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 && OBD_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 if (S_ISDIR(ma->ma_attr.la_mode) &&
543 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
544 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
545 struct obd_export *exp = mdt_info_req(info)->rq_export;
547 /* Only new clients can create remote dir( >= 2.4) and
548 * striped dir(>= 2.6), old client will return -ENOTSUPP
550 if (!mdt_is_dne_client(exp))
553 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
554 if (!mdt_is_striped_client(exp))
557 if (!mdt->mdt_enable_striped_dir)
559 } else if (!mdt->mdt_enable_remote_dir) {
563 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
564 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) >=
568 /* we want rbac roles to have precedence over any other
569 * permission or capability checks
571 if (!uc->uc_rbac_dne_ops ||
572 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
573 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
574 mdt->mdt_enable_remote_dir_gid != -1))
577 /* restripe if later found dir exists, MDS_OPEN_CREAT means
578 * this is create only, don't try restripe.
580 if (mdt->mdt_enable_dir_restripe &&
581 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
582 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
586 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
588 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
590 RETURN(PTR_ERR(parent));
592 if (!mdt_object_exists(parent))
593 GOTO(put_parent, rc = -ENOENT);
595 rc = mdt_check_enc(info, parent);
597 GOTO(put_parent, rc);
599 if (!uc->uc_rbac_fscrypt_admin &&
600 parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
601 GOTO(put_parent, rc = -EPERM);
604 * LU-10235: check if name exists locklessly first to avoid massive
605 * lock recalls on existing directories.
607 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
608 &info->mti_tmp_fid1, 1);
611 GOTO(put_parent, rc = -EEXIST);
613 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
617 /* -ENOENT is expected here */
619 GOTO(put_parent, rc);
621 /* save version of file name for replay, it must be ENOENT here */
622 mdt_enoent_version_save(info, 1);
624 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
626 lh = &info->mti_lh[MDT_LH_PARENT];
627 rc = mdt_parent_lock(info, parent, lh, &rr->rr_name, LCK_PW, false);
629 GOTO(put_parent, rc);
631 if (!mdt_object_remote(parent)) {
632 rc = mdt_version_get_check_save(info, parent, 0);
634 GOTO(unlock_parent, rc);
637 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
638 if (unlikely(IS_ERR(child)))
639 GOTO(unlock_parent, rc = PTR_ERR(child));
641 ma->ma_need = MA_INODE;
644 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
645 OBD_FAIL_MDS_REINT_CREATE_WRITE);
647 /* Version of child will be updated on disk. */
648 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
649 rc = mdt_version_get_check_save(info, child, 2);
653 if (parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD ||
654 (rr->rr_name.ln_namelen == strlen(dot_fscrypt_name) &&
655 strncmp(rr->rr_name.ln_name, dot_fscrypt_name,
656 rr->rr_name.ln_namelen) == 0))
657 child->mot_obj.lo_header->loh_attr |= LOHA_FSCRYPT_MD;
660 * Do not perform lookup sanity check. We know that name does
663 info->mti_spec.sp_cr_lookup = 0;
664 if (mdt_object_remote(parent))
665 info->mti_spec.sp_cr_lookup = 1;
666 info->mti_spec.sp_feat = &dt_directory_features;
668 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
669 mdt_object_child(child), &info->mti_spec, ma);
671 rc = mdt_attr_get_complex(info, child, ma);
677 * On DNE, we need to eliminate dependey between 'mkdir a' and
678 * 'mkdir a/b' if b is a striped directory, to achieve this, two
679 * things are done below:
680 * 1. save child and slaves lock.
681 * 2. if the child is a striped directory, relock parent so to
682 * compare against with COS locks to ensure parent was
685 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
686 struct mdt_lock_handle *lhc;
687 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
690 rc = mdt_object_striped(info, child);
696 if (!mdt_object_remote(parent)) {
697 mdt_object_unlock(info, parent, lh, 1);
698 rc = mdt_parent_lock(info, parent, lh,
699 &rr->rr_name, LCK_PW,
706 lhc = &info->mti_lh[MDT_LH_CHILD];
707 rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
708 MDS_INODELOCK_UPDATE, LCK_PW,
713 mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
716 /* Return fid & attr to client. */
717 if (ma->ma_valid & MA_INODE)
718 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
719 mdt_object_fid(child));
722 mdt_object_put(info->mti_env, child);
724 mdt_object_unlock(info, parent, lh, rc);
726 mdt_object_put(info->mti_env, parent);
730 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
733 struct mdt_lock_handle *lh;
734 int do_vbr = ma->ma_attr.la_valid &
735 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
736 __u64 lockpart = MDS_INODELOCK_UPDATE;
737 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
742 rc = mdt_object_striped(info, mo);
747 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
748 lockpart |= MDS_INODELOCK_PERM;
749 /* Clear xattr cache on clients, so the virtual project ID xattr
750 * can get the new project ID
752 if (ma->ma_attr.la_valid & LA_PROJID)
753 lockpart |= MDS_INODELOCK_XATTR;
755 lh = &info->mti_lh[MDT_LH_PARENT];
756 rc = mdt_object_stripes_lock(info, NULL, mo, lh, einfo, lockpart,
757 LCK_PW, cos_incompat);
761 /* all attrs are packed into mti_attr in unpack_setattr */
762 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
763 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
765 /* VBR: update version if attr changed are important for recovery */
767 /* update on-disk version of changed object */
768 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
769 rc = mdt_version_get_check_save(info, mo, 0);
771 GOTO(out_unlock, rc);
774 /* Ensure constant striping during chown(). See LU-2789. */
775 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
776 mutex_lock(&mo->mot_lov_mutex);
778 /* all attrs are packed into mti_attr in unpack_setattr */
779 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
781 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
782 mutex_unlock(&mo->mot_lov_mutex);
785 GOTO(out_unlock, rc);
786 mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
789 mdt_object_stripes_unlock(info, mo, lh, einfo, rc);
794 * Check HSM flags and add HS_DIRTY flag if relevant.
796 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
797 * and is not RELEASED.
799 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
802 struct lu_ucred *uc = mdt_ucred(info);
803 kernel_cap_t cap_saved;
807 /* If the file was modified, add the dirty flag */
808 ma->ma_need = MA_HSM;
809 rc = mdt_attr_get_complex(info, mo, ma);
811 CERROR("file attribute read error for "DFID": %d.\n",
812 PFID(mdt_object_fid(mo)), rc);
816 /* If an up2date copy exists in the backend, add dirty flag */
817 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
818 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
819 ma->ma_hsm.mh_flags |= HS_DIRTY;
821 /* Bump cap so that closes from non-owner writers can
822 * set the HSM state to dirty.
824 cap_saved = uc->uc_cap;
825 cap_raise(uc->uc_cap, CAP_FOWNER);
826 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
827 uc->uc_cap = cap_saved;
829 CERROR("file attribute change error for "DFID": %d\n",
830 PFID(mdt_object_fid(mo)), rc);
836 static int mdt_reint_setattr(struct mdt_thread_info *info,
837 struct mdt_lock_handle *lhc)
839 struct mdt_device *mdt = info->mti_mdt;
840 struct md_attr *ma = &info->mti_attr;
841 struct mdt_reint_record *rr = &info->mti_rr;
842 struct ptlrpc_request *req = mdt_info_req(info);
843 struct mdt_object *mo;
844 struct mdt_body *repbody;
845 ktime_t kstart = ktime_get();
849 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
850 (unsigned int)ma->ma_attr.la_valid);
852 if (info->mti_dlm_req)
853 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
855 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
857 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
858 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
860 GOTO(out, rc = PTR_ERR(mo));
862 if (!mdt_object_exists(mo))
863 GOTO(out_put, rc = -ENOENT);
865 if (mdt_object_remote(mo))
866 GOTO(out_put, rc = -EREMOTE);
868 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
869 /* revoke lease lock if size is going to be changed */
870 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
871 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
872 atomic_read(&mo->mot_lease_count) > 0)) {
873 down_read(&mo->mot_open_sem);
875 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
876 lhc = &info->mti_lh[MDT_LH_LOCAL];
877 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN,
880 up_read(&mo->mot_open_sem);
884 /* revoke lease lock */
885 mdt_object_unlock(info, mo, lhc, 1);
887 up_read(&mo->mot_open_sem);
890 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
891 /* Check write access for the O_TRUNC case */
892 if (mdt_write_read(mo) < 0)
893 GOTO(out_put, rc = -ETXTBSY);
895 /* LU-10286: compatibility check for FLR.
