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_stripes_unlock(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 *locks = einfo->ei_cbdata;
231 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
234 memset(policy, 0, sizeof(*policy));
235 policy->l_inodebits.bits = einfo->ei_inodebits;
236 mdt_lock_reg_init(lh, einfo->ei_mode);
237 for (i = 0; i < locks->ha_count; i++) {
238 if (test_bit(i, (void *)locks->ha_map))
239 lh->mlh_rreg_lh = locks->ha_handles[i];
241 lh->mlh_reg_lh = locks->ha_handles[i];
242 mdt_object_unlock(mti, NULL, lh, decref);
243 locks->ha_handles[i].cookie = 0ull;
246 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
251 struct mdt_object *obj)
253 struct lu_device *bottom_dev;
254 struct lu_object *bottom_obj;
257 if (!S_ISDIR(obj->mot_header.loh_attr))
260 /* getxattr from bottom obj to avoid reading in shard FIDs */
261 bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
262 bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
263 mdt_object_fid(obj), NULL);
264 if (IS_ERR(bottom_obj))
265 return PTR_ERR(bottom_obj);
267 rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
269 lu_object_put(mti->mti_env, bottom_obj);
271 return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
275 * Lock slave stripes if necessary, the lock handles of slave stripes
276 * will be stored in einfo->ei_cbdata.
278 static int mdt_stripes_lock(struct mdt_thread_info *mti, struct mdt_object *obj,
279 enum ldlm_mode mode, __u64 ibits,
280 struct ldlm_enqueue_info *einfo)
282 union ldlm_policy_data *policy = &mti->mti_policy;
284 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
285 einfo->ei_type = LDLM_IBITS;
286 einfo->ei_mode = mode;
287 einfo->ei_cb_bl = mdt_remote_blocking_ast;
288 einfo->ei_cb_local_bl = mdt_blocking_ast;
289 einfo->ei_cb_cp = ldlm_completion_ast;
290 einfo->ei_enq_slave = 1;
291 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
292 einfo->ei_inodebits = ibits;
293 einfo->ei_req_slot = 1;
294 memset(policy, 0, sizeof(*policy));
295 policy->l_inodebits.bits = ibits;
297 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
301 /** lock object, and stripes if it's a striped directory
303 * object should be local, this is called in operations which modify both object
306 * \param info struct mdt_thread_info
307 * \param parent parent object, if it's NULL, find parent by mdo_lookup()
308 * \param child child object
309 * \param lh lock handle
310 * \param einfo struct ldlm_enqueue_info
311 * \param ibits MDS inode lock bits
312 * \param mode lock mode
313 * \param cos_incompat DNE COS incompatible
315 * \retval 0 on success, -ev on error.
317 int mdt_object_stripes_lock(struct mdt_thread_info *info,
318 struct mdt_object *parent,
319 struct mdt_object *child,
320 struct mdt_lock_handle *lh,
321 struct ldlm_enqueue_info *einfo, __u64 ibits,
322 enum ldlm_mode mode, bool cos_incompat)
327 /* according to the protocol, child should be local, is request sent to
330 if (mdt_object_remote(child)) {
331 CERROR("%s: lock target "DFID", but it is on other MDT: rc = %d\n",
332 mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(child)),
337 memset(einfo, 0, sizeof(*einfo));
338 if (ibits & MDS_INODELOCK_LOOKUP) {
340 rc = mdt_object_check_lock(info, parent, child, lh, ibits,
343 rc = mdt_object_lock(info, child, lh, ibits, mode,
349 rc = mdt_object_striped(info, child);
356 /* lock stripes for striped directory */
357 rc = mdt_stripes_lock(info, child, lh->mlh_reg_mode, ibits, einfo);
358 if (rc == -EIO && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
363 mdt_object_unlock(info, child, lh, rc);
368 void mdt_object_stripes_unlock(struct mdt_thread_info *info,
369 struct mdt_object *obj,
370 struct mdt_lock_handle *lh,
371 struct ldlm_enqueue_info *einfo, int decref)
373 if (einfo->ei_cbdata)
374 mdt_stripes_unlock(info, obj, einfo, decref);
375 mdt_object_unlock(info, obj, lh, decref);
378 static int mdt_restripe(struct mdt_thread_info *info,
379 struct mdt_object *parent,
380 const struct lu_name *lname,
381 const struct lu_fid *tfid,
382 struct md_op_spec *spec,
385 struct mdt_device *mdt = info->mti_mdt;
386 struct lu_fid *fid = &info->mti_tmp_fid2;
387 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
388 struct lmv_user_md *lum = spec->u.sp_ea.eadata;
389 struct lu_ucred *uc = mdt_ucred(info);
390 struct lmv_mds_md_v1 *lmv;
391 struct mdt_object *child;
392 struct mdt_lock_handle *lhp;
393 struct mdt_lock_handle *lhc;
394 struct mdt_body *repbody;
399 /* we want rbac roles to have precedence over any other
400 * permission or capability checks
402 if (!mdt->mdt_enable_dir_restripe && !uc->uc_rbac_dne_ops)
406 lum->lum_hash_type |= cpu_to_le32(LMV_HASH_FLAG_FIXED);
408 rc = mdt_version_get_check_save(info, parent, 0);
412 lhp = &info->mti_lh[MDT_LH_PARENT];
413 rc = mdt_parent_lock(info, parent, lhp, lname, LCK_PW, true);
417 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
419 GOTO(unlock_parent, rc);
421 if (ma->ma_valid & MA_LMV) {
422 /* don't allow restripe if parent dir layout is changing */
423 lmv = &ma->ma_lmv->lmv_md_v1;
424 if (!lmv_is_sane2(lmv))
425 GOTO(unlock_parent, rc = -EBADF);
427 if (lmv_is_layout_changing(lmv))
428 GOTO(unlock_parent, rc = -EBUSY);
432 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
434 GOTO(unlock_parent, rc);
436 child = mdt_object_find(info->mti_env, mdt, fid);
438 GOTO(unlock_parent, rc = PTR_ERR(child));
440 if (!mdt_object_exists(child))
441 GOTO(out_child, rc = -ENOENT);
443 if (mdt_object_remote(child)) {
444 struct mdt_body *repbody;
446 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
448 GOTO(out_child, rc = -EPROTO);
450 repbody->mbo_fid1 = *fid;
451 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
452 GOTO(out_child, rc = -EREMOTE);
455 if (!S_ISDIR(lu_object_attr(&child->mot_obj)))
456 GOTO(out_child, rc = -ENOTDIR);
458 rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
462 /* race with migrate? */
463 if ((ma->ma_valid & MA_LMV) &&
464 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1))
465 GOTO(out_child, rc = -EBUSY);
468 lhc = &info->mti_lh[MDT_LH_CHILD];
469 rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
470 MDS_INODELOCK_FULL, LCK_PW, true);
472 GOTO(unlock_child, rc);
474 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
475 rc = mdt_version_get_check_save(info, child, 1);
477 GOTO(unlock_child, rc);
479 spin_lock(&mdt->mdt_restriper.mdr_lock);
480 if (child->mot_restriping) {
482 spin_unlock(&mdt->mdt_restriper.mdr_lock);
483 GOTO(unlock_child, rc = -EBUSY);
485 child->mot_restriping = 1;
486 spin_unlock(&mdt->mdt_restriper.mdr_lock);
489 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
491 GOTO(restriping_clear, rc);
493 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
495 GOTO(restriping_clear, rc = -EPROTO);
497 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
501 child->mot_restriping = 0;
503 mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
505 mdt_object_put(info->mti_env, child);
507 mdt_object_unlock(info, parent, lhp, rc);
513 * VBR: we save three versions in reply:
514 * 0 - parent. Check that parent version is the same during replay.
515 * 1 - name. Version of 'name' if file exists with the same name or
516 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
517 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
520 static int mdt_create(struct mdt_thread_info *info)
522 struct mdt_device *mdt = info->mti_mdt;
523 struct mdt_object *parent;
524 struct mdt_object *child;
525 struct mdt_lock_handle *lh;
526 struct mdt_body *repbody;
527 struct md_attr *ma = &info->mti_attr;
528 struct mdt_reint_record *rr = &info->mti_rr;
529 struct md_op_spec *spec = &info->mti_spec;
530 struct lu_ucred *uc = mdt_ucred(info);
531 bool restripe = false;
535 DEBUG_REQ(D_INODE, mdt_info_req(info),
536 "Create ("DNAME"->"DFID") in "DFID,
537 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
539 if (!fid_is_md_operative(rr->rr_fid1))
542 /* MDS_OPEN_DEFAULT_LMV means eadata is parent default LMV, which is set
543 * if client maintains inherited default LMV
545 if (S_ISDIR(ma->ma_attr.la_mode) &&
546 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0 &&
547 !(spec->sp_cr_flags & MDS_OPEN_DEFAULT_LMV)) {
548 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
549 struct obd_export *exp = mdt_info_req(info)->rq_export;
551 /* Only new clients can create remote dir( >= 2.4) and
552 * striped dir(>= 2.6), old client will return -ENOTSUPP
554 if (!mdt_is_dne_client(exp))
557 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
558 if (!mdt_is_striped_client(exp))
561 if (!mdt->mdt_enable_striped_dir)
563 } else if (!mdt->mdt_enable_remote_dir) {
567 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
568 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) >=
572 /* we want rbac roles to have precedence over any other
573 * permission or capability checks
575 if (!uc->uc_rbac_dne_ops ||
576 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
577 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
578 mdt->mdt_enable_remote_dir_gid != -1))
581 /* restripe if later found dir exists, MDS_OPEN_CREAT means
582 * this is create only, don't try restripe.
