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->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->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 * VBR: we save three versions in reply:
250 * 0 - parent. Check that parent version is the same during replay.
251 * 1 - name. Version of 'name' if file exists with the same name or
252 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
253 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
256 static int mdt_md_create(struct mdt_thread_info *info)
258 struct mdt_device *mdt = info->mti_mdt;
259 struct mdt_object *parent;
260 struct mdt_object *child;
261 struct mdt_lock_handle *lh;
262 struct mdt_body *repbody;
263 struct md_attr *ma = &info->mti_attr;
264 struct mdt_reint_record *rr = &info->mti_rr;
268 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
270 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
272 if (!fid_is_md_operative(rr->rr_fid1))
275 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
277 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
279 RETURN(PTR_ERR(parent));
281 if (!mdt_object_exists(parent))
282 GOTO(put_parent, rc = -ENOENT);
284 lh = &info->mti_lh[MDT_LH_PARENT];
285 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
286 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE,
289 GOTO(put_parent, rc);
291 if (!mdt_object_remote(parent)) {
292 rc = mdt_version_get_check_save(info, parent, 0);
294 GOTO(unlock_parent, rc);
298 * Check child name version during replay.
299 * During create replay a file may exist with same name.
301 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
302 &info->mti_tmp_fid1, 1);
304 GOTO(unlock_parent, rc = -EEXIST);
306 /* -ENOENT is expected here */
308 GOTO(unlock_parent, rc);
310 /* save version of file name for replay, it must be ENOENT here */
311 mdt_enoent_version_save(info, 1);
313 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
314 if (likely(!IS_ERR(child))) {
315 struct md_object *next = mdt_object_child(parent);
317 if (mdt_object_remote(child)) {
318 struct seq_server_site *ss;
319 struct lu_ucred *uc = mdt_ucred(info);
321 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
323 mdt->mdt_enable_remote_dir_gid &&
324 mdt->mdt_enable_remote_dir_gid != -1) {
325 CERROR("%s: Creating remote dir is only"
326 " permitted for administrator or"
327 " set mdt_enable_remote_dir_gid:"
329 mdt_obd_name(mdt), -EPERM);
330 GOTO(out_put_child, rc = -EPERM);
334 ss = mdt_seq_site(mdt);
335 if (ss->ss_node_id != 0 &&
336 mdt->mdt_enable_remote_dir == 0) {
337 CERROR("%s: remote dir is only permitted on"
339 " mdt.*.enable_remote_dir=1\n",
341 GOTO(out_put_child, rc = -EPERM);
343 if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
344 /* Return -EIO for old client */
345 GOTO(out_put_child, rc = -EIO);
349 ma->ma_need = MA_INODE;
351 /* capa for cross-ref will be stored here */
352 ma->ma_capa = req_capsule_server_get(info->mti_pill,
354 LASSERT(ma->ma_capa);
356 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
357 OBD_FAIL_MDS_REINT_CREATE_WRITE);
359 /* Version of child will be updated on disk. */
360 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
361 rc = mdt_version_get_check_save(info, child, 2);
363 GOTO(out_put_child, rc);
365 /* Let lower layer know current lock mode. */
366 info->mti_spec.sp_cr_mode =
367 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
370 * Do not perform lookup sanity check. We know that name does
373 info->mti_spec.sp_cr_lookup = 0;
374 info->mti_spec.sp_feat = &dt_directory_features;
376 rc = mdo_create(info->mti_env, next, &rr->rr_name,
377 mdt_object_child(child), &info->mti_spec, ma);
379 rc = mdt_attr_get_complex(info, child, ma);
382 /* Return fid & attr to client. */
383 if (ma->ma_valid & MA_INODE)
384 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
385 mdt_object_fid(child));
388 mdt_object_put(info->mti_env, child);
392 mdt_create_pack_capa(info, rc, child, repbody);
394 mdt_object_unlock(info, parent, lh, rc);
396 mdt_object_put(info->mti_env, parent);
400 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
401 struct mdt_object *obj, __u64 ibits,
402 struct ldlm_enqueue_info *einfo)
404 ldlm_policy_data_t *policy = &mti->mti_policy;
408 if (!S_ISDIR(obj->mot_header.loh_attr))
411 memset(policy, 0, sizeof(*policy));
412 policy->l_inodebits.bits = ibits;
414 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
420 * Lock slave stripes if necessary, the lock handles of slave stripes
421 * will be stored in einfo->ei_cbdata.