896 * Please check the comment in mdt_finish_open() for details
898 if (!exp_connect_flr(info->mti_exp) ||
899 !exp_connect_overstriping(info->mti_exp)) {
900 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
901 if (rc < 0 && rc != -ENODATA)
904 if (!exp_connect_flr(info->mti_exp)) {
906 mdt_lmm_is_flr(info->mti_big_lmm))
907 GOTO(out_put, rc = -EOPNOTSUPP);
910 if (!exp_connect_overstriping(info->mti_exp)) {
912 mdt_lmm_is_overstriping(info->mti_big_lmm))
913 GOTO(out_put, rc = -EOPNOTSUPP);
917 /* For truncate, the file size sent from client
918 * is believable, but the blocks are incorrect,
919 * which makes the block size in LSOM attribute
920 * inconsisent with the real block size.
922 rc = mdt_lsom_update(info, mo, true);
927 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
928 if (ma->ma_valid & MA_LOV)
929 GOTO(out_put, rc = -EPROTO);
931 /* MDT supports FMD for regular files due to Data-on-MDT */
932 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
933 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
934 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
937 if (ma->ma_attr.la_valid & LA_MTIME) {
938 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
940 ma->ma_valid |= MA_PFID;
944 rc = mdt_attr_set(info, mo, ma);
947 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
948 (ma->ma_valid & MA_INODE)) {
949 struct lu_buf *buf = &info->mti_buf;
950 struct lu_ucred *uc = mdt_ucred(info);
951 struct mdt_lock_handle *lh;
954 /* reject if either remote or striped dir is disabled */
955 if (ma->ma_valid & MA_LMV) {
956 if (!mdt->mdt_enable_remote_dir ||
957 !mdt->mdt_enable_striped_dir)
958 GOTO(out_put, rc = -EPERM);
960 /* we want rbac roles to have precedence over any other
961 * permission or capability checks
963 if (!uc->uc_rbac_dne_ops ||
964 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
965 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
966 mdt->mdt_enable_remote_dir_gid != -1))
967 GOTO(out_put, rc = -EPERM);
970 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
971 GOTO(out_put, rc = -ENOTDIR);
973 if (ma->ma_attr.la_valid != 0)
974 GOTO(out_put, rc = -EPROTO);
976 lh = &info->mti_lh[MDT_LH_PARENT];
977 if (ma->ma_valid & MA_LOV) {
978 buf->lb_buf = ma->ma_lmm;
979 buf->lb_len = ma->ma_lmm_size;
980 name = XATTR_NAME_LOV;
981 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
984 buf->lb_buf = &ma->ma_lmv->lmv_user_md;
985 buf->lb_len = ma->ma_lmv_size;
986 name = XATTR_NAME_DEFAULT_LMV;
988 if (unlikely(fid_is_root(mdt_object_fid(mo)))) {
989 rc = mdt_object_lock(info, mo, lh,
990 MDS_INODELOCK_XATTR |
991 MDS_INODELOCK_LOOKUP,
994 struct lu_fid *pfid = &info->mti_tmp_fid1;
995 struct lu_name *pname = &info->mti_name;
996 const char dotdot[] = "..";
997 struct mdt_object *pobj;
1000 pname->ln_name = dotdot;
1001 pname->ln_namelen = sizeof(dotdot);
1002 rc = mdo_lookup(info->mti_env,
1003 mdt_object_child(mo), pname,
1008 pobj = mdt_object_find(info->mti_env,
1009 info->mti_mdt, pfid);
1011 GOTO(out_put, rc = PTR_ERR(pobj));
1013 rc = mdt_object_check_lock(info, pobj, mo, lh,
1014 MDS_INODELOCK_XATTR |
1015 MDS_INODELOCK_LOOKUP,
1017 mdt_object_put(info->mti_env, pobj);
1024 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
1027 mdt_object_unlock(info, mo, lh, rc);
1031 GOTO(out_put, rc = -EPROTO);
1034 /* If file data is modified, add the dirty flag */
1035 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1036 rc = mdt_add_dirty_flag(info, mo, ma);
1038 ma->ma_need = MA_INODE;
1040 rc = mdt_attr_get_complex(info, mo, ma);
1044 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1048 mdt_object_put(info->mti_env, mo);
1051 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1052 ktime_us_delta(ktime_get(), kstart));
1054 mdt_client_compatibility(info);
1058 static int mdt_reint_create(struct mdt_thread_info *info,
1059 struct mdt_lock_handle *lhc)
1061 struct ptlrpc_request *req = mdt_info_req(info);
1062 ktime_t kstart = ktime_get();
1066 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1067 RETURN(err_serious(-ESTALE));
1069 if (info->mti_dlm_req)
1070 ldlm_request_cancel(mdt_info_req(info),
1071 info->mti_dlm_req, 0, LATF_SKIP);
1073 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1076 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1086 CERROR("%s: Unsupported mode %o\n",
1087 mdt_obd_name(info->mti_mdt),
1088 info->mti_attr.ma_attr.la_mode);
1089 RETURN(err_serious(-EOPNOTSUPP));
1092 rc = mdt_create(info);
1094 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1095 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1096 ktime_us_delta(ktime_get(), kstart));
1098 /* Special file should stay on the same node as parent*/
1099 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1100 ktime_us_delta(ktime_get(), kstart));
1107 * VBR: save parent version in reply and child version getting by its name.
1108 * Version of child is getting and checking during its lookup. If
1110 static int mdt_reint_unlink(struct mdt_thread_info *info,
1111 struct mdt_lock_handle *lhc)
1113 struct mdt_reint_record *rr = &info->mti_rr;
1114 struct ptlrpc_request *req = mdt_info_req(info);
1115 struct md_attr *ma = &info->mti_attr;
1116 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1117 struct mdt_object *mp;
1118 struct mdt_object *mc;
1119 struct mdt_lock_handle *parent_lh;
1120 struct mdt_lock_handle *child_lh;
1121 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
1122 struct lu_ucred *uc = mdt_ucred(info);
1123 bool cos_incompat = false;
1125 ktime_t kstart = ktime_get();
1129 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1130 PNAME(&rr->rr_name));
1132 if (info->mti_dlm_req)
1133 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1135 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1136 RETURN(err_serious(-ENOENT));
1138 if (!fid_is_md_operative(rr->rr_fid1))
1141 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1143 RETURN(PTR_ERR(mp));
1145 if (mdt_object_remote(mp)) {
1146 cos_incompat = true;
1148 rc = mdt_version_get_check_save(info, mp, 0);
1150 GOTO(put_parent, rc);
1153 if (!uc->uc_rbac_fscrypt_admin &&
1154 mp->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
1155 GOTO(put_parent, rc = -EPERM);
1157 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1158 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1160 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1161 rc = mdt_parent_lock(info, mp, parent_lh, &rr->rr_name, LCK_PW,
1164 GOTO(put_parent, rc);
1166 if (info->mti_spec.sp_rm_entry) {
1167 if (!mdt_is_dne_client(req->rq_export))
1168 /* Return -ENOTSUPP for old client */
1169 GOTO(unlock_parent, rc = -ENOTSUPP);
1171 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1172 GOTO(unlock_parent, rc = -EPERM);
1174 ma->ma_need = MA_INODE;
1176 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1177 NULL, &rr->rr_name, ma, no_name);
1178 GOTO(unlock_parent, rc);
1181 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1182 *child_fid = *rr->rr_fid2;
1184 /* lookup child object along with version checking */
1185 fid_zero(child_fid);
1186 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1189 /* Name might not be able to find during resend of
1190 * remote unlink, considering following case.
1191 * dir_A is a remote directory, the name entry of
1192 * dir_A is on MDT0, the directory is on MDT1,
1194 * 1. client sends unlink req to MDT1.
1195 * 2. MDT1 sends name delete update to MDT0.
1196 * 3. name entry is being deleted in MDT0 synchronously.
1197 * 4. MDT1 is restarted.
1198 * 5. client resends unlink req to MDT1. So it can not
1199 * find the name entry on MDT0 anymore.
1200 * In this case, MDT1 only needs to destory the local
1203 if (mdt_object_remote(mp) && rc == -ENOENT &&
1204 !fid_is_zero(rr->rr_fid2) &&
1205 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1207 *child_fid = *rr->rr_fid2;
1209 GOTO(unlock_parent, rc);
1214 if (!fid_is_md_operative(child_fid))
1215 GOTO(unlock_parent, rc = -EPERM);
1217 /* We will lock the child regardless it is local or remote. No harm. */
1218 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1220 GOTO(unlock_parent, rc = PTR_ERR(mc));
1222 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1223 /* In this case, child fid is embedded in the request, and we do
1224 * not have a proper name as rr_name contains an encoded
1225 * hash. So find name that matches provided hash.
