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, 2013, 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 "mdt_internal.h"
50 #include <lustre_lmv.h>
52 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
55 ma->ma_need = MA_INODE;
59 static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
60 struct mdt_object *object,
61 struct mdt_body *repbody)
65 /* for cross-ref mkdir, mds capa has been fetched from remote obj, then
66 * we won't go to below*/
67 if (repbody->mbo_valid & OBD_MD_FLMDSCAPA)
70 if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa &&
71 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
72 struct lustre_capa *capa;
74 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
76 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
77 rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa,
80 repbody->mbo_valid |= OBD_MD_FLMDSCAPA;
87 * Get version of object by fid.
89 * Return real version or ENOENT_VERSION if object doesn't exist
91 static void mdt_obj_version_get(struct mdt_thread_info *info,
92 struct mdt_object *o, __u64 *version)
95 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
96 !fid_is_obf(mdt_object_fid(o)))
97 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
99 *version = ENOENT_VERSION;
100 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
101 PFID(mdt_object_fid(o)), *version);
105 * Check version is correct.
107 * Should be called only during replay.
109 static int mdt_version_check(struct ptlrpc_request *req,
110 __u64 version, int idx)
112 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
115 if (!exp_connect_vbr(req->rq_export))
118 LASSERT(req_is_replay(req));
119 /** VBR: version is checked always because costs nothing */
120 LASSERT(idx < PTLRPC_NUM_VERSIONS);
121 /** Sanity check for malformed buffers */
122 if (pre_ver == NULL) {
123 CERROR("No versions in request buffer\n");
124 spin_lock(&req->rq_export->exp_lock);
125 req->rq_export->exp_vbr_failed = 1;
126 spin_unlock(&req->rq_export->exp_lock);
128 } else if (pre_ver[idx] != version) {
129 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
130 pre_ver[idx], version);
131 spin_lock(&req->rq_export->exp_lock);
132 req->rq_export->exp_vbr_failed = 1;
133 spin_unlock(&req->rq_export->exp_lock);
140 * Save pre-versions in reply.
142 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
147 if (!exp_connect_vbr(req->rq_export))
150 LASSERT(!req_is_replay(req));
151 LASSERT(req->rq_repmsg != NULL);
152 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
154 reply_ver[idx] = version;
158 * Save enoent version, it is needed when it is obvious that object doesn't
159 * exist, e.g. child during create.
161 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
163 /* save version of file name for replay, it must be ENOENT here */
164 if (!req_is_replay(mdt_info_req(info))) {
165 info->mti_ver[idx] = ENOENT_VERSION;
166 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and save in reply buffer.
173 * Versions are saved in reply only during normal operations not replays.
175 void mdt_version_get_save(struct mdt_thread_info *info,
176 struct mdt_object *mto, int idx)
178 /* don't save versions during replay */
179 if (!req_is_replay(mdt_info_req(info))) {
180 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
181 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
186 * Get version from disk and check it, no save in reply.
188 int mdt_version_get_check(struct mdt_thread_info *info,
189 struct mdt_object *mto, int idx)
191 /* only check versions during replay */
192 if (!req_is_replay(mdt_info_req(info)))
195 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
196 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
200 * Get version from disk and check if recovery or just save.
202 int mdt_version_get_check_save(struct mdt_thread_info *info,
203 struct mdt_object *mto, int idx)
207 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
208 if (req_is_replay(mdt_info_req(info)))
209 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
212 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
217 * Lookup with version checking.
219 * This checks version of 'name'. Many reint functions uses 'name' for child not
220 * FID, therefore we need to get object by name and check its version.
222 int mdt_lookup_version_check(struct mdt_thread_info *info,
223 struct mdt_object *p, const struct lu_name *lname,
224 struct lu_fid *fid, int idx)
228 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
230 /* Check version only during replay */
231 if (!req_is_replay(mdt_info_req(info)))
234 info->mti_ver[idx] = ENOENT_VERSION;
236 struct mdt_object *child;
237 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
238 if (likely(!IS_ERR(child))) {
239 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
240 mdt_object_put(info->mti_env, child);
243 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
244 return vbrc ? vbrc : rc;
249 * mdt_remote_permission: Check whether the remote operation is permitted,
251 * Before we implement async cross-MDT updates (DNE phase 2). There are a few
254 * 1.Only sysadmin can create remote directory and striped directory and
255 * migrate directory now, unless
256 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
257 * 2.Remote directory can only be created on MDT0, unless
258 * lctl set_param mdt.*.enable_remote_dir = 1
259 * 3.Only new clients can access remote dir( >= 2.4) and striped dir(>= 2.6),
260 * old client will return -ENOTSUPP.
262 * XXX these check are only needed for remote synchronization, once async
263 * update is supported, these check will be removed.
265 * param[in]info: execution environment.
266 * param[in]parent: the directory of this operation.
267 * param[in]child: the child of this operation.
269 * retval = 0 remote operation is allowed.
270 * < 0 remote operation is denied.
272 static int mdt_remote_permission(struct mdt_thread_info *info,
273 struct mdt_object *parent,
274 struct mdt_object *child)
276 struct mdt_device *mdt = info->mti_mdt;
277 struct lu_ucred *uc = mdt_ucred(info);
278 struct md_op_spec *spec = &info->mti_spec;
279 struct lu_attr *attr = &info->mti_attr.ma_attr;
280 struct obd_export *exp = mdt_info_req(info)->rq_export;
282 /* Only check create remote directory, striped directory and
284 if (mdt_object_remote(parent) == 0 && mdt_object_remote(child) == 0 &&
285 !(S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
286 spec->u.sp_ea.eadatalen != 0) &&
287 info->mti_rr.rr_opcode != REINT_MIGRATE)
290 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
291 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
292 mdt->mdt_enable_remote_dir_gid != -1)
296 if (mdt->mdt_enable_remote_dir == 0) {
297 struct seq_server_site *ss = mdt_seq_site(mdt);
298 struct lu_seq_range range = { 0 };
301 fld_range_set_type(&range, LU_SEQ_RANGE_MDT);
302 rc = fld_server_lookup(info->mti_env, ss->ss_server_fld,
303 fid_seq(mdt_object_fid(parent)), &range);
307 if (range.lsr_index != 0)
311 if (!mdt_is_dne_client(exp))
314 if (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
315 spec->u.sp_ea.eadatalen != 0 && !mdt_is_striped_client(exp))
322 * VBR: we save three versions in reply:
323 * 0 - parent. Check that parent version is the same during replay.
