc);
1177 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1180 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1183 GOTO(put_child, rc);
1186 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1187 * whether need MA_LOV and MA_COOKIE.
1189 ma->ma_need = MA_INODE;
1192 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1193 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1194 /* save version when object is locked */
1195 mdt_version_get_save(info, mc, 1);
1197 mutex_lock(&mc->mot_lov_mutex);
1199 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1200 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1202 mutex_unlock(&mc->mot_lov_mutex);
1204 GOTO(unlock_child, rc);
1206 if (!lu_object_is_dying(&mc->mot_header)) {
1207 rc = mdt_attr_get_complex(info, mc, ma);
1210 } else if (mdt_dom_check_for_discard(info, mc)) {
1211 mdt_dom_discard_data(info, mc);
1213 mdt_handle_last_unlink(info, mc, ma);
1216 if (ma->ma_valid & MA_INODE) {
1217 switch (ma->ma_attr.la_mode & S_IFMT) {
1219 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1227 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1230 LASSERTF(0, "bad file type %o unlinking\n",
1231 ma->ma_attr.la_mode);
1238 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1240 mdt_object_put(info->mti_env, mc);
1242 mdt_object_unlock(info, mp, parent_lh, rc);
1244 mdt_object_put(info->mti_env, mp);
1245 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1250 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1253 static int mdt_reint_link(struct mdt_thread_info *info,
1254 struct mdt_lock_handle *lhc)
1256 struct mdt_reint_record *rr = &info->mti_rr;
1257 struct ptlrpc_request *req = mdt_info_req(info);
1258 struct md_attr *ma = &info->mti_attr;
1259 struct mdt_object *ms;
1260 struct mdt_object *mp;
1261 struct mdt_lock_handle *lhs;
1262 struct mdt_lock_handle *lhp;
1267 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1268 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1270 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1271 RETURN(err_serious(-ENOENT));
1273 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1274 req->rq_no_reply = 1;
1275 RETURN(err_serious(-ENOENT));
1278 if (info->mti_dlm_req)
1279 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1281 /* Invalid case so return error immediately instead of
1284 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1287 if (!fid_is_md_operative(rr->rr_fid1) ||
1288 !fid_is_md_operative(rr->rr_fid2))
1291 /* step 1: find target parent dir */
1292 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1294 RETURN(PTR_ERR(mp));
1296 rc = mdt_version_get_check_save(info, mp, 0);
1298 GOTO(put_parent, rc);
1300 /* step 2: find source */
1301 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1303 GOTO(put_parent, rc = PTR_ERR(ms));
1305 if (!mdt_object_exists(ms)) {
1306 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1307 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1308 GOTO(put_source, rc = -ENOENT);
1311 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1313 lhp = &info->mti_lh[MDT_LH_PARENT];
1314 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1315 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1318 GOTO(put_source, rc);
1320 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1322 lhs = &info->mti_lh[MDT_LH_CHILD];
1323 mdt_lock_reg_init(lhs, LCK_EX);
1324 rc = mdt_reint_object_lock(info, ms, lhs,
1325 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1328 GOTO(unlock_parent, rc);
1330 /* step 3: link it */
1331 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1332 OBD_FAIL_MDS_REINT_LINK_WRITE);
1334 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1335 rc = mdt_version_get_check_save(info, ms, 1);
1337 GOTO(unlock_source, rc);
1339 /** check target version by name during replay */
1340 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1341 &info->mti_tmp_fid1, 2);
1342 if (rc != 0 && rc != -ENOENT)
1343 GOTO(unlock_source, rc);
1344 /* save version of file name for replay, it must be ENOENT here */
1345 if (!req_is_replay(mdt_info_req(info))) {
1346 if (rc != -ENOENT) {
1347 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1348 PNAME(&rr->rr_name));
1349 GOTO(unlock_source, rc = -EEXIST);
1351 info->mti_ver[2] = ENOENT_VERSION;
1352 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1355 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1356 mdt_object_child(ms), &rr->rr_name, ma);
1359 mdt_counter_incr(req, LPROC_MDT_LINK);
1363 mdt_object_unlock(info, ms, lhs, rc);
1365 mdt_object_unlock(info, mp, lhp, rc);
1367 mdt_object_put(info->mti_env, ms);
1369 mdt_object_put(info->mti_env, mp);
1373 * lock the part of the directory according to the hash of the name
1374 * (lh->mlh_pdo_hash) in parallel directory lock.
1376 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1377 struct mdt_lock_handle *lh,
1378 struct mdt_object *obj, __u64 ibits,
1381 struct ldlm_res_id *res = &info->mti_res_id;
1382 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1383 union ldlm_policy_data *policy = &info->mti_policy;
1384 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1388 * Finish res_id initializing by name hash marking part of
1389 * directory which is taking modification.
1391 LASSERT(lh->mlh_pdo_hash != 0);
1392 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1393 memset(policy, 0, sizeof(*policy));
1394 policy->l_inodebits.bits = ibits;
1396 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1397 dlmflags |= LDLM_FL_COS_INCOMPAT;
1399 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1400 * going to be sent to client. If it is - mdt_intent_policy() path will
1401 * fix it up and turn FL_LOCAL flag off.
1403 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1404 policy, res, dlmflags,
1405 &info->mti_exp->exp_handle.h_cookie);
1410 * Get BFL lock for rename or migrate process.
