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;
792 ktime_t kstart = ktime_get();
796 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
797 (unsigned int)ma->ma_attr.la_valid);
799 if (info->mti_dlm_req)
800 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
802 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
804 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
805 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
807 GOTO(out, rc = PTR_ERR(mo));
809 if (!mdt_object_exists(mo))
810 GOTO(out_put, rc = -ENOENT);
812 if (mdt_object_remote(mo))
813 GOTO(out_put, rc = -EREMOTE);
815 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
816 /* revoke lease lock if size is going to be changed */
817 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
818 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
819 atomic_read(&mo->mot_lease_count) > 0)) {
820 down_read(&mo->mot_open_sem);
822 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
823 lhc = &info->mti_lh[MDT_LH_LOCAL];
824 mdt_lock_reg_init(lhc, LCK_CW);
826 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
828 up_read(&mo->mot_open_sem);
832 /* revoke lease lock */
833 mdt_object_unlock(info, mo, lhc, 1);
835 up_read(&mo->mot_open_sem);
838 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
839 /* Check write access for the O_TRUNC case */
840 if (mdt_write_read(mo) < 0)
841 GOTO(out_put, rc = -ETXTBSY);
843 /* LU-10286: compatibility check for FLR.
844 * Please check the comment in mdt_finish_open() for details
846 if (!exp_connect_flr(info->mti_exp) ||
847 !exp_connect_overstriping(info->mti_exp)) {
848 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
849 if (rc < 0 && rc != -ENODATA)
852 if (!exp_connect_flr(info->mti_exp)) {
854 mdt_lmm_is_flr(info->mti_big_lmm))
855 GOTO(out_put, rc = -EOPNOTSUPP);
858 if (!exp_connect_overstriping(info->mti_exp)) {
860 mdt_lmm_is_overstriping(info->mti_big_lmm))
861 GOTO(out_put, rc = -EOPNOTSUPP);
865 /* For truncate, the file size sent from client
866 * is believable, but the blocks are incorrect,
867 * which makes the block size in LSOM attribute
868 * inconsisent with the real block size.
870 rc = mdt_lsom_update(info, mo, true);
875 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
876 if (ma->ma_valid & MA_LOV)
877 GOTO(out_put, rc = -EPROTO);
879 /* MDT supports FMD for regular files due to Data-on-MDT */
880 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
881 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
882 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
885 rc = mdt_attr_set(info, mo, ma);
888 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
889 (ma->ma_valid & MA_INODE)) {
890 struct lu_buf *buf = &info->mti_buf;
891 struct lu_ucred *uc = mdt_ucred(info);
892 struct mdt_lock_handle *lh;
894 __u64 lockpart = MDS_INODELOCK_XATTR;
896 /* reject if either remote or striped dir is disabled */
897 if (ma->ma_valid & MA_LMV) {
898 if (!mdt->mdt_enable_remote_dir ||
899 !mdt->mdt_enable_striped_dir)
900 GOTO(out_put, rc = -EPERM);
902 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
903 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
904 mdt->mdt_enable_remote_dir_gid != -1)
905 GOTO(out_put, rc = -EPERM);
908 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
909 GOTO(out_put, rc = -ENOTDIR);
911 if (ma->ma_attr.la_valid != 0)
912 GOTO(out_put, rc = -EPROTO);
914 if (ma->ma_valid & MA_LOV) {
915 buf->lb_buf = ma->ma_lmm;
916 buf->lb_len = ma->ma_lmm_size;
917 name = XATTR_NAME_LOV;
919 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
922 buf->lb_len = ma->ma_lmv_size;
923 name = XATTR_NAME_DEFAULT_LMV;
924 /* force client to update dir default layout */
925 lockpart |= MDS_INODELOCK_LOOKUP;
928 lh = &info->mti_lh[MDT_LH_PARENT];
929 mdt_lock_reg_init(lh, LCK_PW);
931 rc = mdt_object_lock(info, mo, lh, lockpart);
935 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
938 mdt_object_unlock(info, mo, lh, rc);
942 GOTO(out_put, rc = -EPROTO);
945 /* If file data is modified, add the dirty flag */
946 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
947 rc = mdt_add_dirty_flag(info, mo, ma);
949 ma->ma_need = MA_INODE;
951 rc = mdt_attr_get_complex(info, mo, ma);
955 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
959 mdt_object_put(info->mti_env, mo);
962 mdt_counter_incr(req, LPROC_MDT_SETATTR,
963 ktime_us_delta(ktime_get(), kstart));
965 mdt_client_compatibility(info);
966 rc2 = mdt_fix_reply(info);
972 static int mdt_reint_create(struct mdt_thread_info *info,
973 struct mdt_lock_handle *lhc)
975 struct ptlrpc_request *req = mdt_info_req(info);
976 ktime_t kstart = ktime_get();
980 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
981 RETURN(err_serious(-ESTALE));
983 if (info->mti_dlm_req)
984 ldlm_request_cancel(mdt_info_req(info),
985 info->mti_dlm_req, 0, LATF_SKIP);
987 if (!lu_name_is_valid(&info->mti_rr.rr_name))
990 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1000 CERROR("%s: Unsupported mode %o\n",
1001 mdt_obd_name(info->mti_mdt),
1002 info->mti_attr.ma_attr.la_mode);
1003 RETURN(err_serious(-EOPNOTSUPP));
1006 rc = mdt_create(info);
1008 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1009 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1010 ktime_us_delta(ktime_get(), kstart));
1012 /* Special file should stay on the same node as parent*/
1013 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1014 ktime_us_delta(ktime_get(), kstart));
1021 * VBR: save parent version in reply and child version getting by its name.
1022 * Version of child is getting and checking during its lookup. If
1024 static int mdt_reint_unlink(struct mdt_thread_info *info,
1025 struct mdt_lock_handle *lhc)
1027 struct mdt_reint_record *rr = &info->mti_rr;
1028 struct ptlrpc_request *req = mdt_info_req(info);
1029 struct md_attr *ma = &info->mti_attr;
1030 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1031 struct mdt_object *mp;
1032 struct mdt_object *mc;
1033 struct mdt_lock_handle *parent_lh;
1034 struct mdt_lock_handle *child_lh;
1035 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1037 bool cos_incompat = false;
1039 ktime_t kstart = ktime_get();
1043 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1044 PNAME(&rr->rr_name));
1046 if (info->mti_dlm_req)
1047 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1049 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1050 RETURN(err_serious(-ENOENT));
1052 if (!fid_is_md_operative(rr->rr_fid1))
1055 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1057 RETURN(PTR_ERR(mp));
1059 if (mdt_object_remote(mp)) {
1060 cos_incompat = true;
1062 rc = mdt_version_get_check_save(info, mp, 0);
1064 GOTO(put_parent, rc);
1068 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1069 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1070 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1073 GOTO(put_parent, rc);
1075 /* lookup child object along with version checking */
1076 fid_zero(child_fid);
1077 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1079 /* Name might not be able to find during resend of
1080 * remote unlink, considering following case.