324 * 1 - name. Version of 'name' if file exists with the same name or
325 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
326 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
329 static int mdt_md_create(struct mdt_thread_info *info)
331 struct mdt_device *mdt = info->mti_mdt;
332 struct mdt_object *parent;
333 struct mdt_object *child;
334 struct mdt_lock_handle *lh;
335 struct mdt_body *repbody;
336 struct md_attr *ma = &info->mti_attr;
337 struct mdt_reint_record *rr = &info->mti_rr;
341 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
343 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
345 if (!fid_is_md_operative(rr->rr_fid1))
348 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
350 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
352 RETURN(PTR_ERR(parent));
354 if (!mdt_object_exists(parent))
355 GOTO(put_parent, rc = -ENOENT);
357 lh = &info->mti_lh[MDT_LH_PARENT];
358 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
359 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE,
362 GOTO(put_parent, rc);
364 if (!mdt_object_remote(parent)) {
365 rc = mdt_version_get_check_save(info, parent, 0);
367 GOTO(unlock_parent, rc);
371 * Check child name version during replay.
372 * During create replay a file may exist with same name.
374 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
375 &info->mti_tmp_fid1, 1);
377 GOTO(unlock_parent, rc = -EEXIST);
379 /* -ENOENT is expected here */
381 GOTO(unlock_parent, rc);
383 /* save version of file name for replay, it must be ENOENT here */
384 mdt_enoent_version_save(info, 1);
386 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
387 if (likely(!IS_ERR(child))) {
388 struct md_object *next = mdt_object_child(parent);
390 rc = mdt_remote_permission(info, parent, child);
392 GOTO(out_put_child, rc);
394 ma->ma_need = MA_INODE;
396 /* capa for cross-ref will be stored here */
397 ma->ma_capa = req_capsule_server_get(info->mti_pill,
399 LASSERT(ma->ma_capa);
401 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
402 OBD_FAIL_MDS_REINT_CREATE_WRITE);
404 /* Version of child will be updated on disk. */
405 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
406 rc = mdt_version_get_check_save(info, child, 2);
408 GOTO(out_put_child, rc);
410 /* Let lower layer know current lock mode. */
411 info->mti_spec.sp_cr_mode =
412 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
415 * Do not perform lookup sanity check. We know that name does
418 info->mti_spec.sp_cr_lookup = 0;
419 info->mti_spec.sp_feat = &dt_directory_features;
421 rc = mdo_create(info->mti_env, next, &rr->rr_name,
422 mdt_object_child(child), &info->mti_spec, ma);
424 rc = mdt_attr_get_complex(info, child, ma);
427 /* Return fid & attr to client. */
428 if (ma->ma_valid & MA_INODE)
429 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
430 mdt_object_fid(child));
433 mdt_create_pack_capa(info, rc, child, repbody);
434 mdt_object_put(info->mti_env, child);
437 mdt_create_pack_capa(info, rc, NULL, repbody);
440 mdt_object_unlock(info, parent, lh, rc);
442 mdt_object_put(info->mti_env, parent);
446 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
447 struct mdt_object *obj, __u64 ibits,
448 struct mdt_lock_handle *s0_lh,
449 struct mdt_object *s0_obj,
450 struct ldlm_enqueue_info *einfo)
452 ldlm_policy_data_t *policy = &mti->mti_policy;
456 if (!S_ISDIR(obj->mot_header.loh_attr))
459 /* Unlock stripe 0 */
460 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
461 LASSERT(s0_obj != NULL);
462 mdt_object_unlock_put(mti, s0_obj, s0_lh, 1);
465 memset(policy, 0, sizeof(*policy));
466 policy->l_inodebits.bits = ibits;
468 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
474 * Lock slave stripes if necessary, the lock handles of slave stripes
475 * will be stored in einfo->ei_cbdata.
477 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
478 ldlm_mode_t mode, __u64 ibits,
479 struct mdt_lock_handle *s0_lh,
480 struct mdt_object **s0_objp,
481 struct ldlm_enqueue_info *einfo)
483 ldlm_policy_data_t *policy = &mti->mti_policy;
484 struct lu_buf *buf = &mti->mti_buf;
485 struct lmv_mds_md_v1 *lmv;
486 struct lu_fid *fid = &mti->mti_tmp_fid1;
490 if (!S_ISDIR(obj->mot_header.loh_attr))
493 buf->lb_buf = mti->mti_xattr_buf;
494 buf->lb_len = sizeof(mti->mti_xattr_buf);
495 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
498 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
500 buf->lb_buf = mti->mti_big_lmm;
501 buf->lb_len = mti->mti_big_lmmsize;
505 if (rc == -ENODATA || rc == -ENOENT)
512 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
515 /* Sigh, 0_stripe and master object are different
516 * object, though they are in the same MDT, to avoid
517 * adding osd_object_lock here, so we will enqueue the
518 * stripe0 lock in MDT0 for now */
519 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
520 *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits);
521 if (IS_ERR(*s0_objp))
522 RETURN(PTR_ERR(*s0_objp));
524 memset(einfo, 0, sizeof(*einfo));
525 einfo->ei_type = LDLM_IBITS;
526 einfo->ei_mode = mode;
527 einfo->ei_cb_bl = mdt_remote_blocking_ast;
528 einfo->ei_cb_cp = ldlm_completion_ast;
529 einfo->ei_enq_slave = 1;
530 memset(policy, 0, sizeof(*policy));
531 policy->l_inodebits.bits = ibits;
533 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
538 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
539 struct md_attr *ma, int flags)
541 struct mdt_lock_handle *lh;
542 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
543 __u64 lockpart = MDS_INODELOCK_UPDATE;
544 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
545 struct mdt_lock_handle *s0_lh;
546 struct mdt_object *s0_obj = NULL;
550 /* attr shouldn't be set on remote object */
551 LASSERT(!mdt_object_remote(mo));
553 lh = &info->mti_lh[MDT_LH_PARENT];
554 mdt_lock_reg_init(lh, LCK_PW);
556 /* Even though the new MDT will grant PERM lock to the old
557 * client, but the old client will almost ignore that during
558 * So it needs to revoke both LOOKUP and PERM lock here, so
559 * both new and old client can cancel the dcache */
560 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
561 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
563 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
567 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
568 mdt_lock_reg_init(s0_lh, LCK_EX);
569 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
571 GOTO(out_unlock, rc);
573 if (mdt_object_exists(mo) == 0)
574 GOTO(out_unlock, rc = -ENOENT);
576 /* all attrs are packed into mti_attr in unpack_setattr */
577 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
578 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
580 /* This is only for set ctime when rename's source is on remote MDS. */
581 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
582 ma->ma_attr_flags |= MDS_VTX_BYPASS;
584 /* VBR: update version if attr changed are important for recovery */
586 /* update on-disk version of changed object */
587 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
588 rc = mdt_version_get_check_save(info, mo, 0);
590 GOTO(out_unlock, rc);
593 /* Ensure constant striping during chown(). See LU-2789. */
594 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
595 mutex_lock(&mo->mot_lov_mutex);
597 /* all attrs are packed into mti_attr in unpack_setattr */
598 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
600 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
601 mutex_unlock(&mo->mot_lov_mutex);
604 GOTO(out_unlock, rc);
608 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
609 mdt_object_unlock(info, mo, lh, rc);
614 * Check HSM flags and add HS_DIRTY flag if relevant.
