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, 2012, 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"
51 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
54 ma->ma_need = MA_INODE;
58 static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
59 struct mdt_object *object,
60 struct mdt_body *repbody)
64 /* for cross-ref mkdir, mds capa has been fetched from remote obj, then
65 * we won't go to below*/
66 if (repbody->valid & OBD_MD_FLMDSCAPA)
69 if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
70 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
71 struct lustre_capa *capa;
73 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
75 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
76 rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa,
79 repbody->valid |= OBD_MD_FLMDSCAPA;
86 * Get version of object by fid.
88 * Return real version or ENOENT_VERSION if object doesn't exist
90 static void mdt_obj_version_get(struct mdt_thread_info *info,
91 struct mdt_object *o, __u64 *version)
94 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
95 !fid_is_obf(mdt_object_fid(o)))
96 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
98 *version = ENOENT_VERSION;
99 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
100 PFID(mdt_object_fid(o)), *version);
104 * Check version is correct.
106 * Should be called only during replay.
108 static int mdt_version_check(struct ptlrpc_request *req,
109 __u64 version, int idx)
111 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
114 if (!exp_connect_vbr(req->rq_export))
117 LASSERT(req_is_replay(req));
118 /** VBR: version is checked always because costs nothing */
119 LASSERT(idx < PTLRPC_NUM_VERSIONS);
120 /** Sanity check for malformed buffers */
121 if (pre_ver == NULL) {
122 CERROR("No versions in request buffer\n");
123 spin_lock(&req->rq_export->exp_lock);
124 req->rq_export->exp_vbr_failed = 1;
125 spin_unlock(&req->rq_export->exp_lock);
127 } else if (pre_ver[idx] != version) {
128 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
129 pre_ver[idx], version);
130 spin_lock(&req->rq_export->exp_lock);
131 req->rq_export->exp_vbr_failed = 1;
132 spin_unlock(&req->rq_export->exp_lock);
139 * Save pre-versions in reply.
141 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
146 if (!exp_connect_vbr(req->rq_export))
149 LASSERT(!req_is_replay(req));
150 LASSERT(req->rq_repmsg != NULL);
151 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
153 reply_ver[idx] = version;
157 * Save enoent version, it is needed when it is obvious that object doesn't
158 * exist, e.g. child during create.
160 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
162 /* save version of file name for replay, it must be ENOENT here */
163 if (!req_is_replay(mdt_info_req(info))) {
164 info->mti_ver[idx] = ENOENT_VERSION;
165 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
170 * Get version from disk and save in reply buffer.
172 * Versions are saved in reply only during normal operations not replays.
174 void mdt_version_get_save(struct mdt_thread_info *info,
175 struct mdt_object *mto, int idx)
177 /* don't save versions during replay */
178 if (!req_is_replay(mdt_info_req(info))) {
179 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
180 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
185 * Get version from disk and check it, no save in reply.
187 int mdt_version_get_check(struct mdt_thread_info *info,
188 struct mdt_object *mto, int idx)
190 /* only check versions during replay */
191 if (!req_is_replay(mdt_info_req(info)))
194 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
195 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
199 * Get version from disk and check if recovery or just save.
201 int mdt_version_get_check_save(struct mdt_thread_info *info,
202 struct mdt_object *mto, int idx)
206 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
207 if (req_is_replay(mdt_info_req(info)))
208 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
211 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
216 * Lookup with version checking.
218 * This checks version of 'name'. Many reint functions uses 'name' for child not
219 * FID, therefore we need to get object by name and check its version.
221 int mdt_lookup_version_check(struct mdt_thread_info *info,
222 struct mdt_object *p, struct lu_name *lname,
223 struct lu_fid *fid, int idx)
227 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
229 /* Check version only during replay */
230 if (!req_is_replay(mdt_info_req(info)))
233 info->mti_ver[idx] = ENOENT_VERSION;
235 struct mdt_object *child;
236 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
237 if (likely(!IS_ERR(child))) {
238 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
239 mdt_object_put(info->mti_env, child);
242 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
243 return vbrc ? vbrc : rc;
248 * VBR: we save three versions in reply:
249 * 0 - parent. Check that parent version is the same during replay.
