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 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
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 lh = &info->mti_lh[MDT_LH_PARENT];
551 mdt_lock_reg_init(lh, LCK_PW);
553 /* Even though the new MDT will grant PERM lock to the old
554 * client, but the old client will almost ignore that during
555 * So it needs to revoke both LOOKUP and PERM lock here, so
556 * both new and old client can cancel the dcache */
557 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
558 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
560 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
564 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
565 mdt_lock_reg_init(s0_lh, LCK_PW);
566 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
568 GOTO(out_unlock, rc);
570 /* all attrs are packed into mti_attr in unpack_setattr */
571 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
572 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
574 /* This is only for set ctime when rename's source is on remote MDS. */
575 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
576 ma->ma_attr_flags |= MDS_VTX_BYPASS;
578 /* VBR: update version if attr changed are important for recovery */
580 /* update on-disk version of changed object */
581 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
582 rc = mdt_version_get_check_save(info, mo, 0);
584 GOTO(out_unlock, rc);
587 /* Ensure constant striping during chown(). See LU-2789. */
588 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
589 mutex_lock(&mo->mot_lov_mutex);
591 /* all attrs are packed into mti_attr in unpack_setattr */
592 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
594 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
595 mutex_unlock(&mo->mot_lov_mutex);
598 GOTO(out_unlock, rc);
602 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
603 mdt_object_unlock(info, mo, lh, rc);
608 * Check HSM flags and add HS_DIRTY flag if relevant.
610 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
611 * and is not RELEASED.
613 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
619 /* If the file was modified, add the dirty flag */
620 ma->ma_need = MA_HSM;
621 rc = mdt_attr_get_complex(info, mo, ma);
623 CERROR("file attribute read error for "DFID": %d.\n",
624 PFID(mdt_object_fid(mo)), rc);
628 /* If an up2date copy exists in the backend, add dirty flag */
629 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
630 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
631 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
633 ma->ma_hsm.mh_flags |= HS_DIRTY;
635 mdt_lock_reg_init(lh, LCK_PW);
636 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
641 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
643 CERROR("file attribute change error for "DFID": %d\n",
644 PFID(mdt_object_fid(mo)), rc);
645 mdt_object_unlock(info, mo, lh, rc);
651 static int mdt_reint_setattr(struct mdt_thread_info *info,
652 struct mdt_lock_handle *lhc)
654 struct md_attr *ma = &info->mti_attr;
655 struct mdt_reint_record *rr = &info->mti_rr;
656 struct ptlrpc_request *req = mdt_info_req(info);
657 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
658 struct mdt_file_data *mfd;
659 struct mdt_object *mo;
660 struct mdt_body *repbody;
664 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
665 (unsigned int)ma->ma_attr.la_valid);
667 if (info->mti_dlm_req)
668 ldlm_request_cancel(req, info->mti_dlm_req, 0);
670 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
671 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
673 GOTO(out, rc = PTR_ERR(mo));
675 if (!mdt_object_exists(mo))
676 GOTO(out_put, rc = -ENOENT);
678 if (mdt_object_remote(mo))
679 GOTO(out_put, rc = -EREMOTE);
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 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 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM))
721 GOTO(out_put, rc = -EPROTO);
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);
735 if (mfd->mfd_mode != MDS_FMODE_SOM ||
736 (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
737 GOTO(out_put, rc = -EPROTO);
739 class_handle_unhash(&mfd->mfd_handle);
740 list_del_init(&mfd->mfd_list);
741 spin_unlock(&med->med_open_lock);
743 mdt_mfd_close(info, mfd);
744 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
745 if (ma->ma_valid & MA_LOV)
746 GOTO(out_put, rc = -EPROTO);
