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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_reint.c
38 * Lustre Metadata Target (mdt) reintegration routines
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Huang Hua <huanghua@clusterfs.com>
44 * Author: Yury Umanets <umka@clusterfs.com>
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include <lprocfs_status.h>
50 #include "mdt_internal.h"
51 #include <lustre_lmv.h>
53 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
56 ma->ma_need = MA_INODE;
61 * Get version of object by fid.
63 * Return real version or ENOENT_VERSION if object doesn't exist
65 static void mdt_obj_version_get(struct mdt_thread_info *info,
66 struct mdt_object *o, __u64 *version)
69 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
70 !fid_is_obf(mdt_object_fid(o)))
71 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
73 *version = ENOENT_VERSION;
74 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
75 PFID(mdt_object_fid(o)), *version);
79 * Check version is correct.
81 * Should be called only during replay.
83 static int mdt_version_check(struct ptlrpc_request *req,
84 __u64 version, int idx)
86 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
89 if (!exp_connect_vbr(req->rq_export))
92 LASSERT(req_is_replay(req));
93 /** VBR: version is checked always because costs nothing */
94 LASSERT(idx < PTLRPC_NUM_VERSIONS);
95 /** Sanity check for malformed buffers */
96 if (pre_ver == NULL) {
97 CERROR("No versions in request buffer\n");
98 spin_lock(&req->rq_export->exp_lock);
99 req->rq_export->exp_vbr_failed = 1;
100 spin_unlock(&req->rq_export->exp_lock);
102 } else if (pre_ver[idx] != version) {
103 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
104 pre_ver[idx], version);
105 spin_lock(&req->rq_export->exp_lock);
106 req->rq_export->exp_vbr_failed = 1;
107 spin_unlock(&req->rq_export->exp_lock);
114 * Save pre-versions in reply.
116 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
121 if (!exp_connect_vbr(req->rq_export))
124 LASSERT(!req_is_replay(req));
125 LASSERT(req->rq_repmsg != NULL);
126 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
128 reply_ver[idx] = version;
132 * Save enoent version, it is needed when it is obvious that object doesn't
133 * exist, e.g. child during create.
135 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
137 /* save version of file name for replay, it must be ENOENT here */
138 if (!req_is_replay(mdt_info_req(info))) {
139 info->mti_ver[idx] = ENOENT_VERSION;
140 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
145 * Get version from disk and save in reply buffer.
147 * Versions are saved in reply only during normal operations not replays.
149 void mdt_version_get_save(struct mdt_thread_info *info,
150 struct mdt_object *mto, int idx)
152 /* don't save versions during replay */
153 if (!req_is_replay(mdt_info_req(info))) {
154 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
155 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
160 * Get version from disk and check it, no save in reply.
162 int mdt_version_get_check(struct mdt_thread_info *info,
163 struct mdt_object *mto, int idx)
165 /* only check versions during replay */
166 if (!req_is_replay(mdt_info_req(info)))
169 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
170 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
174 * Get version from disk and check if recovery or just save.
176 int mdt_version_get_check_save(struct mdt_thread_info *info,
177 struct mdt_object *mto, int idx)
181 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
182 if (req_is_replay(mdt_info_req(info)))
183 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
186 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
191 * Lookup with version checking.
193 * This checks version of 'name'. Many reint functions uses 'name' for child not
194 * FID, therefore we need to get object by name and check its version.
196 static int mdt_lookup_version_check(struct mdt_thread_info *info,
197 struct mdt_object *p,
198 const struct lu_name *lname,
199 struct lu_fid *fid, int idx)
203 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
205 /* Check version only during replay */
206 if (!req_is_replay(mdt_info_req(info)))
209 info->mti_ver[idx] = ENOENT_VERSION;
211 struct mdt_object *child;
212 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
213 if (likely(!IS_ERR(child))) {
214 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
215 mdt_object_put(info->mti_env, child);
218 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
219 return vbrc ? vbrc : rc;
223 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
225 struct lu_ucred *uc = mdt_ucred(info);
226 struct mdt_device *mdt = info->mti_mdt;
228 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
229 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
230 mdt->mdt_enable_remote_dir_gid != -1)
238 * mdt_remote_permission: Check whether the remote operation is permitted,
240 * Only sysadmin can create remote directory / striped directory,
241 * migrate directory and set default stripedEA on directory, unless
243 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
245 * param[in] info: mdt_thread_info.
247 * retval = 0 remote operation is allowed.
248 * < 0 remote operation is denied.
250 static int mdt_remote_permission(struct mdt_thread_info *info)
252 struct md_op_spec *spec = &info->mti_spec;
253 struct lu_attr *attr = &info->mti_attr.ma_attr;
254 struct obd_export *exp = mdt_info_req(info)->rq_export;
257 if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
258 rc = mdt_remote_permission_check(info);
263 if (info->mti_rr.rr_opcode == REINT_CREATE &&
264 (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
265 spec->u.sp_ea.eadatalen != 0)) {
266 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
268 /* Only new clients can create remote dir( >= 2.4) and
269 * striped dir(>= 2.6), old client will return -ENOTSUPP */
270 if (!mdt_is_dne_client(exp))
273 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
274 !mdt_is_striped_client(exp))
277 rc = mdt_remote_permission_check(info);
282 if (info->mti_rr.rr_opcode == REINT_SETATTR) {
283 struct md_attr *ma = &info->mti_attr;
285 if ((ma->ma_valid & MA_LMV)) {
286 rc = mdt_remote_permission_check(info);
296 * VBR: we save three versions in reply:
297 * 0 - parent. Check that parent version is the same during replay.
298 * 1 - name. Version of 'name' if file exists with the same name or
299 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
300 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
303 static int mdt_md_create(struct mdt_thread_info *info)
305 struct mdt_device *mdt = info->mti_mdt;
306 struct mdt_object *parent;
307 struct mdt_object *child;
308 struct mdt_lock_handle *lh;
309 struct mdt_body *repbody;
310 struct md_attr *ma = &info->mti_attr;
311 struct mdt_reint_record *rr = &info->mti_rr;
315 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
317 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
319 if (!fid_is_md_operative(rr->rr_fid1))
322 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
324 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
326 RETURN(PTR_ERR(parent));
328 if (!mdt_object_exists(parent))
329 GOTO(put_parent, rc = -ENOENT);
331 lh = &info->mti_lh[MDT_LH_PARENT];
332 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
333 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
335 GOTO(put_parent, rc);
337 if (!mdt_object_remote(parent)) {
338 rc = mdt_version_get_check_save(info, parent, 0);
340 GOTO(unlock_parent, rc);
344 * Check child name version during replay.