250 * 1 - name. Version of 'name' if file exists with the same name or
251 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
252 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
255 static int mdt_md_create(struct mdt_thread_info *info)
257 struct mdt_device *mdt = info->mti_mdt;
258 struct mdt_object *parent;
259 struct mdt_object *child;
260 struct mdt_lock_handle *lh;
261 struct mdt_body *repbody;
262 struct md_attr *ma = &info->mti_attr;
263 struct mdt_reint_record *rr = &info->mti_rr;
264 struct lu_name *lname;
268 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
269 rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
271 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
274 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
276 lh = &info->mti_lh[MDT_LH_PARENT];
277 mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
279 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
280 MDS_INODELOCK_UPDATE);
282 RETURN(PTR_ERR(parent));
284 rc = mdt_version_get_check_save(info, parent, 0);
286 GOTO(out_put_parent, rc);
289 * Check child name version during replay.
290 * During create replay a file may exist with same name.
292 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
293 rc = mdt_lookup_version_check(info, parent, lname,
294 &info->mti_tmp_fid1, 1);
296 GOTO(out_put_parent, rc = -EEXIST);
298 /* -ENOENT is expected here */
300 GOTO(out_put_parent, rc);
302 /* save version of file name for replay, it must be ENOENT here */
303 mdt_enoent_version_save(info, 1);
305 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
306 if (likely(!IS_ERR(child))) {
307 struct md_object *next = mdt_object_child(parent);
309 if (mdt_object_remote(child)) {
310 struct seq_server_site *ss;
311 struct lu_ucred *uc = mdt_ucred(info);
313 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
315 mdt->mdt_enable_remote_dir_gid &&
316 mdt->mdt_enable_remote_dir_gid != -1) {
317 CERROR("%s: Creating remote dir is only"
318 " permitted for administrator or"
319 " set mdt_enable_remote_dir_gid:"
321 mdt2obd_dev(mdt)->obd_name,
323 GOTO(out_put_child, rc = -EPERM);
327 ss = mdt_seq_site(mdt);
328 if (ss->ss_node_id != 0 &&
329 mdt->mdt_enable_remote_dir == 0) {
330 CERROR("%s: remote dir is only permitted on"
332 " mdt.*.enable_remote_dir=1\n",
333 mdt2obd_dev(mdt)->obd_name);
334 GOTO(out_put_child, rc = -EPERM);
336 if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
337 /* Return -EIO for old client */
338 GOTO(out_put_child, rc = -EIO);
342 ma->ma_need = MA_INODE;
344 /* capa for cross-ref will be stored here */
345 ma->ma_capa = req_capsule_server_get(info->mti_pill,
347 LASSERT(ma->ma_capa);
349 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
350 OBD_FAIL_MDS_REINT_CREATE_WRITE);
352 /* Version of child will be updated on disk. */
353 info->mti_mos = child;
354 rc = mdt_version_get_check_save(info, child, 2);
356 GOTO(out_put_child, rc);
358 /* Let lower layer know current lock mode. */
359 info->mti_spec.sp_cr_mode =
360 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
363 * Do not perform lookup sanity check. We know that name does
366 info->mti_spec.sp_cr_lookup = 0;
367 info->mti_spec.sp_feat = &dt_directory_features;
369 rc = mdo_create(info->mti_env, next, lname,
370 mdt_object_child(child),
371 &info->mti_spec, ma);
373 rc = mdt_attr_get_complex(info, child, ma);
376 /* Return fid & attr to client. */
377 if (ma->ma_valid & MA_INODE)
378 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
379 mdt_object_fid(child));
382 mdt_object_put(info->mti_env, child);
386 mdt_create_pack_capa(info, rc, child, repbody);
388 mdt_object_unlock_put(info, parent, lh, rc);
392 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
393 struct md_attr *ma, int flags)
395 struct mdt_lock_handle *lh;
396 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
397 __u64 lockpart = MDS_INODELOCK_UPDATE;
401 /* attr shouldn't be set on remote object */
402 LASSERT(!mdt_object_remote(mo));
404 lh = &info->mti_lh[MDT_LH_PARENT];
405 mdt_lock_reg_init(lh, LCK_PW);
407 /* Even though the new MDT will grant PERM lock to the old
408 * client, but the old client will almost ignore that during
409 * So it needs to revoke both LOOKUP and PERM lock here, so