1081 * dir_A is a remote directory, the name entry of
1082 * dir_A is on MDT0, the directory is on MDT1,
1084 * 1. client sends unlink req to MDT1.
1085 * 2. MDT1 sends name delete update to MDT0.
1086 * 3. name entry is being deleted in MDT0 synchronously.
1087 * 4. MDT1 is restarted.
1088 * 5. client resends unlink req to MDT1. So it can not
1089 * find the name entry on MDT0 anymore.
1090 * In this case, MDT1 only needs to destory the local
1093 if (mdt_object_remote(mp) && rc == -ENOENT &&
1094 !fid_is_zero(rr->rr_fid2) &&
1095 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1097 *child_fid = *rr->rr_fid2;
1099 GOTO(unlock_parent, rc);
1103 if (!fid_is_md_operative(child_fid))
1104 GOTO(unlock_parent, rc = -EPERM);
1106 /* We will lock the child regardless it is local or remote. No harm. */
1107 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1109 GOTO(unlock_parent, rc = PTR_ERR(mc));
1111 if (!cos_incompat) {
1112 rc = mdt_object_striped(info, mc);
1114 GOTO(put_child, rc);
1118 mdt_object_put(info->mti_env, mc);
1119 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1124 child_lh = &info->mti_lh[MDT_LH_CHILD];
1125 mdt_lock_reg_init(child_lh, LCK_EX);
1126 if (info->mti_spec.sp_rm_entry) {
1127 struct lu_ucred *uc = mdt_ucred(info);
1129 if (!mdt_is_dne_client(req->rq_export))
1130 /* Return -ENOTSUPP for old client */
1131 GOTO(put_child, rc = -ENOTSUPP);
1133 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
1134 GOTO(put_child, rc = -EPERM);
1136 ma->ma_need = MA_INODE;
1138 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1139 NULL, &rr->rr_name, ma, no_name);
1140 GOTO(put_child, rc);
1143 if (mdt_object_remote(mc)) {
1144 struct mdt_body *repbody;
1146 if (!fid_is_zero(rr->rr_fid2)) {
1147 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1148 mdt_obd_name(info->mti_mdt),
1149 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1150 GOTO(put_child, rc = -ENOENT);
1152 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1153 mdt_obd_name(info->mti_mdt),
1154 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1156 if (!mdt_is_dne_client(req->rq_export))
1157 /* Return -ENOTSUPP for old client */
1158 GOTO(put_child, rc = -ENOTSUPP);
1160 /* Revoke the LOOKUP lock of the remote object granted by
1161 * this MDT. Since the unlink will happen on another MDT,
1162 * it will release the LOOKUP lock right away. Then What
1163 * would happen if another client try to grab the LOOKUP
1164 * lock at the same time with unlink XXX
1166 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1167 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1168 LASSERT(repbody != NULL);
1169 repbody->mbo_fid1 = *mdt_object_fid(mc);
1170 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1171 GOTO(unlock_child, rc = -EREMOTE);
1173 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1174 * this now because a running HSM restore on the child (unlink
1175 * victim) will hold the layout lock. See LU-4002.
1177 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1178 if (mdt_object_remote(mp)) {
1179 /* Enqueue lookup lock from parent MDT */
1180 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1181 &child_lh->mlh_rreg_lh,
1182 child_lh->mlh_rreg_mode,
1183 MDS_INODELOCK_LOOKUP, false);
1185 GOTO(put_child, rc);
1187 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1190 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1193 GOTO(put_child, rc);
1196 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1197 * whether need MA_LOV and MA_COOKIE.
1199 ma->ma_need = MA_INODE;
1202 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1203 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1204 /* save version when object is locked */
1205 mdt_version_get_save(info, mc, 1);
1207 mutex_lock(&mc->mot_lov_mutex);
1209 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1210 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1212 mutex_unlock(&mc->mot_lov_mutex);
1214 GOTO(unlock_child, rc);
1216 if (!lu_object_is_dying(&mc->mot_header)) {
1217 rc = mdt_attr_get_complex(info, mc, ma);
1220 } else if (mdt_dom_check_for_discard(info, mc)) {
1221 mdt_dom_discard_data(info, mc);
1223 mdt_handle_last_unlink(info, mc, ma);
1226 if (ma->ma_valid & MA_INODE) {
1227 switch (ma->ma_attr.la_mode & S_IFMT) {
1229 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1230 ktime_us_delta(ktime_get(), kstart));
1238 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1239 ktime_us_delta(ktime_get(), kstart));
1242 LASSERTF(0, "bad file type %o unlinking\n",
1243 ma->ma_attr.la_mode);
1250 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1252 mdt_object_put(info->mti_env, mc);
1254 mdt_object_unlock(info, mp, parent_lh, rc);
1256 mdt_object_put(info->mti_env, mp);
1257 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1262 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1265 static int mdt_reint_link(struct mdt_thread_info *info,
1266 struct mdt_lock_handle *lhc)
1268 struct mdt_reint_record *rr = &info->mti_rr;
1269 struct ptlrpc_request *req = mdt_info_req(info);
1270 struct md_attr *ma = &info->mti_attr;
1271 struct mdt_object *ms;
1272 struct mdt_object *mp;
1273 struct mdt_lock_handle *lhs;
1274 struct mdt_lock_handle *lhp;
1275 ktime_t kstart = ktime_get();
1280 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1281 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1283 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1284 RETURN(err_serious(-ENOENT));
1286 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1287 req->rq_no_reply = 1;
1288 RETURN(err_serious(-ENOENT));
1291 if (info->mti_dlm_req)
1292 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1294 /* Invalid case so return error immediately instead of
1297 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1300 if (!fid_is_md_operative(rr->rr_fid1) ||
1301 !fid_is_md_operative(rr->rr_fid2))
1304 /* step 1: find target parent dir */
1305 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1307 RETURN(PTR_ERR(mp));
1309 rc = mdt_version_get_check_save(info, mp, 0);
1311 GOTO(put_parent, rc);
1313 /* step 2: find source */
1314 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1316 GOTO(put_parent, rc = PTR_ERR(ms));
1318 if (!mdt_object_exists(ms)) {
1319 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1320 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1321 GOTO(put_source, rc = -ENOENT);
1324 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1326 lhp = &info->mti_lh[MDT_LH_PARENT];
1327 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1328 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1331 GOTO(put_source, rc);
1333 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1335 lhs = &info->mti_lh[MDT_LH_CHILD];
1336 mdt_lock_reg_init(lhs, LCK_EX);
1337 rc = mdt_reint_object_lock(info, ms, lhs,
1338 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1341 GOTO(unlock_parent, rc);
1343 /* step 3: link it */
1344 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1345 OBD_FAIL_MDS_REINT_LINK_WRITE);
1347 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1348 rc = mdt_version_get_check_save(info, ms, 1);
1350 GOTO(unlock_source, rc);
1352 /** check target version by name during replay */
1353 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1354 &info->mti_tmp_fid1, 2);
1355 if (rc != 0 && rc != -ENOENT)
1356 GOTO(unlock_source, rc);
1357 /* save version of file name for replay, it must be ENOENT here */
1358 if (!req_is_replay(mdt_info_req(info))) {
1359 if (rc != -ENOENT) {
1360 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1361 PNAME(&rr->rr_name));
1362 GOTO(unlock_source, rc = -EEXIST);
1364 info->mti_ver[2] = ENOENT_VERSION;
1365 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1368 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1369 mdt_object_child(ms), &rr->rr_name, ma);
1372 mdt_counter_incr(req, LPROC_MDT_LINK,
1373 ktime_us_delta(ktime_get(), kstart));
1377 mdt_object_unlock(info, ms, lhs, rc);
1379 mdt_object_unlock(info, mp, lhp, rc);
1381 mdt_object_put(info->mti_env, ms);
1383 mdt_object_put(info->mti_env, mp);
1387 * lock the part of the directory according to the hash of the name
1388 * (lh->mlh_pdo_hash) in parallel directory lock.