345 * During create replay a file may exist with same name.
347 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
348 &info->mti_tmp_fid1, 1);
350 GOTO(unlock_parent, rc = -EEXIST);
352 /* -ENOENT is expected here */
354 GOTO(unlock_parent, rc);
356 /* save version of file name for replay, it must be ENOENT here */
357 mdt_enoent_version_save(info, 1);
359 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
360 if (likely(!IS_ERR(child))) {
361 struct md_object *next = mdt_object_child(parent);
363 rc = mdt_remote_permission(info);
365 GOTO(out_put_child, rc);
367 ma->ma_need = MA_INODE;
370 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
371 OBD_FAIL_MDS_REINT_CREATE_WRITE);
373 /* Version of child will be updated on disk. */
374 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
375 rc = mdt_version_get_check_save(info, child, 2);
377 GOTO(out_put_child, rc);
379 /* Let lower layer know current lock mode. */
380 info->mti_spec.sp_cr_mode =
381 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
384 * Do not perform lookup sanity check. We know that name does
387 info->mti_spec.sp_cr_lookup = 0;
388 info->mti_spec.sp_feat = &dt_directory_features;
390 rc = mdo_create(info->mti_env, next, &rr->rr_name,
391 mdt_object_child(child), &info->mti_spec, ma);
393 rc = mdt_attr_get_complex(info, child, ma);
396 /* Return fid & attr to client. */
397 if (ma->ma_valid & MA_INODE)
398 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
399 mdt_object_fid(child));
402 mdt_object_put(info->mti_env, child);
407 mdt_object_unlock(info, parent, lh, rc);
409 mdt_object_put(info->mti_env, parent);
413 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
414 struct mdt_object *obj, __u64 ibits,
415 struct mdt_lock_handle *s0_lh,
416 struct mdt_object *s0_obj,
417 struct ldlm_enqueue_info *einfo)
419 union ldlm_policy_data *policy = &mti->mti_policy;
423 if (!S_ISDIR(obj->mot_header.loh_attr))
426 /* Unlock stripe 0 */
427 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
428 LASSERT(s0_obj != NULL);
429 mdt_object_unlock_put(mti, s0_obj, s0_lh, 1);
432 memset(policy, 0, sizeof(*policy));
433 policy->l_inodebits.bits = ibits;
435 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
441 * Lock slave stripes if necessary, the lock handles of slave stripes
442 * will be stored in einfo->ei_cbdata.
444 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
445 enum ldlm_mode mode, __u64 ibits,
446 struct mdt_lock_handle *s0_lh,
447 struct mdt_object **s0_objp,
448 struct ldlm_enqueue_info *einfo)
450 union ldlm_policy_data *policy = &mti->mti_policy;
451 struct lu_buf *buf = &mti->mti_buf;
452 struct lmv_mds_md_v1 *lmv;
453 struct lu_fid *fid = &mti->mti_tmp_fid1;
457 if (!S_ISDIR(obj->mot_header.loh_attr))
460 buf->lb_buf = mti->mti_xattr_buf;
461 buf->lb_len = sizeof(mti->mti_xattr_buf);
462 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
465 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
467 buf->lb_buf = mti->mti_big_lmm;
468 buf->lb_len = mti->mti_big_lmmsize;
472 if (rc == -ENODATA || rc == -ENOENT)
479 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
482 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
483 if (!lu_fid_eq(fid, mdt_object_fid(obj))) {
484 /* Except migrating object, whose 0_stripe and master
485 * object are the same object, 0_stripe and master
486 * object are different, though they are in the same
487 * MDT, to avoid adding osd_object_lock here, so we
488 * will enqueue the stripe0 lock in MDT0 for now */
489 *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits);
490 if (IS_ERR(*s0_objp))
491 RETURN(PTR_ERR(*s0_objp));
494 memset(einfo, 0, sizeof(*einfo));
495 einfo->ei_type = LDLM_IBITS;
496 einfo->ei_mode = mode;
497 einfo->ei_cb_bl = mdt_remote_blocking_ast;
498 einfo->ei_cb_local_bl = mdt_blocking_ast;
499 einfo->ei_cb_cp = ldlm_completion_ast;
500 einfo->ei_enq_slave = 1;
501 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
502 memset(policy, 0, sizeof(*policy));
503 policy->l_inodebits.bits = ibits;
505 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
510 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
513 struct mdt_lock_handle *lh;
514 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
515 __u64 lockpart = MDS_INODELOCK_UPDATE;
516 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
517 struct mdt_lock_handle *s0_lh;
518 struct mdt_object *s0_obj = NULL;
522 lh = &info->mti_lh[MDT_LH_PARENT];
523 mdt_lock_reg_init(lh, LCK_PW);
525 /* Even though the new MDT will grant PERM lock to the old
526 * client, but the old client will almost ignore that during
527 * So it needs to revoke both LOOKUP and PERM lock here, so
528 * both new and old client can cancel the dcache */
529 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
530 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
532 rc = mdt_object_lock(info, mo, lh, lockpart);
536 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
537 mdt_lock_reg_init(s0_lh, LCK_PW);
538 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
540 GOTO(out_unlock, rc);
542 /* all attrs are packed into mti_attr in unpack_setattr */
543 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
544 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
546 /* This is only for set ctime when rename's source is on remote MDS. */
547 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
548 ma->ma_attr_flags |= MDS_VTX_BYPASS;
550 /* VBR: update version if attr changed are important for recovery */
552 /* update on-disk version of changed object */
553 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
554 rc = mdt_version_get_check_save(info, mo, 0);
556 GOTO(out_unlock, rc);
559 /* Ensure constant striping during chown(). See LU-2789. */
560 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
561 mutex_lock(&mo->mot_lov_mutex);
563 /* all attrs are packed into mti_attr in unpack_setattr */
564 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
566 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
567 mutex_unlock(&mo->mot_lov_mutex);
570 GOTO(out_unlock, rc);
574 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
575 mdt_object_unlock(info, mo, lh, rc);
580 * Check HSM flags and add HS_DIRTY flag if relevant.
582 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
583 * and is not RELEASED.