410 * both new and old client can cancel the dcache */
411 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
412 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
414 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
418 if (mdt_object_exists(mo) == 0)
419 GOTO(out_unlock, rc = -ENOENT);
421 /* all attrs are packed into mti_attr in unpack_setattr */
422 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
423 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
425 /* This is only for set ctime when rename's source is on remote MDS. */
426 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
427 ma->ma_attr_flags |= MDS_VTX_BYPASS;
429 /* VBR: update version if attr changed are important for recovery */
431 /* update on-disk version of changed object */
433 rc = mdt_version_get_check_save(info, mo, 0);
435 GOTO(out_unlock, rc);
438 /* all attrs are packed into mti_attr in unpack_setattr */
439 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
441 GOTO(out_unlock, rc);
445 mdt_object_unlock(info, mo, lh, rc);
450 * Check HSM flags and add HS_DIRTY flag if relevant.
452 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
453 * and is not RELEASED.
455 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
461 /* If the file was modified, add the dirty flag */
462 ma->ma_need = MA_HSM;
463 rc = mdt_attr_get_complex(info, mo, ma);
465 CERROR("file attribute read error for "DFID": %d.\n",
466 PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
470 /* If an up2date copy exists in the backend, add dirty flag */
471 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
472 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
474 ma->ma_hsm.mh_flags |= HS_DIRTY;
475 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
477 CERROR("file attribute change error for "DFID": %d\n",
478 PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
486 static int mdt_reint_setattr(struct mdt_thread_info *info,
487 struct mdt_lock_handle *lhc)
489 struct md_attr *ma = &info->mti_attr;
490 struct mdt_reint_record *rr = &info->mti_rr;
491 struct ptlrpc_request *req = mdt_info_req(info);
492 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
493 struct mdt_file_data *mfd;
494 struct mdt_object *mo;
495 struct mdt_body *repbody;
499 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
500 (unsigned int)ma->ma_attr.la_valid);
502 if (info->mti_dlm_req)
503 ldlm_request_cancel(req, info->mti_dlm_req, 0);
505 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
508 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
509 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
511 GOTO(out, rc = PTR_ERR(mo));
513 /* start a log jounal handle if needed */
514 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
515 if ((ma->ma_attr.la_valid & LA_SIZE) ||
516 (rr->rr_flags & MRF_OPEN_TRUNC)) {
517 /* Check write access for the O_TRUNC case */
518 if (mdt_write_read(mo) < 0)
519 GOTO(out_put, rc = -ETXTBSY);
521 } else if (info->mti_ioepoch &&
522 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
523 /* Truncate case. IOEpoch is opened. */
524 rc = mdt_write_get(mo);
531 GOTO(out_put, rc = -ENOMEM);
534 mdt_ioepoch_open(info, mo, 0);
535 repbody->ioepoch = mo->mot_ioepoch;
537 mdt_object_get(info->mti_env, mo);
538 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
539 mfd->mfd_object = mo;
540 mfd->mfd_xid = req->rq_xid;
542 spin_lock(&med->med_open_lock);
543 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
544 spin_unlock(&med->med_open_lock);
545 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
548 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
550 /* SOM Attribute update case. Find the proper mfd and update
551 * SOM attributes on the proper object. */
552 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
553 LASSERT(info->mti_ioepoch);
555 spin_lock(&med->med_open_lock);
556 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
558 spin_unlock(&med->med_open_lock);
559 CDEBUG(D_INODE, "no handle for file close: "
560 "fid = "DFID": cookie = "LPX64"\n",
561 PFID(info->mti_rr.rr_fid1),
562 info->mti_ioepoch->handle.cookie);
563 GOTO(out_put, rc = -ESTALE);
565 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