1227 if (!find_name_matching_hash(info, &rr->rr_name,
1229 GOTO(put_child, rc = -ENOENT);
1232 if (!cos_incompat) {
1233 rc = mdt_object_striped(info, mc);
1235 GOTO(put_child, rc);
1239 mdt_object_put(info->mti_env, mc);
1240 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1245 child_lh = &info->mti_lh[MDT_LH_CHILD];
1246 if (mdt_object_remote(mc)) {
1247 struct mdt_body *repbody;
1249 if (!fid_is_zero(rr->rr_fid2)) {
1250 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1251 mdt_obd_name(info->mti_mdt),
1252 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1253 GOTO(put_child, rc = -ENOENT);
1255 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1256 mdt_obd_name(info->mti_mdt),
1257 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1259 if (!mdt_is_dne_client(req->rq_export))
1260 /* Return -ENOTSUPP for old client */
1261 GOTO(put_child, rc = -ENOTSUPP);
1263 /* Revoke the LOOKUP lock of the remote object granted by
1264 * this MDT. Since the unlink will happen on another MDT,
1265 * it will release the LOOKUP lock right away. Then What
1266 * would happen if another client try to grab the LOOKUP
1267 * lock at the same time with unlink XXX
1269 rc = mdt_object_lookup_lock(info, NULL, mc, child_lh, LCK_EX,
1272 GOTO(put_child, rc);
1274 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1275 LASSERT(repbody != NULL);
1276 repbody->mbo_fid1 = *mdt_object_fid(mc);
1277 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1278 GOTO(unlock_child, rc = -EREMOTE);
1280 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1281 * this now because a running HSM restore on the child (unlink
1282 * victim) will hold the layout lock. See LU-4002.
1284 rc = mdt_object_stripes_lock(info, mp, mc, child_lh, einfo,
1285 MDS_INODELOCK_LOOKUP |
1286 MDS_INODELOCK_UPDATE,
1287 LCK_EX, cos_incompat);
1289 GOTO(put_child, rc);
1292 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1293 * whether need MA_LOV and MA_COOKIE.
1295 ma->ma_need = MA_INODE;
1298 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1299 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1300 /* save version when object is locked */
1301 mdt_version_get_save(info, mc, 1);
1303 mutex_lock(&mc->mot_lov_mutex);
1305 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1306 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1308 mutex_unlock(&mc->mot_lov_mutex);
1310 GOTO(unlock_child, rc);
1312 if (!lu_object_is_dying(&mc->mot_header)) {
1313 rc = mdt_attr_get_complex(info, mc, ma);
1316 } else if (mdt_dom_check_for_discard(info, mc)) {
1317 mdt_dom_discard_data(info, mc);
1319 mdt_handle_last_unlink(info, mc, ma);
1322 if (ma->ma_valid & MA_INODE) {
1323 switch (ma->ma_attr.la_mode & S_IFMT) {
1325 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1326 ktime_us_delta(ktime_get(), kstart));
1334 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1335 ktime_us_delta(ktime_get(), kstart));
1338 LASSERTF(0, "bad file type %o unlinking\n",
1339 ma->ma_attr.la_mode);
1346 mdt_object_stripes_unlock(info, mc, child_lh, einfo, rc);
1348 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1349 info->mti_big_buf.lb_buf)
1350 lu_buf_free(&info->mti_big_buf);
1351 mdt_object_put(info->mti_env, mc);
1353 mdt_object_unlock(info, mp, parent_lh, rc);
1355 mdt_object_put(info->mti_env, mp);
1356 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1361 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1364 static int mdt_reint_link(struct mdt_thread_info *info,
1365 struct mdt_lock_handle *lhc)
1367 struct mdt_reint_record *rr = &info->mti_rr;
1368 struct ptlrpc_request *req = mdt_info_req(info);
1369 struct md_attr *ma = &info->mti_attr;
1370 struct mdt_object *ms;
1371 struct mdt_object *mp;
1372 struct mdt_lock_handle *lhs;
1373 struct mdt_lock_handle *lhp;
1374 ktime_t kstart = ktime_get();
1379 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1380 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1382 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1383 RETURN(err_serious(-ENOENT));
1385 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1386 OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1387 req->rq_no_reply = 1;
1388 RETURN(err_serious(-ENOENT));
1391 if (info->mti_dlm_req)
1392 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1394 /* Invalid case so return error immediately instead of
1397 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1400 if (!fid_is_md_operative(rr->rr_fid1) ||
1401 !fid_is_md_operative(rr->rr_fid2))
1404 /* step 1: find target parent dir */
1405 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1407 RETURN(PTR_ERR(mp));
1409 rc = mdt_version_get_check_save(info, mp, 0);
1411 GOTO(put_parent, rc);
1413 rc = mdt_check_enc(info, mp);
1415 GOTO(put_parent, rc);
1417 /* step 2: find source */
1418 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1420 GOTO(put_parent, rc = PTR_ERR(ms));
1422 if (!mdt_object_exists(ms)) {
1423 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1424 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1425 GOTO(put_source, rc = -ENOENT);
1428 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1430 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1432 lhp = &info->mti_lh[MDT_LH_PARENT];
1433 rc = mdt_parent_lock(info, mp, lhp, &rr->rr_name, LCK_PW, cos_incompat);
1435 GOTO(put_source, rc);
1437 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1439 lhs = &info->mti_lh[MDT_LH_CHILD];
1440 rc = mdt_object_lock(info, ms, lhs,
1441 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR, LCK_EX,
1444 GOTO(unlock_parent, rc);
1446 /* step 3: link it */
1447 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1448 OBD_FAIL_MDS_REINT_LINK_WRITE);
1450 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1451 rc = mdt_version_get_check_save(info, ms, 1);
1453 GOTO(unlock_source, rc);
1455 /** check target version by name during replay */
1456 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1457 &info->mti_tmp_fid1, 2);
1458 if (rc != 0 && rc != -ENOENT)
1459 GOTO(unlock_source, rc);
1460 /* save version of file name for replay, it must be ENOENT here */
1461 if (!req_is_replay(mdt_info_req(info))) {
1462 if (rc != -ENOENT) {
1463 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1464 PNAME(&rr->rr_name));
1465 GOTO(unlock_source, rc = -EEXIST);
1467 info->mti_ver[2] = ENOENT_VERSION;
1468 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1471 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1472 mdt_object_child(ms), &rr->rr_name, ma);
1475 mdt_counter_incr(req, LPROC_MDT_LINK,
1476 ktime_us_delta(ktime_get(), kstart));
1480 mdt_object_unlock(info, ms, lhs, rc);
1482 mdt_object_unlock(info, mp, lhp, rc);
1484 mdt_object_put(info->mti_env, ms);
1486 mdt_object_put(info->mti_env, mp);
1491 * Get BFL lock for rename or migrate process.
1493 static int mdt_rename_lock(struct mdt_thread_info *info,
1494 struct mdt_lock_handle *lh)
1496 struct lu_fid *fid = &info->mti_tmp_fid1;
1497 struct mdt_object *obj;
1498 __u64 ibits = MDS_INODELOCK_UPDATE;
1503 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1505 RETURN(PTR_ERR(obj));
1507 mdt_lock_reg_init(lh, LCK_EX);
1508 rc = mdt_object_lock_internal(info, obj, &LUSTRE_BFL_FID, lh,
1509 &ibits, 0, false, false);
1510 mdt_object_put(info->mti_env, obj);
1514 static void mdt_rename_unlock(struct mdt_thread_info *info,
1515 struct mdt_lock_handle *lh)
1518 /* Cancel the single rename lock right away */
1519 mdt_object_unlock(info, NULL, lh, 1);
1523 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1524 const struct lu_fid *fid,
1527 struct mdt_object *dir;
1531 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1535 /* check early, the real version will be saved after locking */
1536 rc = mdt_version_get_check(info, dir, idx);
1540 if (!mdt_object_exists(dir))
1541 GOTO(out_put, rc = -ENOENT);
1543 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1544 GOTO(out_put, rc = -ENOTDIR);
1548 mdt_object_put(info->mti_env, dir);
1553 * lock rename source object.
1555 * Both source and its parent object may be located on remote MDTs, and even on
1556 * different MDTs, which means source object is a remote object on parent.