584 if (mdt->mdt_enable_dir_restripe &&
585 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
586 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
590 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
592 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
594 RETURN(PTR_ERR(parent));
596 if (!mdt_object_exists(parent))
597 GOTO(put_parent, rc = -ENOENT);
599 rc = mdt_check_enc(info, parent);
601 GOTO(put_parent, rc);
603 if (!uc->uc_rbac_fscrypt_admin &&
604 parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
605 GOTO(put_parent, rc = -EPERM);
608 * LU-10235: check if name exists locklessly first to avoid massive
609 * lock recalls on existing directories.
611 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
612 &info->mti_tmp_fid1, 1);
615 GOTO(put_parent, rc = -EEXIST);
617 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
621 /* -ENOENT is expected here */
623 GOTO(put_parent, rc);
625 /* save version of file name for replay, it must be ENOENT here */
626 mdt_enoent_version_save(info, 1);
628 CFS_RACE(OBD_FAIL_MDS_CREATE_RACE);
630 lh = &info->mti_lh[MDT_LH_PARENT];
631 rc = mdt_parent_lock(info, parent, lh, &rr->rr_name, LCK_PW, false);
633 GOTO(put_parent, rc);
635 if (!mdt_object_remote(parent)) {
636 rc = mdt_version_get_check_save(info, parent, 0);
638 GOTO(unlock_parent, rc);
641 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
642 if (unlikely(IS_ERR(child)))
643 GOTO(unlock_parent, rc = PTR_ERR(child));
645 ma->ma_need = MA_INODE;
648 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
649 OBD_FAIL_MDS_REINT_CREATE_WRITE);
651 /* Version of child will be updated on disk. */
652 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
653 rc = mdt_version_get_check_save(info, child, 2);
657 if (parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD ||
658 (rr->rr_name.ln_namelen == strlen(dot_fscrypt_name) &&
659 strncmp(rr->rr_name.ln_name, dot_fscrypt_name,
660 rr->rr_name.ln_namelen) == 0))
661 child->mot_obj.lo_header->loh_attr |= LOHA_FSCRYPT_MD;
664 * Do not perform lookup sanity check. We know that name does
667 info->mti_spec.sp_cr_lookup = 0;
668 if (mdt_object_remote(parent))
669 info->mti_spec.sp_cr_lookup = 1;
670 info->mti_spec.sp_feat = &dt_directory_features;
672 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
673 mdt_object_child(child), &info->mti_spec, ma);
675 rc = mdt_attr_get_complex(info, child, ma);
681 * On DNE, we need to eliminate dependey between 'mkdir a' and
682 * 'mkdir a/b' if b is a striped directory, to achieve this, two
683 * things are done below:
684 * 1. save child and slaves lock.
685 * 2. if the child is a striped directory, relock parent so to
686 * compare against with COS locks to ensure parent was
689 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
690 struct mdt_lock_handle *lhc;
691 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
694 rc = mdt_object_striped(info, child);
700 if (!mdt_object_remote(parent)) {
701 mdt_object_unlock(info, parent, lh, 1);
702 rc = mdt_parent_lock(info, parent, lh,
703 &rr->rr_name, LCK_PW,
710 lhc = &info->mti_lh[MDT_LH_CHILD];
711 rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
712 MDS_INODELOCK_UPDATE, LCK_PW,
717 mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
720 /* Return fid & attr to client. */
721 if (ma->ma_valid & MA_INODE)
722 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
723 mdt_object_fid(child));
726 mdt_object_put(info->mti_env, child);
728 mdt_object_unlock(info, parent, lh, rc);
730 mdt_object_put(info->mti_env, parent);
734 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
737 struct mdt_lock_handle *lh;
738 int do_vbr = ma->ma_attr.la_valid &
739 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
740 __u64 lockpart = MDS_INODELOCK_UPDATE;
741 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
746 rc = mdt_object_striped(info, mo);
751 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
752 lockpart |= MDS_INODELOCK_PERM;
753 /* Clear xattr cache on clients, so the virtual project ID xattr
754 * can get the new project ID
756 if (ma->ma_attr.la_valid & LA_PROJID)
757 lockpart |= MDS_INODELOCK_XATTR;
759 lh = &info->mti_lh[MDT_LH_PARENT];
760 rc = mdt_object_stripes_lock(info, NULL, mo, lh, einfo, lockpart,
761 LCK_PW, cos_incompat);
765 /* all attrs are packed into mti_attr in unpack_setattr */
766 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
767 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
769 /* VBR: update version if attr changed are important for recovery */
771 /* update on-disk version of changed object */
772 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
773 rc = mdt_version_get_check_save(info, mo, 0);
775 GOTO(out_unlock, rc);
778 /* Ensure constant striping during chown(). See LU-2789. */
779 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
780 mutex_lock(&mo->mot_lov_mutex);
782 /* all attrs are packed into mti_attr in unpack_setattr */
783 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
785 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
786 mutex_unlock(&mo->mot_lov_mutex);
789 GOTO(out_unlock, rc);
790 mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
793 mdt_object_stripes_unlock(info, mo, lh, einfo, rc);
798 * Check HSM flags and add HS_DIRTY flag if relevant.
800 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
801 * and is not RELEASED.
803 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
806 struct lu_ucred *uc = mdt_ucred(info);
807 kernel_cap_t cap_saved;
811 /* If the file was modified, add the dirty flag */
812 ma->ma_need = MA_HSM;
813 rc = mdt_attr_get_complex(info, mo, ma);
815 CERROR("file attribute read error for "DFID": %d.\n",
816 PFID(mdt_object_fid(mo)), rc);
820 /* If an up2date copy exists in the backend, add dirty flag */
821 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
822 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
823 ma->ma_hsm.mh_flags |= HS_DIRTY;
825 /* Bump cap so that closes from non-owner writers can
826 * set the HSM state to dirty.
828 cap_saved = uc->uc_cap;
829 cap_raise(uc->uc_cap, CAP_FOWNER);
830 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
831 uc->uc_cap = cap_saved;
833 CERROR("file attribute change error for "DFID": %d\n",
834 PFID(mdt_object_fid(mo)), rc);
840 static int mdt_reint_setattr(struct mdt_thread_info *info,
841 struct mdt_lock_handle *lhc)
843 struct mdt_device *mdt = info->mti_mdt;
844 struct md_attr *ma = &info->mti_attr;
845 struct mdt_reint_record *rr = &info->mti_rr;
846 struct ptlrpc_request *req = mdt_info_req(info);
847 struct mdt_object *mo;
848 struct mdt_body *repbody;
849 ktime_t kstart = ktime_get();
853 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
854 (unsigned int)ma->ma_attr.la_valid);
856 if (info->mti_dlm_req)
857 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
859 CFS_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
861 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
862 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
864 GOTO(out, rc = PTR_ERR(mo));
866 if (!mdt_object_exists(mo))
867 GOTO(out_put, rc = -ENOENT);
869 if (mdt_object_remote(mo))
870 GOTO(out_put, rc = -EREMOTE);
872 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
873 /* revoke lease lock if size is going to be changed */
874 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
875 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
876 atomic_read(&mo->mot_lease_count) > 0)) {
877 down_read(&mo->mot_open_sem);
879 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
880 lhc = &info->mti_lh[MDT_LH_LOCAL];
881 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN,
884 up_read(&mo->mot_open_sem);
888 /* revoke lease lock */
889 mdt_object_unlock(info, mo, lhc, 1);
891 up_read(&mo->mot_open_sem);
894 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
895 /* Check write access for the O_TRUNC case */
896 if (mdt_write_read(mo) < 0)
897 GOTO(out_put, rc = -ETXTBSY);
899 /* LU-10286: compatibility check for FLR.
900 * Please check the comment in mdt_finish_open() for details
902 if (!exp_connect_flr(info->mti_exp) ||
903 !exp_connect_overstriping(info->mti_exp)) {
904 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
905 if (rc < 0 && rc != -ENODATA)
908 if (!exp_connect_flr(info->mti_exp)) {
910 mdt_lmm_is_flr(info->mti_big_lmm))
911 GOTO(out_put, rc = -EOPNOTSUPP);
914 if (!exp_connect_overstriping(info->mti_exp)) {
916 mdt_lmm_is_overstriping(info->mti_big_lmm))
917 GOTO(out_put, rc = -EOPNOTSUPP);
921 /* For truncate, the file size sent from client
922 * is believable, but the blocks are incorrect,
923 * which makes the block size in LSOM attribute
924 * inconsisent with the real block size.