566 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
568 class_handle_unhash(&mfd->mfd_handle);
569 cfs_list_del_init(&mfd->mfd_list);
570 spin_unlock(&med->med_open_lock);
572 mdt_mfd_close(info, mfd);
573 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
574 LASSERT((ma->ma_valid & MA_LOV) == 0);
575 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
578 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
579 struct lu_buf *buf = &info->mti_buf;
580 LASSERT(ma->ma_attr.la_valid == 0);
581 buf->lb_buf = ma->ma_lmm;
582 buf->lb_len = ma->ma_lmm_size;
583 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
584 buf, XATTR_NAME_LOV, 0);
590 /* If file data is modified, add the dirty flag */
591 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
592 rc = mdt_add_dirty_flag(info, mo, ma);
594 ma->ma_need = MA_INODE;
596 rc = mdt_attr_get_complex(info, mo, ma);
600 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
602 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
603 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
604 S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
605 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
606 struct lustre_capa *capa;
608 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
610 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
611 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
614 repbody->valid |= OBD_MD_FLOSSCAPA;
619 mdt_object_put(info->mti_env, mo);
622 mdt_counter_incr(req, LPROC_MDT_SETATTR);
624 mdt_client_compatibility(info);
625 rc2 = mdt_fix_reply(info);
631 static int mdt_reint_create(struct mdt_thread_info *info,
632 struct mdt_lock_handle *lhc)
634 struct ptlrpc_request *req = mdt_info_req(info);
638 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
639 RETURN(err_serious(-ESTALE));
641 if (info->mti_dlm_req)
642 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
644 LASSERT(info->mti_rr.rr_namelen > 0);
645 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
647 mdt_counter_incr(req, LPROC_MDT_MKDIR);
655 /* Special file should stay on the same node as parent. */
656 mdt_counter_incr(req, LPROC_MDT_MKNOD);
659 CERROR("%s: Unsupported mode %o\n",
660 mdt2obd_dev(info->mti_mdt)->obd_name,
661 info->mti_attr.ma_attr.la_mode);
662 RETURN(err_serious(-EOPNOTSUPP));
665 rc = mdt_md_create(info);
670 * VBR: save parent version in reply and child version getting by its name.
671 * Version of child is getting and checking during its lookup. If
673 static int mdt_reint_unlink(struct mdt_thread_info *info,
674 struct mdt_lock_handle *lhc)
676 struct mdt_reint_record *rr = &info->mti_rr;
677 struct ptlrpc_request *req = mdt_info_req(info);
678 struct md_attr *ma = &info->mti_attr;
679 struct lu_fid *child_fid = &info->mti_tmp_fid1;
680 struct mdt_object *mp;
681 struct mdt_object *mc;
682 struct mdt_lock_handle *parent_lh;
683 struct mdt_lock_handle *child_lh;
684 struct lu_name *lname;
689 DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
692 if (info->mti_dlm_req)
693 ldlm_request_cancel(req, info->mti_dlm_req, 0);
695 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
696 RETURN(err_serious(-ENOENT));
698 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
701 * step 1: Found the parent.
703 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
709 parent_lh = &info->mti_lh[MDT_LH_PARENT];
710 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
711 if (mdt_object_remote(mp)) {
712 mdt_lock_reg_init(parent_lh, LCK_EX);
713 rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
714 parent_lh->mlh_rreg_mode,
715 MDS_INODELOCK_UPDATE);
717 GOTO(put_parent, rc);
720 mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
722 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
725 GOTO(put_parent, rc);
727 rc = mdt_version_get_check_save(info, mp, 0);
729 GOTO(unlock_parent, rc);
732 /* step 2: find & lock the child */
733 /* lookup child object along with version checking */
735 rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
737 /* Name might not be able to find during resend of
738 * remote unlink, considering following case.
739 * dir_A is a remote directory, the name entry of
740 * dir_A is on MDT0, the directory is on MDT1,
742 * 1. client sends unlink req to MDT1.