1390 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1391 struct mdt_lock_handle *lh,
1392 struct mdt_object *obj, __u64 ibits,
1395 struct ldlm_res_id *res = &info->mti_res_id;
1396 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1397 union ldlm_policy_data *policy = &info->mti_policy;
1398 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1402 * Finish res_id initializing by name hash marking part of
1403 * directory which is taking modification.
1405 LASSERT(lh->mlh_pdo_hash != 0);
1406 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1407 memset(policy, 0, sizeof(*policy));
1408 policy->l_inodebits.bits = ibits;
1410 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1411 dlmflags |= LDLM_FL_COS_INCOMPAT;
1413 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1414 * going to be sent to client. If it is - mdt_intent_policy() path will
1415 * fix it up and turn FL_LOCAL flag off.
1417 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1418 policy, res, dlmflags,
1419 &info->mti_exp->exp_handle.h_cookie);
1424 * Get BFL lock for rename or migrate process.
1426 static int mdt_rename_lock(struct mdt_thread_info *info,
1427 struct lustre_handle *lh)
1432 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1433 struct lu_fid *fid = &info->mti_tmp_fid1;
1434 struct mdt_object *obj;
1436 /* XXX, right now, it has to use object API to
1437 * enqueue lock cross MDT, so it will enqueue
1438 * rename lock(with LUSTRE_BFL_FID) by root object
1441 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1443 RETURN(PTR_ERR(obj));
1445 rc = mdt_remote_object_lock(info, obj,
1446 &LUSTRE_BFL_FID, lh,
1448 MDS_INODELOCK_UPDATE, false);
1449 mdt_object_put(info->mti_env, obj);
1451 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1452 union ldlm_policy_data *policy = &info->mti_policy;
1453 struct ldlm_res_id *res_id = &info->mti_res_id;
1456 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1457 memset(policy, 0, sizeof(*policy));
1458 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1459 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1460 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1461 LDLM_IBITS, policy, LCK_EX, &flags,
1463 ldlm_completion_ast, NULL, NULL, 0,
1465 &info->mti_exp->exp_handle.h_cookie,
1472 static void mdt_rename_unlock(struct lustre_handle *lh)
1475 LASSERT(lustre_handle_is_used(lh));
1476 /* Cancel the single rename lock right away */
1477 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1481 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1482 const struct lu_fid *fid,
1485 struct mdt_object *dir;
1489 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1493 /* check early, the real version will be saved after locking */
1494 rc = mdt_version_get_check(info, dir, idx);
1498 if (!mdt_object_exists(dir))
1499 GOTO(out_put, rc = -ENOENT);
1501 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1502 GOTO(out_put, rc = -ENOTDIR);
1506 mdt_object_put(info->mti_env, dir);
1511 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1512 * herein we don't really check it, just do revoke.
1514 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1515 struct mdt_object *pobj,
1516 struct mdt_object *obj)
1518 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1521 mdt_lock_handle_init(lh);
1522 mdt_lock_reg_init(lh, LCK_EX);
1524 if (mdt_object_remote(pobj)) {
1525 /* don't bother to check if pobj and obj are on the same MDT. */
1526 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1527 &lh->mlh_rreg_lh, LCK_EX,
1528 MDS_INODELOCK_LOOKUP, false);
1529 } else if (mdt_object_remote(obj)) {
1530 struct ldlm_res_id *res = &info->mti_res_id;
1531 union ldlm_policy_data *policy = &info->mti_policy;
1532 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1533 LDLM_FL_COS_INCOMPAT;
1535 fid_build_reg_res_name(mdt_object_fid(obj), res);
1536 memset(policy, 0, sizeof(*policy));
1537 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1538 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1539 &lh->mlh_reg_lh, LCK_EX, policy, res,
1542 /* do nothing if both are local */
1550 * TODO, currently we don't save this lock because there is no place to
1551 * hold this lock handle, but to avoid race we need to save this lock.
