4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/mdt/mdt_reint.c
33 * Lustre Metadata Target (mdt) reintegration routines
35 * Author: Peter Braam <braam@clusterfs.com>
36 * Author: Andreas Dilger <adilger@clusterfs.com>
37 * Author: Phil Schwan <phil@clusterfs.com>
38 * Author: Huang Hua <huanghua@clusterfs.com>
39 * Author: Yury Umanets <umka@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_MDS
44 #include <lprocfs_status.h>
45 #include "mdt_internal.h"
46 #include <lustre_lmv.h>
47 #include <lustre_crypto.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_user_md *lum = spec->u.sp_ea.eadata;
357 struct lmv_mds_md_v1 *lmv;
358 struct mdt_object *child;
359 struct mdt_lock_handle *lhp;
360 struct mdt_lock_handle *lhc;
361 struct mdt_body *repbody;
365 if (!mdt->mdt_enable_dir_restripe)
369 lum->lum_hash_type |= cpu_to_le32(LMV_HASH_FLAG_FIXED);
371 rc = mdt_version_get_check_save(info, parent, 0);
375 lhp = &info->mti_lh[MDT_LH_PARENT];
376 mdt_lock_pdo_init(lhp, LCK_PW, lname);
377 rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
382 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
384 GOTO(unlock_parent, rc);
386 if (ma->ma_valid & MA_LMV) {
387 /* don't allow restripe if parent dir layout is changing */
388 lmv = &ma->ma_lmv->lmv_md_v1;
389 if (!lmv_is_sane2(lmv))
390 GOTO(unlock_parent, rc = -EBADF);
392 if (lmv_is_layout_changing(lmv))
393 GOTO(unlock_parent, rc = -EBUSY);
397 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
399 GOTO(unlock_parent, rc);
401 child = mdt_object_find(info->mti_env, mdt, fid);
403 GOTO(unlock_parent, rc = PTR_ERR(child));
405 if (!mdt_object_exists(child))
406 GOTO(out_child, rc = -ENOENT);
408 if (mdt_object_remote(child)) {
409 struct mdt_body *repbody;
411 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
413 GOTO(out_child, rc = -EPROTO);
415 repbody->mbo_fid1 = *fid;
416 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
417 GOTO(out_child, rc = -EREMOTE);
420 if (!S_ISDIR(lu_object_attr(&child->mot_obj)))
421 GOTO(out_child, rc = -ENOTDIR);
423 rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
427 /* race with migrate? */
428 if ((ma->ma_valid & MA_LMV) &&
429 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1))
430 GOTO(out_child, rc = -EBUSY);
433 lhc = &info->mti_lh[MDT_LH_CHILD];
434 mdt_lock_reg_init(lhc, LCK_EX);
436 /* enqueue object remote LOOKUP lock */
437 if (mdt_object_remote(parent)) {
438 rc = mdt_remote_object_lock(info, parent, fid,
441 MDS_INODELOCK_LOOKUP, false);
446 rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
449 GOTO(unlock_child, rc);
451 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
452 rc = mdt_version_get_check_save(info, child, 1);
454 GOTO(unlock_child, rc);
456 spin_lock(&mdt->mdt_restriper.mdr_lock);
457 if (child->mot_restriping) {
459 spin_unlock(&mdt->mdt_restriper.mdr_lock);
460 GOTO(unlock_child, rc = -EBUSY);
462 child->mot_restriping = 1;
463 spin_unlock(&mdt->mdt_restriper.mdr_lock);
466 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
468 GOTO(restriping_clear, rc);
470 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
472 GOTO(restriping_clear, rc = -EPROTO);
474 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
478 child->mot_restriping = 0;
480 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
482 mdt_object_put(info->mti_env, child);
484 mdt_object_unlock(info, parent, lhp, rc);
490 * VBR: we save three versions in reply:
491 * 0 - parent. Check that parent version is the same during replay.
492 * 1 - name. Version of 'name' if file exists with the same name or
493 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
494 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
497 static int mdt_create(struct mdt_thread_info *info)
499 struct mdt_device *mdt = info->mti_mdt;
500 struct mdt_object *parent;
501 struct mdt_object *child;
502 struct mdt_lock_handle *lh;
503 struct mdt_body *repbody;
504 struct md_attr *ma = &info->mti_attr;
505 struct mdt_reint_record *rr = &info->mti_rr;
506 struct md_op_spec *spec = &info->mti_spec;
507 bool restripe = false;
511 DEBUG_REQ(D_INODE, mdt_info_req(info),
512 "Create ("DNAME"->"DFID") in "DFID,
513 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
515 if (!fid_is_md_operative(rr->rr_fid1))
518 if (S_ISDIR(ma->ma_attr.la_mode) &&
519 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
520 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
521 struct lu_ucred *uc = mdt_ucred(info);
522 struct obd_export *exp = mdt_info_req(info)->rq_export;
524 /* Only new clients can create remote dir( >= 2.4) and
525 * striped dir(>= 2.6), old client will return -ENOTSUPP
527 if (!mdt_is_dne_client(exp))
530 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
531 if (!mdt_is_striped_client(exp))
534 if (!mdt->mdt_enable_striped_dir)
536 } else if (!mdt->mdt_enable_remote_dir) {
540 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
541 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
545 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
546 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
547 mdt->mdt_enable_remote_dir_gid != -1)
550 /* restripe if later found dir exists, MDS_OPEN_CREAT means
551 * this is create only, don't try restripe.
553 if (mdt->mdt_enable_dir_restripe &&
554 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
555 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
559 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
561 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
563 RETURN(PTR_ERR(parent));
565 if (!mdt_object_exists(parent))
566 GOTO(put_parent, rc = -ENOENT);
569 * LU-10235: check if name exists locklessly first to avoid massive
570 * lock recalls on existing directories.
572 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
573 &info->mti_tmp_fid1, 1);
576 GOTO(put_parent, rc = -EEXIST);
578 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
582 /* -ENOENT is expected here */
584 GOTO(put_parent, rc);
586 /* save version of file name for replay, it must be ENOENT here */
587 mdt_enoent_version_save(info, 1);
589 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
591 lh = &info->mti_lh[MDT_LH_PARENT];
592 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
593 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
595 GOTO(put_parent, rc);
597 if (!mdt_object_remote(parent)) {
598 rc = mdt_version_get_check_save(info, parent, 0);
600 GOTO(unlock_parent, rc);
603 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
604 if (unlikely(IS_ERR(child)))
605 GOTO(unlock_parent, rc = PTR_ERR(child));
607 ma->ma_need = MA_INODE;
610 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
611 OBD_FAIL_MDS_REINT_CREATE_WRITE);
613 /* Version of child will be updated on disk. */
614 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
615 rc = mdt_version_get_check_save(info, child, 2);
620 * Do not perform lookup sanity check. We know that name does
623 info->mti_spec.sp_cr_lookup = 0;
624 info->mti_spec.sp_feat = &dt_directory_features;
626 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
627 mdt_object_child(child), &info->mti_spec, ma);
629 rc = mdt_attr_get_complex(info, child, ma);
635 * On DNE, we need to eliminate dependey between 'mkdir a' and
636 * 'mkdir a/b' if b is a striped directory, to achieve this, two
637 * things are done below:
638 * 1. save child and slaves lock.