743 * 2. MDT1 sends name delete update to MDT0.
744 * 3. name entry is being deleted in MDT0 synchronously.
745 * 4. MDT1 is restarted.
746 * 5. client resends unlink req to MDT1. So it can not
747 * find the name entry on MDT0 anymore.
748 * In this case, MDT1 only needs to destory the local
751 if (mdt_object_remote(mp) && rc == -ENOENT &&
752 !fid_is_zero(rr->rr_fid2) &&
753 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
755 *child_fid = *rr->rr_fid2;
757 GOTO(unlock_parent, rc);
761 if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid))
762 GOTO(unlock_parent, rc = -EPERM);
764 mdt_reint_init_ma(info, ma);
766 /* We will lock the child regardless it is local or remote. No harm. */
767 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
769 GOTO(unlock_parent, rc = PTR_ERR(mc));
771 child_lh = &info->mti_lh[MDT_LH_CHILD];
772 mdt_lock_reg_init(child_lh, LCK_EX);
773 if (mdt_object_remote(mc)) {
774 struct mdt_body *repbody;
776 if (!fid_is_zero(rr->rr_fid2)) {
777 CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
778 mdt2obd_dev(info->mti_mdt)->obd_name,
779 (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
780 GOTO(put_child, rc = -ENOENT);
782 CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
783 mdt2obd_dev(info->mti_mdt)->obd_name,
784 (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
786 if (!mdt_is_dne_client(req->rq_export))
787 /* Return -EIO for old client */
788 GOTO(put_child, rc = -EIO);
790 if (info->mti_spec.sp_rm_entry) {
791 struct lu_ucred *uc = mdt_ucred(info);
793 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
794 CERROR("%s: unlink remote entry is only "
795 "permitted for administrator: rc = %d\n",
796 mdt2obd_dev(info->mti_mdt)->obd_name,
798 GOTO(put_child, rc = -EPERM);
801 ma->ma_need = MA_INODE;
803 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
804 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
805 NULL, lname, ma, no_name);
808 /* Revoke the LOOKUP lock of the remote object granted by
809 * this MDT. Since the unlink will happen on another MDT,
810 * it will release the LOOKUP lock right away. Then What
811 * would happen if another client try to grab the LOOKUP
812 * lock at the same time with unlink XXX */
813 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
815 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
816 LASSERT(repbody != NULL);
817 repbody->fid1 = *mdt_object_fid(mc);
818 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
819 GOTO(unlock_child, rc = -EREMOTE);
820 } else if (info->mti_spec.sp_rm_entry) {
821 CERROR("%s: lfs rmdir should not be used on local dir %s\n",
822 mdt2obd_dev(info->mti_mdt)->obd_name,
823 (char *)rr->rr_name);
824 GOTO(put_child, rc = -EPERM);
827 rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
833 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
834 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
835 /* save version when object is locked */
836 mdt_version_get_save(info, mc, 1);
838 * Now we can only make sure we need MA_INODE, in mdd layer, will check
839 * whether need MA_LOV and MA_COOKIE.