1553 mdt_object_unlock(info, NULL, lh, 1);
1559 * operation may takes locks of linkea, or directory stripes, group them in
1562 struct mdt_sub_lock {
1563 struct mdt_object *msl_obj;
1564 struct mdt_lock_handle msl_lh;
1565 struct list_head msl_linkage;
1568 static void mdt_unlock_list(struct mdt_thread_info *info,
1569 struct list_head *list, int decref)
1571 struct mdt_sub_lock *msl;
1572 struct mdt_sub_lock *tmp;
1574 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1575 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1576 list_del(&msl->msl_linkage);
1581 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1582 struct mdt_object *obj,
1583 struct mdt_lock_handle *lh,
1584 struct ldlm_enqueue_info *einfo,
1585 struct list_head *slave_locks,
1588 if (mdt_object_remote(obj)) {
1589 mdt_unlock_list(info, slave_locks, decref);
1590 mdt_object_unlock(info, obj, lh, decref);
1592 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1597 * lock parents of links, and also check whether total locks don't exceed
1600 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1601 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1602 * \retval -ev negative errno upon error
1604 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1605 struct mdt_object *pobj,
1606 const struct md_attr *ma,
1607 struct mdt_object *obj,
1608 struct mdt_lock_handle *lhp,
1609 struct ldlm_enqueue_info *peinfo,
1610 struct list_head *parent_slave_locks,
1611 struct list_head *link_locks)
1613 struct mdt_device *mdt = info->mti_mdt;
1614 struct lu_buf *buf = &info->mti_big_buf;
1615 struct lu_name *lname = &info->mti_name;
1616 struct linkea_data ldata = { NULL };
1617 bool blocked = false;
1618 int local_lnkp_cnt = 0;
1622 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1625 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1626 if (buf->lb_buf == NULL)
1630 rc = mdt_links_read(info, obj, &ldata);
1632 if (rc == -ENOENT || rc == -ENODATA)
1637 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1638 linkea_next_entry(&ldata)) {
1639 struct mdt_object *lnkp;
1640 struct mdt_sub_lock *msl;
1644 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1647 /* check if it's also linked to parent */
1648 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1649 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1650 PFID(&fid), PNAME(lname));
1651 /* in case link is remote object, revoke LOOKUP lock */
1652 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1658 /* check if it's linked to a stripe of parent */
1659 if (ma->ma_valid & MA_LMV) {
1660 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1661 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1664 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1665 fid_le_to_cpu(stripe_fid,
1666 &lmv->lmv_stripe_fids[j]);
1667 if (lu_fid_eq(stripe_fid, &fid)) {
1668 CDEBUG(D_INFO, "skip stripe "DFID
1669 ", reovke "DNAME"\n",
1670 PFID(&fid), PNAME(lname));
1671 lnkp = mdt_object_find(info->mti_env,
1674 GOTO(out, rc = PTR_ERR(lnkp));
1680 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1682 mdt_object_put(info->mti_env, lnkp);
1687 /* Check if it's already locked */
1688 list_for_each_entry(msl, link_locks, msl_linkage) {
1689 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1691 DFID" was locked, revoke "DNAME"\n",
1692 PFID(&fid), PNAME(lname));
1693 lnkp = msl->msl_obj;
1699 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1703 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1704 PFID(&fid), PNAME(lname));
1706 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1708 CWARN("%s: cannot find obj "DFID": %ld\n",
1709 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1713 if (!mdt_object_exists(lnkp)) {
1714 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1715 PFID(&fid), PNAME(lname));
1716 mdt_object_put(info->mti_env, lnkp);
1720 if (!mdt_object_remote(lnkp))
1725 GOTO(out, rc = -ENOMEM);
1728 * we can't follow parent-child lock order like other MD
1729 * operations, use lock_try here to avoid deadlock, if the lock
1730 * cannot be taken, drop all locks taken, revoke the blocked
1731 * one, and continue processing the remaining entries, and in
1732 * the end of the loop restart from beginning.
1734 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1736 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1737 MDS_INODELOCK_UPDATE, true);
1738 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1740 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1741 PFID(&fid), PNAME(lname));
1743 mdt_unlock_list(info, link_locks, 1);
1744 /* also unlock parent locks to avoid deadlock */
1746 mdt_migrate_object_unlock(info, pobj, lhp,
1753 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1754 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1755 MDS_INODELOCK_UPDATE);
1757 mdt_object_put(info->mti_env, lnkp);
1762 if (mdt_object_remote(lnkp)) {
1763 struct ldlm_lock *lock;
1766 * for remote object, set lock cb_atomic,
1767 * so lock can be released in blocking_ast()
1768 * immediately, then the next lock_try will
1769 * have better chance of success.
1771 lock = ldlm_handle2lock(
1772 &msl->msl_lh.mlh_rreg_lh);
1773 LASSERT(lock != NULL);
1774 lock_res_and_lock(lock);
1775 ldlm_set_atomic_cb(lock);
1776 unlock_res_and_lock(lock);
1777 LDLM_LOCK_PUT(lock);
1780 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1785 INIT_LIST_HEAD(&msl->msl_linkage);
1786 msl->msl_obj = lnkp;
1787 list_add_tail(&msl->msl_linkage, link_locks);
1789 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1793 GOTO(out, rc = -EBUSY);
1798 mdt_unlock_list(info, link_locks, rc);
1799 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1801 * parent may have 3 local objects: master object and 2 stripes
1802 * (if it's being migrated too); source may have 2 local
1803 * objects: master and 1 stripe; target has 1 local object.
1809 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1810 struct mdt_object *obj,
1811 const struct md_attr *ma,
1812 struct list_head *slave_locks)
1814 struct mdt_device *mdt = info->mti_mdt;
1815 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1816 struct lu_fid *fid = &info->mti_tmp_fid1;
1817 struct mdt_object *slave;
1818 struct mdt_sub_lock *msl;
1823 LASSERT(mdt_object_remote(obj));
1824 LASSERT(ma->ma_valid & MA_LMV);
1827 if (!lmv_is_sane(lmv))
1830 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1831 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1833 if (!fid_is_sane(fid))
1836 slave = mdt_object_find(info->mti_env, mdt, fid);
1838 GOTO(out, rc = PTR_ERR(slave));
1842 mdt_object_put(info->mti_env, slave);
1843 GOTO(out, rc = -ENOMEM);
1846 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1847 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1848 MDS_INODELOCK_UPDATE, true);
1851 mdt_object_put(info->mti_env, slave);
1855 INIT_LIST_HEAD(&msl->msl_linkage);
1856 msl->msl_obj = slave;
1857 list_add_tail(&msl->msl_linkage, slave_locks);
1863 mdt_unlock_list(info, slave_locks, rc);
1867 /* lock parent and its stripes */
1868 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1869 struct mdt_object *obj,
1870 const struct md_attr *ma,
1871 struct mdt_lock_handle *lh,
1872 struct ldlm_enqueue_info *einfo,
1873 struct list_head *slave_locks)
1877 if (mdt_object_remote(obj)) {
1878 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1879 &lh->mlh_rreg_lh, LCK_PW,
1880 MDS_INODELOCK_UPDATE, false);
1885 * if obj is remote and striped, lock its stripes explicitly
1886 * because it's not striped in LOD layer on this MDT.
1888 if (ma->ma_valid & MA_LMV) {
1889 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1891 mdt_object_unlock(info, obj, lh, rc);
1894 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1902 * in migration, object may be remote, and we need take full lock of it and its
1903 * stripes if it's directory, besides, object may be a remote object on its
1904 * parent, revoke its LOOKUP lock on where its parent is located.