639 * 2. if the child is a striped directory, relock parent so to
640 * compare against with COS locks to ensure parent was
643 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
644 struct mdt_lock_handle *lhc;
645 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
648 rc = mdt_object_striped(info, child);
654 if (!mdt_object_remote(parent)) {
655 mdt_object_unlock(info, parent, lh, 1);
656 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
657 rc = mdt_reint_object_lock(info, parent, lh,
658 MDS_INODELOCK_UPDATE,
665 lhc = &info->mti_lh[MDT_LH_CHILD];
666 mdt_lock_handle_init(lhc);
667 mdt_lock_reg_init(lhc, LCK_PW);
668 rc = mdt_reint_striped_lock(info, child, lhc,
669 MDS_INODELOCK_UPDATE, einfo,
674 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
677 /* Return fid & attr to client. */
678 if (ma->ma_valid & MA_INODE)
679 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
680 mdt_object_fid(child));
683 mdt_object_put(info->mti_env, child);
685 mdt_object_unlock(info, parent, lh, rc);
687 mdt_object_put(info->mti_env, parent);
691 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
694 struct mdt_lock_handle *lh;
695 int do_vbr = ma->ma_attr.la_valid &
696 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
697 __u64 lockpart = MDS_INODELOCK_UPDATE;
698 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
703 rc = mdt_object_striped(info, mo);
709 lh = &info->mti_lh[MDT_LH_PARENT];
710 mdt_lock_reg_init(lh, LCK_PW);
712 /* Even though the new MDT will grant PERM lock to the old
713 * client, but the old client will almost ignore that during
714 * So it needs to revoke both LOOKUP and PERM lock here, so
715 * both new and old client can cancel the dcache
717 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
718 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
720 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
725 /* all attrs are packed into mti_attr in unpack_setattr */
726 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
727 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
729 /* VBR: update version if attr changed are important for recovery */
731 /* update on-disk version of changed object */
732 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
733 rc = mdt_version_get_check_save(info, mo, 0);
735 GOTO(out_unlock, rc);
738 /* Ensure constant striping during chown(). See LU-2789. */
739 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
740 mutex_lock(&mo->mot_lov_mutex);
742 /* all attrs are packed into mti_attr in unpack_setattr */
743 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
745 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
746 mutex_unlock(&mo->mot_lov_mutex);
749 GOTO(out_unlock, rc);
750 mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
753 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
758 * Check HSM flags and add HS_DIRTY flag if relevant.
760 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
761 * and is not RELEASED.
763 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
766 struct lu_ucred *uc = mdt_ucred(info);
767 kernel_cap_t cap_saved;
771 /* If the file was modified, add the dirty flag */
772 ma->ma_need = MA_HSM;
773 rc = mdt_attr_get_complex(info, mo, ma);
775 CERROR("file attribute read error for "DFID": %d.\n",
776 PFID(mdt_object_fid(mo)), rc);
780 /* If an up2date copy exists in the backend, add dirty flag */
781 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
782 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
783 ma->ma_hsm.mh_flags |= HS_DIRTY;
785 /* Bump cap so that closes from non-owner writers can
786 * set the HSM state to dirty.
788 cap_saved = uc->uc_cap;
789 cap_raise(uc->uc_cap, CAP_FOWNER);
790 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
791 uc->uc_cap = cap_saved;
793 CERROR("file attribute change error for "DFID": %d\n",
794 PFID(mdt_object_fid(mo)), rc);
800 static int mdt_reint_setattr(struct mdt_thread_info *info,
801 struct mdt_lock_handle *lhc)
803 struct mdt_device *mdt = info->mti_mdt;
804 struct md_attr *ma = &info->mti_attr;
805 struct mdt_reint_record *rr = &info->mti_rr;
806 struct ptlrpc_request *req = mdt_info_req(info);
807 struct mdt_object *mo;
808 struct mdt_body *repbody;
809 ktime_t kstart = ktime_get();
813 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
814 (unsigned int)ma->ma_attr.la_valid);
816 if (info->mti_dlm_req)
817 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
819 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
821 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
822 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
824 GOTO(out, rc = PTR_ERR(mo));
826 if (!mdt_object_exists(mo))
827 GOTO(out_put, rc = -ENOENT);
829 if (mdt_object_remote(mo))
830 GOTO(out_put, rc = -EREMOTE);
832 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
833 /* revoke lease lock if size is going to be changed */
834 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
835 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
836 atomic_read(&mo->mot_lease_count) > 0)) {
837 down_read(&mo->mot_open_sem);
839 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
840 lhc = &info->mti_lh[MDT_LH_LOCAL];
841 mdt_lock_reg_init(lhc, LCK_CW);
843 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
845 up_read(&mo->mot_open_sem);
849 /* revoke lease lock */
850 mdt_object_unlock(info, mo, lhc, 1);
852 up_read(&mo->mot_open_sem);
855 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
856 /* Check write access for the O_TRUNC case */
857 if (mdt_write_read(mo) < 0)
858 GOTO(out_put, rc = -ETXTBSY);
860 /* LU-10286: compatibility check for FLR.
861 * Please check the comment in mdt_finish_open() for details
863 if (!exp_connect_flr(info->mti_exp) ||
864 !exp_connect_overstriping(info->mti_exp)) {
865 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
866 if (rc < 0 && rc != -ENODATA)
869 if (!exp_connect_flr(info->mti_exp)) {
871 mdt_lmm_is_flr(info->mti_big_lmm))
872 GOTO(out_put, rc = -EOPNOTSUPP);
875 if (!exp_connect_overstriping(info->mti_exp)) {
877 mdt_lmm_is_overstriping(info->mti_big_lmm))
878 GOTO(out_put, rc = -EOPNOTSUPP);
882 /* For truncate, the file size sent from client
883 * is believable, but the blocks are incorrect,
884 * which makes the block size in LSOM attribute
885 * inconsisent with the real block size.
887 rc = mdt_lsom_update(info, mo, true);
892 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
893 if (ma->ma_valid & MA_LOV)
894 GOTO(out_put, rc = -EPROTO);
896 /* MDT supports FMD for regular files due to Data-on-MDT */
897 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
898 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
899 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
902 if (ma->ma_attr.la_valid & LA_MTIME) {
903 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
905 ma->ma_valid |= MA_PFID;
909 rc = mdt_attr_set(info, mo, ma);
912 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
913 (ma->ma_valid & MA_INODE)) {
914 struct lu_buf *buf = &info->mti_buf;
915 struct lu_ucred *uc = mdt_ucred(info);
916 struct mdt_lock_handle *lh;
918 __u64 lockpart = MDS_INODELOCK_XATTR;
920 /* reject if either remote or striped dir is disabled */
921 if (ma->ma_valid & MA_LMV) {
922 if (!mdt->mdt_enable_remote_dir ||
923 !mdt->mdt_enable_striped_dir)
924 GOTO(out_put, rc = -EPERM);
926 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
927 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
928 mdt->mdt_enable_remote_dir_gid != -1)
929 GOTO(out_put, rc = -EPERM);
932 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
933 GOTO(out_put, rc = -ENOTDIR);
935 if (ma->ma_attr.la_valid != 0)
936 GOTO(out_put, rc = -EPROTO);
938 lh = &info->mti_lh[MDT_LH_PARENT];
939 mdt_lock_reg_init(lh, LCK_PW);
941 if (ma->ma_valid & MA_LOV) {
942 buf->lb_buf = ma->ma_lmm;
943 buf->lb_len = ma->ma_lmm_size;
944 name = XATTR_NAME_LOV;
946 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
947 struct lu_fid *pfid = &info->mti_tmp_fid1;
948 struct lu_name *pname = &info->mti_name;
949 const char dotdot[] = "..";
950 struct mdt_object *pobj;
953 buf->lb_len = ma->ma_lmv_size;
954 name = XATTR_NAME_DEFAULT_LMV;
956 if (fid_is_root(rr->rr_fid1)) {
957 lockpart |= MDS_INODELOCK_LOOKUP;
959 /* force client to update dir default layout */
961 pname->ln_name = dotdot;
962 pname->ln_namelen = sizeof(dotdot);
963 rc = mdo_lookup(info->mti_env,
964 mdt_object_child(mo), pname,
969 pobj = mdt_object_find(info->mti_env, mdt,
972 GOTO(out_put, rc = PTR_ERR(pobj));
974 if (mdt_object_remote(pobj))
975 rc = mdt_remote_object_lock(info, pobj,
977 &lh->mlh_rreg_lh, LCK_EX,
978 MDS_INODELOCK_LOOKUP, false);
980 lockpart |= MDS_INODELOCK_LOOKUP;
982 mdt_object_put(info->mti_env, pobj);
989 rc = mdt_object_lock(info, mo, lh, lockpart);
993 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
996 mdt_object_unlock(info, mo, lh, rc);
1000 GOTO(out_put, rc = -EPROTO);
1003 /* If file data is modified, add the dirty flag */
1004 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1005 rc = mdt_add_dirty_flag(info, mo, ma);
1007 ma->ma_need = MA_INODE;
1009 rc = mdt_attr_get_complex(info, mo, ma);
1013 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1017 mdt_object_put(info->mti_env, mo);
1020 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1021 ktime_us_delta(ktime_get(), kstart));
1023 mdt_client_compatibility(info);
1024 rc2 = mdt_fix_reply(info);
1030 static int mdt_reint_create(struct mdt_thread_info *info,
1031 struct mdt_lock_handle *lhc)
1033 struct ptlrpc_request *req = mdt_info_req(info);
1034 ktime_t kstart = ktime_get();
1038 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1039 RETURN(err_serious(-ESTALE));
1041 if (info->mti_dlm_req)
1042 ldlm_request_cancel(mdt_info_req(info),
1043 info->mti_dlm_req, 0, LATF_SKIP);
1045 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1048 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1058 CERROR("%s: Unsupported mode %o\n",
1059 mdt_obd_name(info->mti_mdt),
1060 info->mti_attr.ma_attr.la_mode);
1061 RETURN(err_serious(-EOPNOTSUPP));
1064 rc = mdt_create(info);
1066 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1067 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1068 ktime_us_delta(ktime_get(), kstart));
1070 /* Special file should stay on the same node as parent*/
1071 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1072 ktime_us_delta(ktime_get(), kstart));
1079 * VBR: save parent version in reply and child version getting by its name.