423 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
424 ldlm_mode_t mode, __u64 ibits,
425 struct ldlm_enqueue_info *einfo)
427 ldlm_policy_data_t *policy = &mti->mti_policy;
431 if (!S_ISDIR(obj->mot_header.loh_attr))
434 memset(einfo, 0, sizeof(*einfo));
435 einfo->ei_type = LDLM_IBITS;
436 einfo->ei_mode = mode;
437 einfo->ei_cb_bl = mdt_remote_blocking_ast;
438 einfo->ei_cb_cp = ldlm_completion_ast;
439 einfo->ei_enq_slave = 1;
440 memset(policy, 0, sizeof(*policy));
441 policy->l_inodebits.bits = ibits;
443 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
448 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
449 struct md_attr *ma, int flags)
451 struct mdt_lock_handle *lh;
452 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
453 __u64 lockpart = MDS_INODELOCK_UPDATE;
454 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
458 /* attr shouldn't be set on remote object */
459 LASSERT(!mdt_object_remote(mo));
461 lh = &info->mti_lh[MDT_LH_PARENT];
462 mdt_lock_reg_init(lh, LCK_PW);
464 /* Even though the new MDT will grant PERM lock to the old
465 * client, but the old client will almost ignore that during
466 * So it needs to revoke both LOOKUP and PERM lock here, so
467 * both new and old client can cancel the dcache */
468 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
469 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
471 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
475 rc = mdt_lock_slaves(info, mo, LCK_EX, lockpart, einfo);
477 GOTO(out_unlock, rc);
479 if (mdt_object_exists(mo) == 0)
480 GOTO(out_unlock, rc = -ENOENT);
482 /* all attrs are packed into mti_attr in unpack_setattr */
483 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
484 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
486 /* This is only for set ctime when rename's source is on remote MDS. */
487 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
488 ma->ma_attr_flags |= MDS_VTX_BYPASS;
490 /* VBR: update version if attr changed are important for recovery */
492 /* update on-disk version of changed object */
493 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
494 rc = mdt_version_get_check_save(info, mo, 0);
496 GOTO(out_unlock, rc);
499 /* Ensure constant striping during chown(). See LU-2789. */
500 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
501 mutex_lock(&mo->mot_lov_mutex);
503 /* all attrs are packed into mti_attr in unpack_setattr */
504 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
506 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
507 mutex_unlock(&mo->mot_lov_mutex);
510 GOTO(out_unlock, rc);
514 mdt_unlock_slaves(info, mo, lockpart, einfo);
515 mdt_object_unlock(info, mo, lh, rc);
520 * Check HSM flags and add HS_DIRTY flag if relevant.
522 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
523 * and is not RELEASED.
525 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
531 /* If the file was modified, add the dirty flag */
532 ma->ma_need = MA_HSM;
533 rc = mdt_attr_get_complex(info, mo, ma);
535 CERROR("file attribute read error for "DFID": %d.\n",
536 PFID(mdt_object_fid(mo)), rc);
540 /* If an up2date copy exists in the backend, add dirty flag */
541 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
542 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
543 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
545 ma->ma_hsm.mh_flags |= HS_DIRTY;
547 mdt_lock_reg_init(lh, LCK_PW);
548 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
553 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
555 CERROR("file attribute change error for "DFID": %d\n",
556 PFID(mdt_object_fid(mo)), rc);
557 mdt_object_unlock(info, mo, lh, rc);
563 static int mdt_reint_setattr(struct mdt_thread_info *info,
564 struct mdt_lock_handle *lhc)
566 struct md_attr *ma = &info->mti_attr;
567 struct mdt_reint_record *rr = &info->mti_rr;
568 struct ptlrpc_request *req = mdt_info_req(info);
569 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
570 struct mdt_file_data *mfd;
571 struct mdt_object *mo;
572 struct mdt_body *repbody;
576 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
577 (unsigned int)ma->ma_attr.la_valid);
579 if (info->mti_dlm_req)
580 ldlm_request_cancel(req, info->mti_dlm_req, 0);
582 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
583 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
585 GOTO(out, rc = PTR_ERR(mo));
587 /* start a log jounal handle if needed */
588 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
589 if ((ma->ma_attr.la_valid & LA_SIZE) ||
590 (rr->rr_flags & MRF_OPEN_TRUNC)) {
591 /* Check write access for the O_TRUNC case */
592 if (mdt_write_read(mo) < 0)
593 GOTO(out_put, rc = -ETXTBSY);
595 } else if (info->mti_ioepoch &&
596 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
597 /* Truncate case. IOEpoch is opened. */
598 rc = mdt_write_get(mo);
602 mfd = mdt_mfd_new(med);
605 GOTO(out_put, rc = -ENOMEM);
608 mdt_ioepoch_open(info, mo, 0);
609 repbody->ioepoch = mo->mot_ioepoch;
611 mdt_object_get(info->mti_env, mo);
612 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
613 mfd->mfd_object = mo;
614 mfd->mfd_xid = req->rq_xid;
616 spin_lock(&med->med_open_lock);
617 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
618 spin_unlock(&med->med_open_lock);
