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"
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 mdt_obd_name(mdt), -EPERM);
322 GOTO(out_put_child, rc = -EPERM);
326 ss = mdt_seq_site(mdt);
327 if (ss->ss_node_id != 0 &&
328 mdt->mdt_enable_remote_dir == 0) {
329 CERROR("%s: remote dir is only permitted on"
331 " mdt.*.enable_remote_dir=1\n",
333 GOTO(out_put_child, rc = -EPERM);
335 if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
336 /* Return -EIO for old client */
337 GOTO(out_put_child, rc = -EIO);
341 ma->ma_need = MA_INODE;
343 /* capa for cross-ref will be stored here */
344 ma->ma_capa = req_capsule_server_get(info->mti_pill,
346 LASSERT(ma->ma_capa);
348 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
349 OBD_FAIL_MDS_REINT_CREATE_WRITE);
351 /* Version of child will be updated on disk. */
352 info->mti_mos = child;
353 rc = mdt_version_get_check_save(info, child, 2);
355 GOTO(out_put_child, rc);
357 /* Let lower layer know current lock mode. */
358 info->mti_spec.sp_cr_mode =
359 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
362 * Do not perform lookup sanity check. We know that name does
365 info->mti_spec.sp_cr_lookup = 0;
366 info->mti_spec.sp_feat = &dt_directory_features;
368 rc = mdo_create(info->mti_env, next, lname,
369 mdt_object_child(child),
370 &info->mti_spec, ma);
372 rc = mdt_attr_get_complex(info, child, ma);
375 /* Return fid & attr to client. */
376 if (ma->ma_valid & MA_INODE)
377 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
378 mdt_object_fid(child));
381 mdt_object_put(info->mti_env, child);
385 mdt_create_pack_capa(info, rc, child, repbody);
387 mdt_object_unlock_put(info, parent, lh, rc);
391 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
392 struct md_attr *ma, int flags)
394 struct mdt_lock_handle *lh;
395 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
396 __u64 lockpart = MDS_INODELOCK_UPDATE;
400 /* attr shouldn't be set on remote object */
401 LASSERT(!mdt_object_remote(mo));
403 lh = &info->mti_lh[MDT_LH_PARENT];
404 mdt_lock_reg_init(lh, LCK_PW);
406 /* Even though the new MDT will grant PERM lock to the old
407 * client, but the old client will almost ignore that during
408 * So it needs to revoke both LOOKUP and PERM lock here, so
409 * both new and old client can cancel the dcache */
410 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
411 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
413 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
417 if (mdt_object_exists(mo) == 0)
418 GOTO(out_unlock, rc = -ENOENT);
420 /* all attrs are packed into mti_attr in unpack_setattr */
421 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
422 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
424 /* This is only for set ctime when rename's source is on remote MDS. */
425 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
426 ma->ma_attr_flags |= MDS_VTX_BYPASS;
428 /* VBR: update version if attr changed are important for recovery */
430 /* update on-disk version of changed object */
432 rc = mdt_version_get_check_save(info, mo, 0);
434 GOTO(out_unlock, rc);
437 /* all attrs are packed into mti_attr in unpack_setattr */
438 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
440 GOTO(out_unlock, rc);
444 mdt_object_unlock(info, mo, lh, rc);
449 * Check HSM flags and add HS_DIRTY flag if relevant.
451 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
452 * and is not RELEASED.