1080 * Version of child is getting and checking during its lookup. If
1082 static int mdt_reint_unlink(struct mdt_thread_info *info,
1083 struct mdt_lock_handle *lhc)
1085 struct mdt_reint_record *rr = &info->mti_rr;
1086 struct ptlrpc_request *req = mdt_info_req(info);
1087 struct md_attr *ma = &info->mti_attr;
1088 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1089 struct mdt_object *mp;
1090 struct mdt_object *mc;
1091 struct mdt_lock_handle *parent_lh;
1092 struct mdt_lock_handle *child_lh;
1093 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1095 bool cos_incompat = false;
1097 ktime_t kstart = ktime_get();
1101 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1102 PNAME(&rr->rr_name));
1104 if (info->mti_dlm_req)
1105 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1107 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1108 RETURN(err_serious(-ENOENT));
1110 if (!fid_is_md_operative(rr->rr_fid1))
1113 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1115 RETURN(PTR_ERR(mp));
1117 if (mdt_object_remote(mp)) {
1118 cos_incompat = true;
1120 rc = mdt_version_get_check_save(info, mp, 0);
1122 GOTO(put_parent, rc);
1125 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1126 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1128 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1129 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1130 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1133 GOTO(put_parent, rc);
1135 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1136 *child_fid = *rr->rr_fid2;
1138 /* lookup child object along with version checking */
1139 fid_zero(child_fid);
1140 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1143 /* Name might not be able to find during resend of
1144 * remote unlink, considering following case.
1145 * dir_A is a remote directory, the name entry of
1146 * dir_A is on MDT0, the directory is on MDT1,
1148 * 1. client sends unlink req to MDT1.
1149 * 2. MDT1 sends name delete update to MDT0.
1150 * 3. name entry is being deleted in MDT0 synchronously.
1151 * 4. MDT1 is restarted.
1152 * 5. client resends unlink req to MDT1. So it can not
1153 * find the name entry on MDT0 anymore.
1154 * In this case, MDT1 only needs to destory the local
1157 if (mdt_object_remote(mp) && rc == -ENOENT &&
1158 !fid_is_zero(rr->rr_fid2) &&
1159 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1161 *child_fid = *rr->rr_fid2;
1163 GOTO(unlock_parent, rc);
1168 if (!fid_is_md_operative(child_fid))
1169 GOTO(unlock_parent, rc = -EPERM);
1171 /* We will lock the child regardless it is local or remote. No harm. */
1172 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1174 GOTO(unlock_parent, rc = PTR_ERR(mc));
1176 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1177 /* In this case, child fid is embedded in the request, and we do
1178 * not have a proper name as rr_name contains an encoded
1179 * hash. So find name that matches provided hash.
1181 if (!find_name_matching_hash(info, &rr->rr_name,
1183 GOTO(put_child, rc = -ENOENT);
1186 if (!cos_incompat) {
1187 rc = mdt_object_striped(info, mc);
1189 GOTO(put_child, rc);
1193 mdt_object_put(info->mti_env, mc);
1194 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1199 child_lh = &info->mti_lh[MDT_LH_CHILD];
1200 mdt_lock_reg_init(child_lh, LCK_EX);
1201 if (info->mti_spec.sp_rm_entry) {
1202 struct lu_ucred *uc = mdt_ucred(info);
1204 if (!mdt_is_dne_client(req->rq_export))
1205 /* Return -ENOTSUPP for old client */
1206 GOTO(put_child, rc = -ENOTSUPP);
1208 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1209 GOTO(put_child, rc = -EPERM);
1211 ma->ma_need = MA_INODE;
1213 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1214 NULL, &rr->rr_name, ma, no_name);
1215 GOTO(put_child, rc);
1218 if (mdt_object_remote(mc)) {
1219 struct mdt_body *repbody;
1221 if (!fid_is_zero(rr->rr_fid2)) {
1222 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1223 mdt_obd_name(info->mti_mdt),
1224 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1225 GOTO(put_child, rc = -ENOENT);
1227 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1228 mdt_obd_name(info->mti_mdt),
1229 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1231 if (!mdt_is_dne_client(req->rq_export))
1232 /* Return -ENOTSUPP for old client */
1233 GOTO(put_child, rc = -ENOTSUPP);
1235 /* Revoke the LOOKUP lock of the remote object granted by
1236 * this MDT. Since the unlink will happen on another MDT,
1237 * it will release the LOOKUP lock right away. Then What
1238 * would happen if another client try to grab the LOOKUP
1239 * lock at the same time with unlink XXX
1241 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1242 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1243 LASSERT(repbody != NULL);
1244 repbody->mbo_fid1 = *mdt_object_fid(mc);
1245 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1246 GOTO(unlock_child, rc = -EREMOTE);
1248 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1249 * this now because a running HSM restore on the child (unlink
1250 * victim) will hold the layout lock. See LU-4002.
1252 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1253 if (mdt_object_remote(mp)) {
1254 /* Enqueue lookup lock from parent MDT */
1255 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1256 &child_lh->mlh_rreg_lh,
1257 child_lh->mlh_rreg_mode,
1258 MDS_INODELOCK_LOOKUP, false);
1260 GOTO(put_child, rc);
1262 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1265 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1268 GOTO(put_child, rc);
1271 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1272 * whether need MA_LOV and MA_COOKIE.