748 rc = mdt_attr_set(info, mo, ma);
751 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
752 struct lu_buf *buf = &info->mti_buf;
754 if (ma->ma_attr.la_valid != 0)
755 GOTO(out_put, rc = -EPROTO);
757 buf->lb_buf = ma->ma_lmm;
758 buf->lb_len = ma->ma_lmm_size;
759 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
760 buf, XATTR_NAME_LOV, 0);
763 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
764 struct lu_buf *buf = &info->mti_buf;
766 if (ma->ma_attr.la_valid != 0)
767 GOTO(out_put, rc = -EPROTO);
769 buf->lb_buf = ma->ma_lmv;
770 buf->lb_len = ma->ma_lmv_size;
771 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
772 buf, XATTR_NAME_DEFAULT_LMV, 0);
776 GOTO(out_put, rc = -EPROTO);
779 /* If file data is modified, add the dirty flag */
780 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
781 rc = mdt_add_dirty_flag(info, mo, ma);
783 ma->ma_need = MA_INODE;
785 rc = mdt_attr_get_complex(info, mo, ma);
789 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
791 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
792 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
793 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
794 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
795 struct lustre_capa *capa;
797 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
799 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
800 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
803 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
808 mdt_object_put(info->mti_env, mo);
811 mdt_counter_incr(req, LPROC_MDT_SETATTR);
813 mdt_client_compatibility(info);
814 rc2 = mdt_fix_reply(info);
820 static int mdt_reint_create(struct mdt_thread_info *info,
821 struct mdt_lock_handle *lhc)
823 struct ptlrpc_request *req = mdt_info_req(info);
827 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
828 RETURN(err_serious(-ESTALE));
830 if (info->mti_dlm_req)
831 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
833 if (!lu_name_is_valid(&info->mti_rr.rr_name))
836 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
838 mdt_counter_incr(req, LPROC_MDT_MKDIR);
846 /* Special file should stay on the same node as parent. */
847 mdt_counter_incr(req, LPROC_MDT_MKNOD);
850 CERROR("%s: Unsupported mode %o\n",
851 mdt_obd_name(info->mti_mdt),
852 info->mti_attr.ma_attr.la_mode);
853 RETURN(err_serious(-EOPNOTSUPP));
856 rc = mdt_md_create(info);
861 * VBR: save parent version in reply and child version getting by its name.
862 * Version of child is getting and checking during its lookup. If
864 static int mdt_reint_unlink(struct mdt_thread_info *info,
865 struct mdt_lock_handle *lhc)
867 struct mdt_reint_record *rr = &info->mti_rr;
868 struct ptlrpc_request *req = mdt_info_req(info);
869 struct md_attr *ma = &info->mti_attr;
870 struct lu_fid *child_fid = &info->mti_tmp_fid1;
871 struct mdt_object *mp;
872 struct mdt_object *mc;
873 struct mdt_lock_handle *parent_lh;
874 struct mdt_lock_handle *child_lh;
875 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
876 struct mdt_lock_handle *s0_lh = NULL;
877 struct mdt_object *s0_obj = NULL;
882 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
883 PNAME(&rr->rr_name));
885 if (info->mti_dlm_req)
886 ldlm_request_cancel(req, info->mti_dlm_req, 0);
888 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
889 RETURN(err_serious(-ENOENT));
891 if (!fid_is_md_operative(rr->rr_fid1))
895 * step 1: Found the parent.
897 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
903 parent_lh = &info->mti_lh[MDT_LH_PARENT];
904 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
905 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
908 GOTO(put_parent, rc);
910 if (!mdt_object_remote(mp)) {
911 rc = mdt_version_get_check_save(info, mp, 0);
913 GOTO(unlock_parent, rc);
916 /* step 2: find & lock the child */
917 /* lookup child object along with version checking */
919 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
921 /* Name might not be able to find during resend of
922 * remote unlink, considering following case.
923 * dir_A is a remote directory, the name entry of
924 * dir_A is on MDT0, the directory is on MDT1,
926 * 1. client sends unlink req to MDT1.
927 * 2. MDT1 sends name delete update to MDT0.
928 * 3. name entry is being deleted in MDT0 synchronously.
929 * 4. MDT1 is restarted.
930 * 5. client resends unlink req to MDT1. So it can not
931 * find the name entry on MDT0 anymore.