619 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
622 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
624 /* SOM Attribute update case. Find the proper mfd and update
625 * SOM attributes on the proper object. */
626 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
627 LASSERT(info->mti_ioepoch);
629 spin_lock(&med->med_open_lock);
630 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
633 spin_unlock(&med->med_open_lock);
634 CDEBUG(D_INODE, "no handle for file close: "
635 "fid = "DFID": cookie = "LPX64"\n",
636 PFID(info->mti_rr.rr_fid1),
637 info->mti_ioepoch->handle.cookie);
638 GOTO(out_put, rc = -ESTALE);
640 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
641 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
643 class_handle_unhash(&mfd->mfd_handle);
644 cfs_list_del_init(&mfd->mfd_list);
645 spin_unlock(&med->med_open_lock);
647 mdt_mfd_close(info, mfd);
648 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
649 LASSERT((ma->ma_valid & MA_LOV) == 0);
650 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
653 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
654 struct lu_buf *buf = &info->mti_buf;
655 LASSERT(ma->ma_attr.la_valid == 0);
656 buf->lb_buf = ma->ma_lmm;
657 buf->lb_len = ma->ma_lmm_size;
658 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
659 buf, XATTR_NAME_LOV, 0);
662 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
663 struct lu_buf *buf = &info->mti_buf;
665 LASSERT(ma->ma_attr.la_valid == 0);
666 buf->lb_buf = ma->ma_lmv;
667 buf->lb_len = ma->ma_lmv_size;
668 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
669 buf, XATTR_NAME_DEFAULT_LMV, 0);
675 /* If file data is modified, add the dirty flag */
676 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
677 rc = mdt_add_dirty_flag(info, mo, ma);
679 ma->ma_need = MA_INODE;
681 rc = mdt_attr_get_complex(info, mo, ma);
685 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
687 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
688 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
689 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
690 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
691 struct lustre_capa *capa;
693 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
695 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
696 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
699 repbody->valid |= OBD_MD_FLOSSCAPA;
704 mdt_object_put(info->mti_env, mo);
707 mdt_counter_incr(req, LPROC_MDT_SETATTR);
709 mdt_client_compatibility(info);
710 rc2 = mdt_fix_reply(info);
716 static int mdt_reint_create(struct mdt_thread_info *info,
717 struct mdt_lock_handle *lhc)
719 struct ptlrpc_request *req = mdt_info_req(info);
723 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
724 RETURN(err_serious(-ESTALE));
726 if (info->mti_dlm_req)
727 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
729 if (!lu_name_is_valid(&info->mti_rr.rr_name))
732 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
734 mdt_counter_incr(req, LPROC_MDT_MKDIR);
742 /* Special file should stay on the same node as parent. */
743 mdt_counter_incr(req, LPROC_MDT_MKNOD);
746 CERROR("%s: Unsupported mode %o\n",
747 mdt_obd_name(info->mti_mdt),
748 info->mti_attr.ma_attr.la_mode);
749 RETURN(err_serious(-EOPNOTSUPP));
752 rc = mdt_md_create(info);
757 * VBR: save parent version in reply and child version getting by its name.
758 * Version of child is getting and checking during its lookup. If
760 static int mdt_reint_unlink(struct mdt_thread_info *info,
761 struct mdt_lock_handle *lhc)
763 struct mdt_reint_record *rr = &info->mti_rr;
764 struct ptlrpc_request *req = mdt_info_req(info);
765 struct md_attr *ma = &info->mti_attr;
766 struct lu_fid *child_fid = &info->mti_tmp_fid1;
767 struct mdt_object *mp;
768 struct mdt_object *mc;
769 struct mdt_lock_handle *parent_lh;
770 struct mdt_lock_handle *child_lh;
771 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
776 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
777 PNAME(&rr->rr_name));
779 if (info->mti_dlm_req)
780 ldlm_request_cancel(req, info->mti_dlm_req, 0);
782 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
783 RETURN(err_serious(-ENOENT));
785 if (!fid_is_md_operative(rr->rr_fid1))
789 * step 1: Found the parent.
791 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
797 parent_lh = &info->mti_lh[MDT_LH_PARENT];
798 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
799 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
802 GOTO(put_parent, rc);
804 if (!mdt_object_remote(mp)) {
805 rc = mdt_version_get_check_save(info, mp, 0);
807 GOTO(unlock_parent, rc);
810 /* step 2: find & lock the child */
811 /* lookup child object along with version checking */
813 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
815 /* Name might not be able to find during resend of
816 * remote unlink, considering following case.
817 * dir_A is a remote directory, the name entry of
818 * dir_A is on MDT0, the directory is on MDT1,
820 * 1. client sends unlink req to MDT1.
821 * 2. MDT1 sends name delete update to MDT0.
822 * 3. name entry is being deleted in MDT0 synchronously.
823 * 4. MDT1 is restarted.
824 * 5. client resends unlink req to MDT1. So it can not
825 * find the name entry on MDT0 anymore.