1412 static int mdt_rename_lock(struct mdt_thread_info *info,
1413 struct lustre_handle *lh)
1418 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1419 struct lu_fid *fid = &info->mti_tmp_fid1;
1420 struct mdt_object *obj;
1422 /* XXX, right now, it has to use object API to
1423 * enqueue lock cross MDT, so it will enqueue
1424 * rename lock(with LUSTRE_BFL_FID) by root object
1427 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1429 RETURN(PTR_ERR(obj));
1431 rc = mdt_remote_object_lock(info, obj,
1432 &LUSTRE_BFL_FID, lh,
1434 MDS_INODELOCK_UPDATE, false);
1435 mdt_object_put(info->mti_env, obj);
1437 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1438 union ldlm_policy_data *policy = &info->mti_policy;
1439 struct ldlm_res_id *res_id = &info->mti_res_id;
1442 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1443 memset(policy, 0, sizeof(*policy));
1444 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1445 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1446 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1447 LDLM_IBITS, policy, LCK_EX, &flags,
1449 ldlm_completion_ast, NULL, NULL, 0,
1451 &info->mti_exp->exp_handle.h_cookie,
1458 static void mdt_rename_unlock(struct lustre_handle *lh)
1461 LASSERT(lustre_handle_is_used(lh));
1462 /* Cancel the single rename lock right away */
1463 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1467 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1468 const struct lu_fid *fid,
1471 struct mdt_object *dir;
1475 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1479 /* check early, the real version will be saved after locking */
1480 rc = mdt_version_get_check(info, dir, idx);
1484 if (!mdt_object_exists(dir))
1485 GOTO(out_put, rc = -ENOENT);
1487 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1488 GOTO(out_put, rc = -ENOTDIR);
1492 mdt_object_put(info->mti_env, dir);
1497 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1498 * herein we don't really check it, just do revoke.
1500 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1501 struct mdt_object *pobj,
1502 struct mdt_object *obj)
1504 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1507 mdt_lock_handle_init(lh);
1508 mdt_lock_reg_init(lh, LCK_EX);
1510 if (mdt_object_remote(pobj)) {
1511 /* don't bother to check if pobj and obj are on the same MDT. */
1512 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1513 &lh->mlh_rreg_lh, LCK_EX,
1514 MDS_INODELOCK_LOOKUP, false);
1515 } else if (mdt_object_remote(obj)) {
1516 struct ldlm_res_id *res = &info->mti_res_id;
1517 union ldlm_policy_data *policy = &info->mti_policy;
1518 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1519 LDLM_FL_COS_INCOMPAT;
1521 fid_build_reg_res_name(mdt_object_fid(obj), res);
1522 memset(policy, 0, sizeof(*policy));
1523 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1524 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1525 &lh->mlh_reg_lh, LCK_EX, policy, res,
1528 /* do nothing if both are local */
1536 * TODO, currently we don't save this lock because there is no place to
1537 * hold this lock handle, but to avoid race we need to save this lock.
1539 mdt_object_unlock(info, NULL, lh, 1);
1545 * operation may takes locks of linkea, or directory stripes, group them in
1548 struct mdt_sub_lock {
1549 struct mdt_object *msl_obj;
1550 struct mdt_lock_handle msl_lh;
1551 struct list_head msl_linkage;
1554 static void mdt_unlock_list(struct mdt_thread_info *info,
1555 struct list_head *list, int decref)
1557 struct mdt_sub_lock *msl;
1558 struct mdt_sub_lock *tmp;
1560 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1561 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1562 list_del(&msl->msl_linkage);
1567 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1568 struct mdt_object *obj,
1569 struct mdt_lock_handle *lh,
1570 struct ldlm_enqueue_info *einfo,
1571 struct list_head *slave_locks,
1574 if (mdt_object_remote(obj)) {
1575 mdt_unlock_list(info, slave_locks, decref);
1576 mdt_object_unlock(info, obj, lh, decref);
1578 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1583 * lock parents of links, and also check whether total locks don't exceed
1586 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1587 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1588 * \retval -ev negative errno upon error
1590 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1591 struct mdt_object *pobj,
1592 const struct md_attr *ma,
1593 struct mdt_object *obj,
1594 struct mdt_lock_handle *lhp,
1595 struct ldlm_enqueue_info *peinfo,
1596 struct list_head *parent_slave_locks,
1597 struct list_head *link_locks)
1599 struct mdt_device *mdt = info->mti_mdt;
1600 struct lu_buf *buf = &info->mti_big_buf;
1601 struct lu_name *lname = &info->mti_name;
1602 struct linkea_data ldata = { NULL };
1603 bool blocked = false;
1604 int local_lnkp_cnt = 0;
1608 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1611 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1612 if (buf->lb_buf == NULL)
1616 rc = mdt_links_read(info, obj, &ldata);
1618 if (rc == -ENOENT || rc == -ENODATA)
1623 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1624 linkea_next_entry(&ldata)) {
1625 struct mdt_object *lnkp;
1626 struct mdt_sub_lock *msl;
1630 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1633 /* check if it's also linked to parent */
1634 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1635 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1636 PFID(&fid), PNAME(lname));
1637 /* in case link is remote object, revoke LOOKUP lock */