1558 * \retval 0 on success
1559 * \retval -ev negative errno upon error
1561 static int mdt_rename_source_lock(struct mdt_thread_info *info,
1562 struct mdt_object *parent,
1563 struct mdt_object *child,
1564 struct mdt_lock_handle *lh,
1565 struct mdt_lock_handle *lh_lookup,
1566 __u64 ibits, bool cos_incompat)
1570 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
1571 /* if @obj is remote object, LOOKUP lock needs to be taken from
1574 rc = mdt_is_remote_object(info, parent, child);
1579 rc = mdt_object_lookup_lock(info, parent, child, lh_lookup,
1580 LCK_EX, cos_incompat);
1584 ibits &= ~MDS_INODELOCK_LOOKUP;
1587 rc = mdt_object_lock(info, child, lh, ibits, LCK_EX, cos_incompat);
1588 if (unlikely(rc && !(ibits & MDS_INODELOCK_LOOKUP)))
1589 mdt_object_unlock(info, NULL, lh_lookup, rc);
1594 static void mdt_rename_source_unlock(struct mdt_thread_info *info,
1595 struct mdt_object *obj,
1596 struct mdt_lock_handle *lh,
1597 struct mdt_lock_handle *lh_lookup,
1600 mdt_object_unlock(info, obj, lh, decref);
1601 mdt_object_unlock(info, NULL, lh_lookup, decref);
1604 /* migration takes UPDATE lock of link parent, and LOOKUP lock of link */
1605 struct mdt_link_lock {
1606 struct mdt_object *mll_obj;
1607 struct mdt_lock_handle mll_lh;
1608 struct list_head mll_linkage;
1611 static inline int mdt_migrate_link_lock_add(struct mdt_thread_info *info,
1612 struct mdt_object *o,
1613 struct mdt_lock_handle *lh,
1614 struct list_head *list)
1616 struct mdt_link_lock *mll;
1622 INIT_LIST_HEAD(&mll->mll_linkage);
1623 mdt_object_get(info->mti_env, o);
1626 memset(lh, 0, sizeof(*lh));
1627 list_add_tail(&mll->mll_linkage, list);
1632 static inline void mdt_migrate_link_lock_del(struct mdt_thread_info *info,
1633 struct mdt_link_lock *mll,
1636 mdt_object_unlock(info, mll->mll_obj, &mll->mll_lh, decref);
1637 mdt_object_put(info->mti_env, mll->mll_obj);
1638 list_del(&mll->mll_linkage);
1642 static void mdt_migrate_links_unlock(struct mdt_thread_info *info,
1643 struct list_head *list, int decref)
1645 struct mdt_link_lock *mll;
1646 struct mdt_link_lock *tmp;
1648 list_for_each_entry_safe(mll, tmp, list, mll_linkage)
1649 mdt_migrate_link_lock_del(info, mll, decref);
1652 /* take link parent UPDATE lock.
1653 * \retval 0 \a lnkp is already locked, no lock taken.
1655 * -ev negative errno.
1657 static int mdt_migrate_link_parent_lock(struct mdt_thread_info *info,
1658 struct mdt_object *lnkp,
1659 struct list_head *update_locks,
1662 const struct lu_fid *fid = mdt_object_fid(lnkp);
1663 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1664 struct mdt_link_lock *entry;
1670 /* check if it's already locked */
1671 list_for_each_entry(entry, update_locks, mll_linkage) {
1672 if (lu_fid_eq(mdt_object_fid(entry->mll_obj), fid)) {
1673 CDEBUG(D_INFO, "skip "DFID" lock\n", PFID(fid));
1678 /* link parent UPDATE lock */
1679 CDEBUG(D_INFO, "lock "DFID"\n", PFID(fid));
1682 /* revoke lock instead of take in *blocked* mode */
1683 rc = mdt_object_lock(info, lnkp, lhl, MDS_INODELOCK_UPDATE,
1688 if (mdt_object_remote(lnkp)) {
1689 struct ldlm_lock *lock;
1692 * for remote object, set lock cb_atomic, so lock can be
1693 * released in blocking_ast() immediately, then the next
1694 * lock_try will have better chance of success.
1696 lock = ldlm_handle2lock(&lhl->mlh_rreg_lh);
1697 LASSERT(lock != NULL);
1698 lock_res_and_lock(lock);
1699 ldlm_set_atomic_cb(lock);
1700 unlock_res_and_lock(lock);
1701 LDLM_LOCK_PUT(lock);
1704 mdt_object_unlock(info, lnkp, lhl, 1);
1709 * we can't follow parent-child lock order like other MD
1710 * operations, use lock_try here to avoid deadlock, if the lock
1711 * cannot be taken, drop all locks taken, revoke the blocked
1712 * one, and continue processing the remaining entries, and in
1713 * the end of the loop restart from beginning.
1715 * don't lock with PDO mode in case two links are under the same
1716 * parent and their hash values are different.
1718 rc = mdt_object_lock_try(info, lnkp, lhl, &ibits, MDS_INODELOCK_UPDATE,
1723 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1724 CDEBUG(D_INFO, "busy lock on "DFID"\n", PFID(fid));
1729 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, update_locks);
1731 mdt_object_unlock(info, lnkp, lhl, 1);
1738 /* take link LOOKUP lock.
1739 * \retval 0 \a lnkp is already locked, no lock taken.
1741 * -ev negative errno.
1743 static int mdt_migrate_link_lock(struct mdt_thread_info *info,
1744 struct mdt_object *lnkp,
1745 struct mdt_object *spobj,
1746 struct mdt_object *obj,
1747 struct list_head *lookup_locks)
1749 const struct lu_fid *fid = mdt_object_fid(lnkp);
1750 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1751 struct mdt_link_lock *entry;
1756 /* check if it's already locked by source */
1757 rc = mdt_fids_different_target(info, fid, mdt_object_fid(spobj));
1759 CDEBUG(D_INFO, "skip lookup lock on source parent "DFID"\n",
1764 /* check if it's already locked by other links */
1765 list_for_each_entry(entry, lookup_locks, mll_linkage) {
1766 rc = mdt_fids_different_target(info, fid,
1767 mdt_object_fid(entry->mll_obj));
1769 CDEBUG(D_INFO, "skip lookup lock on parent "DFID"\n",
1775 rc = mdt_object_lookup_lock(info, lnkp, obj, lhl, LCK_EX, true);
1779 /* don't take local LOOKUP lock, because later we will lock other ibits
1780 * of sobj (which is on local MDT), and lock the same object twice may
1781 * deadlock, just revoke this lock.
1783 if (!mdt_object_remote(lnkp))
1784 GOTO(unlock, rc = 0);
1786 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, lookup_locks);
1792 mdt_object_unlock(info, lnkp, lhl, 1);
1797 * take UPDATE lock of link parents and LOOKUP lock of links, also check whether
1798 * total local lock count exceeds RS_MAX_LOCKS.
1800 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1801 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1802 * \retval -ev negative errno upon error
1804 static int mdt_migrate_links_lock(struct mdt_thread_info *info,
1805 struct mdt_object *spobj,
1806 struct mdt_object *tpobj,
1807 struct mdt_object *obj,
1808 struct mdt_lock_handle *lhsp,
1809 struct mdt_lock_handle *lhtp,
1810 struct list_head *link_locks)
1812 struct mdt_device *mdt = info->mti_mdt;
1813 struct lu_buf *buf = &info->mti_big_buf;
1814 struct lu_name *lname = &info->mti_name;
1815 struct linkea_data ldata = { NULL };
1816 int local_lock_cnt = 0;
1817 bool blocked = false;
1819 struct mdt_object *lnkp;
1821 LIST_HEAD(update_locks);
1822 LIST_HEAD(lookup_locks);
1826 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1829 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1830 if (buf->lb_buf == NULL)
1834 rc = mdt_links_read(info, obj, &ldata);
1836 if (rc == -ENOENT || rc == -ENODATA)
1841 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1842 linkea_next_entry(&ldata)) {
1843 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1846 /* check if link parent is source parent too */
1847 if (lu_fid_eq(mdt_object_fid(spobj), &fid)) {
1849 "skip lock on source parent "DFID"/"DNAME"\n",
1850 PFID(&fid), PNAME(lname));
1854 /* check if link parent is target parent too */
1855 if (tpobj != spobj && lu_fid_eq(mdt_object_fid(tpobj), &fid)) {
1857 "skip lock on target parent "DFID"/"DNAME"\n",
1858 PFID(&fid), PNAME(lname));
1862 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1864 CWARN("%s: cannot find obj "DFID": %ld\n",
1865 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1869 if (!mdt_object_exists(lnkp)) {
1870 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1871 PFID(&fid), PNAME(lname));
1872 mdt_object_put(info->mti_env, lnkp);
1877 rc = mdt_migrate_link_parent_lock(info, lnkp, &update_locks,
1879 if (!saved && blocked) {
1880 /* unlock all locks taken to avoid deadlock */
1881 mdt_migrate_links_unlock(info, &update_locks, 1);
1882 mdt_object_unlock(info, spobj, lhsp, 1);
1884 mdt_object_unlock(info, tpobj, lhtp, 1);
1888 mdt_object_put(info->mti_env, lnkp);
1892 if (rc == 1 && !mdt_object_remote(lnkp))
1895 rc = mdt_migrate_link_lock(info, lnkp, spobj, obj,
1898 mdt_object_put(info->mti_env, lnkp);
1901 if (rc == 1 && !mdt_object_remote(lnkp))
1903 mdt_object_put(info->mti_env, lnkp);
1907 GOTO(out, rc = -EBUSY);
1911 list_splice(&update_locks, link_locks);
1912 list_splice(&lookup_locks, link_locks);
1914 mdt_migrate_links_unlock(info, link_locks, rc);
1915 } else if (local_lock_cnt > RS_MAX_LOCKS - 5) {
1917 * parent may have 3 local objects: master object and 2 stripes
1918 * (if it's being migrated too); source may have 1 local objects
1919 * as regular file; target has 1 local object.