1274 ma->ma_need = MA_INODE;
1277 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1278 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1279 /* save version when object is locked */
1280 mdt_version_get_save(info, mc, 1);
1282 mutex_lock(&mc->mot_lov_mutex);
1284 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1285 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1287 mutex_unlock(&mc->mot_lov_mutex);
1289 GOTO(unlock_child, rc);
1291 if (!lu_object_is_dying(&mc->mot_header)) {
1292 rc = mdt_attr_get_complex(info, mc, ma);
1295 } else if (mdt_dom_check_for_discard(info, mc)) {
1296 mdt_dom_discard_data(info, mc);
1298 mdt_handle_last_unlink(info, mc, ma);
1301 if (ma->ma_valid & MA_INODE) {
1302 switch (ma->ma_attr.la_mode & S_IFMT) {
1304 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1305 ktime_us_delta(ktime_get(), kstart));
1313 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1314 ktime_us_delta(ktime_get(), kstart));
1317 LASSERTF(0, "bad file type %o unlinking\n",
1318 ma->ma_attr.la_mode);
1325 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1327 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1328 info->mti_big_buf.lb_buf)
1329 lu_buf_free(&info->mti_big_buf);
1330 mdt_object_put(info->mti_env, mc);
1332 mdt_object_unlock(info, mp, parent_lh, rc);
1334 mdt_object_put(info->mti_env, mp);
1335 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1340 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1343 static int mdt_reint_link(struct mdt_thread_info *info,
1344 struct mdt_lock_handle *lhc)
1346 struct mdt_reint_record *rr = &info->mti_rr;
1347 struct ptlrpc_request *req = mdt_info_req(info);
1348 struct md_attr *ma = &info->mti_attr;
1349 struct mdt_object *ms;
1350 struct mdt_object *mp;
1351 struct mdt_lock_handle *lhs;
1352 struct mdt_lock_handle *lhp;
1353 ktime_t kstart = ktime_get();
1358 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1359 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1361 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1362 RETURN(err_serious(-ENOENT));
1364 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1365 OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1366 req->rq_no_reply = 1;
1367 RETURN(err_serious(-ENOENT));
1370 if (info->mti_dlm_req)
1371 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1373 /* Invalid case so return error immediately instead of
1376 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1379 if (!fid_is_md_operative(rr->rr_fid1) ||
1380 !fid_is_md_operative(rr->rr_fid2))
1383 /* step 1: find target parent dir */
1384 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1386 RETURN(PTR_ERR(mp));
1388 rc = mdt_version_get_check_save(info, mp, 0);
1390 GOTO(put_parent, rc);
1392 /* step 2: find source */
1393 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1395 GOTO(put_parent, rc = PTR_ERR(ms));
1397 if (!mdt_object_exists(ms)) {
1398 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1399 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1400 GOTO(put_source, rc = -ENOENT);
1403 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1405 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1407 lhp = &info->mti_lh[MDT_LH_PARENT];
1408 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1409 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1412 GOTO(put_source, rc);
1414 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1416 lhs = &info->mti_lh[MDT_LH_CHILD];
1417 mdt_lock_reg_init(lhs, LCK_EX);
1418 rc = mdt_reint_object_lock(info, ms, lhs,
1419 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1422 GOTO(unlock_parent, rc);
1424 /* step 3: link it */
1425 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1426 OBD_FAIL_MDS_REINT_LINK_WRITE);
1428 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1429 rc = mdt_version_get_check_save(info, ms, 1);
1431 GOTO(unlock_source, rc);
1433 /** check target version by name during replay */
1434 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1435 &info->mti_tmp_fid1, 2);
1436 if (rc != 0 && rc != -ENOENT)
1437 GOTO(unlock_source, rc);
1438 /* save version of file name for replay, it must be ENOENT here */
1439 if (!req_is_replay(mdt_info_req(info))) {
1440 if (rc != -ENOENT) {
1441 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1442 PNAME(&rr->rr_name));
1443 GOTO(unlock_source, rc = -EEXIST);
1445 info->mti_ver[2] = ENOENT_VERSION;
1446 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1449 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1450 mdt_object_child(ms), &rr->rr_name, ma);
1453 mdt_counter_incr(req, LPROC_MDT_LINK,
1454 ktime_us_delta(ktime_get(), kstart));
1458 mdt_object_unlock(info, ms, lhs, rc);
1460 mdt_object_unlock(info, mp, lhp, rc);
1462 mdt_object_put(info->mti_env, ms);
1464 mdt_object_put(info->mti_env, mp);
1468 * lock the part of the directory according to the hash of the name
1469 * (lh->mlh_pdo_hash) in parallel directory lock.
1471 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1472 struct mdt_lock_handle *lh,
1473 struct mdt_object *obj, __u64 ibits,
1476 struct ldlm_res_id *res = &info->mti_res_id;
1477 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1478 union ldlm_policy_data *policy = &info->mti_policy;
1479 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1483 * Finish res_id initializing by name hash marking part of
1484 * directory which is taking modification.
1486 LASSERT(lh->mlh_pdo_hash != 0);
1487 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1488 memset(policy, 0, sizeof(*policy));
1489 policy->l_inodebits.bits = ibits;
1491 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1492 dlmflags |= LDLM_FL_COS_INCOMPAT;
1494 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1495 * going to be sent to client. If it is - mdt_intent_policy() path will
1496 * fix it up and turn FL_LOCAL flag off.
1498 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1499 policy, res, dlmflags,
1500 &info->mti_exp->exp_handle.h_cookie);
1505 * Get BFL lock for rename or migrate process.
1507 static int mdt_rename_lock(struct mdt_thread_info *info,
1508 struct lustre_handle *lh)
1513 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1514 struct lu_fid *fid = &info->mti_tmp_fid1;
1515 struct mdt_object *obj;
1517 /* XXX, right now, it has to use object API to
1518 * enqueue lock cross MDT, so it will enqueue
1519 * rename lock(with LUSTRE_BFL_FID) by root object
1522 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1524 RETURN(PTR_ERR(obj));
1526 rc = mdt_remote_object_lock(info, obj,
1527 &LUSTRE_BFL_FID, lh,
1529 MDS_INODELOCK_UPDATE, false);
1530 mdt_object_put(info->mti_env, obj);
1532 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1533 union ldlm_policy_data *policy = &info->mti_policy;
1534 struct ldlm_res_id *res_id = &info->mti_res_id;
1537 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1538 memset(policy, 0, sizeof(*policy));
1539 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1540 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1541 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1542 LDLM_IBITS, policy, LCK_EX, &flags,
1544 ldlm_completion_ast, NULL, NULL, 0,
1546 &info->mti_exp->exp_handle.h_cookie,
1553 static void mdt_rename_unlock(struct lustre_handle *lh)
1556 LASSERT(lustre_handle_is_used(lh));
1557 /* Cancel the single rename lock right away */
1558 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1562 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1563 const struct lu_fid *fid,
1566 struct mdt_object *dir;
1570 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1574 /* check early, the real version will be saved after locking */
1575 rc = mdt_version_get_check(info, dir, idx);
1579 if (!mdt_object_exists(dir))
1580 GOTO(out_put, rc = -ENOENT);
1582 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1583 GOTO(out_put, rc = -ENOTDIR);
1587 mdt_object_put(info->mti_env, dir);
1592 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1593 * herein we don't really check it, just do revoke.
1595 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1596 struct mdt_object *pobj,
1597 struct mdt_object *obj)
1599 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1602 mdt_lock_handle_init(lh);
1603 mdt_lock_reg_init(lh, LCK_EX);
1605 if (mdt_object_remote(pobj)) {
1606 /* don't bother to check if pobj and obj are on the same MDT. */
1607 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1608 &lh->mlh_rreg_lh, LCK_EX,
1609 MDS_INODELOCK_LOOKUP, false);
1610 } else if (mdt_object_remote(obj)) {
1611 struct ldlm_res_id *res = &info->mti_res_id;
1612 union ldlm_policy_data *policy = &info->mti_policy;
1613 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1614 LDLM_FL_COS_INCOMPAT;
1616 fid_build_reg_res_name(mdt_object_fid(obj), res);
1617 memset(policy, 0, sizeof(*policy));
1618 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1619 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1620 &lh->mlh_reg_lh, LCK_EX, policy, res,
1623 /* do nothing if both are local */
1631 * TODO, currently we don't save this lock because there is no place to
1632 * hold this lock handle, but to avoid race we need to save this lock.
1634 mdt_object_unlock(info, NULL, lh, 1);
1640 * operation may takes locks of linkea, or directory stripes, group them in
1643 struct mdt_sub_lock {
1644 struct mdt_object *msl_obj;
1645 struct mdt_lock_handle msl_lh;
1646 struct list_head msl_linkage;
1649 static void mdt_unlock_list(struct mdt_thread_info *info,
1650 struct list_head *list, int decref)
1652 struct mdt_sub_lock *msl;
1653 struct mdt_sub_lock *tmp;
1655 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1656 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1657 list_del(&msl->msl_linkage);
1662 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1663 struct mdt_object *obj,
1664 struct mdt_lock_handle *lh,
1665 struct ldlm_enqueue_info *einfo,
1666 struct list_head *slave_locks,
1669 if (mdt_object_remote(obj)) {
1670 mdt_unlock_list(info, slave_locks, decref);
1671 mdt_object_unlock(info, obj, lh, decref);
1673 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1678 * lock parents of links, and also check whether total locks don't exceed
1681 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1682 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1683 * \retval -ev negative errno upon error
1685 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1686 struct mdt_object *pobj,
1687 const struct md_attr *ma,
1688 struct mdt_object *obj,
1689 struct mdt_lock_handle *lhp,
1690 struct ldlm_enqueue_info *peinfo,
1691 struct list_head *parent_slave_locks,
1692 struct list_head *link_locks)
1694 struct mdt_device *mdt = info->mti_mdt;
1695 struct lu_buf *buf = &info->mti_big_buf;
1696 struct lu_name *lname = &info->mti_name;
1697 struct linkea_data ldata = { NULL };
1698 bool blocked = false;
1699 int local_lnkp_cnt = 0;
1703 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1706 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1707 if (buf->lb_buf == NULL)
1711 rc = mdt_links_read(info, obj, &ldata);
1713 if (rc == -ENOENT || rc == -ENODATA)
1718 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1719 linkea_next_entry(&ldata)) {
1720 struct mdt_object *lnkp;
1721 struct mdt_sub_lock *msl;
1725 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1728 /* check if it's also linked to parent */
1729 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1730 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1731 PFID(&fid), PNAME(lname));
1732 /* in case link is remote object, revoke LOOKUP lock */
1733 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1739 /* check if it's linked to a stripe of parent */
1740 if (ma->ma_valid & MA_LMV) {
1741 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1742 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1745 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1746 fid_le_to_cpu(stripe_fid,
1747 &lmv->lmv_stripe_fids[j]);
1748 if (lu_fid_eq(stripe_fid, &fid)) {
1749 CDEBUG(D_INFO, "skip stripe "DFID
1750 ", reovke "DNAME"\n",
1751 PFID(&fid), PNAME(lname));
1752 lnkp = mdt_object_find(info->mti_env,
1755 GOTO(out, rc = PTR_ERR(lnkp));
1761 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1763 mdt_object_put(info->mti_env, lnkp);
1768 /* Check if it's already locked */
1769 list_for_each_entry(msl, link_locks, msl_linkage) {
1770 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1772 DFID" was locked, revoke "DNAME"\n",
1773 PFID(&fid), PNAME(lname));
1774 lnkp = msl->msl_obj;
1780 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1784 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1785 PFID(&fid), PNAME(lname));
1787 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1789 CWARN("%s: cannot find obj "DFID": %ld\n",
1790 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1794 if (!mdt_object_exists(lnkp)) {
1795 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1796 PFID(&fid), PNAME(lname));
1797 mdt_object_put(info->mti_env, lnkp);
1801 if (!mdt_object_remote(lnkp))
1806 GOTO(out, rc = -ENOMEM);
1809 * we can't follow parent-child lock order like other MD
1810 * operations, use lock_try here to avoid deadlock, if the lock
1811 * cannot be taken, drop all locks taken, revoke the blocked
1812 * one, and continue processing the remaining entries, and in
1813 * the end of the loop restart from beginning.