926 rc = mdt_lsom_update(info, mo, true);
931 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
932 if (ma->ma_valid & MA_LOV)
933 GOTO(out_put, rc = -EPROTO);
935 /* MDT supports FMD for regular files due to Data-on-MDT */
936 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
937 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
938 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
941 if (ma->ma_attr.la_valid & LA_MTIME) {
942 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
944 ma->ma_valid |= MA_PFID;
948 rc = mdt_attr_set(info, mo, ma);
951 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
952 (ma->ma_valid & MA_INODE)) {
953 struct lu_buf *buf = &info->mti_buf;
954 struct lu_ucred *uc = mdt_ucred(info);
955 struct mdt_lock_handle *lh;
958 /* reject if either remote or striped dir is disabled */
959 if (ma->ma_valid & MA_LMV) {
960 if (!mdt->mdt_enable_remote_dir ||
961 !mdt->mdt_enable_striped_dir)
962 GOTO(out_put, rc = -EPERM);
964 /* we want rbac roles to have precedence over any other
965 * permission or capability checks
967 if (!uc->uc_rbac_dne_ops ||
968 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
969 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
970 mdt->mdt_enable_remote_dir_gid != -1))
971 GOTO(out_put, rc = -EPERM);
974 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
975 GOTO(out_put, rc = -ENOTDIR);
977 if (ma->ma_attr.la_valid != 0)
978 GOTO(out_put, rc = -EPROTO);
980 lh = &info->mti_lh[MDT_LH_PARENT];
981 if (ma->ma_valid & MA_LOV) {
982 buf->lb_buf = ma->ma_lmm;
983 buf->lb_len = ma->ma_lmm_size;
984 name = XATTR_NAME_LOV;
985 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
988 buf->lb_buf = &ma->ma_lmv->lmv_user_md;
989 buf->lb_len = ma->ma_lmv_size;
990 name = XATTR_NAME_DEFAULT_LMV;
992 if (unlikely(fid_is_root(mdt_object_fid(mo)))) {
993 rc = mdt_object_lock(info, mo, lh,
994 MDS_INODELOCK_XATTR |
995 MDS_INODELOCK_LOOKUP,
998 struct lu_fid *pfid = &info->mti_tmp_fid1;
999 struct lu_name *pname = &info->mti_name;
1000 const char dotdot[] = "..";
1001 struct mdt_object *pobj;
1004 pname->ln_name = dotdot;
1005 pname->ln_namelen = sizeof(dotdot);
1006 rc = mdo_lookup(info->mti_env,
1007 mdt_object_child(mo), pname,
1012 pobj = mdt_object_find(info->mti_env,
1013 info->mti_mdt, pfid);
1015 GOTO(out_put, rc = PTR_ERR(pobj));
1017 rc = mdt_object_check_lock(info, pobj, mo, lh,
1018 MDS_INODELOCK_XATTR |
1019 MDS_INODELOCK_LOOKUP,
1021 mdt_object_put(info->mti_env, pobj);
1028 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
1031 mdt_object_unlock(info, mo, lh, rc);
1035 GOTO(out_put, rc = -EPROTO);
1038 /* If file data is modified, add the dirty flag */
1039 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1040 rc = mdt_add_dirty_flag(info, mo, ma);
1042 ma->ma_need = MA_INODE;
1044 rc = mdt_attr_get_complex(info, mo, ma);
1048 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1052 mdt_object_put(info->mti_env, mo);
1055 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1056 ktime_us_delta(ktime_get(), kstart));
1058 mdt_client_compatibility(info);
1062 static int mdt_reint_create(struct mdt_thread_info *info,
1063 struct mdt_lock_handle *lhc)
1065 struct ptlrpc_request *req = mdt_info_req(info);
1066 ktime_t kstart = ktime_get();
1070 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1071 RETURN(err_serious(-ESTALE));
1073 if (info->mti_dlm_req)
1074 ldlm_request_cancel(mdt_info_req(info),
1075 info->mti_dlm_req, 0, LATF_SKIP);
1077 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1080 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1090 CERROR("%s: Unsupported mode %o\n",
1091 mdt_obd_name(info->mti_mdt),
1092 info->mti_attr.ma_attr.la_mode);
1093 RETURN(err_serious(-EOPNOTSUPP));
1096 rc = mdt_create(info);
1098 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1099 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1100 ktime_us_delta(ktime_get(), kstart));
1102 /* Special file should stay on the same node as parent*/
1103 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1104 ktime_us_delta(ktime_get(), kstart));
1111 * VBR: save parent version in reply and child version getting by its name.
1112 * Version of child is getting and checking during its lookup. If
1114 static int mdt_reint_unlink(struct mdt_thread_info *info,
1115 struct mdt_lock_handle *lhc)
1117 struct mdt_reint_record *rr = &info->mti_rr;
1118 struct ptlrpc_request *req = mdt_info_req(info);
1119 struct md_attr *ma = &info->mti_attr;
1120 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1121 struct mdt_object *mp;
1122 struct mdt_object *mc;
1123 struct mdt_lock_handle *parent_lh;
1124 struct mdt_lock_handle *child_lh;
1125 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
1126 struct lu_ucred *uc = mdt_ucred(info);
1127 bool cos_incompat = false;
1129 ktime_t kstart = ktime_get();
1133 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1134 PNAME(&rr->rr_name));
1136 if (info->mti_dlm_req)
1137 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1139 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1140 RETURN(err_serious(-ENOENT));
1142 if (!fid_is_md_operative(rr->rr_fid1))
1145 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1147 RETURN(PTR_ERR(mp));
1149 if (mdt_object_remote(mp)) {
1150 cos_incompat = true;
1152 rc = mdt_version_get_check_save(info, mp, 0);
1154 GOTO(put_parent, rc);
1157 if (!uc->uc_rbac_fscrypt_admin &&
1158 mp->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
1159 GOTO(put_parent, rc = -EPERM);
1161 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
1162 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1164 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1165 rc = mdt_parent_lock(info, mp, parent_lh, &rr->rr_name, LCK_PW,
1168 GOTO(put_parent, rc);
1170 if (info->mti_spec.sp_rm_entry) {
1171 if (!mdt_is_dne_client(req->rq_export))
1172 /* Return -ENOTSUPP for old client */
1173 GOTO(unlock_parent, rc = -ENOTSUPP);
1175 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1176 GOTO(unlock_parent, rc = -EPERM);
1178 ma->ma_need = MA_INODE;
1180 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1181 NULL, &rr->rr_name, ma, no_name);
1182 GOTO(unlock_parent, rc);
1185 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1186 *child_fid = *rr->rr_fid2;
1188 /* lookup child object along with version checking */
1189 fid_zero(child_fid);
1190 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1193 /* Name might not be able to find during resend of
1194 * remote unlink, considering following case.
1195 * dir_A is a remote directory, the name entry of
1196 * dir_A is on MDT0, the directory is on MDT1,
1198 * 1. client sends unlink req to MDT1.
1199 * 2. MDT1 sends name delete update to MDT0.
1200 * 3. name entry is being deleted in MDT0 synchronously.
1201 * 4. MDT1 is restarted.
1202 * 5. client resends unlink req to MDT1. So it can not
1203 * find the name entry on MDT0 anymore.
1204 * In this case, MDT1 only needs to destory the local
1207 if (mdt_object_remote(mp) && rc == -ENOENT &&
1208 !fid_is_zero(rr->rr_fid2) &&
1209 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1211 *child_fid = *rr->rr_fid2;
1213 GOTO(unlock_parent, rc);
1218 if (!fid_is_md_operative(child_fid))
1219 GOTO(unlock_parent, rc = -EPERM);
1221 /* We will lock the child regardless it is local or remote. No harm. */
1222 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1224 GOTO(unlock_parent, rc = PTR_ERR(mc));
1226 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1227 /* In this case, child fid is embedded in the request, and we do
1228 * not have a proper name as rr_name contains an encoded
1229 * hash. So find name that matches provided hash.
1231 if (!find_name_matching_hash(info, &rr->rr_name,
1233 GOTO(put_child, rc = -ENOENT);
1236 if (!cos_incompat) {
1237 rc = mdt_object_striped(info, mc);
1239 GOTO(put_child, rc);
1243 mdt_object_put(info->mti_env, mc);
1244 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1249 child_lh = &info->mti_lh[MDT_LH_CHILD];
1250 if (mdt_object_remote(mc)) {
1251 struct mdt_body *repbody;
1253 if (!fid_is_zero(rr->rr_fid2)) {
1254 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1255 mdt_obd_name(info->mti_mdt),
1256 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1257 GOTO(put_child, rc = -ENOENT);
1259 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1260 mdt_obd_name(info->mti_mdt),
1261 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1263 if (!mdt_is_dne_client(req->rq_export))
1264 /* Return -ENOTSUPP for old client */
1265 GOTO(put_child, rc = -ENOTSUPP);
1267 /* Revoke the LOOKUP lock of the remote object granted by
1268 * this MDT. Since the unlink will happen on another MDT,
1269 * it will release the LOOKUP lock right away. Then What
1270 * would happen if another client try to grab the LOOKUP
1271 * lock at the same time with unlink XXX
1273 rc = mdt_object_lookup_lock(info, NULL, mc, child_lh, LCK_EX,
1276 GOTO(put_child, rc);
1278 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1279 LASSERT(repbody != NULL);
1280 repbody->mbo_fid1 = *mdt_object_fid(mc);
1281 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1282 GOTO(unlock_child, rc = -EREMOTE);
1284 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1285 * this now because a running HSM restore on the child (unlink
1286 * victim) will hold the layout lock. See LU-4002.
1288 rc = mdt_object_stripes_lock(info, mp, mc, child_lh, einfo,
1289 MDS_INODELOCK_LOOKUP |
1290 MDS_INODELOCK_UPDATE,
1291 LCK_EX, cos_incompat);
1293 GOTO(put_child, rc);
1296 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1297 * whether need MA_LOV and MA_COOKIE.