826 * In this case, MDT1 only needs to destory the local
829 if (mdt_object_remote(mp) && rc == -ENOENT &&
830 !fid_is_zero(rr->rr_fid2) &&
831 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
833 *child_fid = *rr->rr_fid2;
835 GOTO(unlock_parent, rc);
839 if (!fid_is_md_operative(child_fid))
840 GOTO(unlock_parent, rc = -EPERM);
842 /* We will lock the child regardless it is local or remote. No harm. */
843 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
845 GOTO(unlock_parent, rc = PTR_ERR(mc));
847 child_lh = &info->mti_lh[MDT_LH_CHILD];
848 mdt_lock_reg_init(child_lh, LCK_EX);
849 if (mdt_object_remote(mc)) {
850 struct mdt_body *repbody;
852 if (!fid_is_zero(rr->rr_fid2)) {
853 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
854 mdt_obd_name(info->mti_mdt),
855 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
856 GOTO(put_child, rc = -ENOENT);
858 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
859 mdt_obd_name(info->mti_mdt),
860 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
862 if (!mdt_is_dne_client(req->rq_export))
863 /* Return -EIO for old client */
864 GOTO(put_child, rc = -EIO);
866 if (info->mti_spec.sp_rm_entry) {
867 struct lu_ucred *uc = mdt_ucred(info);
869 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
870 CERROR("%s: unlink remote entry is only "
871 "permitted for administrator: rc = %d\n",
872 mdt_obd_name(info->mti_mdt),
874 GOTO(put_child, rc = -EPERM);
877 ma->ma_need = MA_INODE;
879 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
880 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
881 NULL, &rr->rr_name, ma, no_name);
884 /* Revoke the LOOKUP lock of the remote object granted by
885 * this MDT. Since the unlink will happen on another MDT,
886 * it will release the LOOKUP lock right away. Then What
887 * would happen if another client try to grab the LOOKUP
888 * lock at the same time with unlink XXX */
889 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
891 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
892 LASSERT(repbody != NULL);
893 repbody->fid1 = *mdt_object_fid(mc);
894 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
895 GOTO(unlock_child, rc = -EREMOTE);
896 } else if (info->mti_spec.sp_rm_entry) {
898 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
900 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
904 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
905 * this now because a running HSM restore on the child (unlink
906 * victim) will hold the layout lock. See LU-4002. */
907 rc = mdt_object_lock(info, mc, child_lh,
908 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
913 * Now we can only make sure we need MA_INODE, in mdd layer, will check
914 * whether need MA_LOV and MA_COOKIE.
916 ma->ma_need = MA_INODE;
919 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, einfo);
921 GOTO(unlock_child, rc);
923 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
924 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
925 /* save version when object is locked */
926 mdt_version_get_save(info, mc, 1);
928 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
930 mutex_lock(&mc->mot_lov_mutex);
932 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
933 mdt_object_child(mc), &rr->rr_name, ma, no_name);
935 mutex_unlock(&mc->mot_lov_mutex);
937 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
938 rc = mdt_attr_get_complex(info, mc, ma);
940 mdt_handle_last_unlink(info, mc, ma);
942 if (ma->ma_valid & MA_INODE) {
943 switch (ma->ma_attr.la_mode & S_IFMT) {
945 mdt_counter_incr(req, LPROC_MDT_RMDIR);
953 mdt_counter_incr(req, LPROC_MDT_UNLINK);
956 LASSERTF(0, "bad file type %o unlinking\n",
957 ma->ma_attr.la_mode);
964 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, einfo);
965 mdt_object_unlock(info, mc, child_lh, rc);
967 /* Since we do not need reply md striped dir info to client, so
968 * reset mti_big_lmm_used to avoid confusing mdt_fix_reply */
969 if (info->mti_big_lmm_used)
970 info->mti_big_lmm_used = 0;
972 mdt_object_put(info->mti_env, mc);
974 mdt_object_unlock(info, mp, parent_lh, rc);
976 mdt_object_put(info->mti_env, mp);
982 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
985 static int mdt_reint_link(struct mdt_thread_info *info,
986 struct mdt_lock_handle *lhc)
988 struct mdt_reint_record *rr = &info->mti_rr;
989 struct ptlrpc_request *req = mdt_info_req(info);
990 struct md_attr *ma = &info->mti_attr;
991 struct mdt_object *ms;
992 struct mdt_object *mp;
993 struct mdt_lock_handle *lhs;
994 struct mdt_lock_handle *lhp;
998 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
999 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1001 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1002 RETURN(err_serious(-ENOENT));
1004 if (info->mti_dlm_req)
1005 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1007 /* Invalid case so return error immediately instead of
1009 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1012 if (!fid_is_md_operative(rr->rr_fid1) ||
1013 !fid_is_md_operative(rr->rr_fid2))
1016 /* step 1: find & lock the target parent dir */
1017 lhp = &info->mti_lh[MDT_LH_PARENT];
1018 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1019 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1020 MDS_INODELOCK_UPDATE);
1022 RETURN(PTR_ERR(mp));
1024 rc = mdt_version_get_check_save(info, mp, 0);
1026 GOTO(out_unlock_parent, rc);
1028 /* step 2: find & lock the source */
1029 lhs = &info->mti_lh[MDT_LH_CHILD];
1030 mdt_lock_reg_init(lhs, LCK_EX);
1032 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1034 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1036 if (!mdt_object_exists(ms)) {
1037 mdt_object_put(info->mti_env, ms);
1038 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1039 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1040 GOTO(out_unlock_parent, rc = -ENOENT);
1043 if (mdt_object_remote(ms)) {
1044 mdt_object_put(info->mti_env, ms);
1045 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1046 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1048 GOTO(out_unlock_parent, rc = -EXDEV);
1051 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1052 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1054 mdt_object_put(info->mti_env, ms);
1055 GOTO(out_unlock_parent, rc);
1058 /* step 3: link it */
1059 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1060 OBD_FAIL_MDS_REINT_LINK_WRITE);
1062 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1063 rc = mdt_version_get_check_save(info, ms, 1);
1065 GOTO(out_unlock_child, rc);
1067 /** check target version by name during replay */
1068 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1069 &info->mti_tmp_fid1, 2);
1070 if (rc != 0 && rc != -ENOENT)
1071 GOTO(out_unlock_child, rc);
1072 /* save version of file name for replay, it must be ENOENT here */
1073 if (!req_is_replay(mdt_info_req(info))) {
1074 if (rc != -ENOENT) {
1075 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1076 PNAME(&rr->rr_name));
1077 GOTO(out_unlock_child, rc = -EEXIST);
1079 info->mti_ver[2] = ENOENT_VERSION;
1080 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1083 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1084 mdt_object_child(ms), &rr->rr_name, ma);
1087 mdt_counter_incr(req, LPROC_MDT_LINK);
1091 mdt_object_unlock_put(info, ms, lhs, rc);
1093 mdt_object_unlock_put(info, mp, lhp, rc);
1097 * lock the part of the directory according to the hash of the name
1098 * (lh->mlh_pdo_hash) in parallel directory lock.