616 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
617 * and is not RELEASED.
619 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
625 /* If the file was modified, add the dirty flag */
626 ma->ma_need = MA_HSM;
627 rc = mdt_attr_get_complex(info, mo, ma);
629 CERROR("file attribute read error for "DFID": %d.\n",
630 PFID(mdt_object_fid(mo)), rc);
634 /* If an up2date copy exists in the backend, add dirty flag */
635 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
636 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
637 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
639 ma->ma_hsm.mh_flags |= HS_DIRTY;
641 mdt_lock_reg_init(lh, LCK_PW);
642 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
647 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
649 CERROR("file attribute change error for "DFID": %d\n",
650 PFID(mdt_object_fid(mo)), rc);
651 mdt_object_unlock(info, mo, lh, rc);
657 static int mdt_reint_setattr(struct mdt_thread_info *info,
658 struct mdt_lock_handle *lhc)
660 struct md_attr *ma = &info->mti_attr;
661 struct mdt_reint_record *rr = &info->mti_rr;
662 struct ptlrpc_request *req = mdt_info_req(info);
663 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
664 struct mdt_file_data *mfd;
665 struct mdt_object *mo;
666 struct mdt_body *repbody;
670 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
671 (unsigned int)ma->ma_attr.la_valid);
673 if (info->mti_dlm_req)
674 ldlm_request_cancel(req, info->mti_dlm_req, 0);
676 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
677 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
679 GOTO(out, rc = PTR_ERR(mo));
681 /* start a log jounal handle if needed */
682 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
683 if ((ma->ma_attr.la_valid & LA_SIZE) ||
684 (rr->rr_flags & MRF_OPEN_TRUNC)) {
685 /* Check write access for the O_TRUNC case */
686 if (mdt_write_read(mo) < 0)
687 GOTO(out_put, rc = -ETXTBSY);
689 } else if (info->mti_ioepoch &&
690 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
691 /* Truncate case. IOEpoch is opened. */
692 rc = mdt_write_get(mo);
696 mfd = mdt_mfd_new(med);
699 GOTO(out_put, rc = -ENOMEM);
702 mdt_ioepoch_open(info, mo, 0);
703 repbody->mbo_ioepoch = mo->mot_ioepoch;
705 mdt_object_get(info->mti_env, mo);
706 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
707 mfd->mfd_object = mo;
708 mfd->mfd_xid = req->rq_xid;
710 spin_lock(&med->med_open_lock);
711 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
712 spin_unlock(&med->med_open_lock);
713 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
716 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
718 /* SOM Attribute update case. Find the proper mfd and update
719 * SOM attributes on the proper object. */
720 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
721 LASSERT(info->mti_ioepoch);
723 spin_lock(&med->med_open_lock);
724 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
727 spin_unlock(&med->med_open_lock);
728 CDEBUG(D_INODE, "no handle for file close: "
729 "fid = "DFID": cookie = "LPX64"\n",
730 PFID(info->mti_rr.rr_fid1),
731 info->mti_ioepoch->handle.cookie);
732 GOTO(out_put, rc = -ESTALE);
734 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
735 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
737 class_handle_unhash(&mfd->mfd_handle);
738 cfs_list_del_init(&mfd->mfd_list);
739 spin_unlock(&med->med_open_lock);
741 mdt_mfd_close(info, mfd);
742 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
743 LASSERT((ma->ma_valid & MA_LOV) == 0);
744 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
747 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
748 struct lu_buf *buf = &info->mti_buf;
749 LASSERT(ma->ma_attr.la_valid == 0);
750 buf->lb_buf = ma->ma_lmm;
751 buf->lb_len = ma->ma_lmm_size;
752 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
753 buf, XATTR_NAME_LOV, 0);
756 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
757 struct lu_buf *buf = &info->mti_buf;
759 LASSERT(ma->ma_attr.la_valid == 0);
760 buf->lb_buf = ma->ma_lmv;
761 buf->lb_len = ma->ma_lmv_size;
762 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
763 buf, XATTR_NAME_DEFAULT_LMV, 0);
769 /* If file data is modified, add the dirty flag */
770 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
771 rc = mdt_add_dirty_flag(info, mo, ma);
773 ma->ma_need = MA_INODE;
775 rc = mdt_attr_get_complex(info, mo, ma);
779 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
781 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
782 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
783 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
784 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
785 struct lustre_capa *capa;
787 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
789 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
790 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
793 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
798 mdt_object_put(info->mti_env, mo);
801 mdt_counter_incr(req, LPROC_MDT_SETATTR);
803 mdt_client_compatibility(info);
804 rc2 = mdt_fix_reply(info);
810 static int mdt_reint_create(struct mdt_thread_info *info,
811 struct mdt_lock_handle *lhc)
813 struct ptlrpc_request *req = mdt_info_req(info);
817 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
818 RETURN(err_serious(-ESTALE));
820 if (info->mti_dlm_req)
821 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
823 if (!lu_name_is_valid(&info->mti_rr.rr_name))
826 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
828 mdt_counter_incr(req, LPROC_MDT_MKDIR);
836 /* Special file should stay on the same node as parent. */
837 mdt_counter_incr(req, LPROC_MDT_MKNOD);
840 CERROR("%s: Unsupported mode %o\n",
841 mdt_obd_name(info->mti_mdt),
842 info->mti_attr.ma_attr.la_mode);
843 RETURN(err_serious(-EOPNOTSUPP));
846 rc = mdt_md_create(info);
851 * VBR: save parent version in reply and child version getting by its name.