1299 ma->ma_need = MA_INODE;
1302 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1303 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1304 /* save version when object is locked */
1305 mdt_version_get_save(info, mc, 1);
1307 mutex_lock(&mc->mot_lov_mutex);
1309 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1310 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1312 mutex_unlock(&mc->mot_lov_mutex);
1314 GOTO(unlock_child, rc);
1316 if (!lu_object_is_dying(&mc->mot_header)) {
1317 rc = mdt_attr_get_complex(info, mc, ma);
1320 } else if (mdt_dom_check_for_discard(info, mc)) {
1321 mdt_dom_discard_data(info, mc);
1323 mdt_handle_last_unlink(info, mc, ma);
1326 if (ma->ma_valid & MA_INODE) {
1327 switch (ma->ma_attr.la_mode & S_IFMT) {
1329 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1330 ktime_us_delta(ktime_get(), kstart));
1338 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1339 ktime_us_delta(ktime_get(), kstart));
1342 LASSERTF(0, "bad file type %o unlinking\n",
1343 ma->ma_attr.la_mode);
1350 mdt_object_stripes_unlock(info, mc, child_lh, einfo, rc);
1352 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1353 info->mti_big_buf.lb_buf)
1354 lu_buf_free(&info->mti_big_buf);
1355 mdt_object_put(info->mti_env, mc);
1357 mdt_object_unlock(info, mp, parent_lh, rc);
1359 mdt_object_put(info->mti_env, mp);
1360 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1365 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1368 static int mdt_reint_link(struct mdt_thread_info *info,
1369 struct mdt_lock_handle *lhc)
1371 struct mdt_reint_record *rr = &info->mti_rr;
1372 struct ptlrpc_request *req = mdt_info_req(info);
1373 struct md_attr *ma = &info->mti_attr;
1374 struct mdt_object *ms;
1375 struct mdt_object *mp;
1376 struct mdt_lock_handle *lhs;
1377 struct mdt_lock_handle *lhp;
1378 ktime_t kstart = ktime_get();
1383 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1384 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1386 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1387 RETURN(err_serious(-ENOENT));
1389 if (CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1390 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1391 req->rq_no_reply = 1;
1392 RETURN(err_serious(-ENOENT));
1395 if (info->mti_dlm_req)
1396 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1398 /* Invalid case so return error immediately instead of
1401 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1404 if (!fid_is_md_operative(rr->rr_fid1) ||
1405 !fid_is_md_operative(rr->rr_fid2))
1408 /* step 1: find target parent dir */
1409 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1411 RETURN(PTR_ERR(mp));
1413 rc = mdt_version_get_check_save(info, mp, 0);
1415 GOTO(put_parent, rc);
1417 rc = mdt_check_enc(info, mp);
1419 GOTO(put_parent, rc);
1421 /* step 2: find source */
1422 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1424 GOTO(put_parent, rc = PTR_ERR(ms));
1426 if (!mdt_object_exists(ms)) {
1427 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1428 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1429 GOTO(put_source, rc = -ENOENT);
1432 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1434 CFS_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1436 lhp = &info->mti_lh[MDT_LH_PARENT];
1437 rc = mdt_parent_lock(info, mp, lhp, &rr->rr_name, LCK_PW, cos_incompat);
1439 GOTO(put_source, rc);
1441 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1443 lhs = &info->mti_lh[MDT_LH_CHILD];
1444 rc = mdt_object_lock(info, ms, lhs,
1445 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR, LCK_EX,
1448 GOTO(unlock_parent, rc);
1450 /* step 3: link it */
1451 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1452 OBD_FAIL_MDS_REINT_LINK_WRITE);
1454 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1455 rc = mdt_version_get_check_save(info, ms, 1);
1457 GOTO(unlock_source, rc);
1459 /** check target version by name during replay */
1460 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1461 &info->mti_tmp_fid1, 2);
1462 if (rc != 0 && rc != -ENOENT)
1463 GOTO(unlock_source, rc);
1464 /* save version of file name for replay, it must be ENOENT here */
1465 if (!req_is_replay(mdt_info_req(info))) {
1466 if (rc != -ENOENT) {
1467 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1468 PNAME(&rr->rr_name));
1469 GOTO(unlock_source, rc = -EEXIST);
1471 info->mti_ver[2] = ENOENT_VERSION;
1472 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1475 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1476 mdt_object_child(ms), &rr->rr_name, ma);
1479 mdt_counter_incr(req, LPROC_MDT_LINK,
1480 ktime_us_delta(ktime_get(), kstart));
1484 mdt_object_unlock(info, ms, lhs, rc);
1486 mdt_object_unlock(info, mp, lhp, rc);
1488 mdt_object_put(info->mti_env, ms);
1490 mdt_object_put(info->mti_env, mp);
1495 * Get BFL lock for rename or migrate process.
1497 static int mdt_rename_lock(struct mdt_thread_info *info,
1498 struct mdt_lock_handle *lh)
1500 struct lu_fid *fid = &info->mti_tmp_fid1;
1501 struct mdt_object *obj;
1502 __u64 ibits = MDS_INODELOCK_UPDATE;
1507 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1509 RETURN(PTR_ERR(obj));
1511 mdt_lock_reg_init(lh, LCK_EX);
1512 rc = mdt_object_lock_internal(info, obj, &LUSTRE_BFL_FID, lh,
1513 &ibits, 0, false, false);
1514 mdt_object_put(info->mti_env, obj);
1518 static void mdt_rename_unlock(struct mdt_thread_info *info,
1519 struct mdt_lock_handle *lh)
1522 /* Cancel the single rename lock right away */
1523 mdt_object_unlock(info, NULL, lh, 1);
1527 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1528 const struct lu_fid *fid,
1531 struct mdt_object *dir;
1535 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1539 /* check early, the real version will be saved after locking */
1540 rc = mdt_version_get_check(info, dir, idx);
1544 if (!mdt_object_exists(dir))
1545 GOTO(out_put, rc = -ENOENT);
1547 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1548 GOTO(out_put, rc = -ENOTDIR);
1552 mdt_object_put(info->mti_env, dir);
1557 * lock rename source object.
1559 * Both source and its parent object may be located on remote MDTs, and even on
1560 * different MDTs, which means source object is a remote object on parent.
1562 * \retval 0 on success
1563 * \retval -ev negative errno upon error
1565 static int mdt_rename_source_lock(struct mdt_thread_info *info,
1566 struct mdt_object *parent,
1567 struct mdt_object *child,
1568 struct mdt_lock_handle *lh,
1569 struct mdt_lock_handle *lh_lookup,
1570 __u64 ibits, bool cos_incompat)
1574 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
1575 /* if @obj is remote object, LOOKUP lock needs to be taken from
1578 rc = mdt_is_remote_object(info, parent, child);
1583 rc = mdt_object_lookup_lock(info, parent, child, lh_lookup,
1584 LCK_EX, cos_incompat);
1588 ibits &= ~MDS_INODELOCK_LOOKUP;
1591 rc = mdt_object_lock(info, child, lh, ibits, LCK_EX, cos_incompat);
1592 if (unlikely(rc && !(ibits & MDS_INODELOCK_LOOKUP)))
1593 mdt_object_unlock(info, NULL, lh_lookup, rc);
1598 static void mdt_rename_source_unlock(struct mdt_thread_info *info,
1599 struct mdt_object *obj,
1600 struct mdt_lock_handle *lh,
1601 struct mdt_lock_handle *lh_lookup,
1604 mdt_object_unlock(info, obj, lh, decref);
1605 mdt_object_unlock(info, NULL, lh_lookup, decref);
1608 /* migration takes UPDATE lock of link parent, and LOOKUP lock of link */
1609 struct mdt_link_lock {
1610 struct mdt_object *mll_obj;
1611 struct mdt_lock_handle mll_lh;
1612 struct list_head mll_linkage;
1615 static inline int mdt_migrate_link_lock_add(struct mdt_thread_info *info,
1616 struct mdt_object *o,
1617 struct mdt_lock_handle *lh,
1618 struct list_head *list)
1620 struct mdt_link_lock *mll;
1626 INIT_LIST_HEAD(&mll->mll_linkage);
1627 mdt_object_get(info->mti_env, o);
1630 memset(lh, 0, sizeof(*lh));
1631 list_add_tail(&mll->mll_linkage, list);
1636 static inline void mdt_migrate_link_lock_del(struct mdt_thread_info *info,
1637 struct mdt_link_lock *mll,
1640 mdt_object_unlock(info, mll->mll_obj, &mll->mll_lh, decref);
1641 mdt_object_put(info->mti_env, mll->mll_obj);
1642 list_del(&mll->mll_linkage);
1646 static void mdt_migrate_links_unlock(struct mdt_thread_info *info,
1647 struct list_head *list, int decref)
1649 struct mdt_link_lock *mll;
1650 struct mdt_link_lock *tmp;
1652 list_for_each_entry_safe(mll, tmp, list, mll_linkage)
1653 mdt_migrate_link_lock_del(info, mll, decref);
1656 /* take link parent UPDATE lock.
1657 * \retval 0 \a lnkp is already locked, no lock taken.
1659 * -ev negative errno.
1661 static int mdt_migrate_link_parent_lock(struct mdt_thread_info *info,
1662 struct mdt_object *lnkp,
1663 struct list_head *update_locks,
1666 const struct lu_fid *fid = mdt_object_fid(lnkp);
1667 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1668 struct mdt_link_lock *entry;
1674 /* check if it's already locked */
1675 list_for_each_entry(entry, update_locks, mll_linkage) {
1676 if (lu_fid_eq(mdt_object_fid(entry->mll_obj), fid)) {
1677 CDEBUG(D_INFO, "skip "DFID" lock\n", PFID(fid));
1682 /* link parent UPDATE lock */
1683 CDEBUG(D_INFO, "lock "DFID"\n", PFID(fid));
1686 /* revoke lock instead of take in *blocked* mode */
1687 rc = mdt_object_lock(info, lnkp, lhl, MDS_INODELOCK_UPDATE,
1692 if (mdt_object_remote(lnkp)) {
1693 struct ldlm_lock *lock;
1696 * for remote object, set lock cb_atomic, so lock can be
1697 * released in blocking_ast() immediately, then the next
1698 * lock_try will have better chance of success.
1700 lock = ldlm_handle2lock(&lhl->mlh_rreg_lh);
1701 LASSERT(lock != NULL);
1702 lock_res_and_lock(lock);
1703 ldlm_set_atomic_cb(lock);
1704 unlock_res_and_lock(lock);
1705 LDLM_LOCK_PUT(lock);
1708 mdt_object_unlock(info, lnkp, lhl, 1);
1713 * we can't follow parent-child lock order like other MD
1714 * operations, use lock_try here to avoid deadlock, if the lock
1715 * cannot be taken, drop all locks taken, revoke the blocked
1716 * one, and continue processing the remaining entries, and in
1717 * the end of the loop restart from beginning.
1719 * don't lock with PDO mode in case two links are under the same
1720 * parent and their hash values are different.