585 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
591 /* If the file was modified, add the dirty flag */
592 ma->ma_need = MA_HSM;
593 rc = mdt_attr_get_complex(info, mo, ma);
595 CERROR("file attribute read error for "DFID": %d.\n",
596 PFID(mdt_object_fid(mo)), rc);
600 /* If an up2date copy exists in the backend, add dirty flag */
601 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
602 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
603 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
605 ma->ma_hsm.mh_flags |= HS_DIRTY;
607 mdt_lock_reg_init(lh, LCK_PW);
608 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
612 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
614 CERROR("file attribute change error for "DFID": %d\n",
615 PFID(mdt_object_fid(mo)), rc);
616 mdt_object_unlock(info, mo, lh, rc);
622 static int mdt_reint_setattr(struct mdt_thread_info *info,
623 struct mdt_lock_handle *lhc)
625 struct md_attr *ma = &info->mti_attr;
626 struct mdt_reint_record *rr = &info->mti_rr;
627 struct ptlrpc_request *req = mdt_info_req(info);
628 struct mdt_object *mo;
629 struct mdt_body *repbody;
633 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
634 (unsigned int)ma->ma_attr.la_valid);
636 if (info->mti_dlm_req)
637 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
639 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
640 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
642 GOTO(out, rc = PTR_ERR(mo));
644 if (!mdt_object_exists(mo))
645 GOTO(out_put, rc = -ENOENT);
647 if (mdt_object_remote(mo))
648 GOTO(out_put, rc = -EREMOTE);
650 if ((ma->ma_attr.la_valid & LA_SIZE) ||
651 (rr->rr_flags & MRF_OPEN_TRUNC)) {
652 /* Check write access for the O_TRUNC case */
653 if (mdt_write_read(mo) < 0)
654 GOTO(out_put, rc = -ETXTBSY);
657 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
658 if (ma->ma_valid & MA_LOV)
659 GOTO(out_put, rc = -EPROTO);
661 rc = mdt_attr_set(info, mo, ma);
664 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
665 struct lu_buf *buf = &info->mti_buf;
667 if (ma->ma_attr.la_valid != 0)
668 GOTO(out_put, rc = -EPROTO);
670 buf->lb_buf = ma->ma_lmm;
671 buf->lb_len = ma->ma_lmm_size;
672 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
673 buf, XATTR_NAME_LOV, 0);
676 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
677 struct lu_buf *buf = &info->mti_buf;
678 struct mdt_lock_handle *lh;
680 rc = mdt_remote_permission(info);
684 if (ma->ma_attr.la_valid != 0)
685 GOTO(out_put, rc = -EPROTO);
687 lh = &info->mti_lh[MDT_LH_PARENT];
688 mdt_lock_reg_init(lh, LCK_PW);
690 rc = mdt_object_lock(info, mo, lh,
691 MDS_INODELOCK_XATTR);
695 buf->lb_buf = ma->ma_lmv;
696 buf->lb_len = ma->ma_lmv_size;
697 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
698 buf, XATTR_NAME_DEFAULT_LMV, 0);
700 mdt_object_unlock(info, mo, lh, rc);
704 GOTO(out_put, rc = -EPROTO);
707 /* If file data is modified, add the dirty flag */
708 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
709 rc = mdt_add_dirty_flag(info, mo, ma);
711 ma->ma_need = MA_INODE;
713 rc = mdt_attr_get_complex(info, mo, ma);
717 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
721 mdt_object_put(info->mti_env, mo);
724 mdt_counter_incr(req, LPROC_MDT_SETATTR);
726 mdt_client_compatibility(info);
727 rc2 = mdt_fix_reply(info);
733 static int mdt_reint_create(struct mdt_thread_info *info,
734 struct mdt_lock_handle *lhc)
736 struct ptlrpc_request *req = mdt_info_req(info);
740 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
741 RETURN(err_serious(-ESTALE));
743 if (info->mti_dlm_req)
744 ldlm_request_cancel(mdt_info_req(info),
745 info->mti_dlm_req, 0, LATF_SKIP);
747 if (!lu_name_is_valid(&info->mti_rr.rr_name))
750 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
752 mdt_counter_incr(req, LPROC_MDT_MKDIR);
760 /* Special file should stay on the same node as parent. */
761 mdt_counter_incr(req, LPROC_MDT_MKNOD);
764 CERROR("%s: Unsupported mode %o\n",
765 mdt_obd_name(info->mti_mdt),
766 info->mti_attr.ma_attr.la_mode);
767 RETURN(err_serious(-EOPNOTSUPP));
770 rc = mdt_md_create(info);
775 * VBR: save parent version in reply and child version getting by its name.
776 * Version of child is getting and checking during its lookup. If
778 static int mdt_reint_unlink(struct mdt_thread_info *info,
779 struct mdt_lock_handle *lhc)
781 struct mdt_reint_record *rr = &info->mti_rr;
782 struct ptlrpc_request *req = mdt_info_req(info);
783 struct md_attr *ma = &info->mti_attr;
784 struct lu_fid *child_fid = &info->mti_tmp_fid1;
785 struct mdt_object *mp;
786 struct mdt_object *mc;
787 struct mdt_lock_handle *parent_lh;
788 struct mdt_lock_handle *child_lh;
789 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
790 struct mdt_lock_handle *s0_lh = NULL;
791 struct mdt_object *s0_obj = NULL;
797 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
798 PNAME(&rr->rr_name));
800 if (info->mti_dlm_req)
801 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
803 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
804 RETURN(err_serious(-ENOENT));
806 if (!fid_is_md_operative(rr->rr_fid1))
810 * step 1: Found the parent.
812 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
818 parent_lh = &info->mti_lh[MDT_LH_PARENT];
819 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
820 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE);
822 GOTO(put_parent, rc);
824 if (!mdt_object_remote(mp)) {
825 rc = mdt_version_get_check_save(info, mp, 0);
827 GOTO(unlock_parent, rc);
830 /* step 2: find & lock the child */
831 /* lookup child object along with version checking */
833 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
835 /* Name might not be able to find during resend of
836 * remote unlink, considering following case.
837 * dir_A is a remote directory, the name entry of
838 * dir_A is on MDT0, the directory is on MDT1,
840 * 1. client sends unlink req to MDT1.
841 * 2. MDT1 sends name delete update to MDT0.
842 * 3. name entry is being deleted in MDT0 synchronously.
843 * 4. MDT1 is restarted.
844 * 5. client resends unlink req to MDT1. So it can not
845 * find the name entry on MDT0 anymore.