1815 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1817 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1818 MDS_INODELOCK_UPDATE, true);
1819 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1821 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1822 PFID(&fid), PNAME(lname));
1824 mdt_unlock_list(info, link_locks, 1);
1825 /* also unlock parent locks to avoid deadlock */
1827 mdt_migrate_object_unlock(info, pobj, lhp,
1834 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1835 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1836 MDS_INODELOCK_UPDATE);
1838 mdt_object_put(info->mti_env, lnkp);
1843 if (mdt_object_remote(lnkp)) {
1844 struct ldlm_lock *lock;
1847 * for remote object, set lock cb_atomic,
1848 * so lock can be released in blocking_ast()
1849 * immediately, then the next lock_try will
1850 * have better chance of success.
1852 lock = ldlm_handle2lock(
1853 &msl->msl_lh.mlh_rreg_lh);
1854 LASSERT(lock != NULL);
1855 lock_res_and_lock(lock);
1856 ldlm_set_atomic_cb(lock);
1857 unlock_res_and_lock(lock);
1858 LDLM_LOCK_PUT(lock);
1861 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1866 INIT_LIST_HEAD(&msl->msl_linkage);
1867 msl->msl_obj = lnkp;
1868 list_add_tail(&msl->msl_linkage, link_locks);
1870 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1874 GOTO(out, rc = -EBUSY);
1879 mdt_unlock_list(info, link_locks, rc);
1880 } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1881 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1884 * parent may have 3 local objects: master object and 2 stripes
1885 * (if it's being migrated too); source may have 1 local objects
1886 * as regular file; target has 1 local object.
1887 * Note, source may have 2 local locks if it is directory but it
1888 * can't have hardlinks, so it is not considered here.
1895 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1896 struct mdt_object *obj,
1897 const struct md_attr *ma,
1898 struct list_head *slave_locks)
1900 struct mdt_device *mdt = info->mti_mdt;
1901 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1902 struct lu_fid *fid = &info->mti_tmp_fid1;
1903 struct mdt_object *slave;
1904 struct mdt_sub_lock *msl;
1909 LASSERT(mdt_object_remote(obj));
1910 LASSERT(ma->ma_valid & MA_LMV);
1913 if (!lmv_is_sane(lmv))
1916 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1917 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1919 if (!fid_is_sane(fid))
1922 slave = mdt_object_find(info->mti_env, mdt, fid);
1924 GOTO(out, rc = PTR_ERR(slave));
1928 mdt_object_put(info->mti_env, slave);
1929 GOTO(out, rc = -ENOMEM);
1932 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1933 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1934 MDS_INODELOCK_UPDATE, true);
1937 mdt_object_put(info->mti_env, slave);
1941 INIT_LIST_HEAD(&msl->msl_linkage);
1942 msl->msl_obj = slave;
1943 list_add_tail(&msl->msl_linkage, slave_locks);
1949 mdt_unlock_list(info, slave_locks, rc);
1953 /* lock parent and its stripes */
1954 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1955 struct mdt_object *obj,
1956 const struct md_attr *ma,
1957 struct mdt_lock_handle *lh,
1958 struct ldlm_enqueue_info *einfo,
1959 struct list_head *slave_locks)
1963 if (mdt_object_remote(obj)) {
1964 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1965 &lh->mlh_rreg_lh, LCK_PW,
1966 MDS_INODELOCK_UPDATE, false);
1971 * if obj is remote and striped, lock its stripes explicitly
1972 * because it's not striped in LOD layer on this MDT.
1974 if (ma->ma_valid & MA_LMV) {
1975 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1977 mdt_object_unlock(info, obj, lh, rc);
1980 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1988 * in migration, object may be remote, and we need take full lock of it and its
1989 * stripes if it's directory, besides, object may be a remote object on its
1990 * parent, revoke its LOOKUP lock on where its parent is located.
1992 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1993 struct mdt_object *pobj,
1994 struct mdt_object *obj,
1995 struct mdt_lock_handle *lh,
1996 struct ldlm_enqueue_info *einfo,
1997 struct list_head *slave_locks)
2001 if (mdt_object_remote(obj)) {
2002 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2006 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
2007 &lh->mlh_rreg_lh, LCK_EX,
2008 MDS_INODELOCK_FULL, false);
2013 * if obj is remote and striped, lock its stripes explicitly
2014 * because it's not striped in LOD layer on this MDT.
2016 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
2017 struct md_attr *ma = &info->mti_attr;
2019 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
2021 mdt_object_unlock(info, obj, lh, rc);
2025 if (ma->ma_valid & MA_LMV) {
2026 rc = mdt_lock_remote_slaves(info, obj, ma,
2029 mdt_object_unlock(info, obj, lh, rc);
2033 if (mdt_object_remote(pobj)) {
2034 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2039 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
2047 * lookup source by name, if parent is striped directory, we need to find the
2048 * corresponding stripe where source is located, and then lookup there.
2050 * besides, if parent is migrating too, and file is already in target stripe,
2051 * this should be a redo of 'lfs migrate' on client side.