841 ma->ma_need = MA_INODE;
843 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
845 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
846 mdt_object_child(mc), lname, ma, no_name);
847 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
848 rc = mdt_attr_get_complex(info, mc, ma);
850 mdt_handle_last_unlink(info, mc, ma);
852 if (ma->ma_valid & MA_INODE) {
853 switch (ma->ma_attr.la_mode & S_IFMT) {
855 mdt_counter_incr(req, LPROC_MDT_RMDIR);
863 mdt_counter_incr(req, LPROC_MDT_UNLINK);
866 LASSERTF(0, "bad file type %o unlinking\n",
867 ma->ma_attr.la_mode);
873 mdt_object_unlock(info, mc, child_lh, rc);
875 mdt_object_put(info->mti_env, mc);
877 mdt_object_unlock(info, mp, parent_lh, rc);
879 mdt_object_put(info->mti_env, mp);
885 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
888 static int mdt_reint_link(struct mdt_thread_info *info,
889 struct mdt_lock_handle *lhc)
891 struct mdt_reint_record *rr = &info->mti_rr;
892 struct ptlrpc_request *req = mdt_info_req(info);
893 struct md_attr *ma = &info->mti_attr;
894 struct mdt_object *ms;
895 struct mdt_object *mp;
896 struct mdt_lock_handle *lhs;
897 struct mdt_lock_handle *lhp;
898 struct lu_name *lname;
902 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
903 PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
905 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
906 RETURN(err_serious(-ENOENT));
908 if (info->mti_dlm_req)
909 ldlm_request_cancel(req, info->mti_dlm_req, 0);
911 /* Invalid case so return error immediately instead of
913 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
916 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
917 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
920 /* step 1: find & lock the target parent dir */
921 lhp = &info->mti_lh[MDT_LH_PARENT];
922 mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
924 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
925 MDS_INODELOCK_UPDATE);
929 rc = mdt_version_get_check_save(info, mp, 0);
931 GOTO(out_unlock_parent, rc);
933 /* step 2: find & lock the source */
934 lhs = &info->mti_lh[MDT_LH_CHILD];
935 mdt_lock_reg_init(lhs, LCK_EX);
937 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
939 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
941 if (mdt_object_remote(ms)) {
942 mdt_object_put(info->mti_env, ms);
943 CERROR("Target directory "DFID" is on another MDT\n",
945 GOTO(out_unlock_parent, rc = -EXDEV);
948 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
951 mdt_object_put(info->mti_env, ms);
952 GOTO(out_unlock_parent, rc);
955 /* step 3: link it */
956 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
957 OBD_FAIL_MDS_REINT_LINK_WRITE);
960 rc = mdt_version_get_check_save(info, ms, 1);
962 GOTO(out_unlock_child, rc);
964 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
965 /** check target version by name during replay */
966 rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
967 if (rc != 0 && rc != -ENOENT)
968 GOTO(out_unlock_child, rc);
969 /* save version of file name for replay, it must be ENOENT here */
970 if (!req_is_replay(mdt_info_req(info))) {
971 info->mti_ver[2] = ENOENT_VERSION;
972 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
975 rc = mdo_link(info->mti_env, mdt_object_child(mp),
976 mdt_object_child(ms), lname, ma);
979 mdt_counter_incr(req, LPROC_MDT_LINK);
983 mdt_object_unlock_put(info, ms, lhs, rc);
985 mdt_object_unlock_put(info, mp, lhp, rc);
989 * lock the part of the directory according to the hash of the name
990 * (lh->mlh_pdo_hash) in parallel directory lock.
992 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
993 struct mdt_lock_handle *lh,
994 struct mdt_object *obj, __u64 ibits)
996 struct ldlm_res_id *res_id = &info->mti_res_id;
997 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
998 ldlm_policy_data_t *policy = &info->mti_policy;
1002 * Finish res_id initializing by name hash marking part of
1003 * directory which is taking modification.
1005 LASSERT(lh->mlh_pdo_hash != 0);
1006 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
1007 memset(policy, 0, sizeof(*policy));
1008 policy->l_inodebits.bits = ibits;
1010 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1011 * going to be sent to client. If it is - mdt_intent_policy() path will
1012 * fix it up and turn FL_LOCAL flag off.
1014 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1015 res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1016 &info->mti_exp->exp_handle.h_cookie);
1020 static int mdt_rename_lock(struct mdt_thread_info *info,
1021 struct lustre_handle *lh)
1023 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1024 ldlm_policy_data_t *policy = &info->mti_policy;
1025 struct ldlm_res_id *res_id = &info->mti_res_id;
1030 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1032 memset(policy, 0, sizeof *policy);
1033 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1034 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 4, 53, 0)
1035 /* In phase I, we will not do cross-rename, so local BFL lock would
1038 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1040 * Current node is controller, that is mdt0, where we should
1043 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1044 LCK_EX, &flags, ldlm_blocking_ast,
1045 ldlm_completion_ast, NULL, NULL, 0,
1047 &info->mti_exp->exp_handle.h_cookie,
1050 #warning "Local rename lock is invalid for DNE phase II."