1100 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1101 struct mdt_lock_handle *lh,
1102 struct mdt_object *obj, __u64 ibits)
1104 struct ldlm_res_id *res = &info->mti_res_id;
1105 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1106 ldlm_policy_data_t *policy = &info->mti_policy;
1110 * Finish res_id initializing by name hash marking part of
1111 * directory which is taking modification.
1113 LASSERT(lh->mlh_pdo_hash != 0);
1114 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1115 memset(policy, 0, sizeof(*policy));
1116 policy->l_inodebits.bits = ibits;
1118 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1119 * going to be sent to client. If it is - mdt_intent_policy() path will
1120 * fix it up and turn FL_LOCAL flag off.
1122 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1123 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1124 &info->mti_exp->exp_handle.h_cookie);
1128 enum mdt_rename_lock {
1134 * Get BFL lock for rename or migrate process, right now, it does not support
1135 * cross-MDT rename, so we only need global rename lock during migration.
1137 static int mdt_rename_lock(struct mdt_thread_info *info,
1138 struct lustre_handle *lh,
1139 enum mdt_rename_lock rename_lock)
1141 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1142 ldlm_policy_data_t *policy = &info->mti_policy;
1143 struct ldlm_res_id *res_id = &info->mti_res_id;
1148 /* XXX only do global rename lock for migration */
1149 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 &&
1150 rename_lock == MRL_MIGRATE) {
1151 struct lu_fid *fid = &info->mti_tmp_fid1;
1152 struct mdt_object *obj;
1154 /* XXX, right now, it has to use object API to
1155 * enqueue lock cross MDT, so it will enqueue
1156 * rename lock(with LUSTRE_BFL_FID) by root object */
1158 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1160 RETURN(PTR_ERR(obj));
1162 LASSERT(mdt_object_remote(obj));
1163 rc = mdt_remote_object_lock(info, obj,
1164 &LUSTRE_BFL_FID, lh,
1166 MDS_INODELOCK_UPDATE);
1167 mdt_object_put(info->mti_env, obj);
1169 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1170 memset(policy, 0, sizeof *policy);
1171 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1172 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1173 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1174 LCK_EX, &flags, ldlm_blocking_ast,
1175 ldlm_completion_ast, NULL, NULL, 0,
1177 &info->mti_exp->exp_handle.h_cookie,
1184 static void mdt_rename_unlock(struct lustre_handle *lh)
1187 LASSERT(lustre_handle_is_used(lh));
1188 /* Cancel the single rename lock right away */
1189 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1194 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1195 * target. Source should not be ancestor of target dir. May be other rename
1196 * checks can be moved here later.