852 * Version of child is getting and checking during its lookup. If
854 static int mdt_reint_unlink(struct mdt_thread_info *info,
855 struct mdt_lock_handle *lhc)
857 struct mdt_reint_record *rr = &info->mti_rr;
858 struct ptlrpc_request *req = mdt_info_req(info);
859 struct md_attr *ma = &info->mti_attr;
860 struct lu_fid *child_fid = &info->mti_tmp_fid1;
861 struct mdt_object *mp;
862 struct mdt_object *mc;
863 struct mdt_lock_handle *parent_lh;
864 struct mdt_lock_handle *child_lh;
865 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
866 struct mdt_lock_handle *s0_lh = NULL;
867 struct mdt_object *s0_obj = NULL;
872 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
873 PNAME(&rr->rr_name));
875 if (info->mti_dlm_req)
876 ldlm_request_cancel(req, info->mti_dlm_req, 0);
878 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
879 RETURN(err_serious(-ENOENT));
881 if (!fid_is_md_operative(rr->rr_fid1))
885 * step 1: Found the parent.
887 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
893 parent_lh = &info->mti_lh[MDT_LH_PARENT];
894 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
895 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
898 GOTO(put_parent, rc);
900 if (!mdt_object_remote(mp)) {
901 rc = mdt_version_get_check_save(info, mp, 0);
903 GOTO(unlock_parent, rc);
906 /* step 2: find & lock the child */
907 /* lookup child object along with version checking */
909 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
911 /* Name might not be able to find during resend of
912 * remote unlink, considering following case.
913 * dir_A is a remote directory, the name entry of
914 * dir_A is on MDT0, the directory is on MDT1,
916 * 1. client sends unlink req to MDT1.
917 * 2. MDT1 sends name delete update to MDT0.
918 * 3. name entry is being deleted in MDT0 synchronously.
919 * 4. MDT1 is restarted.
920 * 5. client resends unlink req to MDT1. So it can not
921 * find the name entry on MDT0 anymore.
922 * In this case, MDT1 only needs to destory the local
925 if (mdt_object_remote(mp) && rc == -ENOENT &&
926 !fid_is_zero(rr->rr_fid2) &&
927 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
929 *child_fid = *rr->rr_fid2;
931 GOTO(unlock_parent, rc);
935 if (!fid_is_md_operative(child_fid))
936 GOTO(unlock_parent, rc = -EPERM);
938 /* We will lock the child regardless it is local or remote. No harm. */
939 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
941 GOTO(unlock_parent, rc = PTR_ERR(mc));
943 child_lh = &info->mti_lh[MDT_LH_CHILD];
944 mdt_lock_reg_init(child_lh, LCK_EX);
945 if (mdt_object_remote(mc)) {
946 struct mdt_body *repbody;
948 if (!fid_is_zero(rr->rr_fid2)) {
949 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
950 mdt_obd_name(info->mti_mdt),
951 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
952 GOTO(put_child, rc = -ENOENT);
954 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
955 mdt_obd_name(info->mti_mdt),
956 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
958 if (!mdt_is_dne_client(req->rq_export))
959 /* Return -ENOTSUPP for old client */
960 GOTO(put_child, rc = -ENOTSUPP);
962 if (info->mti_spec.sp_rm_entry) {
963 struct lu_ucred *uc = mdt_ucred(info);
965 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
966 CERROR("%s: unlink remote entry is only "
967 "permitted for administrator: rc = %d\n",
968 mdt_obd_name(info->mti_mdt),
970 GOTO(put_child, rc = -EPERM);
973 ma->ma_need = MA_INODE;
975 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
976 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
977 NULL, &rr->rr_name, ma, no_name);
980 /* Revoke the LOOKUP lock of the remote object granted by
981 * this MDT. Since the unlink will happen on another MDT,
982 * it will release the LOOKUP lock right away. Then What
983 * would happen if another client try to grab the LOOKUP
984 * lock at the same time with unlink XXX */
985 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
987 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
988 LASSERT(repbody != NULL);
989 repbody->mbo_fid1 = *mdt_object_fid(mc);
990 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
991 GOTO(unlock_child, rc = -EREMOTE);
992 } else if (info->mti_spec.sp_rm_entry) {
994 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
996 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
1000 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1001 * this now because a running HSM restore on the child (unlink
1002 * victim) will hold the layout lock. See LU-4002. */
1003 rc = mdt_object_lock(info, mc, child_lh,
1004 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
1007 GOTO(put_child, rc);
1009 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1010 * whether need MA_LOV and MA_COOKIE.