846 * In this case, MDT1 only needs to destory the local
849 if (mdt_object_remote(mp) && rc == -ENOENT &&
850 !fid_is_zero(rr->rr_fid2) &&
851 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
853 *child_fid = *rr->rr_fid2;
855 GOTO(unlock_parent, rc);
859 if (!fid_is_md_operative(child_fid))
860 GOTO(unlock_parent, rc = -EPERM);
862 /* We will lock the child regardless it is local or remote. No harm. */
863 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
865 GOTO(unlock_parent, rc = PTR_ERR(mc));
867 child_lh = &info->mti_lh[MDT_LH_CHILD];
868 mdt_lock_reg_init(child_lh, LCK_EX);
869 if (mdt_object_remote(mc)) {
870 struct mdt_body *repbody;
872 if (!fid_is_zero(rr->rr_fid2)) {
873 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
874 mdt_obd_name(info->mti_mdt),
875 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
876 GOTO(put_child, rc = -ENOENT);
878 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
879 mdt_obd_name(info->mti_mdt),
880 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
882 if (!mdt_is_dne_client(req->rq_export))
883 /* Return -ENOTSUPP for old client */
884 GOTO(put_child, rc = -ENOTSUPP);
886 if (info->mti_spec.sp_rm_entry) {
887 struct lu_ucred *uc = mdt_ucred(info);
889 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
890 CERROR("%s: unlink remote entry is only "
891 "permitted for administrator: rc = %d\n",
892 mdt_obd_name(info->mti_mdt),
894 GOTO(put_child, rc = -EPERM);
897 ma->ma_need = MA_INODE;
899 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
900 NULL, &rr->rr_name, ma, no_name);
903 /* Revoke the LOOKUP lock of the remote object granted by
904 * this MDT. Since the unlink will happen on another MDT,
905 * it will release the LOOKUP lock right away. Then What
906 * would happen if another client try to grab the LOOKUP
907 * lock at the same time with unlink XXX */
908 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
909 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
910 LASSERT(repbody != NULL);
911 repbody->mbo_fid1 = *mdt_object_fid(mc);
912 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
913 GOTO(unlock_child, rc = -EREMOTE);
914 } else if (info->mti_spec.sp_rm_entry) {
916 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
918 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
922 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
923 * this now because a running HSM restore on the child (unlink
924 * victim) will hold the layout lock. See LU-4002. */
925 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
926 if (mdt_object_remote(mp)) {
927 /* Enqueue lookup lock from parent MDT */
928 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
929 &child_lh->mlh_rreg_lh,
930 child_lh->mlh_rreg_mode,
931 MDS_INODELOCK_LOOKUP, false);
935 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
938 rc = mdt_object_lock(info, mc, child_lh, lock_ibits);
942 * Now we can only make sure we need MA_INODE, in mdd layer, will check
943 * whether need MA_LOV and MA_COOKIE.
945 ma->ma_need = MA_INODE;
948 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
949 mdt_lock_reg_init(s0_lh, LCK_EX);
950 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
953 GOTO(unlock_child, rc);
955 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
956 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
957 /* save version when object is locked */
958 mdt_version_get_save(info, mc, 1);
960 mutex_lock(&mc->mot_lov_mutex);
962 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
963 mdt_object_child(mc), &rr->rr_name, ma, no_name);
965 mutex_unlock(&mc->mot_lov_mutex);
967 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
968 rc = mdt_attr_get_complex(info, mc, ma);
970 mdt_handle_last_unlink(info, mc, ma);
972 if (ma->ma_valid & MA_INODE) {
973 switch (ma->ma_attr.la_mode & S_IFMT) {
975 mdt_counter_incr(req, LPROC_MDT_RMDIR);
983 mdt_counter_incr(req, LPROC_MDT_UNLINK);
986 LASSERTF(0, "bad file type %o unlinking\n",
987 ma->ma_attr.la_mode);
994 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
995 mdt_object_unlock(info, mc, child_lh, rc);
997 mdt_object_put(info->mti_env, mc);
999 mdt_object_unlock(info, mp, parent_lh, rc);
1001 mdt_object_put(info->mti_env, mp);
1007 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1010 static int mdt_reint_link(struct mdt_thread_info *info,
1011 struct mdt_lock_handle *lhc)
1013 struct mdt_reint_record *rr = &info->mti_rr;
1014 struct ptlrpc_request *req = mdt_info_req(info);
1015 struct md_attr *ma = &info->mti_attr;
1016 struct mdt_object *ms;
1017 struct mdt_object *mp;
1018 struct mdt_lock_handle *lhs;
1019 struct mdt_lock_handle *lhp;
1023 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1024 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1026 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1027 RETURN(err_serious(-ENOENT));
1029 if (info->mti_dlm_req)
1030 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1032 /* Invalid case so return error immediately instead of
1034 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1037 if (!fid_is_md_operative(rr->rr_fid1) ||
1038 !fid_is_md_operative(rr->rr_fid2))
1041 /* step 1: find & lock the target parent dir */
1042 lhp = &info->mti_lh[MDT_LH_PARENT];
1043 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1044 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1045 MDS_INODELOCK_UPDATE);
1047 RETURN(PTR_ERR(mp));
1049 rc = mdt_version_get_check_save(info, mp, 0);
1051 GOTO(out_unlock_parent, rc);
1053 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1055 /* step 2: find & lock the source */
1056 lhs = &info->mti_lh[MDT_LH_CHILD];
1057 mdt_lock_reg_init(lhs, LCK_EX);
1059 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1061 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1063 if (!mdt_object_exists(ms)) {
1064 mdt_object_put(info->mti_env, ms);
1065 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1066 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1067 GOTO(out_unlock_parent, rc = -ENOENT);
1070 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1071 MDS_INODELOCK_XATTR);
1073 mdt_object_put(info->mti_env, ms);
1074 GOTO(out_unlock_parent, rc);
1077 /* step 3: link it */
1078 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1079 OBD_FAIL_MDS_REINT_LINK_WRITE);
1081 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1082 rc = mdt_version_get_check_save(info, ms, 1);
1084 GOTO(out_unlock_child, rc);
1086 /** check target version by name during replay */
1087 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1088 &info->mti_tmp_fid1, 2);
1089 if (rc != 0 && rc != -ENOENT)
1090 GOTO(out_unlock_child, rc);
1091 /* save version of file name for replay, it must be ENOENT here */
1092 if (!req_is_replay(mdt_info_req(info))) {
1093 if (rc != -ENOENT) {
1094 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1095 PNAME(&rr->rr_name));
1096 GOTO(out_unlock_child, rc = -EEXIST);
1098 info->mti_ver[2] = ENOENT_VERSION;
1099 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1102 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1103 mdt_object_child(ms), &rr->rr_name, ma);
1106 mdt_counter_incr(req, LPROC_MDT_LINK);
1110 mdt_object_unlock_put(info, ms, lhs, rc);
1112 mdt_object_unlock_put(info, mp, lhp, rc);
1116 * lock the part of the directory according to the hash of the name
1117 * (lh->mlh_pdo_hash) in parallel directory lock.