454 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
460 /* If the file was modified, add the dirty flag */
461 ma->ma_need = MA_HSM;
462 rc = mdt_attr_get_complex(info, mo, ma);
464 CERROR("file attribute read error for "DFID": %d.\n",
465 PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
469 /* If an up2date copy exists in the backend, add dirty flag */
470 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
471 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
473 ma->ma_hsm.mh_flags |= HS_DIRTY;
474 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
476 CERROR("file attribute change error for "DFID": %d\n",
477 PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
485 static int mdt_reint_setattr(struct mdt_thread_info *info,
486 struct mdt_lock_handle *lhc)
488 struct md_attr *ma = &info->mti_attr;
489 struct mdt_reint_record *rr = &info->mti_rr;
490 struct ptlrpc_request *req = mdt_info_req(info);
491 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
492 struct mdt_file_data *mfd;
493 struct mdt_object *mo;
494 struct mdt_body *repbody;
498 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
499 (unsigned int)ma->ma_attr.la_valid);
501 if (info->mti_dlm_req)
502 ldlm_request_cancel(req, info->mti_dlm_req, 0);
504 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
505 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
507 GOTO(out, rc = PTR_ERR(mo));
509 /* start a log jounal handle if needed */
510 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
511 if ((ma->ma_attr.la_valid & LA_SIZE) ||
512 (rr->rr_flags & MRF_OPEN_TRUNC)) {
513 /* Check write access for the O_TRUNC case */
514 if (mdt_write_read(mo) < 0)
515 GOTO(out_put, rc = -ETXTBSY);
517 } else if (info->mti_ioepoch &&
518 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
519 /* Truncate case. IOEpoch is opened. */
520 rc = mdt_write_get(mo);
527 GOTO(out_put, rc = -ENOMEM);
530 mdt_ioepoch_open(info, mo, 0);
531 repbody->ioepoch = mo->mot_ioepoch;
533 mdt_object_get(info->mti_env, mo);
534 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
535 mfd->mfd_object = mo;
536 mfd->mfd_xid = req->rq_xid;
538 spin_lock(&med->med_open_lock);
539 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
540 spin_unlock(&med->med_open_lock);
541 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
544 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
546 /* SOM Attribute update case. Find the proper mfd and update
547 * SOM attributes on the proper object. */
548 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
549 LASSERT(info->mti_ioepoch);
551 spin_lock(&med->med_open_lock);
552 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
554 spin_unlock(&med->med_open_lock);
555 CDEBUG(D_INODE, "no handle for file close: "
556 "fid = "DFID": cookie = "LPX64"\n",
557 PFID(info->mti_rr.rr_fid1),
558 info->mti_ioepoch->handle.cookie);
559 GOTO(out_put, rc = -ESTALE);
561 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
562 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
564 class_handle_unhash(&mfd->mfd_handle);
565 cfs_list_del_init(&mfd->mfd_list);
566 spin_unlock(&med->med_open_lock);
568 mdt_mfd_close(info, mfd);
569 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
570 LASSERT((ma->ma_valid & MA_LOV) == 0);
571 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
574 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
575 struct lu_buf *buf = &info->mti_buf;
576 LASSERT(ma->ma_attr.la_valid == 0);
577 buf->lb_buf = ma->ma_lmm;
578 buf->lb_len = ma->ma_lmm_size;
579 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
580 buf, XATTR_NAME_LOV, 0);
586 /* If file data is modified, add the dirty flag */
587 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
588 rc = mdt_add_dirty_flag(info, mo, ma);
590 ma->ma_need = MA_INODE;
592 rc = mdt_attr_get_complex(info, mo, ma);
596 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
598 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
599 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
600 S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
601 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
602 struct lustre_capa *capa;
604 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
606 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
607 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
610 repbody->valid |= OBD_MD_FLOSSCAPA;
615 mdt_object_put(info->mti_env, mo);
618 mdt_counter_incr(req, LPROC_MDT_SETATTR);
620 mdt_client_compatibility(info);
621 rc2 = mdt_fix_reply(info);
627 static int mdt_reint_create(struct mdt_thread_info *info,
628 struct mdt_lock_handle *lhc)
630 struct ptlrpc_request *req = mdt_info_req(info);
634 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
635 RETURN(err_serious(-ESTALE));
637 if (info->mti_dlm_req)
638 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
640 LASSERT(info->mti_rr.rr_namelen > 0);
641 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
643 mdt_counter_incr(req, LPROC_MDT_MKDIR);
651 /* Special file should stay on the same node as parent. */
652 mdt_counter_incr(req, LPROC_MDT_MKNOD);
655 CERROR("%s: Unsupported mode %o\n",
656 mdt_obd_name(info->mti_mdt),
657 info->mti_attr.ma_attr.la_mode);
658 RETURN(err_serious(-EOPNOTSUPP));
661 rc = mdt_md_create(info);
666 * VBR: save parent version in reply and child version getting by its name.
667 * Version of child is getting and checking during its lookup. If
669 static int mdt_reint_unlink(struct mdt_thread_info *info,
670 struct mdt_lock_handle *lhc)
672 struct mdt_reint_record *rr = &info->mti_rr;
673 struct ptlrpc_request *req = mdt_info_req(info);
674 struct md_attr *ma = &info->mti_attr;
675 struct lu_fid *child_fid = &info->mti_tmp_fid1;
676 struct mdt_object *mp;
677 struct mdt_object *mc;
678 struct mdt_lock_handle *parent_lh;
679 struct mdt_lock_handle *child_lh;
680 struct lu_name *lname;
685 DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
688 if (info->mti_dlm_req)
689 ldlm_request_cancel(req, info->mti_dlm_req, 0);
691 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
692 RETURN(err_serious(-ENOENT));
694 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
697 * step 1: Found the parent.
699 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
705 parent_lh = &info->mti_lh[MDT_LH_PARENT];
706 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
707 if (mdt_object_remote(mp)) {
708 mdt_lock_reg_init(parent_lh, LCK_EX);
709 rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
710 parent_lh->mlh_rreg_mode,
711 MDS_INODELOCK_UPDATE);
713 GOTO(put_parent, rc);
716 mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
718 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
721 GOTO(put_parent, rc);
723 rc = mdt_version_get_check_save(info, mp, 0);
725 GOTO(unlock_parent, rc);
728 /* step 2: find & lock the child */
729 /* lookup child object along with version checking */
731 rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
733 /* Name might not be able to find during resend of
734 * remote unlink, considering following case.