2053 static int mdt_migrate_lookup(struct mdt_thread_info *info,
2054 struct mdt_object *pobj,
2055 const struct md_attr *ma,
2056 const struct lu_name *lname,
2057 struct mdt_object **spobj,
2058 struct mdt_object **sobj)
2060 const struct lu_env *env = info->mti_env;
2061 struct lu_fid *fid = &info->mti_tmp_fid1;
2062 struct mdt_object *stripe;
2065 if (ma->ma_valid & MA_LMV) {
2066 /* if parent is striped, lookup on corresponding stripe */
2067 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2069 if (!lmv_is_sane(lmv))
2072 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2077 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2079 stripe = mdt_object_find(env, info->mti_mdt, fid);
2081 return PTR_ERR(stripe);
2084 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2086 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2088 * if parent layout is changeing, and lookup child
2089 * failed on source stripe, lookup again on target
2090 * stripe, if it exists, it means previous migration
2091 * was interrupted, and current file was migrated
2094 mdt_object_put(env, stripe);
2096 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2101 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2103 stripe = mdt_object_find(env, info->mti_mdt, fid);
2105 return PTR_ERR(stripe);
2108 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2109 fid, &info->mti_spec);
2110 mdt_object_put(env, stripe);
2111 return rc ?: -EALREADY;
2113 mdt_object_put(env, stripe);
2118 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2124 mdt_object_get(env, stripe);
2129 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2130 if (IS_ERR(*sobj)) {
2131 mdt_object_put(env, stripe);
2132 rc = PTR_ERR(*sobj);
2140 /* end lease and close file for regular file */
2141 static int mdd_migrate_close(struct mdt_thread_info *info,
2142 struct mdt_object *obj)
2144 struct close_data *data;
2145 struct mdt_body *repbody;
2146 struct ldlm_lock *lease;
2151 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2153 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2157 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2162 lease = ldlm_handle2lock(&data->cd_handle);
2166 /* check if the lease was already canceled */
2167 lock_res_and_lock(lease);
2168 rc = ldlm_is_cancel(lease);
2169 unlock_res_and_lock(lease);
2173 LDLM_DEBUG(lease, DFID" lease broken",
2174 PFID(mdt_object_fid(obj)));
2178 * cancel server side lease, client side counterpart should have been
2179 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2181 ldlm_lock_cancel(lease);
2182 ldlm_reprocess_all(lease->l_resource,
2183 lease->l_policy_data.l_inodebits.bits);
2184 LDLM_LOCK_PUT(lease);
2187 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2188 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2189 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2195 * migrate file in below steps:
2196 * 1. lock parent and its stripes
2197 * 2. lookup source by name
2198 * 3. lock parents of source links if source is not directory
2199 * 4. reject if source is in HSM
2200 * 5. take source open_sem and close file if source is regular file
2201 * 6. lock source and its stripes if it's directory
2202 * 7. lock target so subsequent change to it can trigger COS
2204 * 9. unlock above locks
2205 * 10. sync device if source has links
2207 int mdt_reint_migrate(struct mdt_thread_info *info,
2208 struct mdt_lock_handle *unused)
2210 const struct lu_env *env = info->mti_env;
2211 struct mdt_device *mdt = info->mti_mdt;
2212 struct ptlrpc_request *req = mdt_info_req(info);
2213 struct mdt_reint_record *rr = &info->mti_rr;
2214 struct lu_ucred *uc = mdt_ucred(info);
2215 struct md_attr *ma = &info->mti_attr;
2216 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2217 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2218 struct mdt_object *pobj;
2219 struct mdt_object *spobj = NULL;
2220 struct mdt_object *sobj = NULL;
2221 struct mdt_object *tobj;
2222 struct lustre_handle rename_lh = { 0 };
2223 struct mdt_lock_handle *lhp;
2224 struct mdt_lock_handle *lhs;
2225 struct mdt_lock_handle *lht;
2226 LIST_HEAD(parent_slave_locks);
2227 LIST_HEAD(child_slave_locks);
2228 LIST_HEAD(link_locks);
2229 int lock_retries = 5;
2230 bool open_sem_locked = false;
2231 bool do_sync = false;
2235 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2236 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2238 if (info->mti_dlm_req)
2239 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2241 if (!fid_is_md_operative(rr->rr_fid1) ||
2242 !fid_is_md_operative(rr->rr_fid2))
2245 /* don't allow migrate . or .. */
2246 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2249 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2252 if (uc && !cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2253 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2254 mdt->mdt_enable_remote_dir_gid != -1)
2258 * Note: do not enqueue rename lock for replay request, because
2259 * if other MDT holds rename lock, but being blocked to wait for
2260 * this MDT to finish its recovery, and the failover MDT can not
2261 * get rename lock, which will cause deadlock.
2263 * req is NULL if this is called by directory auto-split.
2265 if (req && !req_is_replay(req)) {
2266 rc = mdt_rename_lock(info, &rename_lh);
2268 CERROR("%s: can't lock FS for rename: rc = %d\n",
2269 mdt_obd_name(info->mti_mdt), rc);
2274 /* pobj is master object of parent */
2275 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2277 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2280 rc = mdt_version_get_check(info, pobj, 0);
2282 GOTO(put_parent, rc);
2285 if (!mdt_object_exists(pobj))
2286 GOTO(put_parent, rc = -ENOENT);
2288 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2289 GOTO(put_parent, rc = -ENOTDIR);
2291 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2293 GOTO(put_parent, rc);
2296 /* lock parent object */
2297 lhp = &info->mti_lh[MDT_LH_PARENT];
2298 mdt_lock_reg_init(lhp, LCK_PW);
2299 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2300 &parent_slave_locks);
2302 GOTO(put_parent, rc);
2305 * spobj is the corresponding stripe against name if pobj is striped
2306 * directory, which is the real parent, and no need to lock, because
2307 * we've taken full lock of pobj.
2309 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2311 GOTO(unlock_parent, rc);
2313 /* lock parents of source links, and revoke LOOKUP lock of links */
2314 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2315 &parent_slave_locks, &link_locks);
2316 if (rc == -EBUSY && lock_retries-- > 0) {
2317 mdt_object_put(env, sobj);
2318 mdt_object_put(env, spobj);
2323 GOTO(put_source, rc);
2326 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2327 * with one request, if total lock count exceeds this limit, we will
2328 * drop all locks after migration, and synchronous device in the end.
2332 /* TODO: DoM migration is not supported, migrate dirent only */
2333 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2334 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2336 GOTO(unlock_links, rc);
2338 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2339 info->mti_spec.sp_migrate_nsonly = 1;
2340 } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2341 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2343 GOTO(unlock_links, rc);
2345 /* race with restripe/auto-split? */
2346 if ((ma->ma_valid & MA_LMV) &&
2347 lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2348 GOTO(unlock_links, rc = -EBUSY);
2351 /* if migration HSM is allowed */
2352 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2353 ma->ma_need = MA_HSM;
2355 rc = mdt_attr_get_complex(info, sobj, ma);
2357 GOTO(unlock_links, rc);
2359 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2360 GOTO(unlock_links, rc = -EOPNOTSUPP);
2363 /* end lease and close file for regular file */
2364 if (info->mti_spec.sp_migrate_close) {
2365 /* try to hold open_sem so that nobody else can open the file */
2366 if (!down_write_trylock(&sobj->mot_open_sem)) {
2368 mdd_migrate_close(info, sobj);
2369 GOTO(unlock_links, rc = -EBUSY);
2371 open_sem_locked = true;
2372 rc = mdd_migrate_close(info, sobj);
2374 GOTO(unlock_open_sem, rc);
2379 lhs = &info->mti_lh[MDT_LH_OLD];
2380 mdt_lock_reg_init(lhs, LCK_EX);
2381 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2382 &child_slave_locks);
2384 GOTO(unlock_open_sem, rc);
2387 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2389 GOTO(unlock_source, rc = PTR_ERR(tobj));
2391 lht = &info->mti_lh[MDT_LH_NEW];
2392 mdt_lock_reg_init(lht, LCK_EX);
2393 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2395 GOTO(put_target, rc);
2397 /* Don't do lookup sanity check. We know name doesn't exist. */
2398 info->mti_spec.sp_cr_lookup = 0;
2399 info->mti_spec.sp_feat = &dt_directory_features;
2401 rc = mdo_migrate(env, mdt_object_child(pobj),
2402 mdt_object_child(sobj), &rr->rr_name,
2403 mdt_object_child(tobj),
2404 &info->mti_spec, ma);
2406 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2407 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2410 mdt_object_unlock(info, tobj, lht, rc);
2412 mdt_object_put(env, tobj);
2414 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2415 &child_slave_locks, rc);
2417 if (open_sem_locked)
2418 up_write(&sobj->mot_open_sem);
2420 /* if we've got too many locks to save into RPC,
2421 * then just commit before the locks are released
2424 mdt_device_sync(env, mdt);
2425 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2427 mdt_object_put(env, sobj);
2428 mdt_object_put(env, spobj);
2430 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2431 &parent_slave_locks, rc);
2433 mdt_object_put(env, pobj);
2435 if (lustre_handle_is_used(&rename_lh))
2436 mdt_rename_unlock(&rename_lh);
2441 static int mdt_object_lock_save(struct mdt_thread_info *info,
2442 struct mdt_object *dir,
2443 struct mdt_lock_handle *lh,
2444 int idx, bool cos_incompat)
2448 /* we lock the target dir if it is local */
2449 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2454 /* get and save correct version after locking */
2455 mdt_version_get_save(info, dir, idx);
2460 * determine lock order of sobj and tobj
2462 * there are two situations we need to lock tobj before sobj:
2463 * 1. sobj is child of tobj
2464 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2465 * larger than that of tobj
2467 * \retval 1 lock tobj before sobj
2468 * \retval 0 lock sobj before tobj
2469 * \retval -ev negative errno upon error
2471 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2472 struct mdt_object *sobj,
2473 struct mdt_object *tobj)
2475 struct md_attr *ma = &info->mti_attr;
2476 struct lu_fid *spfid = &info->mti_tmp_fid1;
2477 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2478 struct lmv_mds_md_v1 *lmv;
2483 /* sobj and tobj are the same */
2487 if (fid_is_root(mdt_object_fid(sobj)))
2490 if (fid_is_root(mdt_object_fid(tobj)))
2493 /* check whether sobj is child of tobj */
2494 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2495 mdt_object_fid(tobj));
2502 /* check whether sobj and tobj are children of the same parent */
2503 rc = mdt_attr_get_pfid(info, sobj, spfid);
2507 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2511 if (!lu_fid_eq(spfid, tpfid))
2514 /* check whether sobj and tobj are sibling stripes */
2515 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2519 if (!(ma->ma_valid & MA_LMV))
2522 lmv = &ma->ma_lmv->lmv_md_v1;
2523 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2525 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2528 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2532 if (!(ma->ma_valid & MA_LMV))
2535 lmv = &ma->ma_lmv->lmv_md_v1;
2536 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2538 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2540 /* check stripe index of sobj and tobj */
2541 if (sindex == tindex)
2544 return sindex < tindex ? 0 : 1;
2548 * lock rename source object.