1119 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1120 struct mdt_lock_handle *lh,
1121 struct mdt_object *obj, __u64 ibits)
1123 struct ldlm_res_id *res = &info->mti_res_id;
1124 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1125 union ldlm_policy_data *policy = &info->mti_policy;
1129 * Finish res_id initializing by name hash marking part of
1130 * directory which is taking modification.
1132 LASSERT(lh->mlh_pdo_hash != 0);
1133 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1134 memset(policy, 0, sizeof(*policy));
1135 policy->l_inodebits.bits = ibits;
1137 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1138 * going to be sent to client. If it is - mdt_intent_policy() path will
1139 * fix it up and turn FL_LOCAL flag off.
1141 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1142 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1143 &info->mti_exp->exp_handle.h_cookie);
1148 * Get BFL lock for rename or migrate process.
1150 static int mdt_rename_lock(struct mdt_thread_info *info,
1151 struct lustre_handle *lh)
1156 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1157 struct lu_fid *fid = &info->mti_tmp_fid1;
1158 struct mdt_object *obj;
1160 /* XXX, right now, it has to use object API to
1161 * enqueue lock cross MDT, so it will enqueue
1162 * rename lock(with LUSTRE_BFL_FID) by root object */
1164 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1166 RETURN(PTR_ERR(obj));
1168 rc = mdt_remote_object_lock(info, obj,
1169 &LUSTRE_BFL_FID, lh,
1171 MDS_INODELOCK_UPDATE, false);
1172 mdt_object_put(info->mti_env, obj);
1174 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1175 union ldlm_policy_data *policy = &info->mti_policy;
1176 struct ldlm_res_id *res_id = &info->mti_res_id;
1179 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1180 memset(policy, 0, sizeof *policy);
1181 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1182 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1183 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1184 LCK_EX, &flags, ldlm_blocking_ast,
1185 ldlm_completion_ast, NULL, NULL, 0,
1187 &info->mti_exp->exp_handle.h_cookie,
1194 static void mdt_rename_unlock(struct lustre_handle *lh)
1197 LASSERT(lustre_handle_is_used(lh));
1198 /* Cancel the single rename lock right away */
1199 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1204 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1205 * target. Source should not be ancestor of target dir. May be other rename
1206 * checks can be moved here later.
1208 static int mdt_is_subdir(struct mdt_thread_info *info,
1209 struct mdt_object *dir,
1210 const struct lu_fid *fid)
1212 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1216 /* If the source and target are in the same directory, they can not
1217 * be parent/child relationship, so subdir check is not needed */
1218 if (lu_fid_eq(&dir_fid, fid))
1221 if (!mdt_object_exists(dir))
1224 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1227 CERROR("%s: failed subdir check in "DFID" for "DFID
1228 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1229 PFID(&dir_fid), PFID(fid), rc);
1230 /* Return EINVAL only if a parent is the @fid */
1234 /* check the found fid */
1235 if (lu_fid_eq(&dir_fid, fid))
1242 /* Update object linkEA */
1243 struct mdt_lock_list {
1244 struct mdt_object *mll_obj;
1245 struct mdt_lock_handle mll_lh;
1246 struct list_head mll_list;
1249 static void mdt_unlock_list(struct mdt_thread_info *info,
1250 struct list_head *list, int rc)
1252 struct mdt_lock_list *mll;
1253 struct mdt_lock_list *mll2;
1255 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1256 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1257 list_del(&mll->mll_list);
1262 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1263 struct mdt_object *obj,
1264 struct mdt_object *pobj,
1265 struct list_head *lock_list)
1267 struct lu_buf *buf = &info->mti_big_buf;
1268 struct linkea_data ldata = { NULL };
1273 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1276 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1277 if (buf->lb_buf == NULL)
1281 rc = mdt_links_read(info, obj, &ldata);
1283 if (rc == -ENOENT || rc == -ENODATA)
1288 LASSERT(ldata.ld_leh != NULL);
1289 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1290 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1291 struct mdt_device *mdt = info->mti_mdt;
1292 struct mdt_object *mdt_pobj;
1293 struct mdt_lock_list *mll;
1294 struct lu_name name;
1297 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1299 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1300 if (IS_ERR(mdt_pobj)) {
1301 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1302 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1306 if (!mdt_object_exists(mdt_pobj)) {
1307 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1308 mdt_obd_name(mdt), PFID(&fid));
1309 mdt_object_put(info->mti_env, mdt_pobj);
1313 /* Check if the object already exists in the list */
1314 list_for_each_entry(mll, lock_list, mll_list) {
1315 if (mll->mll_obj == mdt_pobj) {
1316 mdt_object_put(info->mti_env, mdt_pobj);
1321 if (mdt_pobj == pobj) {
1322 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1323 mdt_obd_name(mdt), PFID(&fid));
1324 mdt_object_put(info->mti_env, mdt_pobj);
1330 mdt_object_put(info->mti_env, mdt_pobj);
1331 GOTO(out, rc = -ENOMEM);
1334 /* Since this needs to lock all of objects in linkea, to avoid
1335 * deadlocks, because it does not follow parent-child order as
1336 * other MDT operation, let's use try_lock here, i.e. it will
1337 * return immediately once there are conflict locks, and return
1338 * EBUSY to client */
1339 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1340 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh,
1341 MDS_INODELOCK_UPDATE);
1343 CDEBUG(D_ERROR, "%s: cannot lock "DFID": rc =%d\n",
1344 mdt_obd_name(mdt), PFID(&fid), rc);
1345 mdt_object_put(info->mti_env, mdt_pobj);
1347 GOTO(out, rc = -EBUSY);
1350 INIT_LIST_HEAD(&mll->mll_list);
1351 mll->mll_obj = mdt_pobj;
1352 list_add_tail(&mll->mll_list, lock_list);
1354 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1360 mdt_unlock_list(info, lock_list, rc);
1364 /* migrate files from one MDT to another MDT */
1365 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1366 struct mdt_lock_handle *lhc)
1368 struct mdt_reint_record *rr = &info->mti_rr;
1369 struct md_attr *ma = &info->mti_attr;