1722 rc = mdt_object_lock_try(info, lnkp, lhl, &ibits, MDS_INODELOCK_UPDATE,
1727 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1728 CDEBUG(D_INFO, "busy lock on "DFID"\n", PFID(fid));
1733 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, update_locks);
1735 mdt_object_unlock(info, lnkp, lhl, 1);
1742 /* take link LOOKUP lock.
1743 * \retval 0 \a lnkp is already locked, no lock taken.
1745 * -ev negative errno.
1747 static int mdt_migrate_link_lock(struct mdt_thread_info *info,
1748 struct mdt_object *lnkp,
1749 struct mdt_object *spobj,
1750 struct mdt_object *obj,
1751 struct list_head *lookup_locks)
1753 const struct lu_fid *fid = mdt_object_fid(lnkp);
1754 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1755 struct mdt_link_lock *entry;
1760 /* check if it's already locked by source */
1761 rc = mdt_fids_different_target(info, fid, mdt_object_fid(spobj));
1763 CDEBUG(D_INFO, "skip lookup lock on source parent "DFID"\n",
1768 /* check if it's already locked by other links */
1769 list_for_each_entry(entry, lookup_locks, mll_linkage) {
1770 rc = mdt_fids_different_target(info, fid,
1771 mdt_object_fid(entry->mll_obj));
1773 CDEBUG(D_INFO, "skip lookup lock on parent "DFID"\n",
1779 rc = mdt_object_lookup_lock(info, lnkp, obj, lhl, LCK_EX, true);
1783 /* don't take local LOOKUP lock, because later we will lock other ibits
1784 * of sobj (which is on local MDT), and lock the same object twice may
1785 * deadlock, just revoke this lock.
1787 if (!mdt_object_remote(lnkp))
1788 GOTO(unlock, rc = 0);
1790 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, lookup_locks);
1796 mdt_object_unlock(info, lnkp, lhl, 1);
1801 * take UPDATE lock of link parents and LOOKUP lock of links, also check whether
1802 * total local lock count exceeds RS_MAX_LOCKS.
1804 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1805 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1806 * \retval -ev negative errno upon error
1808 static int mdt_migrate_links_lock(struct mdt_thread_info *info,
1809 struct mdt_object *spobj,
1810 struct mdt_object *tpobj,
1811 struct mdt_object *obj,
1812 struct mdt_lock_handle *lhsp,
1813 struct mdt_lock_handle *lhtp,
1814 struct list_head *link_locks)
1816 struct mdt_device *mdt = info->mti_mdt;
1817 struct lu_buf *buf = &info->mti_big_buf;
1818 struct lu_name *lname = &info->mti_name;
1819 struct linkea_data ldata = { NULL };
1820 int local_lock_cnt = 0;
1821 bool blocked = false;
1823 struct mdt_object *lnkp;
1825 LIST_HEAD(update_locks);
1826 LIST_HEAD(lookup_locks);
1830 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1833 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1834 if (buf->lb_buf == NULL)
1838 rc = mdt_links_read(info, obj, &ldata);
1840 if (rc == -ENOENT || rc == -ENODATA)
1845 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1846 linkea_next_entry(&ldata)) {
1847 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1850 /* check if link parent is source parent too */
1851 if (lu_fid_eq(mdt_object_fid(spobj), &fid)) {
1853 "skip lock on source parent "DFID"/"DNAME"\n",
1854 PFID(&fid), PNAME(lname));
1858 /* check if link parent is target parent too */
1859 if (tpobj != spobj && lu_fid_eq(mdt_object_fid(tpobj), &fid)) {
1861 "skip lock on target parent "DFID"/"DNAME"\n",
1862 PFID(&fid), PNAME(lname));
1866 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1868 CWARN("%s: cannot find obj "DFID": %ld\n",
1869 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1873 if (!mdt_object_exists(lnkp)) {
1874 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1875 PFID(&fid), PNAME(lname));
1876 mdt_object_put(info->mti_env, lnkp);
1881 rc = mdt_migrate_link_parent_lock(info, lnkp, &update_locks,
1883 if (!saved && blocked) {
1884 /* unlock all locks taken to avoid deadlock */
1885 mdt_migrate_links_unlock(info, &update_locks, 1);
1886 mdt_object_unlock(info, spobj, lhsp, 1);
1888 mdt_object_unlock(info, tpobj, lhtp, 1);
1892 mdt_object_put(info->mti_env, lnkp);
1896 if (rc == 1 && !mdt_object_remote(lnkp))
1899 rc = mdt_migrate_link_lock(info, lnkp, spobj, obj,
1902 mdt_object_put(info->mti_env, lnkp);
1905 if (rc == 1 && !mdt_object_remote(lnkp))
1907 mdt_object_put(info->mti_env, lnkp);
1911 GOTO(out, rc = -EBUSY);
1915 list_splice(&update_locks, link_locks);
1916 list_splice(&lookup_locks, link_locks);
1918 mdt_migrate_links_unlock(info, link_locks, rc);
1919 } else if (local_lock_cnt > RS_MAX_LOCKS - 5) {
1921 * parent may have 3 local objects: master object and 2 stripes
1922 * (if it's being migrated too); source may have 1 local objects
1923 * as regular file; target has 1 local object.
1924 * Note, source may have 2 local locks if it is directory but it
1925 * can't have hardlinks, so it is not considered here.
1927 CDEBUG(D_INFO, "Too many local locks (%d), migrate in sync mode\n",
1935 * lookup source by name, if parent is striped directory, we need to find the
1936 * corresponding stripe where source is located, and then lookup there.
1938 * besides, if parent is migrating too, and file is already in target stripe,
1939 * this should be a redo of 'lfs migrate' on client side.
1941 * \retval 1 tpobj stripe index is less than spobj stripe index
1942 * \retval 0 tpobj stripe index is larger than or equal to spobj stripe index
1943 * \retval -ev negative errno upon error
1945 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1946 struct mdt_object *pobj,
1947 const struct md_attr *ma,
1948 const struct lu_name *lname,
1949 struct mdt_object **spobj,
1950 struct mdt_object **tpobj,
1951 struct mdt_object **sobj)
1953 const struct lu_env *env = info->mti_env;
1954 struct lu_fid *fid = &info->mti_tmp_fid1;
1959 if (ma->ma_valid & MA_LMV) {
1960 /* if parent is striped, lookup on corresponding stripe */
1961 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1962 struct lu_fid *fid2 = &info->mti_tmp_fid2;
1964 if (!lmv_is_sane(lmv))
1967 spindex = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1972 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[spindex]);
1974 *spobj = mdt_object_find(env, info->mti_mdt, fid2);
1975 if (IS_ERR(*spobj)) {
1976 rc = PTR_ERR(*spobj);
1981 if (!mdt_object_exists(*spobj))
1982 GOTO(spobj_put, rc = -ENOENT);
1985 rc = mdo_lookup(env, mdt_object_child(*spobj), lname, fid,
1987 if ((rc == -ENOENT || rc == 0) && lmv_is_layout_changing(lmv)) {
1988 /* fail check here to let top dir migration succeed. */
1989 if (CFS_FAIL_CHECK_RESET(OBD_FAIL_MIGRATE_ENTRIES, 0))
1990 GOTO(spobj_put, rc = -EIO);
1993 * if parent layout is changeing, and lookup child
1994 * failed on source stripe, lookup again on target
1995 * stripe, if it exists, it means previous migration
1996 * was interrupted, and current file was migrated
1999 tpindex = lmv_name_to_stripe_index(lmv, lname->ln_name,
2002 GOTO(spobj_put, rc = tpindex);
2004 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[tpindex]);
2006 *tpobj = mdt_object_find(env, info->mti_mdt, fid2);
2007 if (IS_ERR(*tpobj)) {
2008 rc = PTR_ERR(*tpobj);
2010 GOTO(spobj_put, rc);
2013 if (!mdt_object_exists(*tpobj))
2014 GOTO(tpobj_put, rc = -ENOENT);
2016 if (rc == -ENOENT) {
2018 rc = mdo_lookup(env, mdt_object_child(*tpobj),
2019 lname, fid, &info->mti_spec);
2020 GOTO(tpobj_put, rc = rc ?: -EALREADY);
2023 GOTO(spobj_put, rc);
2027 mdt_object_get(env, *tpobj);
2031 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2038 mdt_object_get(env, pobj);
2039 mdt_object_get(env, pobj);
2042 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2043 if (IS_ERR(*sobj)) {
2044 rc = PTR_ERR(*sobj);
2046 GOTO(tpobj_put, rc);
2049 if (!mdt_object_exists(*sobj))
2050 GOTO(sobj_put, rc = -ENOENT);
2052 return (tpindex < spindex);
2055 mdt_object_put(env, *sobj);
2058 mdt_object_put(env, *tpobj);
2061 mdt_object_put(env, *spobj);
2067 /* end lease and close file for regular file */
2068 static int mdd_migrate_close(struct mdt_thread_info *info,
2069 struct mdt_object *obj)
2071 struct close_data *data;
2072 struct mdt_body *repbody;
2073 struct ldlm_lock *lease;
2078 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2080 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2084 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2089 lease = ldlm_handle2lock(&data->cd_handle);
2093 /* check if the lease was already canceled */
2094 lock_res_and_lock(lease);
2095 rc = ldlm_is_cancel(lease);
2096 unlock_res_and_lock(lease);
2100 LDLM_DEBUG(lease, DFID" lease broken",
2101 PFID(mdt_object_fid(obj)));
2105 * cancel server side lease, client side counterpart should have been
2106 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2108 ldlm_lock_cancel(lease);
2109 ldlm_reprocess_all(lease->l_resource,
2110 lease->l_policy_data.l_inodebits.bits);
2111 LDLM_LOCK_PUT(lease);
2114 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2115 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2116 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2121 /* LFSCK used to clear hash type and MIGRATION flag upon migration failure */
2122 static inline bool lmv_is_failed_migration(const struct lmv_mds_md_v1 *lmv)