1638 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1644 /* check if it's linked to a stripe of parent */
1645 if (ma->ma_valid & MA_LMV) {
1646 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1647 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1650 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1651 fid_le_to_cpu(stripe_fid,
1652 &lmv->lmv_stripe_fids[j]);
1653 if (lu_fid_eq(stripe_fid, &fid)) {
1654 CDEBUG(D_INFO, "skip stripe "DFID
1655 ", reovke "DNAME"\n",
1656 PFID(&fid), PNAME(lname));
1657 lnkp = mdt_object_find(info->mti_env,
1660 GOTO(out, rc = PTR_ERR(lnkp));
1666 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1668 mdt_object_put(info->mti_env, lnkp);
1673 /* Check if it's already locked */
1674 list_for_each_entry(msl, link_locks, msl_linkage) {
1675 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1677 DFID" was locked, revoke "DNAME"\n",
1678 PFID(&fid), PNAME(lname));
1679 lnkp = msl->msl_obj;
1685 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1689 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1690 PFID(&fid), PNAME(lname));
1692 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1694 CWARN("%s: cannot find obj "DFID": %ld\n",
1695 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1699 if (!mdt_object_exists(lnkp)) {
1700 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1701 PFID(&fid), PNAME(lname));
1702 mdt_object_put(info->mti_env, lnkp);
1706 if (!mdt_object_remote(lnkp))
1711 GOTO(out, rc = -ENOMEM);
1714 * we can't follow parent-child lock order like other MD
1715 * operations, use lock_try here to avoid deadlock, if the lock
1716 * cannot be taken, drop all locks taken, revoke the blocked
1717 * one, and continue processing the remaining entries, and in
1718 * the end of the loop restart from beginning.
1720 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1722 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1723 MDS_INODELOCK_UPDATE, true);
1724 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1726 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1727 PFID(&fid), PNAME(lname));
1729 mdt_unlock_list(info, link_locks, 1);
1730 /* also unlock parent locks to avoid deadlock */
1732 mdt_migrate_object_unlock(info, pobj, lhp,
1739 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1740 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1741 MDS_INODELOCK_UPDATE);
1743 mdt_object_put(info->mti_env, lnkp);
1748 if (mdt_object_remote(lnkp)) {
1749 struct ldlm_lock *lock;
1752 * for remote object, set lock cb_atomic,
1753 * so lock can be released in blocking_ast()
1754 * immediately, then the next lock_try will
1755 * have better chance of success.
1757 lock = ldlm_handle2lock(
1758 &msl->msl_lh.mlh_rreg_lh);
1759 LASSERT(lock != NULL);
1760 lock_res_and_lock(lock);
1761 ldlm_set_atomic_cb(lock);
1762 unlock_res_and_lock(lock);
1763 LDLM_LOCK_PUT(lock);
1766 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1771 INIT_LIST_HEAD(&msl->msl_linkage);
1772 msl->msl_obj = lnkp;
1773 list_add_tail(&msl->msl_linkage, link_locks);
1775 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1779 GOTO(out, rc = -EBUSY);
1784 mdt_unlock_list(info, link_locks, rc);
1785 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1787 * parent may have 3 local objects: master object and 2 stripes
1788 * (if it's being migrated too); source may have 2 local
1789 * objects: master and 1 stripe; target has 1 local object.
1795 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1796 struct mdt_object *obj,
1797 const struct md_attr *ma,
1798 struct list_head *slave_locks)
1800 struct mdt_device *mdt = info->mti_mdt;
1801 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1802 struct lu_fid *fid = &info->mti_tmp_fid1;
1803 struct mdt_object *slave;
1804 struct mdt_sub_lock *msl;
1809 LASSERT(mdt_object_remote(obj));
1810 LASSERT(ma->ma_valid & MA_LMV);
1813 if (!lmv_is_sane(lmv))
1816 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1817 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1819 if (!fid_is_sane(fid))
1822 slave = mdt_object_find(info->mti_env, mdt, fid);
1824 GOTO(out, rc = PTR_ERR(slave));
1828 mdt_object_put(info->mti_env, slave);
1829 GOTO(out, rc = -ENOMEM);
1832 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1833 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1834 MDS_INODELOCK_UPDATE, true);
1837 mdt_object_put(info->mti_env, slave);
1841 INIT_LIST_HEAD(&msl->msl_linkage);
1842 msl->msl_obj = slave;
1843 list_add_tail(&msl->msl_linkage, slave_locks);
1849 mdt_unlock_list(info, slave_locks, rc);
1853 /* lock parent and its stripes */
1854 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1855 struct mdt_object *obj,
1856 const struct md_attr *ma,
1857 struct mdt_lock_handle *lh,
1858 struct ldlm_enqueue_info *einfo,
1859 struct list_head *slave_locks)
1863 if (mdt_object_remote(obj)) {
1864 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1865 &lh->mlh_rreg_lh, LCK_PW,
1866 MDS_INODELOCK_UPDATE, false);
1871 * if obj is remote and striped, lock its stripes explicitly
1872 * because it's not striped in LOD layer on this MDT.
1874 if (ma->ma_valid & MA_LMV) {
1875 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1877 mdt_object_unlock(info, obj, lh, rc);
1880 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1888 * in migration, object may be remote, and we need take full lock of it and its
1889 * stripes if it's directory, besides, object may be a remote object on its
1890 * parent, revoke its LOOKUP lock on where its parent is located.