1370 struct mdt_object *msrcdir;
1371 struct mdt_object *mold;
1372 struct mdt_object *mnew = NULL;
1373 struct mdt_lock_handle *lh_dirp;
1374 struct mdt_lock_handle *lh_childp;
1375 struct mdt_lock_handle *lh_tgtp = NULL;
1376 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1377 struct list_head lock_list;
1379 struct ldlm_lock *lease = NULL;
1380 bool lock_open_sem = false;
1384 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1385 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1387 /* 1: lock the source dir. */
1388 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1389 if (IS_ERR(msrcdir)) {
1390 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1391 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1392 (int)PTR_ERR(msrcdir));
1393 RETURN(PTR_ERR(msrcdir));
1396 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1397 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1398 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1399 MDS_INODELOCK_UPDATE);
1401 GOTO(out_put_parent, rc);
1403 if (!mdt_object_remote(msrcdir)) {
1404 rc = mdt_version_get_check_save(info, msrcdir, 0);
1406 GOTO(out_unlock_parent, rc);
1409 /* 2: sanity check and find the object to be migrated. */
1411 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1413 GOTO(out_unlock_parent, rc);
1415 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1416 GOTO(out_unlock_parent, rc = -EINVAL);
1418 if (!fid_is_md_operative(old_fid))
1419 GOTO(out_unlock_parent, rc = -EPERM);
1421 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1423 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1425 if (mdt_object_remote(mold)) {
1426 CERROR("%s: source "DFID" is on the remote MDT\n",
1427 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1428 GOTO(out_put_child, rc = -EREMOTE);
1431 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1432 !mdt_object_remote(msrcdir)) {
1433 CERROR("%s: parent "DFID" is still on the same"
1434 " MDT, which should be migrated first:"
1435 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1436 PFID(mdt_object_fid(msrcdir)), -EPERM);
1437 GOTO(out_put_child, rc = -EPERM);
1440 rc = mdt_remote_permission(info);
1442 GOTO(out_put_child, rc);
1444 /* 3: iterate the linkea of the object and lock all of the objects */
1445 INIT_LIST_HEAD(&lock_list);
1446 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1448 GOTO(out_put_child, rc);
1450 if (info->mti_spec.sp_migrate_close) {
1451 struct close_data *data;
1452 struct mdt_body *repbody;
1453 bool lease_broken = false;
1455 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1457 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1459 GOTO(out_lease, rc = -EPROTO);
1461 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1463 GOTO(out_lease, rc = -EPROTO);
1465 lease = ldlm_handle2lock(&data->cd_handle);
1467 GOTO(out_lease, rc = -ESTALE);
1469 /* try to hold open_sem so that nobody else can open the file */
1470 if (!down_write_trylock(&mold->mot_open_sem)) {
1471 ldlm_lock_cancel(lease);
1472 GOTO(out_lease, rc = -EBUSY);
1475 lock_open_sem = true;
1476 /* Check if the lease open lease has already canceled */
1477 lock_res_and_lock(lease);
1478 lease_broken = ldlm_is_cancel(lease);
1479 unlock_res_and_lock(lease);
1481 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1482 PFID(mdt_object_fid(mold)), lease_broken);
1484 /* Cancel server side lease. Client side counterpart should
1485 * have been cancelled. It's okay to cancel it now as we've
1486 * held mot_open_sem. */
1487 ldlm_lock_cancel(lease);
1490 GOTO(out_lease, rc = -EAGAIN);
1492 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1493 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1494 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1496 GOTO(out_unlock_list, rc);
1499 /* 4: lock of the object migrated object */
1500 lh_childp = &info->mti_lh[MDT_LH_OLD];
1501 mdt_lock_reg_init(lh_childp, LCK_EX);
1502 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1503 MDS_INODELOCK_LAYOUT;
1504 if (mdt_object_remote(msrcdir)) {
1505 /* Enqueue lookup lock from the parent MDT */
1506 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1507 &lh_childp->mlh_rreg_lh,
1508 lh_childp->mlh_rreg_mode,
1509 MDS_INODELOCK_LOOKUP, false);
1511 GOTO(out_unlock_list, rc);
1513 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1516 rc = mdt_object_lock(info, mold, lh_childp, lock_ibits);
1518 GOTO(out_unlock_child, rc);
1520 ma->ma_need = MA_LMV;
1522 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1523 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1524 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1526 GOTO(out_unlock_child, rc);
1528 if ((ma->ma_valid & MA_LMV)) {
1529 struct lmv_mds_md_v1 *lmm1;
1531 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1532 lmm1 = &ma->ma_lmv->lmv_md_v1;
1533 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1534 CERROR("%s: can not migrate striped dir "DFID
1535 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1536 PFID(mdt_object_fid(mold)), -EPERM);
1537 GOTO(out_unlock_child, rc = -EPERM);
1540 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1541 GOTO(out_unlock_child, rc = -EINVAL);
1543 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1544 &lmm1->lmv_stripe_fids[1]);
1546 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1548 if (!mdt_object_remote(mnew)) {
1549 CERROR("%s: "DFID" being migrated is on this MDT:"
1550 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1551 PFID(rr->rr_fid2), -EPERM);
1552 GOTO(out_put_new, rc = -EPERM);
1555 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1556 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1557 rc = mdt_remote_object_lock(info, mnew,
1558 mdt_object_fid(mnew),
1559 &lh_tgtp->mlh_rreg_lh,
1560 lh_tgtp->mlh_rreg_mode,
1561 MDS_INODELOCK_UPDATE, false);
1564 GOTO(out_put_new, rc);
1567 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1570 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1571 if (!mdt_object_remote(mnew)) {
1572 CERROR("%s: Migration "DFID" is on this MDT:"
1573 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1574 PFID(rr->rr_fid2), -EXDEV);
1575 GOTO(out_put_new, rc = -EXDEV);
1580 mdt_reint_init_ma(info, ma);
1582 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1583 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1585 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1586 mdt_object_child(mold), &rr->rr_name,