1012 ma->ma_need = MA_INODE;
1015 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1016 mdt_lock_reg_init(s0_lh, LCK_EX);
1017 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
1020 GOTO(unlock_child, rc);
1022 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1023 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1024 /* save version when object is locked */
1025 mdt_version_get_save(info, mc, 1);
1027 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1029 mutex_lock(&mc->mot_lov_mutex);
1031 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1032 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1034 mutex_unlock(&mc->mot_lov_mutex);
1036 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1037 rc = mdt_attr_get_complex(info, mc, ma);
1039 mdt_handle_last_unlink(info, mc, ma);
1041 if (ma->ma_valid & MA_INODE) {
1042 switch (ma->ma_attr.la_mode & S_IFMT) {
1044 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1052 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1055 LASSERTF(0, "bad file type %o unlinking\n",
1056 ma->ma_attr.la_mode);
1063 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
1064 mdt_object_unlock(info, mc, child_lh, rc);
1066 mdt_object_put(info->mti_env, mc);
1068 mdt_object_unlock(info, mp, parent_lh, rc);
1070 mdt_object_put(info->mti_env, mp);
1076 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1079 static int mdt_reint_link(struct mdt_thread_info *info,
1080 struct mdt_lock_handle *lhc)
1082 struct mdt_reint_record *rr = &info->mti_rr;
1083 struct ptlrpc_request *req = mdt_info_req(info);
1084 struct md_attr *ma = &info->mti_attr;
1085 struct mdt_object *ms;
1086 struct mdt_object *mp;
1087 struct mdt_lock_handle *lhs;
1088 struct mdt_lock_handle *lhp;
1092 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1093 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1095 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1096 RETURN(err_serious(-ENOENT));
1098 if (info->mti_dlm_req)
1099 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1101 /* Invalid case so return error immediately instead of
1103 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1106 if (!fid_is_md_operative(rr->rr_fid1) ||
1107 !fid_is_md_operative(rr->rr_fid2))
1110 /* step 1: find & lock the target parent dir */
1111 lhp = &info->mti_lh[MDT_LH_PARENT];
1112 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1113 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1114 MDS_INODELOCK_UPDATE);
1116 RETURN(PTR_ERR(mp));
1118 rc = mdt_version_get_check_save(info, mp, 0);
1120 GOTO(out_unlock_parent, rc);
1122 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1124 /* step 2: find & lock the source */
1125 lhs = &info->mti_lh[MDT_LH_CHILD];
1126 mdt_lock_reg_init(lhs, LCK_EX);
1128 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1130 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1132 if (!mdt_object_exists(ms)) {
1133 mdt_object_put(info->mti_env, ms);
1134 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1135 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1136 GOTO(out_unlock_parent, rc = -ENOENT);
1139 if (mdt_object_remote(ms)) {
1140 mdt_object_put(info->mti_env, ms);
1141 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1142 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1144 GOTO(out_unlock_parent, rc = -EXDEV);
1147 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1148 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1150 mdt_object_put(info->mti_env, ms);
1151 GOTO(out_unlock_parent, rc);
1154 /* step 3: link it */
1155 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1156 OBD_FAIL_MDS_REINT_LINK_WRITE);
1158 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1159 rc = mdt_version_get_check_save(info, ms, 1);
1161 GOTO(out_unlock_child, rc);
1163 /** check target version by name during replay */
1164 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1165 &info->mti_tmp_fid1, 2);
1166 if (rc != 0 && rc != -ENOENT)
1167 GOTO(out_unlock_child, rc);
1168 /* save version of file name for replay, it must be ENOENT here */
1169 if (!req_is_replay(mdt_info_req(info))) {
1170 if (rc != -ENOENT) {
1171 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1172 PNAME(&rr->rr_name));
1173 GOTO(out_unlock_child, rc = -EEXIST);
1175 info->mti_ver[2] = ENOENT_VERSION;
1176 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1179 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1180 mdt_object_child(ms), &rr->rr_name, ma);
1183 mdt_counter_incr(req, LPROC_MDT_LINK);
1187 mdt_object_unlock_put(info, ms, lhs, rc);
1189 mdt_object_unlock_put(info, mp, lhp, rc);
1193 * lock the part of the directory according to the hash of the name
1194 * (lh->mlh_pdo_hash) in parallel directory lock.
1196 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1197 struct mdt_lock_handle *lh,
1198 struct mdt_object *obj, __u64 ibits)
1200 struct ldlm_res_id *res = &info->mti_res_id;
1201 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1202 ldlm_policy_data_t *policy = &info->mti_policy;
1206 * Finish res_id initializing by name hash marking part of
1207 * directory which is taking modification.
1209 LASSERT(lh->mlh_pdo_hash != 0);
1210 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1211 memset(policy, 0, sizeof(*policy));
1212 policy->l_inodebits.bits = ibits;
1214 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1215 * going to be sent to client. If it is - mdt_intent_policy() path will
1216 * fix it up and turn FL_LOCAL flag off.
1218 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1219 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1220 &info->mti_exp->exp_handle.h_cookie);
1224 enum mdt_rename_lock {
1230 * Get BFL lock for rename or migrate process, right now, it does not support
1231 * cross-MDT rename, so we only need global rename lock during migration.
1233 static int mdt_rename_lock(struct mdt_thread_info *info,
1234 struct lustre_handle *lh,
1235 enum mdt_rename_lock rename_lock)
1237 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1238 ldlm_policy_data_t *policy = &info->mti_policy;
1239 struct ldlm_res_id *res_id = &info->mti_res_id;
1244 /* XXX only do global rename lock for migration */
1245 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 &&
1246 rename_lock == MRL_MIGRATE) {
1247 struct lu_fid *fid = &info->mti_tmp_fid1;
1248 struct mdt_object *obj;
1250 /* XXX, right now, it has to use object API to
1251 * enqueue lock cross MDT, so it will enqueue
1252 * rename lock(with LUSTRE_BFL_FID) by root object */
1254 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1256 RETURN(PTR_ERR(obj));
1258 LASSERT(mdt_object_remote(obj));
1259 rc = mdt_remote_object_lock(info, obj,
1260 &LUSTRE_BFL_FID, lh,
1262 MDS_INODELOCK_UPDATE);
1263 mdt_object_put(info->mti_env, obj);
1265 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1266 memset(policy, 0, sizeof *policy);
1267 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1268 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1269 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1270 LCK_EX, &flags, ldlm_blocking_ast,
1271 ldlm_completion_ast, NULL, NULL, 0,
1273 &info->mti_exp->exp_handle.h_cookie,
1280 static void mdt_rename_unlock(struct lustre_handle *lh)
1283 LASSERT(lustre_handle_is_used(lh));
1284 /* Cancel the single rename lock right away */
1285 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1290 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1291 * target. Source should not be ancestor of target dir. May be other rename
1292 * checks can be moved here later.