1055 static void mdt_rename_unlock(struct lustre_handle *lh)
1058 LASSERT(lustre_handle_is_used(lh));
1059 ldlm_lock_decref(lh, LCK_EX);
1064 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1065 * target. Source should not be ancestor of target dir. May be other rename
1066 * checks can be moved here later.
1068 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1070 struct mdt_reint_record *rr = &info->mti_rr;
1071 struct lu_fid dst_fid = *rr->rr_fid2;
1072 struct mdt_object *dst;
1077 LASSERT(fid_is_sane(&dst_fid));
1078 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1080 rc = mdo_is_subdir(info->mti_env,
1081 mdt_object_child(dst), fid,
1083 mdt_object_put(info->mti_env, dst);
1084 if (rc != -EREMOTE && rc < 0) {
1085 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1087 /* check the found fid */
1088 if (lu_fid_eq(&dst_fid, fid))
1094 } while (rc == -EREMOTE);
1100 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1101 * 2 - src child; 3 - tgt child.
1102 * Update on disk version of src child.
1104 static int mdt_reint_rename(struct mdt_thread_info *info,
1105 struct mdt_lock_handle *lhc)
1107 struct mdt_reint_record *rr = &info->mti_rr;
1108 struct md_attr *ma = &info->mti_attr;
1109 struct ptlrpc_request *req = mdt_info_req(info);
1110 struct mdt_object *msrcdir;
1111 struct mdt_object *mtgtdir;
1112 struct mdt_object *mold;
1113 struct mdt_object *mnew = NULL;
1114 struct mdt_lock_handle *lh_srcdirp;
1115 struct mdt_lock_handle *lh_tgtdirp;
1116 struct mdt_lock_handle *lh_oldp;
1117 struct mdt_lock_handle *lh_newp;
1118 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1119 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1120 struct lustre_handle rename_lh = { 0 };
1121 struct lu_name slname = { 0 };
1122 struct lu_name *lname;
1126 if (info->mti_dlm_req)
1127 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1129 DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s",
1130 PFID(rr->rr_fid1), rr->rr_name,
1131 PFID(rr->rr_fid2), rr->rr_tgt);
1133 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1134 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1137 rc = mdt_rename_lock(info, &rename_lh);
1139 CERROR("Can't lock FS for rename, rc %d\n", rc);
1143 lh_newp = &info->mti_lh[MDT_LH_NEW];
1145 /* step 1: lock the source dir. */
1146 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1147 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name,
1149 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1150 MDS_INODELOCK_UPDATE);
1151 if (IS_ERR(msrcdir))
1152 GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
1154 rc = mdt_version_get_check_save(info, msrcdir, 0);
1156 GOTO(out_unlock_source, rc);
1158 /* step 2: find & lock the target dir. */
1159 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1160 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt,
1162 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1163 mdt_object_get(info->mti_env, msrcdir);
1165 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1166 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1167 MDS_INODELOCK_UPDATE);
1169 GOTO(out_unlock_source, rc);
1170 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1173 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1175 if (IS_ERR(mtgtdir))
1176 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1178 /* check early, the real version will be saved after locking */
1179 rc = mdt_version_get_check(info, mtgtdir, 1);
1181 GOTO(out_put_target, rc);
1183 if (unlikely(mdt_object_remote(mtgtdir))) {
1184 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1185 "on different MDTs\n", PFID(rr->rr_fid1),
1187 GOTO(out_put_target, rc = -EXDEV);
1189 if (likely(mdt_object_exists(mtgtdir))) {
1190 /* we lock the target dir if it is local */
1191 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1192 MDS_INODELOCK_UPDATE,
1195 GOTO(out_put_target, rc);
1196 /* get and save correct version after locking */
1197 mdt_version_get_save(info, mtgtdir, 1);
1199 GOTO(out_put_target, rc = -ESTALE);
1204 /* step 3: find & lock the old object. */