932 * In this case, MDT1 only needs to destory the local
935 if (mdt_object_remote(mp) && rc == -ENOENT &&
936 !fid_is_zero(rr->rr_fid2) &&
937 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
939 *child_fid = *rr->rr_fid2;
941 GOTO(unlock_parent, rc);
945 if (!fid_is_md_operative(child_fid))
946 GOTO(unlock_parent, rc = -EPERM);
948 /* We will lock the child regardless it is local or remote. No harm. */
949 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
951 GOTO(unlock_parent, rc = PTR_ERR(mc));
953 child_lh = &info->mti_lh[MDT_LH_CHILD];
954 mdt_lock_reg_init(child_lh, LCK_EX);
955 if (mdt_object_remote(mc)) {
956 struct mdt_body *repbody;
958 if (!fid_is_zero(rr->rr_fid2)) {
959 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
960 mdt_obd_name(info->mti_mdt),
961 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
962 GOTO(put_child, rc = -ENOENT);
964 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
965 mdt_obd_name(info->mti_mdt),
966 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
968 if (!mdt_is_dne_client(req->rq_export))
969 /* Return -ENOTSUPP for old client */
970 GOTO(put_child, rc = -ENOTSUPP);
972 if (info->mti_spec.sp_rm_entry) {
973 struct lu_ucred *uc = mdt_ucred(info);
975 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
976 CERROR("%s: unlink remote entry is only "
977 "permitted for administrator: rc = %d\n",
978 mdt_obd_name(info->mti_mdt),
980 GOTO(put_child, rc = -EPERM);
983 ma->ma_need = MA_INODE;
985 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
986 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
987 NULL, &rr->rr_name, ma, no_name);
990 /* Revoke the LOOKUP lock of the remote object granted by
991 * this MDT. Since the unlink will happen on another MDT,
992 * it will release the LOOKUP lock right away. Then What
993 * would happen if another client try to grab the LOOKUP
994 * lock at the same time with unlink XXX */
995 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
997 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
998 LASSERT(repbody != NULL);
999 repbody->mbo_fid1 = *mdt_object_fid(mc);
1000 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1001 GOTO(unlock_child, rc = -EREMOTE);
1002 } else if (info->mti_spec.sp_rm_entry) {
1004 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
1006 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
1007 GOTO(put_child, rc);
1010 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1011 * this now because a running HSM restore on the child (unlink
1012 * victim) will hold the layout lock. See LU-4002. */
1013 rc = mdt_object_lock(info, mc, child_lh,
1014 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
1017 GOTO(put_child, rc);
1019 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1020 * whether need MA_LOV and MA_COOKIE.
1022 ma->ma_need = MA_INODE;
1025 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1026 mdt_lock_reg_init(s0_lh, LCK_EX);
1027 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
1030 GOTO(unlock_child, rc);
1032 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1033 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1034 /* save version when object is locked */
1035 mdt_version_get_save(info, mc, 1);
1037 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1039 mutex_lock(&mc->mot_lov_mutex);
1041 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1042 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1044 mutex_unlock(&mc->mot_lov_mutex);
1046 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1047 rc = mdt_attr_get_complex(info, mc, ma);
1049 mdt_handle_last_unlink(info, mc, ma);
1051 if (ma->ma_valid & MA_INODE) {
1052 switch (ma->ma_attr.la_mode & S_IFMT) {
1054 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1062 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1065 LASSERTF(0, "bad file type %o unlinking\n",
1066 ma->ma_attr.la_mode);
1073 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
1074 mdt_object_unlock(info, mc, child_lh, rc);
1076 mdt_object_put(info->mti_env, mc);
1078 mdt_object_unlock(info, mp, parent_lh, rc);
1080 mdt_object_put(info->mti_env, mp);
1086 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1089 static int mdt_reint_link(struct mdt_thread_info *info,
1090 struct mdt_lock_handle *lhc)
1092 struct mdt_reint_record *rr = &info->mti_rr;
1093 struct ptlrpc_request *req = mdt_info_req(info);
1094 struct md_attr *ma = &info->mti_attr;
1095 struct mdt_object *ms;
1096 struct mdt_object *mp;
1097 struct mdt_lock_handle *lhs;
1098 struct mdt_lock_handle *lhp;
1102 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1103 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1105 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1106 RETURN(err_serious(-ENOENT));
1108 if (info->mti_dlm_req)
1109 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1111 /* Invalid case so return error immediately instead of
1113 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1116 if (!fid_is_md_operative(rr->rr_fid1) ||
1117 !fid_is_md_operative(rr->rr_fid2))
1120 /* step 1: find & lock the target parent dir */
1121 lhp = &info->mti_lh[MDT_LH_PARENT];