1920 * Note, source may have 2 local locks if it is directory but it
1921 * can't have hardlinks, so it is not considered here.
1923 CDEBUG(D_INFO, "Too many local locks (%d), migrate in sync mode\n",
1931 * lookup source by name, if parent is striped directory, we need to find the
1932 * corresponding stripe where source is located, and then lookup there.
1934 * besides, if parent is migrating too, and file is already in target stripe,
1935 * this should be a redo of 'lfs migrate' on client side.
1937 * \retval 1 tpobj stripe index is less than spobj stripe index
1938 * \retval 0 tpobj stripe index is larger than or equal to spobj stripe index
1939 * \retval -ev negative errno upon error
1941 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1942 struct mdt_object *pobj,
1943 const struct md_attr *ma,
1944 const struct lu_name *lname,
1945 struct mdt_object **spobj,
1946 struct mdt_object **tpobj,
1947 struct mdt_object **sobj)
1949 const struct lu_env *env = info->mti_env;
1950 struct lu_fid *fid = &info->mti_tmp_fid1;
1955 if (ma->ma_valid & MA_LMV) {
1956 /* if parent is striped, lookup on corresponding stripe */
1957 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1958 struct lu_fid *fid2 = &info->mti_tmp_fid2;
1960 if (!lmv_is_sane(lmv))
1963 spindex = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1968 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[spindex]);
1970 *spobj = mdt_object_find(env, info->mti_mdt, fid2);
1971 if (IS_ERR(*spobj)) {
1972 rc = PTR_ERR(*spobj);
1977 if (!mdt_object_exists(*spobj))
1978 GOTO(spobj_put, rc = -ENOENT);
1981 rc = mdo_lookup(env, mdt_object_child(*spobj), lname, fid,
1983 if ((rc == -ENOENT || rc == 0) && lmv_is_layout_changing(lmv)) {
1984 /* fail check here to let top dir migration succeed. */
1985 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MIGRATE_ENTRIES, 0))
1986 GOTO(spobj_put, rc = -EIO);
1989 * if parent layout is changeing, and lookup child
1990 * failed on source stripe, lookup again on target
1991 * stripe, if it exists, it means previous migration
1992 * was interrupted, and current file was migrated
1995 tpindex = lmv_name_to_stripe_index(lmv, lname->ln_name,
1998 GOTO(spobj_put, rc = tpindex);
2000 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[tpindex]);
2002 *tpobj = mdt_object_find(env, info->mti_mdt, fid2);
2003 if (IS_ERR(*tpobj)) {
2004 rc = PTR_ERR(*tpobj);
2006 GOTO(spobj_put, rc);
2009 if (!mdt_object_exists(*tpobj))
2010 GOTO(tpobj_put, rc = -ENOENT);
2012 if (rc == -ENOENT) {
2014 rc = mdo_lookup(env, mdt_object_child(*tpobj),
2015 lname, fid, &info->mti_spec);
2016 GOTO(tpobj_put, rc = rc ?: -EALREADY);
2019 GOTO(spobj_put, rc);
2023 mdt_object_get(env, *tpobj);
2027 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2034 mdt_object_get(env, pobj);
2035 mdt_object_get(env, pobj);
2038 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2039 if (IS_ERR(*sobj)) {
2040 rc = PTR_ERR(*sobj);
2042 GOTO(tpobj_put, rc);
2045 if (!mdt_object_exists(*sobj))
2046 GOTO(sobj_put, rc = -ENOENT);
2048 return (tpindex < spindex);
2051 mdt_object_put(env, *sobj);
2054 mdt_object_put(env, *tpobj);
2057 mdt_object_put(env, *spobj);
2063 /* end lease and close file for regular file */
2064 static int mdd_migrate_close(struct mdt_thread_info *info,
2065 struct mdt_object *obj)
2067 struct close_data *data;
2068 struct mdt_body *repbody;
2069 struct ldlm_lock *lease;
2074 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2076 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2080 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2085 lease = ldlm_handle2lock(&data->cd_handle);
2089 /* check if the lease was already canceled */
2090 lock_res_and_lock(lease);
2091 rc = ldlm_is_cancel(lease);
2092 unlock_res_and_lock(lease);
2096 LDLM_DEBUG(lease, DFID" lease broken",
2097 PFID(mdt_object_fid(obj)));
2101 * cancel server side lease, client side counterpart should have been
2102 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2104 ldlm_lock_cancel(lease);
2105 ldlm_reprocess_all(lease->l_resource,
2106 lease->l_policy_data.l_inodebits.bits);
2107 LDLM_LOCK_PUT(lease);
2110 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2111 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2112 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2117 /* LFSCK used to clear hash type and MIGRATION flag upon migration failure */
2118 static inline bool lmv_is_failed_migration(const struct lmv_mds_md_v1 *lmv)
2120 return le32_to_cpu(lmv->lmv_hash_type) ==
2121 (LMV_HASH_TYPE_UNKNOWN | LMV_HASH_FLAG_BAD_TYPE) &&
2122 lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_migrate_hash)) &&
2123 le32_to_cpu(lmv->lmv_migrate_offset) > 0 &&
2124 le32_to_cpu(lmv->lmv_migrate_offset) <
2125 le32_to_cpu(lmv->lmv_stripe_count);
2129 * migrate file in below steps:
2130 * 1. lock source and target stripes
2131 * 2. lookup source by name
2132 * 3. lock parents of source links if source is not directory
2133 * 4. reject if source is in HSM
2134 * 5. take source open_sem and close file if source is regular file
2135 * 6. lock source, and its stripes if it's directory
2137 * 8. lock target so subsequent change to it can trigger COS
2138 * 9. unlock above locks
2139 * 10. sync device if source has too many links
2141 int mdt_reint_migrate(struct mdt_thread_info *info,
2142 struct mdt_lock_handle *unused)
2144 const struct lu_env *env = info->mti_env;
2145 struct mdt_device *mdt = info->mti_mdt;
2146 struct ptlrpc_request *req = mdt_info_req(info);
2147 struct mdt_reint_record *rr = &info->mti_rr;
2148 struct lu_ucred *uc = mdt_ucred(info);
2149 struct md_attr *ma = &info->mti_attr;
2150 struct mdt_object *pobj;
2151 struct mdt_object *spobj;
2152 struct mdt_object *tpobj;
2153 struct mdt_object *sobj;
2154 struct mdt_object *tobj;
2155 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2156 struct mdt_lock_handle *lhsp;
2157 struct mdt_lock_handle *lhtp;
2158 struct mdt_lock_handle *lhs;
2159 struct mdt_lock_handle *lhl;
2160 LIST_HEAD(link_locks);
2161 int lock_retries = 5;
2162 bool reverse = false;
2163 bool open_sem_locked = false;
2164 bool do_sync = false;
2165 bool is_plain_dir = false;
2169 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2170 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2172 if (info->mti_dlm_req)
2173 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2175 if (!fid_is_md_operative(rr->rr_fid1) ||
2176 !fid_is_md_operative(rr->rr_fid2))
2179 /* don't allow migrate . or .. */
2180 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2183 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2186 /* we want rbac roles to have precedence over any other
2187 * permission or capability checks
2189 if (uc && (!uc->uc_rbac_dne_ops ||
2190 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2191 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2192 mdt->mdt_enable_remote_dir_gid != -1)))
2196 * Note: do not enqueue rename lock for replay request, because
2197 * if other MDT holds rename lock, but being blocked to wait for
2198 * this MDT to finish its recovery, and the failover MDT can not
2199 * get rename lock, which will cause deadlock.
2201 * req is NULL if this is called by directory auto-split.