1892 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1893 struct mdt_object *pobj,
1894 struct mdt_object *obj,
1895 struct mdt_lock_handle *lh,
1896 struct ldlm_enqueue_info *einfo,
1897 struct list_head *slave_locks)
1901 if (mdt_object_remote(obj)) {
1902 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1906 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1907 &lh->mlh_rreg_lh, LCK_EX,
1908 MDS_INODELOCK_FULL, false);
1913 * if obj is remote and striped, lock its stripes explicitly
1914 * because it's not striped in LOD layer on this MDT.
1916 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1917 struct md_attr *ma = &info->mti_attr;
1919 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1921 mdt_object_unlock(info, obj, lh, rc);
1925 if (ma->ma_valid & MA_LMV) {
1926 rc = mdt_lock_remote_slaves(info, obj, ma,
1929 mdt_object_unlock(info, obj, lh, rc);
1933 if (mdt_object_remote(pobj)) {
1934 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1939 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1947 * lookup source by name, if parent is striped directory, we need to find the
1948 * corresponding stripe where source is located, and then lookup there.
1950 * besides, if parent is migrating too, and file is already in target stripe,
1951 * this should be a redo of 'lfs migrate' on client side.
1953 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1954 struct mdt_object *pobj,
1955 const struct md_attr *ma,
1956 const struct lu_name *lname,
1957 struct mdt_object **spobj,
1958 struct mdt_object **sobj)
1960 const struct lu_env *env = info->mti_env;
1961 struct lu_fid *fid = &info->mti_tmp_fid1;
1962 struct mdt_object *stripe;
1965 if (ma->ma_valid & MA_LMV) {
1966 /* if parent is striped, lookup on corresponding stripe */
1967 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1969 if (!lmv_is_sane(lmv))
1972 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1977 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1979 stripe = mdt_object_find(env, info->mti_mdt, fid);
1981 return PTR_ERR(stripe);
1984 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1986 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
1988 * if parent layout is changeing, and lookup child
1989 * failed on source stripe, lookup again on target
1990 * stripe, if it exists, it means previous migration
1991 * was interrupted, and current file was migrated
1994 mdt_object_put(env, stripe);
1996 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2001 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2003 stripe = mdt_object_find(env, info->mti_mdt, fid);
2005 return PTR_ERR(stripe);
2008 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2009 fid, &info->mti_spec);
2010 mdt_object_put(env, stripe);
2011 return rc ?: -EALREADY;
2013 mdt_object_put(env, stripe);
2018 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2024 mdt_object_get(env, stripe);
2029 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2030 if (IS_ERR(*sobj)) {
2031 mdt_object_put(env, stripe);
2032 rc = PTR_ERR(*sobj);
2040 /* end lease and close file for regular file */
2041 static int mdd_migrate_close(struct mdt_thread_info *info,
2042 struct mdt_object *obj)
2044 struct close_data *data;
2045 struct mdt_body *repbody;
2046 struct ldlm_lock *lease;
2051 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2053 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2057 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2062 lease = ldlm_handle2lock(&data->cd_handle);
2066 /* check if the lease was already canceled */
2067 lock_res_and_lock(lease);
2068 rc = ldlm_is_cancel(lease);
2069 unlock_res_and_lock(lease);
2073 LDLM_DEBUG(lease, DFID" lease broken",
2074 PFID(mdt_object_fid(obj)));
2078 * cancel server side lease, client side counterpart should have been
2079 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2081 ldlm_lock_cancel(lease);
2082 ldlm_reprocess_all(lease->l_resource, lease);
2083 LDLM_LOCK_PUT(lease);
2086 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2087 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2088 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2094 * migrate file in below steps:
2095 * 1. lock parent and its stripes
2096 * 2. lookup source by name
2097 * 3. lock parents of source links if source is not directory
2098 * 4. reject if source is in HSM
2099 * 5. take source open_sem and close file if source is regular file
2100 * 6. lock source and its stripes if it's directory
2101 * 7. lock target so subsequent change to it can trigger COS
2103 * 9. unlock above locks
2104 * 10. sync device if source has links
2106 int mdt_reint_migrate(struct mdt_thread_info *info,
2107 struct mdt_lock_handle *unused)
2109 const struct lu_env *env = info->mti_env;
2110 struct mdt_device *mdt = info->mti_mdt;
2111 struct ptlrpc_request *req = mdt_info_req(info);
2112 struct mdt_reint_record *rr = &info->mti_rr;
2113 struct lu_ucred *uc = mdt_ucred(info);
2114 struct md_attr *ma = &info->mti_attr;
2115 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2116 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2117 struct mdt_object *pobj;
2118 struct mdt_object *spobj = NULL;
2119 struct mdt_object *sobj = NULL;
2120 struct mdt_object *tobj;
2121 struct lustre_handle rename_lh = { 0 };
2122 struct mdt_lock_handle *lhp;
2123 struct mdt_lock_handle *lhs;
2124 struct mdt_lock_handle *lht;
2125 LIST_HEAD(parent_slave_locks);
2126 LIST_HEAD(child_slave_locks);
2127 LIST_HEAD(link_locks);
2128 int lock_retries = 5;
2129 bool open_sem_locked = false;
2130 bool do_sync = false;
2134 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2135 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2137 if (info->mti_dlm_req)
2138 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2140 if (!fid_is_md_operative(rr->rr_fid1) ||
2141 !fid_is_md_operative(rr->rr_fid2))
2144 /* don't allow migrate . or .. */
2145 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2148 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2151 if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2152 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2153 mdt->mdt_enable_remote_dir_gid != -1)
2157 * Note: do not enqueue rename lock for replay request, because
2158 * if other MDT holds rename lock, but being blocked to wait for
2159 * this MDT to finish its recovery, and the failover MDT can not
2160 * get rename lock, which will cause deadlock.