1122 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1123 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1124 MDS_INODELOCK_UPDATE);
1126 RETURN(PTR_ERR(mp));
1128 rc = mdt_version_get_check_save(info, mp, 0);
1130 GOTO(out_unlock_parent, rc);
1132 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1134 /* step 2: find & lock the source */
1135 lhs = &info->mti_lh[MDT_LH_CHILD];
1136 mdt_lock_reg_init(lhs, LCK_EX);
1138 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1140 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1142 if (!mdt_object_exists(ms)) {
1143 mdt_object_put(info->mti_env, ms);
1144 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1145 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1146 GOTO(out_unlock_parent, rc = -ENOENT);
1149 if (mdt_object_remote(ms)) {
1150 mdt_object_put(info->mti_env, ms);
1151 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1152 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1154 GOTO(out_unlock_parent, rc = -EXDEV);
1157 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1158 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1160 mdt_object_put(info->mti_env, ms);
1161 GOTO(out_unlock_parent, rc);
1164 /* step 3: link it */
1165 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1166 OBD_FAIL_MDS_REINT_LINK_WRITE);
1168 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1169 rc = mdt_version_get_check_save(info, ms, 1);
1171 GOTO(out_unlock_child, rc);
1173 /** check target version by name during replay */
1174 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1175 &info->mti_tmp_fid1, 2);
1176 if (rc != 0 && rc != -ENOENT)
1177 GOTO(out_unlock_child, rc);
1178 /* save version of file name for replay, it must be ENOENT here */
1179 if (!req_is_replay(mdt_info_req(info))) {
1180 if (rc != -ENOENT) {
1181 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1182 PNAME(&rr->rr_name));
1183 GOTO(out_unlock_child, rc = -EEXIST);
1185 info->mti_ver[2] = ENOENT_VERSION;
1186 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1189 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1190 mdt_object_child(ms), &rr->rr_name, ma);
1193 mdt_counter_incr(req, LPROC_MDT_LINK);
1197 mdt_object_unlock_put(info, ms, lhs, rc);
1199 mdt_object_unlock_put(info, mp, lhp, rc);
1203 * lock the part of the directory according to the hash of the name
1204 * (lh->mlh_pdo_hash) in parallel directory lock.
1206 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1207 struct mdt_lock_handle *lh,
1208 struct mdt_object *obj, __u64 ibits)
1210 struct ldlm_res_id *res = &info->mti_res_id;
1211 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1212 ldlm_policy_data_t *policy = &info->mti_policy;
1216 * Finish res_id initializing by name hash marking part of
1217 * directory which is taking modification.
1219 LASSERT(lh->mlh_pdo_hash != 0);
1220 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1221 memset(policy, 0, sizeof(*policy));
1222 policy->l_inodebits.bits = ibits;
1224 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1225 * going to be sent to client. If it is - mdt_intent_policy() path will
1226 * fix it up and turn FL_LOCAL flag off.
1228 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1229 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1230 &info->mti_exp->exp_handle.h_cookie);
1234 enum mdt_rename_lock {
1240 * Get BFL lock for rename or migrate process, right now, it does not support
1241 * cross-MDT rename, so we only need global rename lock during migration.
1243 static int mdt_rename_lock(struct mdt_thread_info *info,
1244 struct lustre_handle *lh,
1245 enum mdt_rename_lock rename_lock)
1247 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1248 ldlm_policy_data_t *policy = &info->mti_policy;
1249 struct ldlm_res_id *res_id = &info->mti_res_id;
1254 /* XXX only do global rename lock for migration */
1255 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 &&
1256 rename_lock == MRL_MIGRATE) {
1257 struct lu_fid *fid = &info->mti_tmp_fid1;
1258 struct mdt_object *obj;
1260 /* XXX, right now, it has to use object API to
1261 * enqueue lock cross MDT, so it will enqueue
1262 * rename lock(with LUSTRE_BFL_FID) by root object */
1264 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1266 RETURN(PTR_ERR(obj));
1268 LASSERT(mdt_object_remote(obj));
1269 rc = mdt_remote_object_lock(info, obj,
1270 &LUSTRE_BFL_FID, lh,
1272 MDS_INODELOCK_UPDATE);
1273 mdt_object_put(info->mti_env, obj);
1275 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1276 memset(policy, 0, sizeof *policy);
1277 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1278 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1279 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1280 LCK_EX, &flags, ldlm_blocking_ast,
1281 ldlm_completion_ast, NULL, NULL, 0,
1283 &info->mti_exp->exp_handle.h_cookie,
1290 static void mdt_rename_unlock(struct lustre_handle *lh)
1293 LASSERT(lustre_handle_is_used(lh));
1294 /* Cancel the single rename lock right away */
1295 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1300 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1301 * target. Source should not be ancestor of target dir. May be other rename
1302 * checks can be moved here later.