2124 return le32_to_cpu(lmv->lmv_hash_type) ==
2125 (LMV_HASH_TYPE_UNKNOWN | LMV_HASH_FLAG_BAD_TYPE) &&
2126 lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_migrate_hash)) &&
2127 le32_to_cpu(lmv->lmv_migrate_offset) > 0 &&
2128 le32_to_cpu(lmv->lmv_migrate_offset) <
2129 le32_to_cpu(lmv->lmv_stripe_count);
2133 * migrate file in below steps:
2134 * 1. lock source and target stripes
2135 * 2. lookup source by name
2136 * 3. lock parents of source links if source is not directory
2137 * 4. reject if source is in HSM
2138 * 5. take source open_sem and close file if source is regular file
2139 * 6. lock source, and its stripes if it's directory
2141 * 8. lock target so subsequent change to it can trigger COS
2142 * 9. unlock above locks
2143 * 10. sync device if source has too many links
2145 int mdt_reint_migrate(struct mdt_thread_info *info,
2146 struct mdt_lock_handle *unused)
2148 const struct lu_env *env = info->mti_env;
2149 struct mdt_device *mdt = info->mti_mdt;
2150 struct ptlrpc_request *req = mdt_info_req(info);
2151 struct mdt_reint_record *rr = &info->mti_rr;
2152 struct lu_ucred *uc = mdt_ucred(info);
2153 struct md_attr *ma = &info->mti_attr;
2154 struct mdt_object *pobj;
2155 struct mdt_object *spobj;
2156 struct mdt_object *tpobj;
2157 struct mdt_object *sobj;
2158 struct mdt_object *tobj;
2159 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2160 struct mdt_lock_handle *lhsp;
2161 struct mdt_lock_handle *lhtp;
2162 struct mdt_lock_handle *lhs;
2163 struct mdt_lock_handle *lhl;
2164 LIST_HEAD(link_locks);
2165 int lock_retries = 5;
2166 bool reverse = false;
2167 bool open_sem_locked = false;
2168 bool do_sync = false;
2169 bool is_plain_dir = false;
2173 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2174 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2176 if (info->mti_dlm_req)
2177 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2179 if (!fid_is_md_operative(rr->rr_fid1) ||
2180 !fid_is_md_operative(rr->rr_fid2))
2183 /* don't allow migrate . or .. */
2184 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2187 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2190 /* we want rbac roles to have precedence over any other
2191 * permission or capability checks
2193 if (uc && (!uc->uc_rbac_dne_ops ||
2194 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2195 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2196 mdt->mdt_enable_remote_dir_gid != -1)))
2200 * Note: do not enqueue rename lock for replay request, because
2201 * if other MDT holds rename lock, but being blocked to wait for
2202 * this MDT to finish its recovery, and the failover MDT can not
2203 * get rename lock, which will cause deadlock.
2205 * req is NULL if this is called by directory auto-split.
2207 if (req && !req_is_replay(req)) {
2208 rc = mdt_rename_lock(info, rename_lh);
2210 CERROR("%s: can't lock FS for rename: rc = %d\n",
2211 mdt_obd_name(info->mti_mdt), rc);
2216 /* pobj is master object of parent */
2217 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2219 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2222 rc = mdt_version_get_check(info, pobj, 0);
2224 GOTO(put_parent, rc);
2227 if (!mdt_object_exists(pobj))
2228 GOTO(put_parent, rc = -ENOENT);
2230 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2231 GOTO(put_parent, rc = -ENOTDIR);
2233 rc = mdt_check_enc(info, pobj);
2235 GOTO(put_parent, rc);
2237 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2239 GOTO(put_parent, rc);
2241 if (CFS_FAIL_CHECK(OBD_FAIL_MIGRATE_BAD_HASH) &&
2242 (ma->ma_valid & MA_LMV) &&
2243 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1)) {
2244 struct lu_buf *buf = &info->mti_buf;
2245 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2246 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2248 lmv->lmv_hash_type = cpu_to_le32(LMV_HASH_TYPE_UNKNOWN |
2249 LMV_HASH_FLAG_BAD_TYPE);
2250 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2252 buf->lb_len = sizeof(*lmv);
2253 rc = mo_xattr_set(env, mdt_object_child(pobj), buf,
2254 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2255 mo_invalidate(env, mdt_object_child(pobj));
2256 GOTO(put_parent, rc);
2259 /* @spobj is the parent stripe of @sobj if @pobj is striped directory,
2260 * if @pobj is migrating too, tpobj is the target parent stripe.
2262 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &tpobj,
2265 GOTO(put_parent, rc);
2268 /* parent unchanged, this happens in dir restripe */
2269 if (info->mti_spec.sp_migrate_nsonly && spobj == tpobj)
2270 GOTO(put_source, rc = -EALREADY);
2275 lhsp = &info->mti_lh[MDT_LH_PARENT];
2276 lhtp = &info->mti_lh[MDT_LH_CHILD];
2277 /* lock spobj and tpobj in stripe index order */
2279 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name, LCK_PW,
2282 GOTO(put_source, rc);
2284 LASSERT(spobj != tpobj);
2285 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW,
2288 GOTO(unlock_parent, rc);
2290 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW,
2293 GOTO(put_source, rc);
2295 if (tpobj != spobj) {
2296 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name,
2299 GOTO(unlock_parent, rc);
2303 /* if inode is not migrated, or is dir, no need to lock links */
2304 if (!info->mti_spec.sp_migrate_nsonly &&
2305 !S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2306 /* lock link parents, and take LOOKUP lock of links */
2307 rc = mdt_migrate_links_lock(info, spobj, tpobj, sobj, lhsp,
2309 if (rc == -EBUSY && lock_retries-- > 0) {
2310 LASSERT(list_empty(&link_locks));
2315 GOTO(put_source, rc);
2318 * RS_MAX_LOCKS is the limit of number of locks that can be
2319 * saved along with one request, if total lock count exceeds
2320 * this limit, we will drop all locks after migration, and
2321 * trigger commit in the end.
2327 lhs = &info->mti_lh[MDT_LH_OLD];
2328 lhl = &info->mti_lh[MDT_LH_LOOKUP];
2329 rc = mdt_rename_source_lock(info, spobj, sobj, lhs, lhl,
2330 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR |
2331 MDS_INODELOCK_OPEN, true);
2333 GOTO(unlock_links, rc);
2335 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2336 /* TODO: DoM migration is not supported, migrate dirent only */
2337 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2339 GOTO(unlock_source, rc);
2341 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2342 info->mti_spec.sp_migrate_nsonly = 1;
2343 } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2344 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2346 GOTO(unlock_source, rc);
2348 if (!(ma->ma_valid & MA_LMV))
2349 is_plain_dir = true;
2350 else if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2351 /* race with restripe/auto-split */
2352 GOTO(unlock_source, rc = -EBUSY);
2353 else if (lmv_is_failed_migration(&ma->ma_lmv->lmv_md_v1)) {
2354 struct lu_buf *buf = &info->mti_buf;
2355 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2356 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2358 /* migration failed before, and LFSCK cleared hash type
2359 * and flags, fake it to resume migration.
2361 lmv->lmv_hash_type =
2362 cpu_to_le32(LMV_HASH_TYPE_FNV_1A_64 |
2363 LMV_HASH_FLAG_MIGRATION |
2364 LMV_HASH_FLAG_BAD_TYPE |
2365 LMV_HASH_FLAG_FIXED);
2366 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2368 buf->lb_len = sizeof(*lmv);
2369 rc = mo_xattr_set(env, mdt_object_child(sobj), buf,
2370 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2371 mo_invalidate(env, mdt_object_child(sobj));
2372 GOTO(unlock_source, rc = -EALREADY);
2376 /* if migration HSM is allowed */
2377 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2378 ma->ma_need = MA_HSM;
2380 rc = mdt_attr_get_complex(info, sobj, ma);
2382 GOTO(unlock_source, rc);
2384 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2385 GOTO(unlock_source, rc = -EOPNOTSUPP);
2388 /* end lease and close file for regular file */
2389 if (info->mti_spec.sp_migrate_close) {
2390 /* try to hold open_sem so that nobody else can open the file */
2391 if (!down_write_trylock(&sobj->mot_open_sem)) {
2393 mdd_migrate_close(info, sobj);
2394 GOTO(unlock_source, rc = -EBUSY);
2396 open_sem_locked = true;
2397 rc = mdd_migrate_close(info, sobj);
2398 if (rc && rc != -ESTALE)
2399 GOTO(unlock_open_sem, rc);
2403 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2405 GOTO(unlock_open_sem, rc = PTR_ERR(tobj));
2407 /* Don't do lookup sanity check. We know name doesn't exist. */
2408 info->mti_spec.sp_cr_lookup = 0;
2409 info->mti_spec.sp_feat = &dt_directory_features;
2411 rc = mdo_migrate(env, mdt_object_child(spobj),
2412 mdt_object_child(tpobj), mdt_object_child(sobj),
2413 mdt_object_child(tobj), &rr->rr_name,
2414 &info->mti_spec, ma);
2416 GOTO(put_target, rc);
2418 /* save target locks for directory */
2419 if (S_ISDIR(lu_object_attr(&sobj->mot_obj)) &&
2420 !info->mti_spec.sp_migrate_nsonly) {
2421 struct mdt_lock_handle *lht = &info->mti_lh[MDT_LH_NEW];
2422 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
2424 /* in case sobj becomes a stripe of tobj, unlock sobj here,
2425 * otherwise stripes lock may deadlock.