2550 * Both source and source parent may be remote, and source may be a remote
2551 * object on source parent, to avoid overriding lock handle, store remote
2552 * LOOKUP lock separately in @lhr.
2554 * \retval 0 on success
2555 * \retval -ev negative errno upon error
2557 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2558 struct mdt_object *parent,
2559 struct mdt_object *child,
2560 struct mdt_lock_handle *lhc,
2561 struct mdt_lock_handle *lhr,
2567 rc = mdt_is_remote_object(info, parent, child);
2572 /* enqueue remote LOOKUP lock from the parent MDT */
2573 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2575 if (mdt_object_remote(parent)) {
2576 rc = mdt_remote_object_lock(info, parent,
2577 mdt_object_fid(child),
2584 LASSERT(mdt_object_remote(child));
2585 rc = mdt_object_local_lock(info, child, lhr,
2586 &rmt_ibits, 0, true);
2591 ibits &= ~MDS_INODELOCK_LOOKUP;
2594 if (mdt_object_remote(child)) {
2595 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2602 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2607 mdt_object_unlock(info, child, lhr, rc);
2613 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2614 * 2 - srcdir child; 3 - tgtdir child.
2615 * Update on disk version of srcdir child.
2617 static int mdt_reint_rename(struct mdt_thread_info *info,
2618 struct mdt_lock_handle *unused)
2620 struct mdt_device *mdt = info->mti_mdt;
2621 struct mdt_reint_record *rr = &info->mti_rr;
2622 struct md_attr *ma = &info->mti_attr;
2623 struct ptlrpc_request *req = mdt_info_req(info);
2624 struct mdt_object *msrcdir = NULL;
2625 struct mdt_object *mtgtdir = NULL;
2626 struct mdt_object *mold;
2627 struct mdt_object *mnew = NULL;
2628 struct lustre_handle rename_lh = { 0 };
2629 struct mdt_lock_handle *lh_srcdirp;
2630 struct mdt_lock_handle *lh_tgtdirp;
2631 struct mdt_lock_handle *lh_oldp = NULL;
2632 struct mdt_lock_handle *lh_rmt = NULL;
2633 struct mdt_lock_handle *lh_newp = NULL;
2634 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2635 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2637 bool reverse = false, discard = false;
2639 ktime_t kstart = ktime_get();
2643 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2644 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2645 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2647 if (info->mti_dlm_req)
2648 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2650 if (!fid_is_md_operative(rr->rr_fid1) ||
2651 !fid_is_md_operative(rr->rr_fid2))
2654 /* find both parents. */
2655 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2656 if (IS_ERR(msrcdir))
2657 RETURN(PTR_ERR(msrcdir));
2659 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2661 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2663 mdt_object_get(info->mti_env, mtgtdir);
2665 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2666 if (IS_ERR(mtgtdir))
2667 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2671 * Note: do not enqueue rename lock for replay request, because
2672 * if other MDT holds rename lock, but being blocked to wait for
2673 * this MDT to finish its recovery, and the failover MDT can not
2674 * get rename lock, which will cause deadlock.
2676 if (!req_is_replay(req)) {
2678 * Normally rename RPC is handled on the MDT with the target
2679 * directory (if target exists, it's on the MDT with the
2680 * target), if the source directory is remote, it's a hint that
2681 * source is remote too (this may not be true, but it won't
2682 * cause any issue), return -EXDEV early to avoid taking
2685 if (!mdt->mdt_enable_remote_rename &&
2686 mdt_object_remote(msrcdir))
2687 GOTO(out_put_tgtdir, rc = -EXDEV);
2689 /* This might be further relaxed in the future for regular file
2690 * renames in different source and target parents. Start with
2691 * only same-directory renames for simplicity and because this
2692 * is by far the most the common use case.
2694 if (msrcdir != mtgtdir) {
2695 rc = mdt_rename_lock(info, &rename_lh);
2697 CERROR("%s: cannot lock for rename: rc = %d\n",
2698 mdt_obd_name(mdt), rc);
2699 GOTO(out_put_tgtdir, rc);
2702 CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2703 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2704 PNAME(&rr->rr_name));
2708 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2710 GOTO(out_unlock_rename, rc);
2714 /* source needs to be looked up after locking source parent, otherwise
2715 * this rename may race with unlink source, and cause rename hang, see
2716 * sanityn.sh 55b, so check parents first, if later we found source is
2717 * remote, relock parents.
2719 cos_incompat = (mdt_object_remote(msrcdir) ||
2720 mdt_object_remote(mtgtdir));
2722 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2724 /* lock parents in the proper order. */
2725 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2726 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2728 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2729 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2731 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2732 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2735 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2738 GOTO(out_unlock_rename, rc);
2740 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2742 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2745 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2746 GOTO(out_unlock_rename, rc);
2749 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2752 GOTO(out_unlock_rename, rc);
2754 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2756 if (mtgtdir != msrcdir) {
2757 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2759 } else if (!mdt_object_remote(mtgtdir) &&
2760 lh_srcdirp->mlh_pdo_hash !=
2761 lh_tgtdirp->mlh_pdo_hash) {
2762 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2763 MDS_INODELOCK_UPDATE,
2765 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2768 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2769 GOTO(out_unlock_rename, rc);
2773 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2774 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2776 /* find mold object. */
2778 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2780 GOTO(out_unlock_parents, rc);
2782 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2783 GOTO(out_unlock_parents, rc = -EINVAL);
2785 if (!fid_is_md_operative(old_fid))
2786 GOTO(out_unlock_parents, rc = -EPERM);
2788 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2790 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2792 if (!mdt_object_exists(mold)) {
2793 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2795 "object does not exist");
2796 GOTO(out_put_old, rc = -ENOENT);
2799 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2800 GOTO(out_put_old, rc = -EXDEV);
2802 /* Check if @mtgtdir is subdir of @mold, before locking child
2803 * to avoid reverse locking.
2805 if (mtgtdir != msrcdir) {
2806 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2811 GOTO(out_put_old, rc);
2815 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2816 /* save version after locking */
2817 mdt_version_get_save(info, mold, 2);
2819 if (!cos_incompat && mdt_object_remote(mold)) {
2820 cos_incompat = true;
2821 mdt_object_put(info->mti_env, mold);
2822 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2823 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2827 /* find mnew object:
2828 * mnew target object may not exist now
2829 * lookup with version checking
2832 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2835 /* the new_fid should have been filled at this moment */
2836 if (lu_fid_eq(old_fid, new_fid))
2837 GOTO(out_put_old, rc);
2839 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2840 lu_fid_eq(new_fid, rr->rr_fid2))
2841 GOTO(out_put_old, rc = -EINVAL);
2843 if (!fid_is_md_operative(new_fid))
2844 GOTO(out_put_old, rc = -EPERM);
2846 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2848 GOTO(out_put_old, rc = PTR_ERR(mnew));
2850 if (!mdt_object_exists(mnew)) {
2851 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2853 "object does not exist");
2854 GOTO(out_put_new, rc = -ENOENT);
2857 if (mdt_object_remote(mnew)) {
2858 struct mdt_body *repbody;
2860 /* Always send rename req to the target child MDT */
2861 repbody = req_capsule_server_get(info->mti_pill,
2863 LASSERT(repbody != NULL);
2864 repbody->mbo_fid1 = *new_fid;
2865 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2866 GOTO(out_put_new, rc = -EXDEV);
2868 /* Before locking the target dir, check we do not replace
2869 * a dir with a non-dir, otherwise it may deadlock with
2870 * link op which tries to create a link in this dir
2871 * back to this non-dir.