1304 static int mdt_is_subdir(struct mdt_thread_info *info,
1305 struct mdt_object *dir,
1306 const struct lu_fid *fid)
1308 struct lu_fid dir_fid = dir->mot_header.loh_fid;
1312 /* If the source and target are in the same directory, they can not
1313 * be parent/child relationship, so subdir check is not needed */
1314 if (lu_fid_eq(&dir_fid, fid))
1317 if (!mdt_object_exists(dir))
1320 rc = mdo_is_subdir(info->mti_env, mdt_object_child(dir),
1323 CERROR("%s: failed subdir check in "DFID" for "DFID
1324 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1325 PFID(&dir_fid), PFID(fid), rc);
1326 /* Return EINVAL only if a parent is the @fid */
1330 /* check the found fid */
1331 if (lu_fid_eq(&dir_fid, fid))
1338 /* Update object linkEA */
1339 struct mdt_lock_list {
1340 struct mdt_object *mll_obj;
1341 struct mdt_lock_handle mll_lh;
1342 struct list_head mll_list;
1345 static void mdt_unlock_list(struct mdt_thread_info *info,
1346 struct list_head *list, int rc)
1348 struct mdt_lock_list *mll;
1349 struct mdt_lock_list *mll2;
1351 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1352 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1353 list_del(&mll->mll_list);
1358 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1359 struct mdt_object *obj,
1360 struct mdt_object *pobj,
1361 struct list_head *lock_list)
1363 struct lu_buf *buf = &info->mti_big_buf;
1364 struct linkea_data ldata = { 0 };
1369 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1372 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1373 if (buf->lb_buf == NULL)
1377 rc = mdt_links_read(info, obj, &ldata);
1379 if (rc == -ENOENT || rc == -ENODATA)
1384 LASSERT(ldata.ld_leh != NULL);
1385 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1386 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1387 struct mdt_device *mdt = info->mti_mdt;
1388 struct mdt_object *mdt_pobj;
1389 struct mdt_lock_list *mll;
1390 struct lu_name name;
1393 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1395 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1397 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1398 if (IS_ERR(mdt_pobj)) {
1399 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1400 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1404 if (!mdt_object_exists(mdt_pobj)) {
1405 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1406 mdt_obd_name(mdt), PFID(&fid));
1407 mdt_object_put(info->mti_env, mdt_pobj);
1411 if (mdt_pobj == pobj) {
1412 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1413 mdt_obd_name(mdt), PFID(&fid));
1414 mdt_object_put(info->mti_env, mdt_pobj);
1420 mdt_object_put(info->mti_env, mdt_pobj);
1421 GOTO(out, rc = -ENOMEM);
1424 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1425 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1426 MDS_INODELOCK_UPDATE,
1429 CERROR("%s: cannot lock "DFID": rc =%d\n",
1430 mdt_obd_name(mdt), PFID(&fid), rc);
1431 mdt_object_put(info->mti_env, mdt_pobj);
1436 INIT_LIST_HEAD(&mll->mll_list);
1437 mll->mll_obj = mdt_pobj;
1438 list_add_tail(&mll->mll_list, lock_list);
1442 mdt_unlock_list(info, lock_list, rc);
1446 /* migrate files from one MDT to another MDT */
1447 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1448 struct mdt_lock_handle *lhc)
1450 struct mdt_reint_record *rr = &info->mti_rr;
1451 struct md_attr *ma = &info->mti_attr;
1452 struct mdt_object *msrcdir;
1453 struct mdt_object *mold;
1454 struct mdt_object *mnew = NULL;
1455 struct mdt_lock_handle *lh_dirp;
1456 struct mdt_lock_handle *lh_childp;
1457 struct mdt_lock_handle *lh_tgtp = NULL;
1458 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1459 struct list_head lock_list;
1463 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1464 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1466 /* 1: lock the source dir. */
1467 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1468 if (IS_ERR(msrcdir)) {
1469 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1470 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1471 (int)PTR_ERR(msrcdir));
1472 RETURN(PTR_ERR(msrcdir));
1475 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1476 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1477 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1478 MDS_INODELOCK_UPDATE,
1481 GOTO(out_put_parent, rc);
1483 if (!mdt_object_remote(msrcdir)) {
1484 rc = mdt_version_get_check_save(info, msrcdir, 0);
1486 GOTO(out_unlock_parent, rc);
1489 /* 2: sanity check and find the object to be migrated. */
1491 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1493 GOTO(out_unlock_parent, rc);