1198 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1200 struct mdt_reint_record *rr = &info->mti_rr;
1201 struct lu_fid dst_fid = *rr->rr_fid2;
1202 struct mdt_object *dst;
1206 /* If the source and target are in the same directory, they can not
1207 * be parent/child relationship, so subdir check is not needed */
1208 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1212 LASSERT(fid_is_sane(&dst_fid));
1213 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1215 rc = mdo_is_subdir(info->mti_env,
1216 mdt_object_child(dst), fid,
1218 mdt_object_put(info->mti_env, dst);
1219 if (rc != -EREMOTE && rc < 0) {
1220 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1222 /* check the found fid */
1223 if (lu_fid_eq(&dst_fid, fid))
1229 } while (rc == -EREMOTE);
1234 /* Update object linkEA */
1235 struct mdt_lock_list {
1236 struct mdt_object *mll_obj;
1237 struct mdt_lock_handle mll_lh;
1238 struct list_head mll_list;
1241 static void mdt_unlock_list(struct mdt_thread_info *info,
1242 struct list_head *list, int rc)
1244 struct mdt_lock_list *mll;
1245 struct mdt_lock_list *mll2;
1247 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1248 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1249 list_del(&mll->mll_list);
1254 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1255 struct mdt_object *obj,
1256 struct mdt_object *pobj,
1257 struct list_head *lock_list)
1259 struct lu_buf *buf = &info->mti_big_buf;
1260 struct linkea_data ldata = { 0 };
1265 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1268 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1269 if (buf->lb_buf == NULL)
1273 rc = mdt_links_read(info, obj, &ldata);
1275 if (rc == -ENOENT || rc == -ENODATA)
1280 LASSERT(ldata.ld_leh != NULL);
1281 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1282 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1283 struct mdt_device *mdt = info->mti_mdt;
1284 struct mdt_object *mdt_pobj;
1285 struct mdt_lock_list *mll;
1286 struct lu_name name;
1289 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1291 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1293 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1294 if (IS_ERR(mdt_pobj)) {
1295 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1296 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1300 if (!mdt_object_exists(mdt_pobj)) {
1301 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1302 mdt_obd_name(mdt), PFID(&fid));
1303 mdt_object_put(info->mti_env, mdt_pobj);
1307 if (mdt_pobj == pobj) {
1308 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1309 mdt_obd_name(mdt), PFID(&fid));
1310 mdt_object_put(info->mti_env, mdt_pobj);
1316 mdt_object_put(info->mti_env, mdt_pobj);
1317 GOTO(out, rc = -ENOMEM);
1320 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1321 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1322 MDS_INODELOCK_UPDATE,
1325 CERROR("%s: cannot lock "DFID": rc =%d\n",
1326 mdt_obd_name(mdt), PFID(&fid), rc);
1327 mdt_object_put(info->mti_env, mdt_pobj);
1332 CFS_INIT_LIST_HEAD(&mll->mll_list);
1333 mll->mll_obj = mdt_pobj;
1334 list_add_tail(&mll->mll_list, lock_list);
1338 mdt_unlock_list(info, lock_list, rc);
1342 /* migrate files from one MDT to another MDT */
1343 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1344 struct mdt_lock_handle *lhc)
1346 struct mdt_reint_record *rr = &info->mti_rr;
1347 struct md_attr *ma = &info->mti_attr;
1348 struct mdt_object *msrcdir;
1349 struct mdt_object *mold;
1350 struct mdt_object *mnew = NULL;
1351 struct mdt_lock_handle *lh_dirp;
1352 struct mdt_lock_handle *lh_childp;
1353 struct mdt_lock_handle *lh_tgtp = NULL;
1354 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1355 struct list_head lock_list;
1359 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1360 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1361 /* 1: lock the source dir. */
1362 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1363 if (IS_ERR(msrcdir)) {
1364 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1365 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1366 (int)PTR_ERR(msrcdir));
1367 RETURN(PTR_ERR(msrcdir));
1370 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1371 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1372 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1373 MDS_INODELOCK_UPDATE,
1376 GOTO(out_put_parent, rc);
1378 if (!mdt_object_remote(msrcdir)) {
1379 rc = mdt_version_get_check_save(info, msrcdir, 0);
1381 GOTO(out_unlock_parent, rc);
1384 /* 2: sanity check and find the object to be migrated. */
1386 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1388 GOTO(out_unlock_parent, rc);
1390 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1391 GOTO(out_unlock_parent, rc = -EINVAL);
1393 if (!fid_is_md_operative(old_fid))
1394 GOTO(out_unlock_parent, rc = -EPERM);
1396 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1398 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1400 if (mdt_object_remote(mold)) {
1401 CERROR("%s: source "DFID" is on the remote MDT\n",
1402 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1403 GOTO(out_put_child, rc = -EREMOTE);
1406 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1407 !mdt_object_remote(msrcdir)) {
1408 CERROR("%s: parent "DFID" is still on the same"
1409 " MDT, which should be migrated first:"
1410 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1411 PFID(mdt_object_fid(msrcdir)), -EPERM);
1412 GOTO(out_put_child, rc = -EPERM);
1415 /* 3: iterate the linkea of the object and lock all of the objects */
1416 CFS_INIT_LIST_HEAD(&lock_list);
1417 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1419 GOTO(out_put_child, rc);
1421 /* 4: lock of the object migrated object */
1422 lh_childp = &info->mti_lh[MDT_LH_OLD];
1423 mdt_lock_reg_init(lh_childp, LCK_EX);