1906 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1907 struct mdt_object *pobj,
1908 struct mdt_object *obj,
1909 struct mdt_lock_handle *lh,
1910 struct ldlm_enqueue_info *einfo,
1911 struct list_head *slave_locks)
1915 if (mdt_object_remote(obj)) {
1916 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1920 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1921 &lh->mlh_rreg_lh, LCK_EX,
1922 MDS_INODELOCK_FULL, false);
1927 * if obj is remote and striped, lock its stripes explicitly
1928 * because it's not striped in LOD layer on this MDT.
1930 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1931 struct md_attr *ma = &info->mti_attr;
1933 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1935 mdt_object_unlock(info, obj, lh, rc);
1939 if (ma->ma_valid & MA_LMV) {
1940 rc = mdt_lock_remote_slaves(info, obj, ma,
1943 mdt_object_unlock(info, obj, lh, rc);
1947 if (mdt_object_remote(pobj)) {
1948 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1953 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1961 * lookup source by name, if parent is striped directory, we need to find the
1962 * corresponding stripe where source is located, and then lookup there.
1964 * besides, if parent is migrating too, and file is already in target stripe,
1965 * this should be a redo of 'lfs migrate' on client side.
1967 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1968 struct mdt_object *pobj,
1969 const struct md_attr *ma,
1970 const struct lu_name *lname,
1971 struct mdt_object **spobj,
1972 struct mdt_object **sobj)
1974 const struct lu_env *env = info->mti_env;
1975 struct lu_fid *fid = &info->mti_tmp_fid1;
1976 struct mdt_object *stripe;
1979 if (ma->ma_valid & MA_LMV) {
1980 /* if parent is striped, lookup on corresponding stripe */
1981 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1983 if (!lmv_is_sane(lmv))
1986 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1991 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1993 stripe = mdt_object_find(env, info->mti_mdt, fid);
1995 return PTR_ERR(stripe);
1998 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2000 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2002 * if parent layout is changeing, and lookup child
2003 * failed on source stripe, lookup again on target
2004 * stripe, if it exists, it means previous migration
2005 * was interrupted, and current file was migrated
2008 mdt_object_put(env, stripe);
2010 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2015 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2017 stripe = mdt_object_find(env, info->mti_mdt, fid);
2019 return PTR_ERR(stripe);
2022 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2023 fid, &info->mti_spec);
2024 mdt_object_put(env, stripe);
2025 return rc ?: -EALREADY;
2027 mdt_object_put(env, stripe);
2032 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2038 mdt_object_get(env, stripe);
2043 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2044 if (IS_ERR(*sobj)) {
2045 mdt_object_put(env, stripe);
2046 rc = PTR_ERR(*sobj);
2054 /* end lease and close file for regular file */
2055 static int mdd_migrate_close(struct mdt_thread_info *info,
2056 struct mdt_object *obj)
2058 struct close_data *data;
2059 struct mdt_body *repbody;
2060 struct ldlm_lock *lease;
2065 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2067 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2071 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2076 lease = ldlm_handle2lock(&data->cd_handle);
2080 /* check if the lease was already canceled */
2081 lock_res_and_lock(lease);
2082 rc = ldlm_is_cancel(lease);
2083 unlock_res_and_lock(lease);
2087 LDLM_DEBUG(lease, DFID" lease broken",
2088 PFID(mdt_object_fid(obj)));
2092 * cancel server side lease, client side counterpart should have been
2093 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2095 ldlm_lock_cancel(lease);
2096 ldlm_reprocess_all(lease->l_resource, lease);
2097 LDLM_LOCK_PUT(lease);
2100 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2101 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2102 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2108 * migrate file in below steps:
2109 * 1. lock parent and its stripes
2110 * 2. lookup source by name
2111 * 3. lock parents of source links if source is not directory
2112 * 4. reject if source is in HSM
2113 * 5. take source open_sem and close file if source is regular file
2114 * 6. lock source and its stripes if it's directory
2115 * 7. lock target so subsequent change to it can trigger COS
2117 * 9. unlock above locks
2118 * 10. sync device if source has links
2120 int mdt_reint_migrate(struct mdt_thread_info *info,
2121 struct mdt_lock_handle *unused)
2123 const struct lu_env *env = info->mti_env;
2124 struct mdt_device *mdt = info->mti_mdt;
2125 struct ptlrpc_request *req = mdt_info_req(info);
2126 struct mdt_reint_record *rr = &info->mti_rr;
2127 struct lu_ucred *uc = mdt_ucred(info);
2128 struct md_attr *ma = &info->mti_attr;
2129 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2130 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2131 struct mdt_object *pobj;
2132 struct mdt_object *spobj = NULL;
2133 struct mdt_object *sobj = NULL;
2134 struct mdt_object *tobj;
2135 struct lustre_handle rename_lh = { 0 };
2136 struct mdt_lock_handle *lhp;
2137 struct mdt_lock_handle *lhs;
2138 struct mdt_lock_handle *lht;
2139 LIST_HEAD(parent_slave_locks);
2140 LIST_HEAD(child_slave_locks);
2141 LIST_HEAD(link_locks);
2142 int lock_retries = 5;
2143 bool open_sem_locked = false;
2144 bool do_sync = false;
2148 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2149 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2151 if (info->mti_dlm_req)
2152 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2154 if (!fid_is_md_operative(rr->rr_fid1) ||
2155 !fid_is_md_operative(rr->rr_fid2))
2158 /* don't allow migrate . or .. */
2159 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2162 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2165 if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2166 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2167 mdt->mdt_enable_remote_dir_gid != -1)
2171 * Note: do not enqueue rename lock for replay request, because
2172 * if other MDT holds rename lock, but being blocked to wait for
2173 * this MDT to finish its recovery, and the failover MDT can not
2174 * get rename lock, which will cause deadlock.
2176 * req is NULL if this is called by directory auto-split.
2178 if (req && !req_is_replay(req)) {
2179 rc = mdt_rename_lock(info, &rename_lh);
2181 CERROR("%s: can't lock FS for rename: rc = %d\n",
2182 mdt_obd_name(info->mti_mdt), rc);
2187 /* pobj is master object of parent */
2188 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2190 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2193 rc = mdt_version_get_check(info, pobj, 0);
2195 GOTO(put_parent, rc);
2198 if (!mdt_object_exists(pobj))
2199 GOTO(put_parent, rc = -ENOENT);
2201 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2202 GOTO(put_parent, rc = -ENOTDIR);
2204 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2206 GOTO(put_parent, rc);
2209 /* lock parent object */
2210 lhp = &info->mti_lh[MDT_LH_PARENT];
2211 mdt_lock_reg_init(lhp, LCK_PW);
2212 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2213 &parent_slave_locks);
2215 GOTO(put_parent, rc);
2218 * spobj is the corresponding stripe against name if pobj is striped
2219 * directory, which is the real parent, and no need to lock, because
2220 * we've taken full lock of pobj.