2203 if (req && !req_is_replay(req)) {
2204 rc = mdt_rename_lock(info, rename_lh);
2206 CERROR("%s: can't lock FS for rename: rc = %d\n",
2207 mdt_obd_name(info->mti_mdt), rc);
2212 /* pobj is master object of parent */
2213 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2215 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2218 rc = mdt_version_get_check(info, pobj, 0);
2220 GOTO(put_parent, rc);
2223 if (!mdt_object_exists(pobj))
2224 GOTO(put_parent, rc = -ENOENT);
2226 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2227 GOTO(put_parent, rc = -ENOTDIR);
2229 rc = mdt_check_enc(info, pobj);
2231 GOTO(put_parent, rc);
2233 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2235 GOTO(put_parent, rc);
2237 if (CFS_FAIL_CHECK(OBD_FAIL_MIGRATE_BAD_HASH) &&
2238 (ma->ma_valid & MA_LMV) &&
2239 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1)) {
2240 struct lu_buf *buf = &info->mti_buf;
2241 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2242 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2244 lmv->lmv_hash_type = cpu_to_le32(LMV_HASH_TYPE_UNKNOWN |
2245 LMV_HASH_FLAG_BAD_TYPE);
2246 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2248 buf->lb_len = sizeof(*lmv);
2249 rc = mo_xattr_set(env, mdt_object_child(pobj), buf,
2250 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2251 mo_invalidate(env, mdt_object_child(pobj));
2252 GOTO(put_parent, rc);
2255 /* @spobj is the parent stripe of @sobj if @pobj is striped directory,
2256 * if @pobj is migrating too, tpobj is the target parent stripe.
2258 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &tpobj,
2261 GOTO(put_parent, rc);
2264 /* parent unchanged, this happens in dir restripe */
2265 if (info->mti_spec.sp_migrate_nsonly && spobj == tpobj)
2266 GOTO(put_source, rc = -EALREADY);
2271 lhsp = &info->mti_lh[MDT_LH_PARENT];
2272 lhtp = &info->mti_lh[MDT_LH_CHILD];
2273 /* lock spobj and tpobj in stripe index order */
2275 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name, LCK_PW,
2278 GOTO(put_source, rc);
2280 LASSERT(spobj != tpobj);
2281 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW,
2284 GOTO(unlock_parent, rc);
2286 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW,
2289 GOTO(put_source, rc);
2291 if (tpobj != spobj) {
2292 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name,
2295 GOTO(unlock_parent, rc);
2299 /* if inode is not migrated, or is dir, no need to lock links */
2300 if (!info->mti_spec.sp_migrate_nsonly &&
2301 !S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2302 /* lock link parents, and take LOOKUP lock of links */
2303 rc = mdt_migrate_links_lock(info, spobj, tpobj, sobj, lhsp,
2305 if (rc == -EBUSY && lock_retries-- > 0) {
2306 LASSERT(list_empty(&link_locks));
2311 GOTO(put_source, rc);
2314 * RS_MAX_LOCKS is the limit of number of locks that can be
2315 * saved along with one request, if total lock count exceeds
2316 * this limit, we will drop all locks after migration, and
2317 * trigger commit in the end.
2323 lhs = &info->mti_lh[MDT_LH_OLD];
2324 lhl = &info->mti_lh[MDT_LH_LOOKUP];
2325 rc = mdt_rename_source_lock(info, spobj, sobj, lhs, lhl,
2326 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR |
2327 MDS_INODELOCK_OPEN, true);
2329 GOTO(unlock_links, rc);
2331 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2332 /* TODO: DoM migration is not supported, migrate dirent only */
2333 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2335 GOTO(unlock_source, rc);
2337 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2338 info->mti_spec.sp_migrate_nsonly = 1;
2339 } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2340 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2342 GOTO(unlock_source, rc);
2344 if (!(ma->ma_valid & MA_LMV))
2345 is_plain_dir = true;
2346 else if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2347 /* race with restripe/auto-split */
2348 GOTO(unlock_source, rc = -EBUSY);
2349 else if (lmv_is_failed_migration(&ma->ma_lmv->lmv_md_v1)) {
2350 struct lu_buf *buf = &info->mti_buf;
2351 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2352 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2354 /* migration failed before, and LFSCK cleared hash type
2355 * and flags, fake it to resume migration.
2357 lmv->lmv_hash_type =
2358 cpu_to_le32(LMV_HASH_TYPE_FNV_1A_64 |
2359 LMV_HASH_FLAG_MIGRATION |
2360 LMV_HASH_FLAG_BAD_TYPE |
2361 LMV_HASH_FLAG_FIXED);
2362 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2364 buf->lb_len = sizeof(*lmv);
2365 rc = mo_xattr_set(env, mdt_object_child(sobj), buf,
2366 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2367 mo_invalidate(env, mdt_object_child(sobj));
2368 GOTO(unlock_source, rc = -EALREADY);
2372 /* if migration HSM is allowed */
2373 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2374 ma->ma_need = MA_HSM;
2376 rc = mdt_attr_get_complex(info, sobj, ma);
2378 GOTO(unlock_source, rc);
2380 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2381 GOTO(unlock_source, rc = -EOPNOTSUPP);
2384 /* end lease and close file for regular file */
2385 if (info->mti_spec.sp_migrate_close) {
2386 /* try to hold open_sem so that nobody else can open the file */
2387 if (!down_write_trylock(&sobj->mot_open_sem)) {
2389 mdd_migrate_close(info, sobj);
2390 GOTO(unlock_source, rc = -EBUSY);
2392 open_sem_locked = true;
2393 rc = mdd_migrate_close(info, sobj);
2394 if (rc && rc != -ESTALE)
2395 GOTO(unlock_open_sem, rc);
2399 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2401 GOTO(unlock_open_sem, rc = PTR_ERR(tobj));
2403 /* Don't do lookup sanity check. We know name doesn't exist. */
2404 info->mti_spec.sp_cr_lookup = 0;
2405 info->mti_spec.sp_feat = &dt_directory_features;
2407 rc = mdo_migrate(env, mdt_object_child(spobj),
2408 mdt_object_child(tpobj), mdt_object_child(sobj),
2409 mdt_object_child(tobj), &rr->rr_name,
2410 &info->mti_spec, ma);
2412 GOTO(put_target, rc);
2414 /* save target locks for directory */
2415 if (S_ISDIR(lu_object_attr(&sobj->mot_obj)) &&
2416 !info->mti_spec.sp_migrate_nsonly) {
2417 struct mdt_lock_handle *lht = &info->mti_lh[MDT_LH_NEW];
2418 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
2420 /* in case sobj becomes a stripe of tobj, unlock sobj here,
2421 * otherwise stripes lock may deadlock.
2424 mdt_rename_source_unlock(info, sobj, lhs, lhl, 1);
2426 rc = mdt_object_stripes_lock(info, tpobj, tobj, lht, einfo,
2427 MDS_INODELOCK_UPDATE, LCK_PW,
2430 GOTO(put_target, rc);
2432 mdt_object_stripes_unlock(info, tobj, lht, einfo, 0);
2435 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2436 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2440 mdt_object_put(env, tobj);
2442 if (open_sem_locked)
2443 up_write(&sobj->mot_open_sem);
2445 mdt_rename_source_unlock(info, sobj, lhs, lhl, rc);
2447 /* if we've got too many locks to save into RPC,
2448 * then just commit before the locks are released
2451 mdt_device_sync(env, mdt);
2452 mdt_migrate_links_unlock(info, &link_locks, do_sync ? 1 : rc);
2454 mdt_object_unlock(info, spobj, lhsp, rc);
2455 mdt_object_unlock(info, tpobj, lhtp, rc);
2457 mdt_object_put(env, sobj);
2458 mdt_object_put(env, spobj);
2459 mdt_object_put(env, tpobj);
2461 mo_invalidate(env, mdt_object_child(pobj));
2462 mdt_object_put(env, pobj);
2464 mdt_rename_unlock(info, rename_lh);
2467 CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
2468 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
2469 PNAME(&rr->rr_name), rc);
2475 * determine lock order of sobj and tobj
2477 * there are two situations we need to lock tobj before sobj:
2478 * 1. sobj is child of tobj
2479 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2480 * larger than that of tobj
2482 * \retval 1 lock tobj before sobj
2483 * \retval 0 lock sobj before tobj
2484 * \retval -ev negative errno upon error
2486 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2487 struct mdt_object *sobj,
2488 struct mdt_object *tobj)
2490 struct md_attr *ma = &info->mti_attr;
2491 struct lu_fid *spfid = &info->mti_tmp_fid1;
2492 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2493 struct lmv_mds_md_v1 *lmv;
2498 /* sobj and tobj are the same */
2502 if (fid_is_root(mdt_object_fid(sobj)))
2505 if (fid_is_root(mdt_object_fid(tobj)))
2508 /* check whether sobj is child of tobj */
2509 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2510 mdt_object_fid(tobj));
2517 /* check whether sobj and tobj are children of the same parent */
2518 rc = mdt_attr_get_pfid(info, sobj, spfid);
2522 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2526 if (!lu_fid_eq(spfid, tpfid))
2529 /* check whether sobj and tobj are sibling stripes */
2530 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2534 if (!(ma->ma_valid & MA_LMV))
2537 lmv = &ma->ma_lmv->lmv_md_v1;
2538 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2540 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2543 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2547 if (!(ma->ma_valid & MA_LMV))
2550 lmv = &ma->ma_lmv->lmv_md_v1;
2551 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2553 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2555 /* check stripe index of sobj and tobj */
2556 if (sindex == tindex)
2559 return sindex < tindex ? 0 : 1;
2562 /* Helper function for mdt_reint_rename so we don't need to opencode
2563 * two different order lockings
2565 static int mdt_lock_two_dirs(struct mdt_thread_info *info,
2566 struct mdt_object *mfirstdir,
2567 struct mdt_lock_handle *lh_firstdirp,
2568 const struct lu_name *firstname,
2569 struct mdt_object *mseconddir,
2570 struct mdt_lock_handle *lh_seconddirp,
2571 const struct lu_name *secondname,
2576 rc = mdt_parent_lock(info, mfirstdir, lh_firstdirp, firstname, LCK_PW,
2581 mdt_version_get_save(info, mfirstdir, 0);
2582 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2584 if (mfirstdir != mseconddir) {
2585 rc = mdt_parent_lock(info, mseconddir, lh_seconddirp,
2586 secondname, LCK_PW, cos_incompat);
2587 } else if (!mdt_object_remote(mseconddir)) {
2588 if (lh_firstdirp->mlh_pdo_hash !=
2589 lh_seconddirp->mlh_pdo_hash) {
2590 rc = mdt_object_pdo_lock(info, mseconddir,
2591 lh_seconddirp, secondname,
2592 LCK_PW, false, cos_incompat);
2593 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2596 mdt_version_get_save(info, mseconddir, 1);
2599 mdt_object_unlock(info, mfirstdir, lh_firstdirp, rc);
2605 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2606 * 2 - srcdir child; 3 - tgtdir child.