2428 mdt_rename_source_unlock(info, sobj, lhs, lhl, 1);
2430 rc = mdt_object_stripes_lock(info, tpobj, tobj, lht, einfo,
2431 MDS_INODELOCK_UPDATE, LCK_PW,
2434 GOTO(put_target, rc);
2436 mdt_object_stripes_unlock(info, tobj, lht, einfo, 0);
2439 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2440 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2444 mdt_object_put(env, tobj);
2446 if (open_sem_locked)
2447 up_write(&sobj->mot_open_sem);
2449 mdt_rename_source_unlock(info, sobj, lhs, lhl, rc);
2451 /* if we've got too many locks to save into RPC,
2452 * then just commit before the locks are released
2455 mdt_device_sync(env, mdt);
2456 mdt_migrate_links_unlock(info, &link_locks, do_sync ? 1 : rc);
2458 mdt_object_unlock(info, spobj, lhsp, rc);
2459 mdt_object_unlock(info, tpobj, lhtp, rc);
2461 mdt_object_put(env, sobj);
2462 mdt_object_put(env, spobj);
2463 mdt_object_put(env, tpobj);
2465 mo_invalidate(env, mdt_object_child(pobj));
2466 mdt_object_put(env, pobj);
2468 mdt_rename_unlock(info, rename_lh);
2471 CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
2472 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
2473 PNAME(&rr->rr_name), rc);
2479 * determine lock order of sobj and tobj
2481 * there are two situations we need to lock tobj before sobj:
2482 * 1. sobj is child of tobj
2483 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2484 * larger than that of tobj
2486 * \retval 1 lock tobj before sobj
2487 * \retval 0 lock sobj before tobj
2488 * \retval -ev negative errno upon error
2490 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2491 struct mdt_object *sobj,
2492 struct mdt_object *tobj)
2494 struct md_attr *ma = &info->mti_attr;
2495 struct lu_fid *spfid = &info->mti_tmp_fid1;
2496 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2497 struct lmv_mds_md_v1 *lmv;
2502 /* sobj and tobj are the same */
2506 if (fid_is_root(mdt_object_fid(sobj)))
2509 if (fid_is_root(mdt_object_fid(tobj)))
2512 /* check whether sobj is child of tobj */
2513 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2514 mdt_object_fid(tobj));
2521 /* check whether sobj and tobj are children of the same parent */
2522 rc = mdt_attr_get_pfid(info, sobj, spfid);
2526 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2530 if (!lu_fid_eq(spfid, tpfid))
2533 /* check whether sobj and tobj are sibling stripes */
2534 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2538 if (!(ma->ma_valid & MA_LMV))
2541 lmv = &ma->ma_lmv->lmv_md_v1;
2542 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2544 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2547 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2551 if (!(ma->ma_valid & MA_LMV))
2554 lmv = &ma->ma_lmv->lmv_md_v1;
2555 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2557 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2559 /* check stripe index of sobj and tobj */
2560 if (sindex == tindex)
2563 return sindex < tindex ? 0 : 1;
2566 /* Helper function for mdt_reint_rename so we don't need to opencode
2567 * two different order lockings
2569 static int mdt_lock_two_dirs(struct mdt_thread_info *info,
2570 struct mdt_object *mfirstdir,
2571 struct mdt_lock_handle *lh_firstdirp,
2572 const struct lu_name *firstname,
2573 struct mdt_object *mseconddir,
2574 struct mdt_lock_handle *lh_seconddirp,
2575 const struct lu_name *secondname,
2580 rc = mdt_parent_lock(info, mfirstdir, lh_firstdirp, firstname, LCK_PW,
2585 mdt_version_get_save(info, mfirstdir, 0);
2586 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2588 if (mfirstdir != mseconddir) {
2589 rc = mdt_parent_lock(info, mseconddir, lh_seconddirp,
2590 secondname, LCK_PW, cos_incompat);
2591 } else if (!mdt_object_remote(mseconddir)) {
2592 if (lh_firstdirp->mlh_pdo_hash !=
2593 lh_seconddirp->mlh_pdo_hash) {
2594 rc = mdt_object_pdo_lock(info, mseconddir,
2595 lh_seconddirp, secondname,
2596 LCK_PW, false, cos_incompat);
2597 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2600 mdt_version_get_save(info, mseconddir, 1);
2603 mdt_object_unlock(info, mfirstdir, lh_firstdirp, rc);
2609 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2610 * 2 - srcdir child; 3 - tgtdir child.
2611 * Update on disk version of srcdir child.
2613 static int mdt_reint_rename(struct mdt_thread_info *info,
2614 struct mdt_lock_handle *unused)
2616 struct mdt_device *mdt = info->mti_mdt;
2617 struct mdt_reint_record *rr = &info->mti_rr;
2618 struct md_attr *ma = &info->mti_attr;
2619 struct ptlrpc_request *req = mdt_info_req(info);
2620 struct mdt_object *msrcdir = NULL;
2621 struct mdt_object *mtgtdir = NULL;
2622 struct mdt_object *mold;
2623 struct mdt_object *mnew = NULL;
2624 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2625 struct mdt_lock_handle *lh_srcdirp;
2626 struct mdt_lock_handle *lh_tgtdirp;
2627 struct mdt_lock_handle *lh_oldp = NULL;
2628 struct mdt_lock_handle *lh_lookup = NULL;
2629 struct mdt_lock_handle *lh_newp = NULL;
2630 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2631 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2632 struct lu_ucred *uc = mdt_ucred(info);
2633 bool reverse = false, discard = false;
2635 ktime_t kstart = ktime_get();
2636 enum mdt_stat_idx msi = 0;
2640 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2641 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2642 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2644 if (info->mti_dlm_req)
2645 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2647 if (!fid_is_md_operative(rr->rr_fid1) ||
2648 !fid_is_md_operative(rr->rr_fid2))
2651 /* find both parents. */
2652 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2653 if (IS_ERR(msrcdir))
2654 RETURN(PTR_ERR(msrcdir));
2656 rc = mdt_check_enc(info, msrcdir);
2658 GOTO(out_put_srcdir, rc);
2660 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2662 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2664 mdt_object_get(info->mti_env, mtgtdir);
2666 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2667 if (IS_ERR(mtgtdir))
2668 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2671 rc = mdt_check_enc(info, mtgtdir);
2673 GOTO(out_put_tgtdir, rc);
2675 if (!uc->uc_rbac_fscrypt_admin &&
2676 mtgtdir->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
2677 GOTO(out_put_tgtdir, rc = -EPERM);
2680 * Note: do not enqueue rename lock for replay request, because
2681 * if other MDT holds rename lock, but being blocked to wait for
2682 * this MDT to finish its recovery, and the failover MDT can not
2683 * get rename lock, which will cause deadlock.
2685 if (!req_is_replay(req)) {
2686 bool remote = mdt_object_remote(msrcdir);
2689 * Normally rename RPC is handled on the MDT with the target
2690 * directory (if target exists, it's on the MDT with the
2691 * target), if the source directory is remote, it's a hint that
2692 * source is remote too (this may not be true, but it won't
2693 * cause any issue), return -EXDEV early to avoid taking
2696 if (!mdt->mdt_enable_remote_rename && remote)
2697 GOTO(out_put_tgtdir, rc = -EXDEV);
2699 /* This might be further relaxed in the future for regular file
2700 * renames in different source and target parents. Start with
2701 * only same-directory renames for simplicity and because this
2702 * is by far the most the common use case.
2704 * Striped directories should be considered "remote".
2706 if (msrcdir != mtgtdir || remote ||
2707 (S_ISDIR(ma->ma_attr.la_mode) &&
2708 !mdt->mdt_enable_parallel_rename_dir) ||
2709 (!S_ISDIR(ma->ma_attr.la_mode) &&
2710 !mdt->mdt_enable_parallel_rename_file)) {
2711 rc = mdt_rename_lock(info, rename_lh);
2713 CERROR("%s: cannot lock for rename: rc = %d\n",
2714 mdt_obd_name(mdt), rc);
2715 GOTO(out_put_tgtdir, rc);
2718 if (S_ISDIR(ma->ma_attr.la_mode))
2719 msi = LPROC_MDT_RENAME_PAR_DIR;
2721 msi = LPROC_MDT_RENAME_PAR_FILE;
2724 "%s: samedir parallel rename "DFID"/"DNAME"\n",
2725 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2726 PNAME(&rr->rr_name));
2730 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2732 GOTO(out_unlock_rename, rc);
2735 /* source needs to be looked up after locking source parent, otherwise
2736 * this rename may race with unlink source, and cause rename hang, see
2737 * sanityn.sh 55b, so check parents first, if later we found source is
2738 * remote, relock parents.
2740 cos_incompat = (mdt_object_remote(msrcdir) ||
2741 mdt_object_remote(mtgtdir));
2743 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2745 /* lock parents in the proper order. */
2746 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2747 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2749 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
2750 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2752 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2753 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2755 /* In case of same dir local rename we must sort by the hash,
2756 * otherwise a lock deadlock is possible when renaming
2757 * a to b and b to a at the same time LU-15285
2759 if (!mdt_object_remote(mtgtdir) && mtgtdir == msrcdir)
2760 reverse = lh_srcdirp->mlh_pdo_hash > lh_tgtdirp->mlh_pdo_hash;
2761 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)))
2765 rc = mdt_lock_two_dirs(info, mtgtdir, lh_tgtdirp,
2766 &rr->rr_tgt_name, msrcdir, lh_srcdirp,
2767 &rr->rr_name, cos_incompat);
2769 rc = mdt_lock_two_dirs(info, msrcdir, lh_srcdirp, &rr->rr_name,
2770 mtgtdir, lh_tgtdirp, &rr->rr_tgt_name,
2774 GOTO(out_unlock_rename, rc);
2776 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2777 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2779 /* find mold object. */
2781 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2783 GOTO(out_unlock_parents, rc);
2785 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2786 GOTO(out_unlock_parents, rc = -EINVAL);
2788 if (!fid_is_md_operative(old_fid))
2789 GOTO(out_unlock_parents, rc = -EPERM);
2791 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2793 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2795 if (!mdt_object_exists(mold)) {
2796 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2798 "object does not exist");
2799 GOTO(out_put_old, rc = -ENOENT);
2802 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2803 GOTO(out_put_old, rc = -EXDEV);
2805 /* Check if @mtgtdir is subdir of @mold, before locking child
2806 * to avoid reverse locking.