2222 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2224 GOTO(unlock_parent, rc);
2226 /* lock parents of source links, and revoke LOOKUP lock of links */
2227 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2228 &parent_slave_locks, &link_locks);
2229 if (rc == -EBUSY && lock_retries-- > 0) {
2230 mdt_object_put(env, sobj);
2231 mdt_object_put(env, spobj);
2236 GOTO(put_source, rc);
2239 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2240 * with one request, if total lock count exceeds this limit, we will
2241 * drop all locks after migration, and synchronous device in the end.
2245 /* TODO: DoM migration is not supported, migrate dirent only */
2246 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2247 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2249 GOTO(unlock_links, rc);
2251 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2252 info->mti_spec.sp_migrate_nsonly = 1;
2255 /* if migration HSM is allowed */
2256 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2257 ma->ma_need = MA_HSM;
2259 rc = mdt_attr_get_complex(info, sobj, ma);
2261 GOTO(unlock_links, rc);
2263 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2264 GOTO(unlock_links, rc = -EOPNOTSUPP);
2267 /* end lease and close file for regular file */
2268 if (info->mti_spec.sp_migrate_close) {
2269 /* try to hold open_sem so that nobody else can open the file */
2270 if (!down_write_trylock(&sobj->mot_open_sem)) {
2272 mdd_migrate_close(info, sobj);
2273 GOTO(unlock_links, rc = -EBUSY);
2275 open_sem_locked = true;
2276 rc = mdd_migrate_close(info, sobj);
2278 GOTO(unlock_open_sem, rc);
2283 lhs = &info->mti_lh[MDT_LH_OLD];
2284 mdt_lock_reg_init(lhs, LCK_EX);
2285 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2286 &child_slave_locks);
2288 GOTO(unlock_open_sem, rc);
2291 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2293 GOTO(unlock_source, rc = PTR_ERR(tobj));
2295 lht = &info->mti_lh[MDT_LH_NEW];
2296 mdt_lock_reg_init(lht, LCK_EX);
2297 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2299 GOTO(put_target, rc);
2301 /* Don't do lookup sanity check. We know name doesn't exist. */
2302 info->mti_spec.sp_cr_lookup = 0;
2303 info->mti_spec.sp_feat = &dt_directory_features;
2305 rc = mdo_migrate(env, mdt_object_child(pobj),
2306 mdt_object_child(sobj), &rr->rr_name,
2307 mdt_object_child(tobj),
2308 &info->mti_spec, ma);
2310 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2311 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2314 mdt_object_unlock(info, tobj, lht, rc);
2316 mdt_object_put(env, tobj);
2318 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2319 &child_slave_locks, rc);
2321 if (open_sem_locked)
2322 up_write(&sobj->mot_open_sem);
2324 /* if we've got too many locks to save into RPC,
2325 * then just commit before the locks are released
2328 mdt_device_sync(env, mdt);
2329 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2331 mdt_object_put(env, sobj);
2332 mdt_object_put(env, spobj);
2334 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2335 &parent_slave_locks, rc);
2337 mdt_object_put(env, pobj);
2339 if (lustre_handle_is_used(&rename_lh))
2340 mdt_rename_unlock(&rename_lh);
2345 static int mdt_object_lock_save(struct mdt_thread_info *info,
2346 struct mdt_object *dir,
2347 struct mdt_lock_handle *lh,
2348 int idx, bool cos_incompat)
2352 /* we lock the target dir if it is local */
2353 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2358 /* get and save correct version after locking */
2359 mdt_version_get_save(info, dir, idx);
2364 * determine lock order of sobj and tobj
2366 * there are two situations we need to lock tobj before sobj:
2367 * 1. sobj is child of tobj
2368 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2369 * larger than that of tobj
2371 * \retval 1 lock tobj before sobj
2372 * \retval 0 lock sobj before tobj
2373 * \retval -ev negative errno upon error
2375 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2376 struct mdt_object *sobj,
2377 struct mdt_object *tobj)
2379 struct md_attr *ma = &info->mti_attr;
2380 struct lu_fid *spfid = &info->mti_tmp_fid1;
2381 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2382 struct lmv_mds_md_v1 *lmv;
2387 /* sobj and tobj are the same */
2391 if (fid_is_root(mdt_object_fid(sobj)))
2394 if (fid_is_root(mdt_object_fid(tobj)))
2397 /* check whether sobj is child of tobj */
2398 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2399 mdt_object_fid(tobj));
2406 /* check whether sobj and tobj are children of the same parent */
2407 rc = mdt_attr_get_pfid(info, sobj, spfid);
2411 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2415 if (!lu_fid_eq(spfid, tpfid))
2418 /* check whether sobj and tobj are sibling stripes */
2419 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2423 if (!(ma->ma_valid & MA_LMV))
2426 lmv = &ma->ma_lmv->lmv_md_v1;
2427 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2429 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2432 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2436 if (!(ma->ma_valid & MA_LMV))
2439 lmv = &ma->ma_lmv->lmv_md_v1;
2440 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2442 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2444 /* check stripe index of sobj and tobj */
2445 if (sindex == tindex)
2448 return sindex < tindex ? 0 : 1;
2452 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2453 * 2 - srcdir child; 3 - tgtdir child.
2454 * Update on disk version of srcdir child.
2456 static int mdt_reint_rename(struct mdt_thread_info *info,
2457 struct mdt_lock_handle *unused)
2459 struct mdt_device *mdt = info->mti_mdt;
2460 struct mdt_reint_record *rr = &info->mti_rr;
2461 struct md_attr *ma = &info->mti_attr;
2462 struct ptlrpc_request *req = mdt_info_req(info);
2463 struct mdt_object *msrcdir = NULL;
2464 struct mdt_object *mtgtdir = NULL;
2465 struct mdt_object *mold;
2466 struct mdt_object *mnew = NULL;
2467 struct lustre_handle rename_lh = { 0 };
2468 struct mdt_lock_handle *lh_srcdirp;
2469 struct mdt_lock_handle *lh_tgtdirp;
2470 struct mdt_lock_handle *lh_oldp = NULL;
2471 struct mdt_lock_handle *lh_newp = NULL;
2472 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2473 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2475 bool reverse = false, discard = false;
2477 ktime_t kstart = ktime_get();
2481 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2482 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2483 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2485 if (info->mti_dlm_req)
2486 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2488 if (!fid_is_md_operative(rr->rr_fid1) ||
2489 !fid_is_md_operative(rr->rr_fid2))
2492 /* find both parents. */
2493 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2494 if (IS_ERR(msrcdir))
2495 RETURN(PTR_ERR(msrcdir));
2497 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2499 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2501 mdt_object_get(info->mti_env, mtgtdir);
2503 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2504 if (IS_ERR(mtgtdir))
2505 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2509 * Note: do not enqueue rename lock for replay request, because
2510 * if other MDT holds rename lock, but being blocked to wait for
2511 * this MDT to finish its recovery, and the failover MDT can not
2512 * get rename lock, which will cause deadlock.