1424 rc = mdt_object_lock(info, mold, lh_childp,
1425 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1426 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1428 GOTO(out_unlock_list, rc);
1430 ma->ma_need = MA_LMV;
1432 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1433 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1434 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1436 GOTO(out_unlock_list, rc);
1438 if ((ma->ma_valid & MA_LMV)) {
1439 struct lmv_mds_md_v1 *lmm1;
1441 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1442 lmm1 = &ma->ma_lmv->lmv_md_v1;
1443 if (lmm1->lmv_magic != LMV_MAGIC_MIGRATE) {
1444 CERROR("%s: can not migrate striped dir "DFID
1445 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1446 PFID(mdt_object_fid(mold)), -EPERM);
1447 GOTO(out_unlock_child, rc = -EPERM);
1450 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1451 GOTO(out_unlock_child, rc = -EINVAL);
1453 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1454 &lmm1->lmv_stripe_fids[1]);
1456 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1458 if (!mdt_object_remote(mnew)) {
1459 CERROR("%s: "DFID" being migrated is on this MDT:"
1460 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1461 PFID(rr->rr_fid2), -EPERM);
1462 GOTO(out_put_new, rc = -EPERM);
1465 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1466 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1467 rc = mdt_remote_object_lock(info, mnew,
1468 mdt_object_fid(mnew),
1469 &lh_tgtp->mlh_rreg_lh,
1470 lh_tgtp->mlh_rreg_mode,
1471 MDS_INODELOCK_UPDATE);
1474 GOTO(out_put_new, rc);
1477 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1480 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1481 if (!mdt_object_remote(mnew)) {
1482 CERROR("%s: Migration "DFID" is on this MDT:"
1483 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1484 PFID(rr->rr_fid2), -EXDEV);
1485 GOTO(out_put_new, rc = -EXDEV);
1490 mdt_reint_init_ma(info, ma);
1492 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1493 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1495 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1496 mdt_object_child(mold), &rr->rr_name,
1497 mdt_object_child(mnew), ma);
1499 GOTO(out_unlock_new, rc);
1501 if (lh_tgtp != NULL)
1502 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1505 mdt_object_put(info->mti_env, mnew);
1507 mdt_object_unlock(info, mold, lh_childp, rc);
1509 mdt_unlock_list(info, &lock_list, rc);
1511 mdt_object_put(info->mti_env, mold);
1513 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1515 mdt_object_put(info->mti_env, msrcdir);
1521 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1522 * 2 - src child; 3 - tgt child.
1523 * Update on disk version of src child.
1526 * For DNE phase I, only these renames are allowed
1527 * mv src_p/src_c tgt_p/tgt_c
1528 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1529 * 2. src_p and tgt_p are same directory, and tgt_c does not
1530 * exists. In this case, all of modification will happen
1531 * in the MDT where ithesource parent is, only one remote
1532 * update is needed, i.e. set c_time/m_time on the child.
1533 * And tgt_c will be still in the same MDT as the original
1536 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1537 struct mdt_lock_handle *lhc)
1539 struct mdt_reint_record *rr = &info->mti_rr;
1540 struct md_attr *ma = &info->mti_attr;
1541 struct ptlrpc_request *req = mdt_info_req(info);
1542 struct mdt_object *msrcdir;
1543 struct mdt_object *mtgtdir;
1544 struct mdt_object *mold;
1545 struct mdt_object *mnew = NULL;
1546 struct mdt_lock_handle *lh_srcdirp;
1547 struct mdt_lock_handle *lh_tgtdirp;
1548 struct mdt_lock_handle *lh_oldp = NULL;
1549 struct mdt_lock_handle *lh_newp = NULL;
1550 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1551 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1555 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1556 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1557 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1559 /* step 1: lock the source dir. */
1560 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1561 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1562 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1563 MDS_INODELOCK_UPDATE);
1564 if (IS_ERR(msrcdir))
1565 RETURN(PTR_ERR(msrcdir));
1567 rc = mdt_version_get_check_save(info, msrcdir, 0);
1569 GOTO(out_unlock_source, rc);
1571 /* step 2: find & lock the target dir. */
1572 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1573 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1574 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1575 mdt_object_get(info->mti_env, msrcdir);
1577 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1578 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1579 MDS_INODELOCK_UPDATE);
1581 GOTO(out_unlock_source, rc);
1582 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1585 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1587 if (IS_ERR(mtgtdir))
1588 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1590 /* check early, the real version will be saved after locking */
1591 rc = mdt_version_get_check(info, mtgtdir, 1);
1593 GOTO(out_put_target, rc);
1595 if (unlikely(mdt_object_remote(mtgtdir))) {
1596 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1597 "on different MDTs\n", PFID(rr->rr_fid1),
1599 GOTO(out_put_target, rc = -EXDEV);
1601 if (likely(mdt_object_exists(mtgtdir))) {
1602 /* we lock the target dir if it is local */
1603 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1604 MDS_INODELOCK_UPDATE,
1607 GOTO(out_put_target, rc);
1608 /* get and save correct version after locking */
1609 mdt_version_get_save(info, mtgtdir, 1);
1611 GOTO(out_put_target, rc = -ESTALE);
1616 /* step 3: find & lock the old object. */