2162 * req is NULL if this is called by directory auto-split.
2164 if (req && !req_is_replay(req)) {
2165 rc = mdt_rename_lock(info, &rename_lh);
2167 CERROR("%s: can't lock FS for rename: rc = %d\n",
2168 mdt_obd_name(info->mti_mdt), rc);
2173 /* pobj is master object of parent */
2174 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2176 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2179 rc = mdt_version_get_check(info, pobj, 0);
2181 GOTO(put_parent, rc);
2184 if (!mdt_object_exists(pobj))
2185 GOTO(put_parent, rc = -ENOENT);
2187 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2188 GOTO(put_parent, rc = -ENOTDIR);
2190 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2192 GOTO(put_parent, rc);
2195 /* lock parent object */
2196 lhp = &info->mti_lh[MDT_LH_PARENT];
2197 mdt_lock_reg_init(lhp, LCK_PW);
2198 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2199 &parent_slave_locks);
2201 GOTO(put_parent, rc);
2204 * spobj is the corresponding stripe against name if pobj is striped
2205 * directory, which is the real parent, and no need to lock, because
2206 * we've taken full lock of pobj.
2208 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2210 GOTO(unlock_parent, rc);
2212 /* lock parents of source links, and revoke LOOKUP lock of links */
2213 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2214 &parent_slave_locks, &link_locks);
2215 if (rc == -EBUSY && lock_retries-- > 0) {
2216 mdt_object_put(env, sobj);
2217 mdt_object_put(env, spobj);
2222 GOTO(put_source, rc);
2225 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2226 * with one request, if total lock count exceeds this limit, we will
2227 * drop all locks after migration, and synchronous device in the end.
2231 /* TODO: DoM migration is not supported, migrate dirent only */
2232 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2233 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2235 GOTO(unlock_links, rc);
2237 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2238 info->mti_spec.sp_migrate_nsonly = 1;
2241 /* if migration HSM is allowed */
2242 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2243 ma->ma_need = MA_HSM;
2245 rc = mdt_attr_get_complex(info, sobj, ma);
2247 GOTO(unlock_links, rc);
2249 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2250 GOTO(unlock_links, rc = -EOPNOTSUPP);
2253 /* end lease and close file for regular file */
2254 if (info->mti_spec.sp_migrate_close) {
2255 /* try to hold open_sem so that nobody else can open the file */
2256 if (!down_write_trylock(&sobj->mot_open_sem)) {
2258 mdd_migrate_close(info, sobj);
2259 GOTO(unlock_links, rc = -EBUSY);
2261 open_sem_locked = true;
2262 rc = mdd_migrate_close(info, sobj);
2264 GOTO(unlock_open_sem, rc);
2269 lhs = &info->mti_lh[MDT_LH_OLD];
2270 mdt_lock_reg_init(lhs, LCK_EX);
2271 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2272 &child_slave_locks);
2274 GOTO(unlock_open_sem, rc);
2277 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2279 GOTO(unlock_source, rc = PTR_ERR(tobj));
2281 lht = &info->mti_lh[MDT_LH_NEW];
2282 mdt_lock_reg_init(lht, LCK_EX);
2283 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2285 GOTO(put_target, rc);
2287 /* Don't do lookup sanity check. We know name doesn't exist. */
2288 info->mti_spec.sp_cr_lookup = 0;
2289 info->mti_spec.sp_feat = &dt_directory_features;
2291 rc = mdo_migrate(env, mdt_object_child(pobj),
2292 mdt_object_child(sobj), &rr->rr_name,
2293 mdt_object_child(tobj),
2294 &info->mti_spec, ma);
2296 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2297 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2300 mdt_object_unlock(info, tobj, lht, rc);
2302 mdt_object_put(env, tobj);
2304 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2305 &child_slave_locks, rc);
2307 if (open_sem_locked)
2308 up_write(&sobj->mot_open_sem);
2310 /* if we've got too many locks to save into RPC,
2311 * then just commit before the locks are released
2314 mdt_device_sync(env, mdt);
2315 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2317 mdt_object_put(env, sobj);
2318 mdt_object_put(env, spobj);
2320 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2321 &parent_slave_locks, rc);
2323 mdt_object_put(env, pobj);
2325 if (lustre_handle_is_used(&rename_lh))
2326 mdt_rename_unlock(&rename_lh);
2331 static int mdt_object_lock_save(struct mdt_thread_info *info,
2332 struct mdt_object *dir,
2333 struct mdt_lock_handle *lh,
2334 int idx, bool cos_incompat)
2338 /* we lock the target dir if it is local */
2339 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2344 /* get and save correct version after locking */
2345 mdt_version_get_save(info, dir, idx);
2350 * determine lock order of sobj and tobj
2352 * there are two situations we need to lock tobj before sobj:
2353 * 1. sobj is child of tobj
2354 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2355 * larger than that of tobj
2357 * \retval 1 lock tobj before sobj
2358 * \retval 0 lock sobj before tobj
2359 * \retval -ev negative errno upon error
2361 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2362 struct mdt_object *sobj,
2363 struct mdt_object *tobj)
2365 struct md_attr *ma = &info->mti_attr;
2366 struct lu_fid *spfid = &info->mti_tmp_fid1;
2367 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2368 struct lmv_mds_md_v1 *lmv;
2373 /* sobj and tobj are the same */
2377 if (fid_is_root(mdt_object_fid(sobj)))
2380 if (fid_is_root(mdt_object_fid(tobj)))
2383 /* check whether sobj is child of tobj */
2384 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2385 mdt_object_fid(tobj));
2392 /* check whether sobj and tobj are children of the same parent */
2393 rc = mdt_attr_get_pfid(info, sobj, spfid);
2397 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2401 if (!lu_fid_eq(spfid, tpfid))