2514 if (!req_is_replay(req)) {
2516 * Normally rename RPC is handled on the MDT with the target
2517 * directory (if target exists, it's on the MDT with the
2518 * target), if the source directory is remote, it's a hint that
2519 * source is remote too (this may not be true, but it won't
2520 * cause any issue), return -EXDEV early to avoid taking
2523 if (!mdt->mdt_enable_remote_rename &&
2524 mdt_object_remote(msrcdir))
2525 GOTO(out_put_tgtdir, rc = -EXDEV);
2527 rc = mdt_rename_lock(info, &rename_lh);
2529 CERROR("%s: can't lock FS for rename: rc = %d\n",
2530 mdt_obd_name(mdt), rc);
2531 GOTO(out_put_tgtdir, rc);
2535 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2537 GOTO(out_unlock_rename, rc);
2541 /* source needs to be looked up after locking source parent, otherwise
2542 * this rename may race with unlink source, and cause rename hang, see
2543 * sanityn.sh 55b, so check parents first, if later we found source is
2544 * remote, relock parents.
2546 cos_incompat = (mdt_object_remote(msrcdir) ||
2547 mdt_object_remote(mtgtdir));
2549 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2551 /* lock parents in the proper order. */
2552 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2553 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2556 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2557 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2560 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2563 GOTO(out_unlock_rename, rc);
2565 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2567 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2570 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2571 GOTO(out_unlock_rename, rc);
2574 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2577 GOTO(out_unlock_rename, rc);
2579 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2581 if (mtgtdir != msrcdir) {
2582 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2584 } else if (!mdt_object_remote(mtgtdir) &&
2585 lh_srcdirp->mlh_pdo_hash !=
2586 lh_tgtdirp->mlh_pdo_hash) {
2587 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2588 MDS_INODELOCK_UPDATE,
2590 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2593 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2594 GOTO(out_unlock_rename, rc);
2598 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2599 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2601 /* find mold object. */
2603 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2605 GOTO(out_unlock_parents, rc);
2607 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2608 GOTO(out_unlock_parents, rc = -EINVAL);
2610 if (!fid_is_md_operative(old_fid))
2611 GOTO(out_unlock_parents, rc = -EPERM);
2613 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2615 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2617 if (!mdt_object_exists(mold)) {
2618 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2620 "object does not exist");
2621 GOTO(out_put_old, rc = -ENOENT);
2624 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2625 GOTO(out_put_old, rc = -EXDEV);
2627 /* Check if @mtgtdir is subdir of @mold, before locking child
2628 * to avoid reverse locking.
2630 if (mtgtdir != msrcdir) {
2631 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2636 GOTO(out_put_old, rc);
2640 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2641 /* save version after locking */
2642 mdt_version_get_save(info, mold, 2);
2644 if (!cos_incompat && mdt_object_remote(mold)) {
2645 cos_incompat = true;
2646 mdt_object_put(info->mti_env, mold);
2647 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2648 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2652 /* find mnew object:
2653 * mnew target object may not exist now
2654 * lookup with version checking
2657 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2660 /* the new_fid should have been filled at this moment */
2661 if (lu_fid_eq(old_fid, new_fid))
2662 GOTO(out_put_old, rc);
2664 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2665 lu_fid_eq(new_fid, rr->rr_fid2))
2666 GOTO(out_put_old, rc = -EINVAL);
2668 if (!fid_is_md_operative(new_fid))
2669 GOTO(out_put_old, rc = -EPERM);
2671 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2673 GOTO(out_put_old, rc = PTR_ERR(mnew));
2675 if (!mdt_object_exists(mnew)) {
2676 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2678 "object does not exist");
2679 GOTO(out_put_new, rc = -ENOENT);
2682 if (mdt_object_remote(mnew)) {
2683 struct mdt_body *repbody;
2685 /* Always send rename req to the target child MDT */
2686 repbody = req_capsule_server_get(info->mti_pill,
2688 LASSERT(repbody != NULL);
2689 repbody->mbo_fid1 = *new_fid;
2690 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2691 GOTO(out_put_new, rc = -EXDEV);
2693 /* Before locking the target dir, check we do not replace
2694 * a dir with a non-dir, otherwise it may deadlock with
2695 * link op which tries to create a link in this dir
2696 * back to this non-dir.
2698 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2699 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2700 GOTO(out_put_new, rc = -EISDIR);
2702 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2703 mdt_lock_reg_init(lh_oldp, LCK_EX);
2704 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2705 if (mdt_object_remote(msrcdir)) {
2706 /* Enqueue lookup lock from the parent MDT */
2707 rc = mdt_remote_object_lock(info, msrcdir,
2708 mdt_object_fid(mold),
2709 &lh_oldp->mlh_rreg_lh,
2710 lh_oldp->mlh_rreg_mode,
2711 MDS_INODELOCK_LOOKUP,
2714 GOTO(out_put_new, rc);
2716 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2719 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2722 GOTO(out_unlock_old, rc);
2724 /* Check if @msrcdir is subdir of @mnew, before locking child
2725 * to avoid reverse locking.
2727 if (mtgtdir != msrcdir) {
2728 rc = mdo_is_subdir(info->mti_env,
2729 mdt_object_child(msrcdir), new_fid);
2733 GOTO(out_unlock_old, rc);
2737 /* We used to acquire MDS_INODELOCK_FULL here but we
2738 * can't do this now because a running HSM restore on
2739 * the rename onto victim will hold the layout
2740 * lock. See LU-4002.