2873 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2874 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2875 GOTO(out_put_new, rc = -EISDIR);
2877 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2878 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2879 mdt_lock_reg_init(lh_oldp, LCK_EX);
2880 mdt_lock_reg_init(lh_rmt, LCK_EX);
2881 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2882 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2883 lh_rmt, lock_ibits, cos_incompat);
2885 GOTO(out_put_new, rc);
2887 /* Check if @msrcdir is subdir of @mnew, before locking child
2888 * to avoid reverse locking.
2890 if (mtgtdir != msrcdir) {
2891 rc = mdo_is_subdir(info->mti_env,
2892 mdt_object_child(msrcdir), new_fid);
2896 GOTO(out_unlock_old, rc);
2900 /* We used to acquire MDS_INODELOCK_FULL here but we
2901 * can't do this now because a running HSM restore on
2902 * the rename onto victim will hold the layout
2903 * lock. See LU-4002.
2906 lh_newp = &info->mti_lh[MDT_LH_NEW];
2907 mdt_lock_reg_init(lh_newp, LCK_EX);
2908 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2909 if (mdt_object_remote(mtgtdir)) {
2910 rc = mdt_remote_object_lock(info, mtgtdir,
2911 mdt_object_fid(mnew),
2912 &lh_newp->mlh_rreg_lh,
2913 lh_newp->mlh_rreg_mode,
2914 MDS_INODELOCK_LOOKUP,
2917 GOTO(out_unlock_old, rc);
2919 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2921 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2924 GOTO(out_unlock_new, rc);
2926 /* get and save version after locking */
2927 mdt_version_get_save(info, mnew, 3);
2928 } else if (rc != -ENOENT) {
2929 GOTO(out_put_old, rc);
2931 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2932 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2933 mdt_lock_reg_init(lh_oldp, LCK_EX);
2934 mdt_lock_reg_init(lh_rmt, LCK_EX);
2935 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2936 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2937 lh_rmt, lock_ibits, cos_incompat);
2939 GOTO(out_put_old, rc);
2941 mdt_enoent_version_save(info, 3);
2944 /* step 5: rename it */
2945 mdt_reint_init_ma(info, ma);
2947 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2948 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2951 mutex_lock(&mnew->mot_lov_mutex);
2953 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2954 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2955 mnew != NULL ? mdt_object_child(mnew) : NULL,
2956 &rr->rr_tgt_name, ma);
2959 mutex_unlock(&mnew->mot_lov_mutex);
2961 /* handle last link of tgt object */
2963 mdt_counter_incr(req, LPROC_MDT_RENAME,
2964 ktime_us_delta(ktime_get(), kstart));
2966 mdt_handle_last_unlink(info, mnew, ma);
2967 discard = mdt_dom_check_for_discard(info, mnew);
2969 mdt_rename_counter_tally(info, info->mti_mdt, req,
2971 ktime_us_delta(ktime_get(), kstart));
2977 mdt_object_unlock(info, mnew, lh_newp, rc);
2979 mdt_object_unlock(info, NULL, lh_rmt, rc);
2980 mdt_object_unlock(info, mold, lh_oldp, rc);
2982 if (mnew && !discard)
2983 mdt_object_put(info->mti_env, mnew);
2985 mdt_object_put(info->mti_env, mold);
2987 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2988 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2990 if (lustre_handle_is_used(&rename_lh))
2991 mdt_rename_unlock(&rename_lh);
2993 mdt_object_put(info->mti_env, mtgtdir);
2995 mdt_object_put(info->mti_env, msrcdir);
2997 /* The DoM discard can be done right in the place above where it is
2998 * assigned, meanwhile it is done here after rename unlock due to
2999 * compatibility with old clients, for them the discard blocks
3000 * the main thread until completion. Check LU-11359 for details.
3003 mdt_dom_discard_data(info, mnew);
3004 mdt_object_put(info->mti_env, mnew);
3006 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
3010 static int mdt_reint_resync(struct mdt_thread_info *info,
3011 struct mdt_lock_handle *lhc)
3013 struct mdt_reint_record *rr = &info->mti_rr;
3014 struct ptlrpc_request *req = mdt_info_req(info);
3015 struct md_attr *ma = &info->mti_attr;
3016 struct mdt_object *mo;
3017 struct ldlm_lock *lease;
3018 struct mdt_body *repbody;
3019 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
3024 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
3026 if (info->mti_dlm_req)
3027 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
3029 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
3031 GOTO(out, rc = PTR_ERR(mo));
3033 if (!mdt_object_exists(mo))
3034 GOTO(out_obj, rc = -ENOENT);
3036 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
3037 GOTO(out_obj, rc = -EINVAL);
3039 if (mdt_object_remote(mo))
3040 GOTO(out_obj, rc = -EREMOTE);
3042 lease = ldlm_handle2lock(rr->rr_lease_handle);
3044 GOTO(out_obj, rc = -ESTALE);
3046 /* It's really necessary to grab open_sem and check if the lease lock
3047 * has been lost. There would exist a concurrent writer coming in and
3048 * generating some dirty data in memory cache, the writeback would fail
3049 * after the layout version is increased by MDS_REINT_RESYNC RPC.
3051 if (!down_write_trylock(&mo->mot_open_sem))
3052 GOTO(out_put_lease, rc = -EBUSY);
3054 lock_res_and_lock(lease);
3055 lease_broken = ldlm_is_cancel(lease);
3056 unlock_res_and_lock(lease);
3058 GOTO(out_unlock, rc = -EBUSY);
3060 /* the file has yet opened by anyone else after we took the lease. */
3061 layout.mlc_opc = MD_LAYOUT_RESYNC;
3062 lhc = &info->mti_lh[MDT_LH_LOCAL];
3063 rc = mdt_layout_change(info, mo, lhc, &layout);
3065 GOTO(out_unlock, rc);
3067 mdt_object_unlock(info, mo, lhc, 0);
3069 ma->ma_need = MA_INODE;
3071 rc = mdt_attr_get_complex(info, mo, ma);
3073 GOTO(out_unlock, rc);
3075 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3076 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3080 up_write(&mo->mot_open_sem);
3082 LDLM_LOCK_PUT(lease);
3084 mdt_object_put(info->mti_env, mo);
3086 mdt_client_compatibility(info);
3087 rc2 = mdt_fix_reply(info);
3093 struct mdt_reinter {
3094 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3095 enum lprocfs_extra_opc mr_extra_opc;
3098 static const struct mdt_reinter mdt_reinters[] = {
3100 .mr_handler = &mdt_reint_setattr,
3101 .mr_extra_opc = MDS_REINT_SETATTR,
3104 .mr_handler = &mdt_reint_create,
3105 .mr_extra_opc = MDS_REINT_CREATE,
3108 .mr_handler = &mdt_reint_link,
3109 .mr_extra_opc = MDS_REINT_LINK,
3112 .mr_handler = &mdt_reint_unlink,
3113 .mr_extra_opc = MDS_REINT_UNLINK,
3116 .mr_handler = &mdt_reint_rename,
3117 .mr_extra_opc = MDS_REINT_RENAME,
3120 .mr_handler = &mdt_reint_open,
3121 .mr_extra_opc = MDS_REINT_OPEN,
3123 [REINT_SETXATTR] = {
3124 .mr_handler = &mdt_reint_setxattr,
3125 .mr_extra_opc = MDS_REINT_SETXATTR,
3128 .mr_handler = &mdt_reint_unlink,
3129 .mr_extra_opc = MDS_REINT_UNLINK,
3132 .mr_handler = &mdt_reint_migrate,
3133 .mr_extra_opc = MDS_REINT_RENAME,
3136 .mr_handler = &mdt_reint_resync,
3137 .mr_extra_opc = MDS_REINT_RESYNC,
3141 int mdt_reint_rec(struct mdt_thread_info *info,
3142 struct mdt_lock_handle *lhc)
3144 const struct mdt_reinter *mr;
3148 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3151 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3152 if (mr->mr_handler == NULL)
3155 rc = (*mr->mr_handler)(info, lhc);
3157 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3158 PTLRPC_LAST_CNTR + mr->mr_extra_opc);