1294 static int mdt_is_subdir(struct mdt_thread_info *info,
1295 struct mdt_object *dir,
1296 const struct lu_fid *fid)
1298 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1302 /* If the source and target are in the same directory, they can not
1303 * be parent/child relationship, so subdir check is not needed */
1304 if (lu_fid_eq(&dir_fid, fid))
1307 if (!mdt_object_exists(dir))
1310 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1313 CERROR("%s: failed subdir check in "DFID" for "DFID
1314 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1315 PFID(&dir_fid), PFID(fid), rc);
1316 /* Return EINVAL only if a parent is the @fid */
1320 /* check the found fid */
1321 if (lu_fid_eq(&dir_fid, fid))
1328 /* Update object linkEA */
1329 struct mdt_lock_list {
1330 struct mdt_object *mll_obj;
1331 struct mdt_lock_handle mll_lh;
1332 struct list_head mll_list;
1335 static void mdt_unlock_list(struct mdt_thread_info *info,
1336 struct list_head *list, int rc)
1338 struct mdt_lock_list *mll;
1339 struct mdt_lock_list *mll2;
1341 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1342 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1343 list_del(&mll->mll_list);
1348 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1349 struct mdt_object *obj,
1350 struct mdt_object *pobj,
1351 struct list_head *lock_list)
1353 struct lu_buf *buf = &info->mti_big_buf;
1354 struct linkea_data ldata = { 0 };
1359 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1362 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1363 if (buf->lb_buf == NULL)
1367 rc = mdt_links_read(info, obj, &ldata);
1369 if (rc == -ENOENT || rc == -ENODATA)
1374 LASSERT(ldata.ld_leh != NULL);
1375 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1376 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1377 struct mdt_device *mdt = info->mti_mdt;
1378 struct mdt_object *mdt_pobj;
1379 struct mdt_lock_list *mll;
1380 struct lu_name name;
1383 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1385 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1387 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1388 if (IS_ERR(mdt_pobj)) {
1389 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1390 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1394 if (!mdt_object_exists(mdt_pobj)) {
1395 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1396 mdt_obd_name(mdt), PFID(&fid));
1397 mdt_object_put(info->mti_env, mdt_pobj);
1401 if (mdt_pobj == pobj) {
1402 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1403 mdt_obd_name(mdt), PFID(&fid));
1404 mdt_object_put(info->mti_env, mdt_pobj);
1410 mdt_object_put(info->mti_env, mdt_pobj);
1411 GOTO(out, rc = -ENOMEM);
1414 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1415 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1416 MDS_INODELOCK_UPDATE,
1419 CERROR("%s: cannot lock "DFID": rc =%d\n",
1420 mdt_obd_name(mdt), PFID(&fid), rc);
1421 mdt_object_put(info->mti_env, mdt_pobj);
1426 CFS_INIT_LIST_HEAD(&mll->mll_list);
1427 mll->mll_obj = mdt_pobj;
1428 list_add_tail(&mll->mll_list, lock_list);
1432 mdt_unlock_list(info, lock_list, rc);
1436 /* migrate files from one MDT to another MDT */
1437 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1438 struct mdt_lock_handle *lhc)
1440 struct mdt_reint_record *rr = &info->mti_rr;
1441 struct md_attr *ma = &info->mti_attr;
1442 struct mdt_object *msrcdir;
1443 struct mdt_object *mold;
1444 struct mdt_object *mnew = NULL;
1445 struct mdt_lock_handle *lh_dirp;
1446 struct mdt_lock_handle *lh_childp;
1447 struct mdt_lock_handle *lh_tgtp = NULL;
1448 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1449 struct list_head lock_list;
1453 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1454 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1456 /* 1: lock the source dir. */
1457 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1458 if (IS_ERR(msrcdir)) {
1459 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1460 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1461 (int)PTR_ERR(msrcdir));
1462 RETURN(PTR_ERR(msrcdir));
1465 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1466 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1467 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1468 MDS_INODELOCK_UPDATE,
1471 GOTO(out_put_parent, rc);
1473 if (!mdt_object_remote(msrcdir)) {
1474 rc = mdt_version_get_check_save(info, msrcdir, 0);
1476 GOTO(out_unlock_parent, rc);
1479 /* 2: sanity check and find the object to be migrated. */
1481 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1483 GOTO(out_unlock_parent, rc);
1485 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1486 GOTO(out_unlock_parent, rc = -EINVAL);
1488 if (!fid_is_md_operative(old_fid))
1489 GOTO(out_unlock_parent, rc = -EPERM);
1491 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1493 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1495 if (mdt_object_remote(mold)) {
1496 CERROR("%s: source "DFID" is on the remote MDT\n",
1497 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1498 GOTO(out_put_child, rc = -EREMOTE);
1501 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1502 !mdt_object_remote(msrcdir)) {
1503 CERROR("%s: parent "DFID" is still on the same"
1504 " MDT, which should be migrated first:"
1505 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1506 PFID(mdt_object_fid(msrcdir)), -EPERM);
1507 GOTO(out_put_child, rc = -EPERM);
1510 rc = mdt_remote_permission(info, msrcdir, mold);
1512 GOTO(out_put_child, rc);
1514 /* 3: iterate the linkea of the object and lock all of the objects */
1515 CFS_INIT_LIST_HEAD(&lock_list);
1516 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1518 GOTO(out_put_child, rc);
1520 /* 4: lock of the object migrated object */
1521 lh_childp = &info->mti_lh[MDT_LH_OLD];
1522 mdt_lock_reg_init(lh_childp, LCK_EX);
1523 rc = mdt_object_lock(info, mold, lh_childp,