2404 /* check whether sobj and tobj are sibling stripes */
2405 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2409 if (!(ma->ma_valid & MA_LMV))
2412 lmv = &ma->ma_lmv->lmv_md_v1;
2413 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2415 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2418 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2422 if (!(ma->ma_valid & MA_LMV))
2425 lmv = &ma->ma_lmv->lmv_md_v1;
2426 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2428 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2430 /* check stripe index of sobj and tobj */
2431 if (sindex == tindex)
2434 return sindex < tindex ? 0 : 1;
2438 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2439 * 2 - srcdir child; 3 - tgtdir child.
2440 * Update on disk version of srcdir child.
2442 static int mdt_reint_rename(struct mdt_thread_info *info,
2443 struct mdt_lock_handle *unused)
2445 struct mdt_device *mdt = info->mti_mdt;
2446 struct mdt_reint_record *rr = &info->mti_rr;
2447 struct md_attr *ma = &info->mti_attr;
2448 struct ptlrpc_request *req = mdt_info_req(info);
2449 struct mdt_object *msrcdir = NULL;
2450 struct mdt_object *mtgtdir = NULL;
2451 struct mdt_object *mold;
2452 struct mdt_object *mnew = NULL;
2453 struct lustre_handle rename_lh = { 0 };
2454 struct mdt_lock_handle *lh_srcdirp;
2455 struct mdt_lock_handle *lh_tgtdirp;
2456 struct mdt_lock_handle *lh_oldp = NULL;
2457 struct mdt_lock_handle *lh_newp = NULL;
2458 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2459 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2461 bool reverse = false, discard = false;
2466 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2467 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2468 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2470 if (info->mti_dlm_req)
2471 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2473 if (!fid_is_md_operative(rr->rr_fid1) ||
2474 !fid_is_md_operative(rr->rr_fid2))
2477 /* find both parents. */
2478 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2479 if (IS_ERR(msrcdir))
2480 RETURN(PTR_ERR(msrcdir));
2482 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2484 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2486 mdt_object_get(info->mti_env, mtgtdir);
2488 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2489 if (IS_ERR(mtgtdir))
2490 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2494 * Note: do not enqueue rename lock for replay request, because
2495 * if other MDT holds rename lock, but being blocked to wait for
2496 * this MDT to finish its recovery, and the failover MDT can not
2497 * get rename lock, which will cause deadlock.
2499 if (!req_is_replay(req)) {
2501 * Normally rename RPC is handled on the MDT with the target
2502 * directory (if target exists, it's on the MDT with the
2503 * target), if the source directory is remote, it's a hint that
2504 * source is remote too (this may not be true, but it won't
2505 * cause any issue), return -EXDEV early to avoid taking
2508 if (!mdt->mdt_enable_remote_rename &&
2509 mdt_object_remote(msrcdir))
2510 GOTO(out_put_tgtdir, rc = -EXDEV);
2512 rc = mdt_rename_lock(info, &rename_lh);
2514 CERROR("%s: can't lock FS for rename: rc = %d\n",
2515 mdt_obd_name(mdt), rc);
2516 GOTO(out_put_tgtdir, rc);
2520 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2522 GOTO(out_unlock_rename, rc);
2526 /* source needs to be looked up after locking source parent, otherwise
2527 * this rename may race with unlink source, and cause rename hang, see
2528 * sanityn.sh 55b, so check parents first, if later we found source is
2529 * remote, relock parents.
2531 cos_incompat = (mdt_object_remote(msrcdir) ||
2532 mdt_object_remote(mtgtdir));
2534 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2536 /* lock parents in the proper order. */
2537 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2538 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2541 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2542 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2545 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2548 GOTO(out_unlock_rename, rc);
2550 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2552 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2555 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2556 GOTO(out_unlock_rename, rc);
2559 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2562 GOTO(out_unlock_rename, rc);
2564 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2566 if (mtgtdir != msrcdir) {
2567 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2569 } else if (!mdt_object_remote(mtgtdir) &&
2570 lh_srcdirp->mlh_pdo_hash !=
2571 lh_tgtdirp->mlh_pdo_hash) {
2572 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2573 MDS_INODELOCK_UPDATE,
2575 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2578 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2579 GOTO(out_unlock_rename, rc);
2583 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2584 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2586 /* find mold object. */
2588 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2590 GOTO(out_unlock_parents, rc);
2592 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2593 GOTO(out_unlock_parents, rc = -EINVAL);
2595 if (!fid_is_md_operative(old_fid))
2596 GOTO(out_unlock_parents, rc = -EPERM);
2598 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2600 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2602 if (!mdt_object_exists(mold)) {
2603 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2605 "object does not exist");
2606 GOTO(out_put_old, rc = -ENOENT);
2609 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2610 GOTO(out_put_old, rc = -EXDEV);
2612 /* Check if @mtgtdir is subdir of @mold, before locking child
2613 * to avoid reverse locking.