1495 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1496 GOTO(out_unlock_parent, rc = -EINVAL);
1498 if (!fid_is_md_operative(old_fid))
1499 GOTO(out_unlock_parent, rc = -EPERM);
1501 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1503 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1505 if (mdt_object_remote(mold)) {
1506 CERROR("%s: source "DFID" is on the remote MDT\n",
1507 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1508 GOTO(out_put_child, rc = -EREMOTE);
1511 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1512 !mdt_object_remote(msrcdir)) {
1513 CERROR("%s: parent "DFID" is still on the same"
1514 " MDT, which should be migrated first:"
1515 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1516 PFID(mdt_object_fid(msrcdir)), -EPERM);
1517 GOTO(out_put_child, rc = -EPERM);
1520 rc = mdt_remote_permission(info, msrcdir, mold);
1522 GOTO(out_put_child, rc);
1524 /* 3: iterate the linkea of the object and lock all of the objects */
1525 INIT_LIST_HEAD(&lock_list);
1526 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1528 GOTO(out_put_child, rc);
1530 /* 4: lock of the object migrated object */
1531 lh_childp = &info->mti_lh[MDT_LH_OLD];
1532 mdt_lock_reg_init(lh_childp, LCK_EX);
1533 rc = mdt_object_lock(info, mold, lh_childp,
1534 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1535 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1537 GOTO(out_unlock_list, rc);
1539 ma->ma_need = MA_LMV;
1541 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1542 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1543 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1545 GOTO(out_unlock_list, rc);
1547 if ((ma->ma_valid & MA_LMV)) {
1548 struct lmv_mds_md_v1 *lmm1;
1550 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1551 lmm1 = &ma->ma_lmv->lmv_md_v1;
1552 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1553 CERROR("%s: can not migrate striped dir "DFID
1554 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1555 PFID(mdt_object_fid(mold)), -EPERM);
1556 GOTO(out_unlock_child, rc = -EPERM);
1559 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1560 GOTO(out_unlock_child, rc = -EINVAL);
1562 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1563 &lmm1->lmv_stripe_fids[1]);
1565 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1567 if (!mdt_object_remote(mnew)) {
1568 CERROR("%s: "DFID" being migrated is on this MDT:"
1569 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1570 PFID(rr->rr_fid2), -EPERM);
1571 GOTO(out_put_new, rc = -EPERM);
1574 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1575 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1576 rc = mdt_remote_object_lock(info, mnew,
1577 mdt_object_fid(mnew),
1578 &lh_tgtp->mlh_rreg_lh,
1579 lh_tgtp->mlh_rreg_mode,
1580 MDS_INODELOCK_UPDATE);
1583 GOTO(out_put_new, rc);
1586 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1589 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1590 if (!mdt_object_remote(mnew)) {
1591 CERROR("%s: Migration "DFID" is on this MDT:"
1592 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1593 PFID(rr->rr_fid2), -EXDEV);
1594 GOTO(out_put_new, rc = -EXDEV);
1599 mdt_reint_init_ma(info, ma);
1601 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1602 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1604 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1605 mdt_object_child(mold), &rr->rr_name,
1606 mdt_object_child(mnew), ma);
1608 GOTO(out_unlock_new, rc);
1610 if (lh_tgtp != NULL)
1611 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1614 mdt_object_put(info->mti_env, mnew);
1616 mdt_object_unlock(info, mold, lh_childp, rc);
1618 mdt_unlock_list(info, &lock_list, rc);
1620 mdt_object_put(info->mti_env, mold);
1622 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1624 mdt_object_put(info->mti_env, msrcdir);
1629 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1630 const struct lu_fid *fid,
1633 struct mdt_object *dir;
1637 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1641 /* check early, the real version will be saved after locking */
1642 rc = mdt_version_get_check(info, dir, idx);
1648 mdt_object_put(info->mti_env, dir);
1652 static int mdt_object_lock_save(struct mdt_thread_info *info,
1653 struct mdt_object *dir,
1654 struct mdt_lock_handle *lh,
1659 /* we lock the target dir if it is local */
1660 rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1665 /* get and save correct version after locking */
1666 mdt_version_get_save(info, dir, idx);
1671 static int mdt_rename_parents_lock(struct mdt_thread_info *info,
1672 struct mdt_object **srcp,
1673 struct mdt_object **tgtp)
1675 struct mdt_reint_record *rr = &info->mti_rr;