735 * dir_A is a remote directory, the name entry of
736 * dir_A is on MDT0, the directory is on MDT1,
738 * 1. client sends unlink req to MDT1.
739 * 2. MDT1 sends name delete update to MDT0.
740 * 3. name entry is being deleted in MDT0 synchronously.
741 * 4. MDT1 is restarted.
742 * 5. client resends unlink req to MDT1. So it can not
743 * find the name entry on MDT0 anymore.
744 * In this case, MDT1 only needs to destory the local
747 if (mdt_object_remote(mp) && rc == -ENOENT &&
748 !fid_is_zero(rr->rr_fid2) &&
749 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
751 *child_fid = *rr->rr_fid2;
753 GOTO(unlock_parent, rc);
757 if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid))
758 GOTO(unlock_parent, rc = -EPERM);
760 mdt_reint_init_ma(info, ma);
762 /* We will lock the child regardless it is local or remote. No harm. */
763 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
765 GOTO(unlock_parent, rc = PTR_ERR(mc));
767 child_lh = &info->mti_lh[MDT_LH_CHILD];
768 mdt_lock_reg_init(child_lh, LCK_EX);
769 if (mdt_object_remote(mc)) {
770 struct mdt_body *repbody;
772 if (!fid_is_zero(rr->rr_fid2)) {
773 CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
774 mdt_obd_name(info->mti_mdt),
775 (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
776 GOTO(put_child, rc = -ENOENT);
778 CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
779 mdt_obd_name(info->mti_mdt),
780 (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
782 if (!mdt_is_dne_client(req->rq_export))
783 /* Return -EIO for old client */
784 GOTO(put_child, rc = -EIO);
786 if (info->mti_spec.sp_rm_entry) {
787 struct lu_ucred *uc = mdt_ucred(info);
789 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
790 CERROR("%s: unlink remote entry is only "
791 "permitted for administrator: rc = %d\n",
792 mdt_obd_name(info->mti_mdt),
794 GOTO(put_child, rc = -EPERM);
797 ma->ma_need = MA_INODE;
799 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
800 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
801 NULL, lname, ma, no_name);
804 /* Revoke the LOOKUP lock of the remote object granted by
805 * this MDT. Since the unlink will happen on another MDT,
806 * it will release the LOOKUP lock right away. Then What
807 * would happen if another client try to grab the LOOKUP
808 * lock at the same time with unlink XXX */
809 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
811 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
812 LASSERT(repbody != NULL);
813 repbody->fid1 = *mdt_object_fid(mc);
814 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
815 GOTO(unlock_child, rc = -EREMOTE);
816 } else if (info->mti_spec.sp_rm_entry) {
818 CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n",
819 mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc);
823 rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
829 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
830 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
831 /* save version when object is locked */
832 mdt_version_get_save(info, mc, 1);
834 * Now we can only make sure we need MA_INODE, in mdd layer, will check
835 * whether need MA_LOV and MA_COOKIE.
837 ma->ma_need = MA_INODE;
839 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
841 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
842 mdt_object_child(mc), lname, ma, no_name);
843 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
844 rc = mdt_attr_get_complex(info, mc, ma);
846 mdt_handle_last_unlink(info, mc, ma);
848 if (ma->ma_valid & MA_INODE) {
849 switch (ma->ma_attr.la_mode & S_IFMT) {
851 mdt_counter_incr(req, LPROC_MDT_RMDIR);
859 mdt_counter_incr(req, LPROC_MDT_UNLINK);
862 LASSERTF(0, "bad file type %o unlinking\n",
863 ma->ma_attr.la_mode);
869 mdt_object_unlock(info, mc, child_lh, rc);
871 mdt_object_put(info->mti_env, mc);
873 mdt_object_unlock(info, mp, parent_lh, rc);
875 mdt_object_put(info->mti_env, mp);
881 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
884 static int mdt_reint_link(struct mdt_thread_info *info,
885 struct mdt_lock_handle *lhc)
887 struct mdt_reint_record *rr = &info->mti_rr;
888 struct ptlrpc_request *req = mdt_info_req(info);
889 struct md_attr *ma = &info->mti_attr;
890 struct mdt_object *ms;
891 struct mdt_object *mp;
892 struct mdt_lock_handle *lhs;
893 struct mdt_lock_handle *lhp;
894 struct lu_name *lname;
898 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
899 PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
901 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
902 RETURN(err_serious(-ENOENT));
904 if (info->mti_dlm_req)
905 ldlm_request_cancel(req, info->mti_dlm_req, 0);
907 /* Invalid case so return error immediately instead of
909 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
912 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
913 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