2615 if (mtgtdir != msrcdir) {
2616 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2621 GOTO(out_put_old, rc);
2625 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2626 /* save version after locking */
2627 mdt_version_get_save(info, mold, 2);
2629 if (!cos_incompat && mdt_object_remote(mold)) {
2630 cos_incompat = true;
2631 mdt_object_put(info->mti_env, mold);
2632 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2633 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2637 /* find mnew object:
2638 * mnew target object may not exist now
2639 * lookup with version checking
2642 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2645 /* the new_fid should have been filled at this moment */
2646 if (lu_fid_eq(old_fid, new_fid))
2647 GOTO(out_put_old, rc);
2649 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2650 lu_fid_eq(new_fid, rr->rr_fid2))
2651 GOTO(out_put_old, rc = -EINVAL);
2653 if (!fid_is_md_operative(new_fid))
2654 GOTO(out_put_old, rc = -EPERM);
2656 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2658 GOTO(out_put_old, rc = PTR_ERR(mnew));
2660 if (!mdt_object_exists(mnew)) {
2661 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2663 "object does not exist");
2664 GOTO(out_put_new, rc = -ENOENT);
2667 if (mdt_object_remote(mnew)) {
2668 struct mdt_body *repbody;
2670 /* Always send rename req to the target child MDT */
2671 repbody = req_capsule_server_get(info->mti_pill,
2673 LASSERT(repbody != NULL);
2674 repbody->mbo_fid1 = *new_fid;
2675 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2676 GOTO(out_put_new, rc = -EXDEV);
2678 /* Before locking the target dir, check we do not replace
2679 * a dir with a non-dir, otherwise it may deadlock with
2680 * link op which tries to create a link in this dir
2681 * back to this non-dir.
2683 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2684 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2685 GOTO(out_put_new, rc = -EISDIR);
2687 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2688 mdt_lock_reg_init(lh_oldp, LCK_EX);
2689 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2690 if (mdt_object_remote(msrcdir)) {
2691 /* Enqueue lookup lock from the parent MDT */
2692 rc = mdt_remote_object_lock(info, msrcdir,
2693 mdt_object_fid(mold),
2694 &lh_oldp->mlh_rreg_lh,
2695 lh_oldp->mlh_rreg_mode,
2696 MDS_INODELOCK_LOOKUP,
2699 GOTO(out_put_new, rc);
2701 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2704 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2707 GOTO(out_unlock_old, rc);
2709 /* Check if @msrcdir is subdir of @mnew, before locking child
2710 * to avoid reverse locking.
2712 if (mtgtdir != msrcdir) {
2713 rc = mdo_is_subdir(info->mti_env,
2714 mdt_object_child(msrcdir), new_fid);
2718 GOTO(out_unlock_old, rc);
2722 /* We used to acquire MDS_INODELOCK_FULL here but we
2723 * can't do this now because a running HSM restore on
2724 * the rename onto victim will hold the layout
2725 * lock. See LU-4002.
2728 lh_newp = &info->mti_lh[MDT_LH_NEW];
2729 mdt_lock_reg_init(lh_newp, LCK_EX);
2730 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2731 MDS_INODELOCK_LOOKUP |
2732 MDS_INODELOCK_UPDATE,
2735 GOTO(out_unlock_old, rc);
2737 /* get and save version after locking */
2738 mdt_version_get_save(info, mnew, 3);
2739 } else if (rc != -EREMOTE && rc != -ENOENT) {
2740 GOTO(out_put_old, rc);
2742 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2743 mdt_lock_reg_init(lh_oldp, LCK_EX);
2744 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2745 if (mdt_object_remote(msrcdir)) {
2746 /* Enqueue lookup lock from the parent MDT */
2747 rc = mdt_remote_object_lock(info, msrcdir,
2748 mdt_object_fid(mold),
2749 &lh_oldp->mlh_rreg_lh,
2750 lh_oldp->mlh_rreg_mode,
2751 MDS_INODELOCK_LOOKUP,
2754 GOTO(out_put_old, rc);
2756 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2759 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2762 GOTO(out_unlock_old, rc);
2764 mdt_enoent_version_save(info, 3);
2767 /* step 5: rename it */
2768 mdt_reint_init_ma(info, ma);
2770 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2771 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2774 mutex_lock(&mnew->mot_lov_mutex);
2776 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2777 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2778 mnew != NULL ? mdt_object_child(mnew) : NULL,
2779 &rr->rr_tgt_name, ma);
2782 mutex_unlock(&mnew->mot_lov_mutex);
2784 /* handle last link of tgt object */
2786 mdt_counter_incr(req, LPROC_MDT_RENAME);
2788 mdt_handle_last_unlink(info, mnew, ma);
2789 discard = mdt_dom_check_for_discard(info, mnew);
2791 mdt_rename_counter_tally(info, info->mti_mdt, req,
2797 mdt_object_unlock(info, mnew, lh_newp, rc);
2799 mdt_object_unlock(info, mold, lh_oldp, rc);
2801 if (mnew && !discard)
2802 mdt_object_put(info->mti_env, mnew);
2804 mdt_object_put(info->mti_env, mold);
2806 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2807 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2809 if (lustre_handle_is_used(&rename_lh))
2810 mdt_rename_unlock(&rename_lh);
2812 mdt_object_put(info->mti_env, mtgtdir);
2814 mdt_object_put(info->mti_env, msrcdir);
2816 /* The DoM discard can be done right in the place above where it is
2817 * assigned, meanwhile it is done here after rename unlock due to
2818 * compatibility with old clients, for them the discard blocks
2819 * the main thread until completion. Check LU-11359 for details.