1205 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1206 mdt_name_copy(&slname, lname);
1208 rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2);
1210 GOTO(out_unlock_target, rc);
1212 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1213 GOTO(out_unlock_target, rc = -EINVAL);
1215 if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid))
1216 GOTO(out_unlock_target, rc = -EPERM);
1218 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1220 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1221 if (mdt_object_remote(mold)) {
1222 mdt_object_put(info->mti_env, mold);
1223 CDEBUG(D_INFO, "Source child "DFID" is on another MDT\n",
1225 GOTO(out_unlock_target, rc = -EXDEV);
1228 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1229 mdt_lock_reg_init(lh_oldp, LCK_EX);
1230 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
1233 mdt_object_put(info->mti_env, mold);
1234 GOTO(out_unlock_target, rc);
1237 info->mti_mos = mold;
1238 /* save version after locking */
1239 mdt_version_get_save(info, mold, 2);
1240 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1242 /* step 4: find & lock the new object. */
1243 /* new target object may not exist now */
1244 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
1245 /* lookup with version checking */
1247 rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3);
1249 /* the new_fid should have been filled at this moment */
1250 if (lu_fid_eq(old_fid, new_fid))
1251 GOTO(out_unlock_old, rc);
1253 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1254 lu_fid_eq(new_fid, rr->rr_fid2))
1255 GOTO(out_unlock_old, rc = -EINVAL);
1257 if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid))
1258 GOTO(out_unlock_old, rc = -EPERM);
1260 mdt_lock_reg_init(lh_newp, LCK_EX);
1261 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1263 GOTO(out_unlock_old, rc = PTR_ERR(mnew));
1265 if (mdt_object_remote(mnew)) {
1266 mdt_object_put(info->mti_env, mnew);
1267 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1269 GOTO(out_unlock_old, rc = -EXDEV);
1272 rc = mdt_object_lock(info, mnew, lh_newp,
1273 MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
1275 mdt_object_put(info->mti_env, mnew);
1276 GOTO(out_unlock_old, rc);
1278 /* get and save version after locking */
1279 mdt_version_get_save(info, mnew, 3);
1280 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1281 } else if (rc != -EREMOTE && rc != -ENOENT) {
1282 GOTO(out_unlock_old, rc);
1284 mdt_enoent_version_save(info, 3);
1287 /* step 5: rename it */
1288 mdt_reint_init_ma(info, ma);
1290 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1291 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1294 /* Check if @dst is subdir of @src. */
1295 rc = mdt_rename_sanity(info, old_fid);
1297 GOTO(out_unlock_new, rc);
1299 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1300 mdt_object_child(mtgtdir), old_fid, &slname,
1301 (mnew ? mdt_object_child(mnew) : NULL),
1304 /* handle last link of tgt object */
1306 mdt_counter_incr(req, LPROC_MDT_RENAME);
1308 mdt_handle_last_unlink(info, mnew, ma);
1310 mdt_rename_counter_tally(info, info->mti_mdt, req,
1317 mdt_object_unlock_put(info, mnew, lh_newp, rc);
1319 mdt_object_unlock_put(info, mold, lh_oldp, rc);
1321 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1323 mdt_object_put(info->mti_env, mtgtdir);
1325 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1327 if (lustre_handle_is_used(&rename_lh))
1328 mdt_rename_unlock(&rename_lh);
1332 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1333 struct mdt_lock_handle *lhc);
1335 static mdt_reinter reinters[REINT_MAX] = {
1336 [REINT_SETATTR] = mdt_reint_setattr,
1337 [REINT_CREATE] = mdt_reint_create,
1338 [REINT_LINK] = mdt_reint_link,
1339 [REINT_UNLINK] = mdt_reint_unlink,
1340 [REINT_RENAME] = mdt_reint_rename,
1341 [REINT_OPEN] = mdt_reint_open,
1342 [REINT_SETXATTR] = mdt_reint_setxattr,
1343 [REINT_RMENTRY] = mdt_reint_unlink
1346 int mdt_reint_rec(struct mdt_thread_info *info,
1347 struct mdt_lock_handle *lhc)
1352 rc = reinters[info->mti_rr.rr_opcode](info, lhc);