1524 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1525 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1527 GOTO(out_unlock_list, rc);
1529 ma->ma_need = MA_LMV;
1531 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1532 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1533 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1535 GOTO(out_unlock_list, rc);
1537 if ((ma->ma_valid & MA_LMV)) {
1538 struct lmv_mds_md_v1 *lmm1;
1540 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1541 lmm1 = &ma->ma_lmv->lmv_md_v1;
1542 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1543 CERROR("%s: can not migrate striped dir "DFID
1544 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1545 PFID(mdt_object_fid(mold)), -EPERM);
1546 GOTO(out_unlock_child, rc = -EPERM);
1549 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1550 GOTO(out_unlock_child, rc = -EINVAL);
1552 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1553 &lmm1->lmv_stripe_fids[1]);
1555 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1557 if (!mdt_object_remote(mnew)) {
1558 CERROR("%s: "DFID" being migrated is on this MDT:"
1559 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1560 PFID(rr->rr_fid2), -EPERM);
1561 GOTO(out_put_new, rc = -EPERM);
1564 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1565 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1566 rc = mdt_remote_object_lock(info, mnew,
1567 mdt_object_fid(mnew),
1568 &lh_tgtp->mlh_rreg_lh,
1569 lh_tgtp->mlh_rreg_mode,
1570 MDS_INODELOCK_UPDATE);
1573 GOTO(out_put_new, rc);
1576 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1579 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1580 if (!mdt_object_remote(mnew)) {
1581 CERROR("%s: Migration "DFID" is on this MDT:"
1582 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1583 PFID(rr->rr_fid2), -EXDEV);
1584 GOTO(out_put_new, rc = -EXDEV);
1589 mdt_reint_init_ma(info, ma);
1591 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1592 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1594 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1595 mdt_object_child(mold), &rr->rr_name,
1596 mdt_object_child(mnew), ma);
1598 GOTO(out_unlock_new, rc);
1600 if (lh_tgtp != NULL)
1601 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1604 mdt_object_put(info->mti_env, mnew);
1606 mdt_object_unlock(info, mold, lh_childp, rc);
1608 mdt_unlock_list(info, &lock_list, rc);
1610 mdt_object_put(info->mti_env, mold);
1612 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1614 mdt_object_put(info->mti_env, msrcdir);
1619 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1620 const struct lu_fid *fid,
1623 struct mdt_object *dir;
1627 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1631 /* check early, the real version will be saved after locking */
1632 rc = mdt_version_get_check(info, dir, idx);
1638 mdt_object_put(info->mti_env, dir);
1642 static int mdt_object_lock_save(struct mdt_thread_info *info,
1643 struct mdt_object *dir,
1644 struct mdt_lock_handle *lh,
1649 /* we lock the target dir if it is local */
1650 rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1655 /* get and save correct version after locking */
1656 mdt_version_get_save(info, dir, idx);
1661 static int mdt_rename_parents_lock(struct mdt_thread_info *info,
1662 struct mdt_object **srcp,
1663 struct mdt_object **tgtp)
1665 struct mdt_reint_record *rr = &info->mti_rr;
1666 const struct lu_fid *fid_src = rr->rr_fid1;
1667 const struct lu_fid *fid_tgt = rr->rr_fid2;
1668 struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT];
1669 struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
1670 struct mdt_object *src;
1671 struct mdt_object *tgt;
1676 /* find both parents. */
1677 src = mdt_object_find_check(info, fid_src, 0);
1679 RETURN(PTR_ERR(src));
1681 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1683 if (lu_fid_eq(fid_src, fid_tgt)) {
1685 mdt_object_get(info->mti_env, tgt);
1687 /* Check if the @src is not a child of the @tgt, otherwise a
1688 * reverse locking must take place. */
1689 rc = mdt_is_subdir(info, src, fid_tgt);
1693 GOTO(err_src_put, rc);
1695 tgt = mdt_object_find_check(info, fid_tgt, 1);
1697 GOTO(err_src_put, rc = PTR_ERR(tgt));
1699 if (unlikely(mdt_object_remote(tgt))) {
1700 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1701 "on different MDTs\n", PFID(fid_src),
1703 GOTO(err_tgt_put, rc = -EXDEV);
1707 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1709 /* lock parents in the proper order. */
1711 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1713 GOTO(err_tgt_put, rc);
1715 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1717 rc = mdt_object_lock_save(info, src, lh_src, 0);
1719 rc = mdt_object_lock_save(info, src, lh_src, 0);
1721 GOTO(err_tgt_put, rc);
1723 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1726 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1727 else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
1728 rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
1729 MDS_INODELOCK_UPDATE);
1730 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1734 GOTO(err_unlock, rc);
1736 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1743 /* The order does not matter as the handle is checked inside,
1744 * as well as not used handle. */
1745 mdt_object_unlock(info, src, lh_src, rc);
1746 mdt_object_unlock(info, tgt, lh_tgt, rc);
1748 mdt_object_put(info->mti_env, tgt);
1750 mdt_object_put(info->mti_env, src);
1755 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1756 * 2 - src child; 3 - tgt child.
1757 * Update on disk version of src child.
1760 * For DNE phase I, only these renames are allowed
1761 * mv src_p/src_c tgt_p/tgt_c
1762 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1763 * 2. src_p and tgt_p are same directory, and tgt_c does not
1764 * exists. In this case, all of modification will happen
1765 * in the MDT where ithesource parent is, only one remote
1766 * update is needed, i.e. set c_time/m_time on the child.