1676 const struct lu_fid *fid_src = rr->rr_fid1;
1677 const struct lu_fid *fid_tgt = rr->rr_fid2;
1678 struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT];
1679 struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
1680 struct mdt_object *src;
1681 struct mdt_object *tgt;
1686 /* find both parents. */
1687 src = mdt_object_find_check(info, fid_src, 0);
1689 RETURN(PTR_ERR(src));
1691 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1693 if (lu_fid_eq(fid_src, fid_tgt)) {
1695 mdt_object_get(info->mti_env, tgt);
1697 /* Check if the @src is not a child of the @tgt, otherwise a
1698 * reverse locking must take place. */
1699 rc = mdt_is_subdir(info, src, fid_tgt);
1703 GOTO(err_src_put, rc);
1705 tgt = mdt_object_find_check(info, fid_tgt, 1);
1707 GOTO(err_src_put, rc = PTR_ERR(tgt));
1709 if (unlikely(mdt_object_remote(tgt))) {
1710 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1711 "on different MDTs\n", PFID(fid_src),
1713 GOTO(err_tgt_put, rc = -EXDEV);
1717 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1719 /* lock parents in the proper order. */
1721 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1723 GOTO(err_tgt_put, rc);
1725 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1727 rc = mdt_object_lock_save(info, src, lh_src, 0);
1729 rc = mdt_object_lock_save(info, src, lh_src, 0);
1731 GOTO(err_tgt_put, rc);
1733 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1736 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1737 else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
1738 rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
1739 MDS_INODELOCK_UPDATE);
1740 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1744 GOTO(err_unlock, rc);
1746 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1753 /* The order does not matter as the handle is checked inside,
1754 * as well as not used handle. */
1755 mdt_object_unlock(info, src, lh_src, rc);
1756 mdt_object_unlock(info, tgt, lh_tgt, rc);
1758 mdt_object_put(info->mti_env, tgt);
1760 mdt_object_put(info->mti_env, src);
1765 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1766 * 2 - src child; 3 - tgt child.
1767 * Update on disk version of src child.
1770 * For DNE phase I, only these renames are allowed
1771 * mv src_p/src_c tgt_p/tgt_c
1772 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1773 * 2. src_p and tgt_p are same directory, and tgt_c does not
1774 * exists. In this case, all of modification will happen
1775 * in the MDT where ithesource parent is, only one remote
1776 * update is needed, i.e. set c_time/m_time on the child.
1777 * And tgt_c will be still in the same MDT as the original
1780 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1781 struct mdt_lock_handle *lhc)
1783 struct mdt_reint_record *rr = &info->mti_rr;
1784 struct md_attr *ma = &info->mti_attr;
1785 struct ptlrpc_request *req = mdt_info_req(info);
1786 struct mdt_object *msrcdir = NULL;
1787 struct mdt_object *mtgtdir = NULL;
1788 struct mdt_object *mold;
1789 struct mdt_object *mnew = NULL;
1790 struct mdt_lock_handle *lh_srcdirp;
1791 struct mdt_lock_handle *lh_tgtdirp;
1792 struct mdt_lock_handle *lh_oldp = NULL;
1793 struct mdt_lock_handle *lh_newp = NULL;
1794 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1795 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1799 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1800 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1801 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1803 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1804 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1805 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1806 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1808 /* step 1&2: lock the source and target dirs. */
1809 rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
1813 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1815 /* step 3: find & lock the old object. */
1817 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1819 GOTO(out_unlock_parents, rc);
1821 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1822 GOTO(out_unlock_parents, rc = -EINVAL);
1824 if (!fid_is_md_operative(old_fid))
1825 GOTO(out_unlock_parents, rc = -EPERM);
1827 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1829 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1831 /* Check if @mtgtdir is subdir of @mold, before locking child
1832 * to avoid reverse locking. */
1833 rc = mdt_is_subdir(info, mtgtdir, old_fid);
1835 GOTO(out_put_old, rc);
1837 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1838 /* save version after locking */