1587 mdt_object_child(mnew), ma);
1589 GOTO(out_unlock_new, rc);
1592 if (lh_tgtp != NULL)
1593 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1596 mdt_object_put(info->mti_env, mnew);
1598 mdt_object_unlock(info, mold, lh_childp, rc);
1600 mdt_unlock_list(info, &lock_list, rc);
1601 if (lease != NULL) {
1602 ldlm_reprocess_all(lease->l_resource);
1603 LDLM_LOCK_PUT(lease);
1607 up_write(&mold->mot_open_sem);
1609 mdt_object_put(info->mti_env, mold);
1611 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1613 mdt_object_put(info->mti_env, msrcdir);
1618 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1619 const struct lu_fid *fid,
1622 struct mdt_object *dir;
1626 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1630 /* check early, the real version will be saved after locking */
1631 rc = mdt_version_get_check(info, dir, idx);
1637 mdt_object_put(info->mti_env, dir);
1641 static int mdt_object_lock_save(struct mdt_thread_info *info,
1642 struct mdt_object *dir,
1643 struct mdt_lock_handle *lh,
1648 /* we lock the target dir if it is local */
1649 rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE);
1653 /* get and save correct version after locking */
1654 mdt_version_get_save(info, dir, idx);
1659 static int mdt_rename_parents_lock(struct mdt_thread_info *info,
1660 struct mdt_object **srcp,
1661 struct mdt_object **tgtp)
1663 struct mdt_reint_record *rr = &info->mti_rr;
1664 const struct lu_fid *fid_src = rr->rr_fid1;
1665 const struct lu_fid *fid_tgt = rr->rr_fid2;
1666 struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT];
1667 struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
1668 struct mdt_object *src;
1669 struct mdt_object *tgt;
1674 /* find both parents. */
1675 src = mdt_object_find_check(info, fid_src, 0);
1677 RETURN(PTR_ERR(src));
1679 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1681 if (lu_fid_eq(fid_src, fid_tgt)) {
1683 mdt_object_get(info->mti_env, tgt);
1685 /* Check if the @src is not a child of the @tgt, otherwise a
1686 * reverse locking must take place. */
1687 rc = mdt_is_subdir(info, src, fid_tgt);
1691 GOTO(err_src_put, rc);
1693 tgt = mdt_object_find_check(info, fid_tgt, 1);
1695 GOTO(err_src_put, rc = PTR_ERR(tgt));
1698 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1700 /* lock parents in the proper order. */
1702 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1704 GOTO(err_tgt_put, rc);
1706 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1708 rc = mdt_object_lock_save(info, src, lh_src, 0);
1710 rc = mdt_object_lock_save(info, src, lh_src, 0);
1712 GOTO(err_tgt_put, rc);
1714 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1717 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1718 else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
1719 rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
1720 MDS_INODELOCK_UPDATE);
1721 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1725 GOTO(err_unlock, rc);
1727 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1734 /* The order does not matter as the handle is checked inside,
1735 * as well as not used handle. */
1736 mdt_object_unlock(info, src, lh_src, rc);
1737 mdt_object_unlock(info, tgt, lh_tgt, rc);
1739 mdt_object_put(info->mti_env, tgt);
1741 mdt_object_put(info->mti_env, src);
1746 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1747 * 2 - src child; 3 - tgt child.
1748 * Update on disk version of src child.
1751 * For DNE phase I, only these renames are allowed
1752 * mv src_p/src_c tgt_p/tgt_c
1753 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1754 * 2. src_p and tgt_p are same directory, and tgt_c does not
1755 * exists. In this case, all of modification will happen
1756 * in the MDT where ithesource parent is, only one remote
1757 * update is needed, i.e. set c_time/m_time on the child.
1758 * And tgt_c will be still in the same MDT as the original
1761 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1762 struct mdt_lock_handle *lhc)
1764 struct mdt_reint_record *rr = &info->mti_rr;
1765 struct md_attr *ma = &info->mti_attr;
1766 struct ptlrpc_request *req = mdt_info_req(info);
1767 struct mdt_object *msrcdir = NULL;
1768 struct mdt_object *mtgtdir = NULL;
1769 struct mdt_object *mold;
1770 struct mdt_object *mnew = NULL;
1771 struct mdt_lock_handle *lh_srcdirp;
1772 struct mdt_lock_handle *lh_tgtdirp;
1773 struct mdt_lock_handle *lh_oldp = NULL;
1774 struct mdt_lock_handle *lh_newp = NULL;
1775 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1776 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1781 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1782 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1783 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1785 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1786 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1787 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1788 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1790 /* step 1&2: lock the source and target dirs. */
1791 rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
1795 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1797 /* step 3: find & lock the old object. */
1799 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1801 GOTO(out_unlock_parents, rc);
1803 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1804 GOTO(out_unlock_parents, rc = -EINVAL);
1806 if (!fid_is_md_operative(old_fid))
1807 GOTO(out_unlock_parents, rc = -EPERM);
1809 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1811 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1813 /* Check if @mtgtdir is subdir of @mold, before locking child
1814 * to avoid reverse locking. */
1815 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1817 GOTO(out_put_old, rc);
1819 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1820 /* save version after locking */
1821 mdt_version_get_save(info, mold, 2);
1823 /* step 4: find & lock the new object. */
1824 /* new target object may not exist now */
1825 /* lookup with version checking */
1827 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1830 /* the new_fid should have been filled at this moment */
1831 if (lu_fid_eq(old_fid, new_fid))
1832 GOTO(out_put_old, rc);
1834 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1835 lu_fid_eq(new_fid, rr->rr_fid2))
1836 GOTO(out_put_old, rc = -EINVAL);
1838 if (!fid_is_md_operative(new_fid))