1618 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1620 GOTO(out_unlock_target, rc);
1622 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1623 GOTO(out_unlock_target, rc = -EINVAL);
1625 if (!fid_is_md_operative(old_fid))
1626 GOTO(out_unlock_target, rc = -EPERM);
1628 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1630 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1632 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1633 /* save version after locking */
1634 mdt_version_get_save(info, mold, 2);
1635 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1637 /* step 4: find & lock the new object. */
1638 /* new target object may not exist now */
1639 /* lookup with version checking */
1641 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1644 /* the new_fid should have been filled at this moment */
1645 if (lu_fid_eq(old_fid, new_fid))
1646 GOTO(out_put_old, rc);
1648 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1649 lu_fid_eq(new_fid, rr->rr_fid2))
1650 GOTO(out_put_old, rc = -EINVAL);
1652 if (!fid_is_md_operative(new_fid))
1653 GOTO(out_put_old, rc = -EPERM);
1655 if (mdt_object_remote(mold)) {
1656 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1658 GOTO(out_put_old, rc = -EXDEV);
1661 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1663 GOTO(out_put_old, rc = PTR_ERR(mnew));
1665 if (mdt_object_remote(mnew)) {
1666 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1668 GOTO(out_put_new, rc = -EXDEV);
1671 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1672 mdt_lock_reg_init(lh_oldp, LCK_EX);
1673 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1674 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1676 GOTO(out_put_new, rc);
1678 /* We used to acquire MDS_INODELOCK_FULL here but we
1679 * can't do this now because a running HSM restore on
1680 * the rename onto victim will hold the layout
1681 * lock. See LU-4002. */
1683 lh_newp = &info->mti_lh[MDT_LH_NEW];
1684 mdt_lock_reg_init(lh_newp, LCK_EX);
1685 rc = mdt_object_lock(info, mnew, lh_newp,
1686 MDS_INODELOCK_LOOKUP |
1687 MDS_INODELOCK_UPDATE,
1690 GOTO(out_unlock_old, rc);
1692 /* get and save version after locking */
1693 mdt_version_get_save(info, mnew, 3);
1694 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1695 } else if (rc != -EREMOTE && rc != -ENOENT) {
1696 GOTO(out_put_old, rc);
1698 /* If mnew does not exist and mold are remote directory,
1699 * it only allows rename if they are under same directory */
1700 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1701 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1703 GOTO(out_put_old, rc = -EXDEV);
1706 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1707 mdt_lock_reg_init(lh_oldp, LCK_EX);
1708 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1709 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1711 GOTO(out_put_old, rc);
1713 mdt_enoent_version_save(info, 3);
1716 /* step 5: rename it */
1717 mdt_reint_init_ma(info, ma);
1719 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1720 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1722 /* Check if @dst is subdir of @src. */
1723 rc = mdt_rename_sanity(info, old_fid);
1725 GOTO(out_unlock_new, rc);
1728 mutex_lock(&mnew->mot_lov_mutex);
1730 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1731 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1732 mnew != NULL ? mdt_object_child(mnew) : NULL,
1733 &rr->rr_tgt_name, ma);
1736 mutex_unlock(&mnew->mot_lov_mutex);
1738 /* handle last link of tgt object */
1740 mdt_counter_incr(req, LPROC_MDT_RENAME);
1742 mdt_handle_last_unlink(info, mnew, ma);
1744 mdt_rename_counter_tally(info, info->mti_mdt, req,
1751 mdt_object_unlock(info, mnew, lh_newp, rc);
1753 mdt_object_unlock(info, mold, lh_oldp, rc);
1756 mdt_object_put(info->mti_env, mnew);
1758 mdt_object_put(info->mti_env, mold);
1760 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1762 mdt_object_put(info->mti_env, mtgtdir);
1764 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1768 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1769 struct mdt_lock_handle *lhc,
1770 enum mdt_rename_lock rename_lock)
1772 struct mdt_reint_record *rr = &info->mti_rr;
1773 struct ptlrpc_request *req = mdt_info_req(info);
1774 struct lustre_handle rename_lh = { 0 };
1778 if (info->mti_dlm_req)
1779 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1781 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1782 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1785 rc = mdt_rename_lock(info, &rename_lh, rename_lock);
1787 CERROR("%s: can't lock FS for rename: rc = %d\n",
1788 mdt_obd_name(info->mti_mdt), rc);
1792 if (rename_lock == MRL_RENAME)
1793 rc = mdt_reint_rename_internal(info, lhc);
1795 rc = mdt_reint_migrate_internal(info, lhc);
1797 if (lustre_handle_is_used(&rename_lh))
1798 mdt_rename_unlock(&rename_lh);
1803 static int mdt_reint_rename(struct mdt_thread_info *info,
1804 struct mdt_lock_handle *lhc)
1806 return mdt_reint_rename_or_migrate(info, lhc, MRL_RENAME);
1809 static int mdt_reint_migrate(struct mdt_thread_info *info,
1810 struct mdt_lock_handle *lhc)
1812 return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE);
1815 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1816 struct mdt_lock_handle *lhc);
1818 static mdt_reinter reinters[REINT_MAX] = {
1819 [REINT_SETATTR] = mdt_reint_setattr,
1820 [REINT_CREATE] = mdt_reint_create,
1821 [REINT_LINK] = mdt_reint_link,
1822 [REINT_UNLINK] = mdt_reint_unlink,
1823 [REINT_RENAME] = mdt_reint_rename,
1824 [REINT_OPEN] = mdt_reint_open,
1825 [REINT_SETXATTR] = mdt_reint_setxattr,
1826 [REINT_RMENTRY] = mdt_reint_unlink,
1827 [REINT_MIGRATE] = mdt_reint_migrate,
1830 int mdt_reint_rec(struct mdt_thread_info *info,
1831 struct mdt_lock_handle *lhc)
1836 rc = reinters[info->mti_rr.rr_opcode](info, lhc);