2607 * Update on disk version of srcdir child.
2609 static int mdt_reint_rename(struct mdt_thread_info *info,
2610 struct mdt_lock_handle *unused)
2612 struct mdt_device *mdt = info->mti_mdt;
2613 struct mdt_reint_record *rr = &info->mti_rr;
2614 struct md_attr *ma = &info->mti_attr;
2615 struct ptlrpc_request *req = mdt_info_req(info);
2616 struct mdt_object *msrcdir = NULL;
2617 struct mdt_object *mtgtdir = NULL;
2618 struct mdt_object *mold;
2619 struct mdt_object *mnew = NULL;
2620 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2621 struct mdt_lock_handle *lh_srcdirp;
2622 struct mdt_lock_handle *lh_tgtdirp;
2623 struct mdt_lock_handle *lh_oldp = NULL;
2624 struct mdt_lock_handle *lh_lookup = NULL;
2625 struct mdt_lock_handle *lh_newp = NULL;
2626 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2627 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2628 struct lu_ucred *uc = mdt_ucred(info);
2629 bool reverse = false, discard = false;
2631 ktime_t kstart = ktime_get();
2632 enum mdt_stat_idx msi = 0;
2636 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2637 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2638 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2640 if (info->mti_dlm_req)
2641 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2643 if (!fid_is_md_operative(rr->rr_fid1) ||
2644 !fid_is_md_operative(rr->rr_fid2))
2647 /* find both parents. */
2648 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2649 if (IS_ERR(msrcdir))
2650 RETURN(PTR_ERR(msrcdir));
2652 rc = mdt_check_enc(info, msrcdir);
2654 GOTO(out_put_srcdir, rc);
2656 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2658 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2660 mdt_object_get(info->mti_env, mtgtdir);
2662 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2663 if (IS_ERR(mtgtdir))
2664 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2667 rc = mdt_check_enc(info, mtgtdir);
2669 GOTO(out_put_tgtdir, rc);
2671 if (!uc->uc_rbac_fscrypt_admin &&
2672 mtgtdir->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
2673 GOTO(out_put_tgtdir, rc = -EPERM);
2676 * Note: do not enqueue rename lock for replay request, because
2677 * if other MDT holds rename lock, but being blocked to wait for
2678 * this MDT to finish its recovery, and the failover MDT can not
2679 * get rename lock, which will cause deadlock.
2681 if (!req_is_replay(req)) {
2682 bool remote = mdt_object_remote(msrcdir);
2685 * Normally rename RPC is handled on the MDT with the target
2686 * directory (if target exists, it's on the MDT with the
2687 * target), if the source directory is remote, it's a hint that
2688 * source is remote too (this may not be true, but it won't
2689 * cause any issue), return -EXDEV early to avoid taking
2692 if (!mdt->mdt_enable_remote_rename && remote)
2693 GOTO(out_put_tgtdir, rc = -EXDEV);
2695 /* This might be further relaxed in the future for regular file
2696 * renames in different source and target parents. Start with
2697 * only same-directory renames for simplicity and because this
2698 * is by far the most the common use case.
2700 * Striped directories should be considered "remote".
2702 if (msrcdir != mtgtdir || remote ||
2703 (S_ISDIR(ma->ma_attr.la_mode) &&
2704 !mdt->mdt_enable_parallel_rename_dir) ||
2705 (!S_ISDIR(ma->ma_attr.la_mode) &&
2706 !mdt->mdt_enable_parallel_rename_file)) {
2707 rc = mdt_rename_lock(info, rename_lh);
2709 CERROR("%s: cannot lock for rename: rc = %d\n",
2710 mdt_obd_name(mdt), rc);
2711 GOTO(out_put_tgtdir, rc);
2714 if (S_ISDIR(ma->ma_attr.la_mode))
2715 msi = LPROC_MDT_RENAME_PAR_DIR;
2717 msi = LPROC_MDT_RENAME_PAR_FILE;
2720 "%s: samedir parallel rename "DFID"/"DNAME"\n",
2721 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2722 PNAME(&rr->rr_name));
2726 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2728 GOTO(out_unlock_rename, rc);
2731 /* source needs to be looked up after locking source parent, otherwise
2732 * this rename may race with unlink source, and cause rename hang, see
2733 * sanityn.sh 55b, so check parents first, if later we found source is
2734 * remote, relock parents.
2736 cos_incompat = (mdt_object_remote(msrcdir) ||
2737 mdt_object_remote(mtgtdir));
2739 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2741 /* lock parents in the proper order. */
2742 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2743 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2745 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2746 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2748 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2749 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2751 /* In case of same dir local rename we must sort by the hash,
2752 * otherwise a lock deadlock is possible when renaming
2753 * a to b and b to a at the same time LU-15285
2755 if (!mdt_object_remote(mtgtdir) && mtgtdir == msrcdir)
2756 reverse = lh_srcdirp->mlh_pdo_hash > lh_tgtdirp->mlh_pdo_hash;
2757 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)))
2761 rc = mdt_lock_two_dirs(info, mtgtdir, lh_tgtdirp,
2762 &rr->rr_tgt_name, msrcdir, lh_srcdirp,
2763 &rr->rr_name, cos_incompat);
2765 rc = mdt_lock_two_dirs(info, msrcdir, lh_srcdirp, &rr->rr_name,
2766 mtgtdir, lh_tgtdirp, &rr->rr_tgt_name,
2770 GOTO(out_unlock_rename, rc);
2772 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2773 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2775 /* find mold object. */
2777 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2779 GOTO(out_unlock_parents, rc);
2781 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2782 GOTO(out_unlock_parents, rc = -EINVAL);
2784 if (!fid_is_md_operative(old_fid))
2785 GOTO(out_unlock_parents, rc = -EPERM);
2787 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2789 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2791 if (!mdt_object_exists(mold)) {
2792 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2794 "object does not exist");
2795 GOTO(out_put_old, rc = -ENOENT);
2798 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2799 GOTO(out_put_old, rc = -EXDEV);
2801 /* Check if @mtgtdir is subdir of @mold, before locking child
2802 * to avoid reverse locking.