1839 mdt_version_get_save(info, mold, 2);
1840 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1842 /* step 4: find & lock the new object. */
1843 /* new target object may not exist now */
1844 /* lookup with version checking */
1846 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1849 /* the new_fid should have been filled at this moment */
1850 if (lu_fid_eq(old_fid, new_fid))
1851 GOTO(out_put_old, rc);
1853 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1854 lu_fid_eq(new_fid, rr->rr_fid2))
1855 GOTO(out_put_old, rc = -EINVAL);
1857 if (!fid_is_md_operative(new_fid))
1858 GOTO(out_put_old, rc = -EPERM);
1860 if (mdt_object_remote(mold)) {
1861 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1863 GOTO(out_put_old, rc = -EXDEV);
1866 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1868 GOTO(out_put_old, rc = PTR_ERR(mnew));
1870 if (mdt_object_remote(mnew)) {
1871 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1873 GOTO(out_put_new, rc = -EXDEV);
1876 /* Before locking the target dir, check we do not replace
1877 * a dir with a non-dir, otherwise it may deadlock with
1878 * link op which tries to create a link in this dir
1879 * back to this non-dir. */
1880 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
1881 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
1882 GOTO(out_put_new, rc = -EISDIR);
1884 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1885 mdt_lock_reg_init(lh_oldp, LCK_EX);
1886 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1887 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1889 GOTO(out_put_new, rc);
1891 /* Check if @msrcdir is subdir of @mnew, before locking child
1892 * to avoid reverse locking. */
1893 rc = mdt_is_subdir(info, msrcdir, new_fid);
1895 GOTO(out_unlock_old, rc);
1897 /* We used to acquire MDS_INODELOCK_FULL here but we
1898 * can't do this now because a running HSM restore on
1899 * the rename onto victim will hold the layout
1900 * lock. See LU-4002. */
1902 lh_newp = &info->mti_lh[MDT_LH_NEW];
1903 mdt_lock_reg_init(lh_newp, LCK_EX);
1904 rc = mdt_object_lock(info, mnew, lh_newp,
1905 MDS_INODELOCK_LOOKUP |
1906 MDS_INODELOCK_UPDATE,
1909 GOTO(out_unlock_old, rc);
1911 /* get and save version after locking */
1912 mdt_version_get_save(info, mnew, 3);
1913 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1914 } else if (rc != -EREMOTE && rc != -ENOENT) {
1915 GOTO(out_put_old, rc);
1917 /* If mnew does not exist and mold are remote directory,
1918 * it only allows rename if they are under same directory */
1919 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1920 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1922 GOTO(out_put_old, rc = -EXDEV);
1925 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1926 mdt_lock_reg_init(lh_oldp, LCK_EX);
1927 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1928 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
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,
1980 enum mdt_rename_lock rename_lock)
1982 struct mdt_reint_record *rr = &info->mti_rr;
1983 struct ptlrpc_request *req = mdt_info_req(info);
1984 struct lustre_handle rename_lh = { 0 };
1988 if (info->mti_dlm_req)
1989 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1991 if (!fid_is_md_operative(rr->rr_fid1) ||
1992 !fid_is_md_operative(rr->rr_fid2))
1995 rc = mdt_rename_lock(info, &rename_lh, rename_lock);
1997 CERROR("%s: can't lock FS for rename: rc = %d\n",
1998 mdt_obd_name(info->mti_mdt), rc);
2002 if (rename_lock == MRL_RENAME)
2003 rc = mdt_reint_rename_internal(info, lhc);
2005 rc = mdt_reint_migrate_internal(info, lhc);
2007 if (lustre_handle_is_used(&rename_lh))
2008 mdt_rename_unlock(&rename_lh);
2013 static int mdt_reint_rename(struct mdt_thread_info *info,
2014 struct mdt_lock_handle *lhc)
2016 return mdt_reint_rename_or_migrate(info, lhc, MRL_RENAME);
2019 static int mdt_reint_migrate(struct mdt_thread_info *info,
2020 struct mdt_lock_handle *lhc)
2022 return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE);
2025 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
2026 struct mdt_lock_handle *lhc);
2028 static mdt_reinter reinters[REINT_MAX] = {
2029 [REINT_SETATTR] = mdt_reint_setattr,
2030 [REINT_CREATE] = mdt_reint_create,
2031 [REINT_LINK] = mdt_reint_link,
2032 [REINT_UNLINK] = mdt_reint_unlink,
2033 [REINT_RENAME] = mdt_reint_rename,
2034 [REINT_OPEN] = mdt_reint_open,
2035 [REINT_SETXATTR] = mdt_reint_setxattr,
2036 [REINT_RMENTRY] = mdt_reint_unlink,
2037 [REINT_MIGRATE] = mdt_reint_migrate,
2040 int mdt_reint_rec(struct mdt_thread_info *info,
2041 struct mdt_lock_handle *lhc)
2046 rc = reinters[info->mti_rr.rr_opcode](info, lhc);