2808 if (mtgtdir != msrcdir) {
2809 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2814 GOTO(out_put_old, rc);
2818 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2819 /* save version after locking */
2820 mdt_version_get_save(info, mold, 2);
2822 if (!cos_incompat && mdt_object_remote(mold)) {
2823 cos_incompat = true;
2824 mdt_object_put(info->mti_env, mold);
2825 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2826 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2830 /* find mnew object:
2831 * mnew target object may not exist now
2832 * lookup with version checking
2835 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2838 /* the new_fid should have been filled at this moment */
2839 if (lu_fid_eq(old_fid, new_fid))
2840 GOTO(out_put_old, rc);
2842 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2843 lu_fid_eq(new_fid, rr->rr_fid2))
2844 GOTO(out_put_old, rc = -EINVAL);
2846 if (!fid_is_md_operative(new_fid))
2847 GOTO(out_put_old, rc = -EPERM);
2849 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2851 GOTO(out_put_old, rc = PTR_ERR(mnew));
2853 if (!mdt_object_exists(mnew)) {
2854 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2856 "object does not exist");
2857 GOTO(out_put_new, rc = -ENOENT);
2860 if (mdt_object_remote(mnew)) {
2861 struct mdt_body *repbody;
2863 /* Always send rename req to the target child MDT */
2864 repbody = req_capsule_server_get(info->mti_pill,
2866 LASSERT(repbody != NULL);
2867 repbody->mbo_fid1 = *new_fid;
2868 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2869 GOTO(out_put_new, rc = -EXDEV);
2871 /* Before locking the target dir, check we do not replace
2872 * a dir with a non-dir, otherwise it may deadlock with
2873 * link op which tries to create a link in this dir
2874 * back to this non-dir.
2876 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2877 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2878 GOTO(out_put_new, rc = -EISDIR);
2880 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2881 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2882 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2884 MDS_INODELOCK_LOOKUP |
2885 MDS_INODELOCK_XATTR, cos_incompat);
2887 GOTO(out_put_new, rc);
2889 /* Check if @msrcdir is subdir of @mnew, before locking child
2890 * to avoid reverse locking.
2892 if (mtgtdir != msrcdir) {
2893 rc = mdo_is_subdir(info->mti_env,
2894 mdt_object_child(msrcdir), new_fid);
2898 GOTO(out_unlock_old, rc);
2902 /* We used to acquire MDS_INODELOCK_FULL here but we
2903 * can't do this now because a running HSM restore on
2904 * the rename onto victim will hold the layout
2905 * lock. See LU-4002.
2908 lh_newp = &info->mti_lh[MDT_LH_NEW];
2909 rc = mdt_object_check_lock(info, mtgtdir, mnew, lh_newp,
2910 MDS_INODELOCK_LOOKUP |
2911 MDS_INODELOCK_UPDATE, LCK_EX,
2914 GOTO(out_unlock_new, rc);
2916 /* get and save version after locking */
2917 mdt_version_get_save(info, mnew, 3);
2918 } else if (rc != -ENOENT) {
2919 GOTO(out_put_old, rc);
2921 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2922 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2923 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2925 MDS_INODELOCK_LOOKUP |
2926 MDS_INODELOCK_XATTR, cos_incompat);
2928 GOTO(out_put_old, rc);
2930 mdt_enoent_version_save(info, 3);
2933 /* step 5: rename it */
2934 mdt_reint_init_ma(info, ma);
2936 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2937 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2940 mutex_lock(&mnew->mot_lov_mutex);
2942 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2943 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2944 mnew != NULL ? mdt_object_child(mnew) : NULL,
2945 &rr->rr_tgt_name, ma);
2948 mutex_unlock(&mnew->mot_lov_mutex);
2950 /* handle last link of tgt object */
2953 mdt_handle_last_unlink(info, mnew, ma);
2954 discard = mdt_dom_check_for_discard(info, mnew);
2956 mdt_rename_counter_tally(info, info->mti_mdt, req,
2957 msrcdir, mtgtdir, msi,
2958 ktime_us_delta(ktime_get(), kstart));
2964 mdt_object_unlock(info, mnew, lh_newp, rc);
2966 mdt_object_unlock(info, NULL, lh_lookup, rc);
2967 mdt_object_unlock(info, mold, lh_oldp, rc);
2969 if (mnew && !discard)
2970 mdt_object_put(info->mti_env, mnew);
2972 mdt_object_put(info->mti_env, mold);
2974 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2975 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2977 mdt_rename_unlock(info, rename_lh);
2979 mdt_object_put(info->mti_env, mtgtdir);
2981 mdt_object_put(info->mti_env, msrcdir);
2983 /* The DoM discard can be done right in the place above where it is
2984 * assigned, meanwhile it is done here after rename unlock due to
2985 * compatibility with old clients, for them the discard blocks
2986 * the main thread until completion. Check LU-11359 for details.
2989 mdt_dom_discard_data(info, mnew);
2990 mdt_object_put(info->mti_env, mnew);
2992 CFS_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
2996 static int mdt_reint_resync(struct mdt_thread_info *info,
2997 struct mdt_lock_handle *lhc)
2999 struct mdt_reint_record *rr = &info->mti_rr;
3000 struct ptlrpc_request *req = mdt_info_req(info);
3001 struct md_attr *ma = &info->mti_attr;
3002 struct mdt_object *mo;
3003 struct ldlm_lock *lease;
3004 struct mdt_body *repbody;
3005 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
3010 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
3012 if (info->mti_dlm_req)
3013 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
3015 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
3017 GOTO(out, rc = PTR_ERR(mo));
3019 if (!mdt_object_exists(mo))
3020 GOTO(out_obj, rc = -ENOENT);
3022 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
3023 GOTO(out_obj, rc = -EINVAL);
3025 if (mdt_object_remote(mo))
3026 GOTO(out_obj, rc = -EREMOTE);
3028 lease = ldlm_handle2lock(rr->rr_lease_handle);
3030 GOTO(out_obj, rc = -ESTALE);
3032 /* It's really necessary to grab open_sem and check if the lease lock
3033 * has been lost. There would exist a concurrent writer coming in and
3034 * generating some dirty data in memory cache, the writeback would fail
3035 * after the layout version is increased by MDS_REINT_RESYNC RPC.
3037 if (!down_write_trylock(&mo->mot_open_sem))
3038 GOTO(out_put_lease, rc = -EBUSY);
3040 lock_res_and_lock(lease);
3041 lease_broken = ldlm_is_cancel(lease);
3042 unlock_res_and_lock(lease);
3044 GOTO(out_unlock, rc = -EBUSY);
3046 /* the file has yet opened by anyone else after we took the lease. */
3047 layout.mlc_opc = MD_LAYOUT_RESYNC;
3048 lhc = &info->mti_lh[MDT_LH_LOCAL];
3049 rc = mdt_layout_change(info, mo, lhc, &layout);
3051 GOTO(out_unlock, rc);
3053 mdt_object_unlock(info, mo, lhc, 0);
3055 ma->ma_need = MA_INODE;
3057 rc = mdt_attr_get_complex(info, mo, ma);
3059 GOTO(out_unlock, rc);
3061 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3062 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3066 up_write(&mo->mot_open_sem);
3068 LDLM_LOCK_PUT(lease);
3070 mdt_object_put(info->mti_env, mo);
3072 mdt_client_compatibility(info);
3076 struct mdt_reinter {
3077 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3078 enum lprocfs_extra_opc mr_extra_opc;
3081 static const struct mdt_reinter mdt_reinters[] = {
3083 .mr_handler = &mdt_reint_setattr,
3084 .mr_extra_opc = MDS_REINT_SETATTR,
3087 .mr_handler = &mdt_reint_create,
3088 .mr_extra_opc = MDS_REINT_CREATE,
3091 .mr_handler = &mdt_reint_link,
3092 .mr_extra_opc = MDS_REINT_LINK,
3095 .mr_handler = &mdt_reint_unlink,
3096 .mr_extra_opc = MDS_REINT_UNLINK,
3099 .mr_handler = &mdt_reint_rename,
3100 .mr_extra_opc = MDS_REINT_RENAME,
3103 .mr_handler = &mdt_reint_open,
3104 .mr_extra_opc = MDS_REINT_OPEN,
3106 [REINT_SETXATTR] = {
3107 .mr_handler = &mdt_reint_setxattr,
3108 .mr_extra_opc = MDS_REINT_SETXATTR,
3111 .mr_handler = &mdt_reint_unlink,
3112 .mr_extra_opc = MDS_REINT_UNLINK,
3115 .mr_handler = &mdt_reint_migrate,
3116 .mr_extra_opc = MDS_REINT_RENAME,
3119 .mr_handler = &mdt_reint_resync,
3120 .mr_extra_opc = MDS_REINT_RESYNC,
3124 int mdt_reint_rec(struct mdt_thread_info *info,
3125 struct mdt_lock_handle *lhc)
3127 const struct mdt_reinter *mr;
3131 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3134 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3135 if (mr->mr_handler == NULL)
3138 rc = (*mr->mr_handler)(info, lhc);
3140 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3141 PTLRPC_LAST_CNTR + mr->mr_extra_opc);