2804 if (mtgtdir != msrcdir) {
2805 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2810 GOTO(out_put_old, rc);
2814 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2815 /* save version after locking */
2816 mdt_version_get_save(info, mold, 2);
2818 if (!cos_incompat && mdt_object_remote(mold)) {
2819 cos_incompat = true;
2820 mdt_object_put(info->mti_env, mold);
2821 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2822 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2826 /* find mnew object:
2827 * mnew target object may not exist now
2828 * lookup with version checking
2831 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2834 /* the new_fid should have been filled at this moment */
2835 if (lu_fid_eq(old_fid, new_fid))
2836 GOTO(out_put_old, rc);
2838 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2839 lu_fid_eq(new_fid, rr->rr_fid2))
2840 GOTO(out_put_old, rc = -EINVAL);
2842 if (!fid_is_md_operative(new_fid))
2843 GOTO(out_put_old, rc = -EPERM);
2845 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2847 GOTO(out_put_old, rc = PTR_ERR(mnew));
2849 if (!mdt_object_exists(mnew)) {
2850 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2852 "object does not exist");
2853 GOTO(out_put_new, rc = -ENOENT);
2856 if (mdt_object_remote(mnew)) {
2857 struct mdt_body *repbody;
2859 /* Always send rename req to the target child MDT */
2860 repbody = req_capsule_server_get(info->mti_pill,
2862 LASSERT(repbody != NULL);
2863 repbody->mbo_fid1 = *new_fid;
2864 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2865 GOTO(out_put_new, rc = -EXDEV);
2867 /* Before locking the target dir, check we do not replace
2868 * a dir with a non-dir, otherwise it may deadlock with
2869 * link op which tries to create a link in this dir
2870 * back to this non-dir.
2872 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2873 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2874 GOTO(out_put_new, rc = -EISDIR);
2876 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2877 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2878 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2880 MDS_INODELOCK_LOOKUP |
2881 MDS_INODELOCK_XATTR, cos_incompat);
2883 GOTO(out_put_new, rc);
2885 /* Check if @msrcdir is subdir of @mnew, before locking child
2886 * to avoid reverse locking.
2888 if (mtgtdir != msrcdir) {
2889 rc = mdo_is_subdir(info->mti_env,
2890 mdt_object_child(msrcdir), new_fid);
2894 GOTO(out_unlock_old, rc);
2898 /* We used to acquire MDS_INODELOCK_FULL here but we
2899 * can't do this now because a running HSM restore on
2900 * the rename onto victim will hold the layout
2901 * lock. See LU-4002.
2904 lh_newp = &info->mti_lh[MDT_LH_NEW];
2905 rc = mdt_object_check_lock(info, mtgtdir, mnew, lh_newp,
2906 MDS_INODELOCK_LOOKUP |
2907 MDS_INODELOCK_UPDATE, LCK_EX,
2910 GOTO(out_unlock_new, rc);
2912 /* get and save version after locking */
2913 mdt_version_get_save(info, mnew, 3);
2914 } else if (rc != -ENOENT) {
2915 GOTO(out_put_old, rc);
2917 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2918 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2919 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2921 MDS_INODELOCK_LOOKUP |
2922 MDS_INODELOCK_XATTR, cos_incompat);
2924 GOTO(out_put_old, rc);
2926 mdt_enoent_version_save(info, 3);
2929 /* step 5: rename it */
2930 mdt_reint_init_ma(info, ma);
2932 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2933 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2936 mutex_lock(&mnew->mot_lov_mutex);
2938 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2939 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2940 mnew != NULL ? mdt_object_child(mnew) : NULL,
2941 &rr->rr_tgt_name, ma);
2944 mutex_unlock(&mnew->mot_lov_mutex);
2946 /* handle last link of tgt object */
2949 mdt_handle_last_unlink(info, mnew, ma);
2950 discard = mdt_dom_check_for_discard(info, mnew);
2952 mdt_rename_counter_tally(info, info->mti_mdt, req,
2953 msrcdir, mtgtdir, msi,
2954 ktime_us_delta(ktime_get(), kstart));
2960 mdt_object_unlock(info, mnew, lh_newp, rc);
2962 mdt_object_unlock(info, NULL, lh_lookup, rc);
2963 mdt_object_unlock(info, mold, lh_oldp, rc);
2965 if (mnew && !discard)
2966 mdt_object_put(info->mti_env, mnew);
2968 mdt_object_put(info->mti_env, mold);
2970 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2971 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2973 mdt_rename_unlock(info, rename_lh);
2975 mdt_object_put(info->mti_env, mtgtdir);
2977 mdt_object_put(info->mti_env, msrcdir);
2979 /* The DoM discard can be done right in the place above where it is
2980 * assigned, meanwhile it is done here after rename unlock due to
2981 * compatibility with old clients, for them the discard blocks
2982 * the main thread until completion. Check LU-11359 for details.
2985 mdt_dom_discard_data(info, mnew);
2986 mdt_object_put(info->mti_env, mnew);
2988 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
2992 static int mdt_reint_resync(struct mdt_thread_info *info,
2993 struct mdt_lock_handle *lhc)
2995 struct mdt_reint_record *rr = &info->mti_rr;
2996 struct ptlrpc_request *req = mdt_info_req(info);
2997 struct md_attr *ma = &info->mti_attr;
2998 struct mdt_object *mo;
2999 struct ldlm_lock *lease;
3000 struct mdt_body *repbody;
3001 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
3006 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
3008 if (info->mti_dlm_req)
3009 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
3011 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
3013 GOTO(out, rc = PTR_ERR(mo));
3015 if (!mdt_object_exists(mo))
3016 GOTO(out_obj, rc = -ENOENT);
3018 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
3019 GOTO(out_obj, rc = -EINVAL);
3021 if (mdt_object_remote(mo))
3022 GOTO(out_obj, rc = -EREMOTE);
3024 lease = ldlm_handle2lock(rr->rr_lease_handle);
3026 GOTO(out_obj, rc = -ESTALE);
3028 /* It's really necessary to grab open_sem and check if the lease lock
3029 * has been lost. There would exist a concurrent writer coming in and
3030 * generating some dirty data in memory cache, the writeback would fail
3031 * after the layout version is increased by MDS_REINT_RESYNC RPC.
3033 if (!down_write_trylock(&mo->mot_open_sem))
3034 GOTO(out_put_lease, rc = -EBUSY);
3036 lock_res_and_lock(lease);
3037 lease_broken = ldlm_is_cancel(lease);
3038 unlock_res_and_lock(lease);
3040 GOTO(out_unlock, rc = -EBUSY);
3042 /* the file has yet opened by anyone else after we took the lease. */
3043 layout.mlc_opc = MD_LAYOUT_RESYNC;
3044 lhc = &info->mti_lh[MDT_LH_LOCAL];
3045 rc = mdt_layout_change(info, mo, lhc, &layout);
3047 GOTO(out_unlock, rc);
3049 mdt_object_unlock(info, mo, lhc, 0);
3051 ma->ma_need = MA_INODE;
3053 rc = mdt_attr_get_complex(info, mo, ma);
3055 GOTO(out_unlock, rc);
3057 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3058 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3062 up_write(&mo->mot_open_sem);
3064 LDLM_LOCK_PUT(lease);
3066 mdt_object_put(info->mti_env, mo);
3068 mdt_client_compatibility(info);
3072 struct mdt_reinter {
3073 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3074 enum lprocfs_extra_opc mr_extra_opc;
3077 static const struct mdt_reinter mdt_reinters[] = {
3079 .mr_handler = &mdt_reint_setattr,
3080 .mr_extra_opc = MDS_REINT_SETATTR,
3083 .mr_handler = &mdt_reint_create,
3084 .mr_extra_opc = MDS_REINT_CREATE,
3087 .mr_handler = &mdt_reint_link,
3088 .mr_extra_opc = MDS_REINT_LINK,
3091 .mr_handler = &mdt_reint_unlink,
3092 .mr_extra_opc = MDS_REINT_UNLINK,
3095 .mr_handler = &mdt_reint_rename,
3096 .mr_extra_opc = MDS_REINT_RENAME,
3099 .mr_handler = &mdt_reint_open,
3100 .mr_extra_opc = MDS_REINT_OPEN,
3102 [REINT_SETXATTR] = {
3103 .mr_handler = &mdt_reint_setxattr,
3104 .mr_extra_opc = MDS_REINT_SETXATTR,
3107 .mr_handler = &mdt_reint_unlink,
3108 .mr_extra_opc = MDS_REINT_UNLINK,
3111 .mr_handler = &mdt_reint_migrate,
3112 .mr_extra_opc = MDS_REINT_RENAME,
3115 .mr_handler = &mdt_reint_resync,
3116 .mr_extra_opc = MDS_REINT_RESYNC,
3120 int mdt_reint_rec(struct mdt_thread_info *info,
3121 struct mdt_lock_handle *lhc)
3123 const struct mdt_reinter *mr;
3127 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3130 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3131 if (mr->mr_handler == NULL)
3134 rc = (*mr->mr_handler)(info, lhc);
3136 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3137 PTLRPC_LAST_CNTR + mr->mr_extra_opc);