1839 GOTO(out_put_old, rc = -EPERM);
1841 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1843 GOTO(out_put_old, rc = PTR_ERR(mnew));
1845 if (mdt_object_remote(mnew)) {
1846 struct mdt_body *repbody;
1848 /* Always send rename req to the target child MDT */
1849 repbody = req_capsule_server_get(info->mti_pill,
1851 LASSERT(repbody != NULL);
1852 repbody->mbo_fid1 = *new_fid;
1853 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1854 GOTO(out_put_old, rc = -EXDEV);
1856 /* Before locking the target dir, check we do not replace
1857 * a dir with a non-dir, otherwise it may deadlock with
1858 * link op which tries to create a link in this dir
1859 * back to this non-dir. */
1860 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
1861 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
1862 GOTO(out_put_new, rc = -EISDIR);
1864 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1865 mdt_lock_reg_init(lh_oldp, LCK_EX);
1867 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
1868 if (mdt_object_remote(msrcdir)) {
1869 /* Enqueue lookup lock from the parent MDT */
1870 rc = mdt_remote_object_lock(info, msrcdir,
1871 mdt_object_fid(mold),
1872 &lh_oldp->mlh_rreg_lh,
1873 lh_oldp->mlh_rreg_mode,
1874 MDS_INODELOCK_LOOKUP,
1877 GOTO(out_put_new, rc);
1879 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1882 rc = mdt_object_lock(info, mold, lh_oldp, lock_ibits);
1884 GOTO(out_unlock_old, rc);
1886 /* Check if @msrcdir is subdir of @mnew, before locking child
1887 * to avoid reverse locking. */
1888 rc = mdt_is_subdir(info, msrcdir, new_fid);
1890 GOTO(out_unlock_old, rc);
1892 /* We used to acquire MDS_INODELOCK_FULL here but we
1893 * can't do this now because a running HSM restore on
1894 * the rename onto victim will hold the layout
1895 * lock. See LU-4002. */
1897 lh_newp = &info->mti_lh[MDT_LH_NEW];
1898 mdt_lock_reg_init(lh_newp, LCK_EX);
1899 rc = mdt_object_lock(info, mnew, lh_newp,
1900 MDS_INODELOCK_LOOKUP |
1901 MDS_INODELOCK_UPDATE);
1903 GOTO(out_unlock_old, rc);
1905 /* get and save version after locking */
1906 mdt_version_get_save(info, mnew, 3);
1907 } else if (rc != -EREMOTE && rc != -ENOENT) {
1908 GOTO(out_put_old, rc);
1910 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1911 mdt_lock_reg_init(lh_oldp, LCK_EX);
1913 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
1914 if (mdt_object_remote(msrcdir)) {
1915 /* Enqueue lookup lock from the parent MDT */
1916 rc = mdt_remote_object_lock(info, msrcdir,
1917 mdt_object_fid(mold),
1918 &lh_oldp->mlh_rreg_lh,
1919 lh_oldp->mlh_rreg_mode,
1920 MDS_INODELOCK_LOOKUP,
1923 GOTO(out_put_new, rc);
1925 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1928 rc = mdt_object_lock(info, mold, lh_oldp, lock_ibits);
1930 GOTO(out_put_old, rc);
1932 mdt_enoent_version_save(info, 3);
1935 /* step 5: rename it */
1936 mdt_reint_init_ma(info, ma);
1938 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1939 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1942 mutex_lock(&mnew->mot_lov_mutex);
1944 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1945 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1946 mnew != NULL ? mdt_object_child(mnew) : NULL,
1947 &rr->rr_tgt_name, ma);
1950 mutex_unlock(&mnew->mot_lov_mutex);
1952 /* handle last link of tgt object */
1954 mdt_counter_incr(req, LPROC_MDT_RENAME);
1956 mdt_handle_last_unlink(info, mnew, ma);
1958 mdt_rename_counter_tally(info, info->mti_mdt, req,
1964 mdt_object_unlock(info, mnew, lh_newp, rc);
1966 mdt_object_unlock(info, mold, lh_oldp, rc);
1969 mdt_object_put(info->mti_env, mnew);
1971 mdt_object_put(info->mti_env, mold);
1973 mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc);
1974 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1978 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1979 struct mdt_lock_handle *lhc, bool rename)
1981 struct mdt_reint_record *rr = &info->mti_rr;
1982 struct ptlrpc_request *req = mdt_info_req(info);
1983 struct lustre_handle rename_lh = { 0 };
1987 if (info->mti_dlm_req)
1988 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1990 if (!fid_is_md_operative(rr->rr_fid1) ||
1991 !fid_is_md_operative(rr->rr_fid2))
1994 /* Note: do not enqueue rename lock for replay request, because
1995 * if other MDT holds rename lock, but being blocked to wait for
1996 * this MDT to finish its recovery, and the failover MDT can not
1997 * get rename lock, which will cause deadlock. */
1998 if (!req_is_replay(req)) {
1999 rc = mdt_rename_lock(info, &rename_lh);
2001 CERROR("%s: can't lock FS for rename: rc = %d\n",
2002 mdt_obd_name(info->mti_mdt), rc);
2008 rc = mdt_reint_rename_internal(info, lhc);
2010 rc = mdt_reint_migrate_internal(info, lhc);
2012 if (lustre_handle_is_used(&rename_lh))
2013 mdt_rename_unlock(&rename_lh);
2018 static int mdt_reint_rename(struct mdt_thread_info *info,
2019 struct mdt_lock_handle *lhc)
2021 return mdt_reint_rename_or_migrate(info, lhc, true);
2024 static int mdt_reint_migrate(struct mdt_thread_info *info,
2025 struct mdt_lock_handle *lhc)
2027 return mdt_reint_rename_or_migrate(info, lhc, false);
2030 struct mdt_reinter {
2031 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2032 enum lprocfs_extra_opc mr_extra_opc;
2035 static const struct mdt_reinter mdt_reinters[] = {
2037 .mr_handler = &mdt_reint_setattr,
2038 .mr_extra_opc = MDS_REINT_SETATTR,
2041 .mr_handler = &mdt_reint_create,
2042 .mr_extra_opc = MDS_REINT_CREATE,
2045 .mr_handler = &mdt_reint_link,
2046 .mr_extra_opc = MDS_REINT_LINK,
2049 .mr_handler = &mdt_reint_unlink,
2050 .mr_extra_opc = MDS_REINT_UNLINK,
2053 .mr_handler = &mdt_reint_rename,
2054 .mr_extra_opc = MDS_REINT_RENAME,
2057 .mr_handler = &mdt_reint_open,
2058 .mr_extra_opc = MDS_REINT_OPEN,
2060 [REINT_SETXATTR] = {
2061 .mr_handler = &mdt_reint_setxattr,
2062 .mr_extra_opc = MDS_REINT_SETXATTR,
2065 .mr_handler = &mdt_reint_unlink,
2066 .mr_extra_opc = MDS_REINT_UNLINK,
2069 .mr_handler = &mdt_reint_migrate,
2070 .mr_extra_opc = MDS_REINT_RENAME,
2074 int mdt_reint_rec(struct mdt_thread_info *info,
2075 struct mdt_lock_handle *lhc)
2077 const struct mdt_reinter *mr;
2081 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2084 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2085 if (mr->mr_handler == NULL)
2088 rc = (*mr->mr_handler)(info, lhc);
2090 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2091 PTLRPC_LAST_CNTR + mr->mr_extra_opc);