2822 mdt_dom_discard_data(info, mnew);
2823 mdt_object_put(info->mti_env, mnew);
2828 static int mdt_reint_resync(struct mdt_thread_info *info,
2829 struct mdt_lock_handle *lhc)
2831 struct mdt_reint_record *rr = &info->mti_rr;
2832 struct ptlrpc_request *req = mdt_info_req(info);
2833 struct md_attr *ma = &info->mti_attr;
2834 struct mdt_object *mo;
2835 struct ldlm_lock *lease;
2836 struct mdt_body *repbody;
2837 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2842 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2844 if (info->mti_dlm_req)
2845 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2847 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2849 GOTO(out, rc = PTR_ERR(mo));
2851 if (!mdt_object_exists(mo))
2852 GOTO(out_obj, rc = -ENOENT);
2854 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2855 GOTO(out_obj, rc = -EINVAL);
2857 if (mdt_object_remote(mo))
2858 GOTO(out_obj, rc = -EREMOTE);
2860 lease = ldlm_handle2lock(rr->rr_lease_handle);
2862 GOTO(out_obj, rc = -ESTALE);
2864 /* It's really necessary to grab open_sem and check if the lease lock
2865 * has been lost. There would exist a concurrent writer coming in and
2866 * generating some dirty data in memory cache, the writeback would fail
2867 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2869 if (!down_write_trylock(&mo->mot_open_sem))
2870 GOTO(out_put_lease, rc = -EBUSY);
2872 lock_res_and_lock(lease);
2873 lease_broken = ldlm_is_cancel(lease);
2874 unlock_res_and_lock(lease);
2876 GOTO(out_unlock, rc = -EBUSY);
2878 /* the file has yet opened by anyone else after we took the lease. */
2879 layout.mlc_opc = MD_LAYOUT_RESYNC;
2880 lhc = &info->mti_lh[MDT_LH_LOCAL];
2881 rc = mdt_layout_change(info, mo, lhc, &layout);
2883 GOTO(out_unlock, rc);
2885 mdt_object_unlock(info, mo, lhc, 0);
2887 ma->ma_need = MA_INODE;
2889 rc = mdt_attr_get_complex(info, mo, ma);
2891 GOTO(out_unlock, rc);
2893 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2894 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2898 up_write(&mo->mot_open_sem);
2900 LDLM_LOCK_PUT(lease);
2902 mdt_object_put(info->mti_env, mo);
2904 mdt_client_compatibility(info);
2905 rc2 = mdt_fix_reply(info);
2911 struct mdt_reinter {
2912 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2913 enum lprocfs_extra_opc mr_extra_opc;
2916 static const struct mdt_reinter mdt_reinters[] = {
2918 .mr_handler = &mdt_reint_setattr,
2919 .mr_extra_opc = MDS_REINT_SETATTR,
2922 .mr_handler = &mdt_reint_create,
2923 .mr_extra_opc = MDS_REINT_CREATE,
2926 .mr_handler = &mdt_reint_link,
2927 .mr_extra_opc = MDS_REINT_LINK,
2930 .mr_handler = &mdt_reint_unlink,
2931 .mr_extra_opc = MDS_REINT_UNLINK,
2934 .mr_handler = &mdt_reint_rename,
2935 .mr_extra_opc = MDS_REINT_RENAME,
2938 .mr_handler = &mdt_reint_open,
2939 .mr_extra_opc = MDS_REINT_OPEN,
2941 [REINT_SETXATTR] = {
2942 .mr_handler = &mdt_reint_setxattr,
2943 .mr_extra_opc = MDS_REINT_SETXATTR,
2946 .mr_handler = &mdt_reint_unlink,
2947 .mr_extra_opc = MDS_REINT_UNLINK,
2950 .mr_handler = &mdt_reint_migrate,
2951 .mr_extra_opc = MDS_REINT_RENAME,
2954 .mr_handler = &mdt_reint_resync,
2955 .mr_extra_opc = MDS_REINT_RESYNC,
2959 int mdt_reint_rec(struct mdt_thread_info *info,
2960 struct mdt_lock_handle *lhc)
2962 const struct mdt_reinter *mr;
2966 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2969 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2970 if (mr->mr_handler == NULL)
2973 rc = (*mr->mr_handler)(info, lhc);
2975 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2976 PTLRPC_LAST_CNTR + mr->mr_extra_opc);