2743 lh_newp = &info->mti_lh[MDT_LH_NEW];
2744 mdt_lock_reg_init(lh_newp, LCK_EX);
2745 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2746 MDS_INODELOCK_LOOKUP |
2747 MDS_INODELOCK_UPDATE,
2750 GOTO(out_unlock_old, rc);
2752 /* get and save version after locking */
2753 mdt_version_get_save(info, mnew, 3);
2754 } else if (rc != -EREMOTE && rc != -ENOENT) {
2755 GOTO(out_put_old, rc);
2757 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2758 mdt_lock_reg_init(lh_oldp, LCK_EX);
2759 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2760 if (mdt_object_remote(msrcdir)) {
2761 /* Enqueue lookup lock from the parent MDT */
2762 rc = mdt_remote_object_lock(info, msrcdir,
2763 mdt_object_fid(mold),
2764 &lh_oldp->mlh_rreg_lh,
2765 lh_oldp->mlh_rreg_mode,
2766 MDS_INODELOCK_LOOKUP,
2769 GOTO(out_put_old, rc);
2771 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2774 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2777 GOTO(out_unlock_old, rc);
2779 mdt_enoent_version_save(info, 3);
2782 /* step 5: rename it */
2783 mdt_reint_init_ma(info, ma);
2785 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2786 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2789 mutex_lock(&mnew->mot_lov_mutex);
2791 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2792 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2793 mnew != NULL ? mdt_object_child(mnew) : NULL,
2794 &rr->rr_tgt_name, ma);
2797 mutex_unlock(&mnew->mot_lov_mutex);
2799 /* handle last link of tgt object */
2801 mdt_counter_incr(req, LPROC_MDT_RENAME,
2802 ktime_us_delta(ktime_get(), kstart));
2804 mdt_handle_last_unlink(info, mnew, ma);
2805 discard = mdt_dom_check_for_discard(info, mnew);
2807 mdt_rename_counter_tally(info, info->mti_mdt, req,
2809 ktime_us_delta(ktime_get(), kstart));
2814 mdt_object_unlock(info, mnew, lh_newp, rc);
2816 mdt_object_unlock(info, mold, lh_oldp, rc);
2818 if (mnew && !discard)
2819 mdt_object_put(info->mti_env, mnew);
2821 mdt_object_put(info->mti_env, mold);
2823 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2824 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2826 if (lustre_handle_is_used(&rename_lh))
2827 mdt_rename_unlock(&rename_lh);
2829 mdt_object_put(info->mti_env, mtgtdir);
2831 mdt_object_put(info->mti_env, msrcdir);
2833 /* The DoM discard can be done right in the place above where it is
2834 * assigned, meanwhile it is done here after rename unlock due to
2835 * compatibility with old clients, for them the discard blocks
2836 * the main thread until completion. Check LU-11359 for details.
2839 mdt_dom_discard_data(info, mnew);
2840 mdt_object_put(info->mti_env, mnew);
2845 static int mdt_reint_resync(struct mdt_thread_info *info,
2846 struct mdt_lock_handle *lhc)
2848 struct mdt_reint_record *rr = &info->mti_rr;
2849 struct ptlrpc_request *req = mdt_info_req(info);
2850 struct md_attr *ma = &info->mti_attr;
2851 struct mdt_object *mo;
2852 struct ldlm_lock *lease;
2853 struct mdt_body *repbody;
2854 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2859 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2861 if (info->mti_dlm_req)
2862 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2864 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2866 GOTO(out, rc = PTR_ERR(mo));
2868 if (!mdt_object_exists(mo))
2869 GOTO(out_obj, rc = -ENOENT);
2871 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2872 GOTO(out_obj, rc = -EINVAL);
2874 if (mdt_object_remote(mo))
2875 GOTO(out_obj, rc = -EREMOTE);
2877 lease = ldlm_handle2lock(rr->rr_lease_handle);
2879 GOTO(out_obj, rc = -ESTALE);
2881 /* It's really necessary to grab open_sem and check if the lease lock
2882 * has been lost. There would exist a concurrent writer coming in and
2883 * generating some dirty data in memory cache, the writeback would fail
2884 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2886 if (!down_write_trylock(&mo->mot_open_sem))
2887 GOTO(out_put_lease, rc = -EBUSY);
2889 lock_res_and_lock(lease);
2890 lease_broken = ldlm_is_cancel(lease);
2891 unlock_res_and_lock(lease);
2893 GOTO(out_unlock, rc = -EBUSY);
2895 /* the file has yet opened by anyone else after we took the lease. */
2896 layout.mlc_opc = MD_LAYOUT_RESYNC;
2897 lhc = &info->mti_lh[MDT_LH_LOCAL];
2898 rc = mdt_layout_change(info, mo, lhc, &layout);
2900 GOTO(out_unlock, rc);
2902 mdt_object_unlock(info, mo, lhc, 0);
2904 ma->ma_need = MA_INODE;
2906 rc = mdt_attr_get_complex(info, mo, ma);
2908 GOTO(out_unlock, rc);
2910 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2911 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2915 up_write(&mo->mot_open_sem);
2917 LDLM_LOCK_PUT(lease);
2919 mdt_object_put(info->mti_env, mo);
2921 mdt_client_compatibility(info);
2922 rc2 = mdt_fix_reply(info);
2928 struct mdt_reinter {
2929 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2930 enum lprocfs_extra_opc mr_extra_opc;
2933 static const struct mdt_reinter mdt_reinters[] = {
2935 .mr_handler = &mdt_reint_setattr,
2936 .mr_extra_opc = MDS_REINT_SETATTR,
2939 .mr_handler = &mdt_reint_create,
2940 .mr_extra_opc = MDS_REINT_CREATE,
2943 .mr_handler = &mdt_reint_link,
2944 .mr_extra_opc = MDS_REINT_LINK,
2947 .mr_handler = &mdt_reint_unlink,
2948 .mr_extra_opc = MDS_REINT_UNLINK,
2951 .mr_handler = &mdt_reint_rename,
2952 .mr_extra_opc = MDS_REINT_RENAME,
2955 .mr_handler = &mdt_reint_open,
2956 .mr_extra_opc = MDS_REINT_OPEN,
2958 [REINT_SETXATTR] = {
2959 .mr_handler = &mdt_reint_setxattr,
2960 .mr_extra_opc = MDS_REINT_SETXATTR,
2963 .mr_handler = &mdt_reint_unlink,
2964 .mr_extra_opc = MDS_REINT_UNLINK,
2967 .mr_handler = &mdt_reint_migrate,
2968 .mr_extra_opc = MDS_REINT_RENAME,
2971 .mr_handler = &mdt_reint_resync,
2972 .mr_extra_opc = MDS_REINT_RESYNC,
2976 int mdt_reint_rec(struct mdt_thread_info *info,
2977 struct mdt_lock_handle *lhc)
2979 const struct mdt_reinter *mr;
2983 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2986 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2987 if (mr->mr_handler == NULL)
2990 rc = (*mr->mr_handler)(info, lhc);
2992 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2993 PTLRPC_LAST_CNTR + mr->mr_extra_opc);