1767 * And tgt_c will be still in the same MDT as the original
1770 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1771 struct mdt_lock_handle *lhc)
1773 struct mdt_reint_record *rr = &info->mti_rr;
1774 struct md_attr *ma = &info->mti_attr;
1775 struct ptlrpc_request *req = mdt_info_req(info);
1776 struct mdt_object *msrcdir = NULL;
1777 struct mdt_object *mtgtdir = NULL;
1778 struct mdt_object *mold;
1779 struct mdt_object *mnew = NULL;
1780 struct mdt_lock_handle *lh_srcdirp;
1781 struct mdt_lock_handle *lh_tgtdirp;
1782 struct mdt_lock_handle *lh_oldp = NULL;
1783 struct mdt_lock_handle *lh_newp = NULL;
1784 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1785 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1789 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1790 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1791 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1793 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1794 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1795 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1796 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1798 /* step 1&2: lock the source and target dirs. */
1799 rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
1803 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1805 /* step 3: find & lock the old object. */
1807 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1809 GOTO(out_unlock_parents, rc);
1811 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1812 GOTO(out_unlock_parents, rc = -EINVAL);
1814 if (!fid_is_md_operative(old_fid))
1815 GOTO(out_unlock_parents, rc = -EPERM);
1817 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1819 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1821 /* Check if @mtgtdir is subdir of @mold, before locking child
1822 * to avoid reverse locking. */
1823 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1825 GOTO(out_put_old, rc);
1827 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1828 /* save version after locking */
1829 mdt_version_get_save(info, mold, 2);
1830 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1832 /* step 4: find & lock the new object. */
1833 /* new target object may not exist now */
1834 /* lookup with version checking */
1836 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1839 /* the new_fid should have been filled at this moment */
1840 if (lu_fid_eq(old_fid, new_fid))
1841 GOTO(out_put_old, rc);
1843 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1844 lu_fid_eq(new_fid, rr->rr_fid2))
1845 GOTO(out_put_old, rc = -EINVAL);
1847 if (!fid_is_md_operative(new_fid))
1848 GOTO(out_put_old, rc = -EPERM);
1850 if (mdt_object_remote(mold)) {
1851 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1853 GOTO(out_put_old, rc = -EXDEV);
1856 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1858 GOTO(out_put_old, rc = PTR_ERR(mnew));
1860 if (mdt_object_remote(mnew)) {
1861 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1863 GOTO(out_put_new, rc = -EXDEV);
1866 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
1867 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
1868 GOTO(out_put_new, rc = -EISDIR);
1870 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1871 mdt_lock_reg_init(lh_oldp, LCK_EX);
1872 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1873 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1875 GOTO(out_put_new, rc);
1877 /* Check if @msrcdir is subdir of @mnew, before locking child
1878 * to avoid reverse locking. */
1879 rc = mdt_is_subdir(info, msrcdir, new_fid);
1881 GOTO(out_unlock_old, rc);
1883 /* We used to acquire MDS_INODELOCK_FULL here but we
1884 * can't do this now because a running HSM restore on
1885 * the rename onto victim will hold the layout
1886 * lock. See LU-4002. */
1888 lh_newp = &info->mti_lh[MDT_LH_NEW];
1889 mdt_lock_reg_init(lh_newp, LCK_EX);
1890 rc = mdt_object_lock(info, mnew, lh_newp,
1891 MDS_INODELOCK_LOOKUP |
1892 MDS_INODELOCK_UPDATE,
1895 GOTO(out_unlock_old, rc);
1897 /* get and save version after locking */
1898 mdt_version_get_save(info, mnew, 3);
1899 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1900 } else if (rc != -EREMOTE && rc != -ENOENT) {
1901 GOTO(out_put_old, rc);
1903 /* If mnew does not exist and mold are remote directory,
1904 * it only allows rename if they are under same directory */
1905 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1906 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1908 GOTO(out_put_old, rc = -EXDEV);
1911 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1912 mdt_lock_reg_init(lh_oldp, LCK_EX);
1913 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1914 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1916 GOTO(out_put_old, rc);
1918 mdt_enoent_version_save(info, 3);
1921 /* step 5: rename it */
1922 mdt_reint_init_ma(info, ma);
1924 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1925 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1928 mutex_lock(&mnew->mot_lov_mutex);
1930 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1931 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1932 mnew != NULL ? mdt_object_child(mnew) : NULL,
1933 &rr->rr_tgt_name, ma);
1936 mutex_unlock(&mnew->mot_lov_mutex);
1938 /* handle last link of tgt object */
1940 mdt_counter_incr(req, LPROC_MDT_RENAME);
1942 mdt_handle_last_unlink(info, mnew, ma);
1944 mdt_rename_counter_tally(info, info->mti_mdt, req,
1950 mdt_object_unlock(info, mnew, lh_newp, rc);
1952 mdt_object_unlock(info, mold, lh_oldp, rc);
1955 mdt_object_put(info->mti_env, mnew);
1957 mdt_object_put(info->mti_env, mold);
1959 mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc);
1960 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1964 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1965 struct mdt_lock_handle *lhc,
1966 enum mdt_rename_lock rename_lock)
1968 struct mdt_reint_record *rr = &info->mti_rr;
1969 struct ptlrpc_request *req = mdt_info_req(info);
1970 struct lustre_handle rename_lh = { 0 };
1974 if (info->mti_dlm_req)
1975 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1977 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1978 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1981 rc = mdt_rename_lock(info, &rename_lh, rename_lock);
1983 CERROR("%s: can't lock FS for rename: rc = %d\n",
1984 mdt_obd_name(info->mti_mdt), rc);
1988 if (rename_lock == MRL_RENAME)
1989 rc = mdt_reint_rename_internal(info, lhc);
1991 rc = mdt_reint_migrate_internal(info, lhc);
1993 if (lustre_handle_is_used(&rename_lh))
1994 mdt_rename_unlock(&rename_lh);
1999 static int mdt_reint_rename(struct mdt_thread_info *info,
2000 struct mdt_lock_handle *lhc)
2002 return mdt_reint_rename_or_migrate(info, lhc, MRL_RENAME);
2005 static int mdt_reint_migrate(struct mdt_thread_info *info,
2006 struct mdt_lock_handle *lhc)
2008 return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE);
2011 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
2012 struct mdt_lock_handle *lhc);
2014 static mdt_reinter reinters[REINT_MAX] = {
2015 [REINT_SETATTR] = mdt_reint_setattr,
2016 [REINT_CREATE] = mdt_reint_create,
2017 [REINT_LINK] = mdt_reint_link,
2018 [REINT_UNLINK] = mdt_reint_unlink,
2019 [REINT_RENAME] = mdt_reint_rename,
2020 [REINT_OPEN] = mdt_reint_open,
2021 [REINT_SETXATTR] = mdt_reint_setxattr,
2022 [REINT_RMENTRY] = mdt_reint_unlink,
2023 [REINT_MIGRATE] = mdt_reint_migrate,
2026 int mdt_reint_rec(struct mdt_thread_info *info,
2027 struct mdt_lock_handle *lhc)
2032 rc = reinters[info->mti_rr.rr_opcode](info, lhc);