916 /* step 1: find & lock the target parent dir */
917 lhp = &info->mti_lh[MDT_LH_PARENT];
918 mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
920 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
921 MDS_INODELOCK_UPDATE);
925 rc = mdt_version_get_check_save(info, mp, 0);
927 GOTO(out_unlock_parent, rc);
929 /* step 2: find & lock the source */
930 lhs = &info->mti_lh[MDT_LH_CHILD];
931 mdt_lock_reg_init(lhs, LCK_EX);
933 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
935 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
937 if (mdt_object_remote(ms)) {
938 mdt_object_put(info->mti_env, ms);
939 CERROR("Target directory "DFID" is on another MDT\n",
941 GOTO(out_unlock_parent, rc = -EXDEV);
944 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
947 mdt_object_put(info->mti_env, ms);
948 GOTO(out_unlock_parent, rc);
951 /* step 3: link it */
952 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
953 OBD_FAIL_MDS_REINT_LINK_WRITE);
956 rc = mdt_version_get_check_save(info, ms, 1);
958 GOTO(out_unlock_child, rc);
960 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
961 /** check target version by name during replay */
962 rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
963 if (rc != 0 && rc != -ENOENT)
964 GOTO(out_unlock_child, rc);
965 /* save version of file name for replay, it must be ENOENT here */
966 if (!req_is_replay(mdt_info_req(info))) {
967 info->mti_ver[2] = ENOENT_VERSION;
968 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
971 rc = mdo_link(info->mti_env, mdt_object_child(mp),
972 mdt_object_child(ms), lname, ma);
975 mdt_counter_incr(req, LPROC_MDT_LINK);
979 mdt_object_unlock_put(info, ms, lhs, rc);
981 mdt_object_unlock_put(info, mp, lhp, rc);
985 * lock the part of the directory according to the hash of the name
986 * (lh->mlh_pdo_hash) in parallel directory lock.
988 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
989 struct mdt_lock_handle *lh,
990 struct mdt_object *obj, __u64 ibits)
992 struct ldlm_res_id *res_id = &info->mti_res_id;
993 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
994 ldlm_policy_data_t *policy = &info->mti_policy;
998 * Finish res_id initializing by name hash marking part of
999 * directory which is taking modification.
1001 LASSERT(lh->mlh_pdo_hash != 0);
1002 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
1003 memset(policy, 0, sizeof(*policy));
1004 policy->l_inodebits.bits = ibits;
1006 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1007 * going to be sent to client. If it is - mdt_intent_policy() path will
1008 * fix it up and turn FL_LOCAL flag off.
1010 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1011 res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1012 &info->mti_exp->exp_handle.h_cookie);
1016 static int mdt_rename_lock(struct mdt_thread_info *info,
1017 struct lustre_handle *lh)
1019 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1020 ldlm_policy_data_t *policy = &info->mti_policy;
1021 struct ldlm_res_id *res_id = &info->mti_res_id;
1026 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1028 memset(policy, 0, sizeof *policy);
1029 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1030 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 4, 53, 0)
1031 /* In phase I, we will not do cross-rename, so local BFL lock would
1034 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1036 * Current node is controller, that is mdt0, where we should
1039 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1040 LCK_EX, &flags, ldlm_blocking_ast,
1041 ldlm_completion_ast, NULL, NULL, 0,
1043 &info->mti_exp->exp_handle.h_cookie,
1046 #warning "Local rename lock is invalid for DNE phase II."
1051 static void mdt_rename_unlock(struct lustre_handle *lh)
1054 LASSERT(lustre_handle_is_used(lh));
1055 ldlm_lock_decref(lh, LCK_EX);
1060 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1061 * target. Source should not be ancestor of target dir. May be other rename
1062 * checks can be moved here later.
1064 static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
1066 struct mdt_reint_record *rr = &info->mti_rr;
1067 struct lu_fid dst_fid = *rr->rr_fid2;
1068 struct mdt_object *dst;
1073 LASSERT(fid_is_sane(&dst_fid));
1074 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1076 rc = mdo_is_subdir(info->mti_env,
1077 mdt_object_child(dst), fid,
1079 mdt_object_put(info->mti_env, dst);
1080 if (rc != -EREMOTE && rc < 0) {
1081 CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
1083 /* check the found fid */
1084 if (lu_fid_eq(&dst_fid, fid))
1090 } while (rc == -EREMOTE);
1096 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1097 * 2 - src child; 3 - tgt child.
1098 * Update on disk version of src child.
1101 * For DNE phase I, only these renames are allowed
1102 * mv src_p/src_c tgt_p/tgt_c
1103 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1104 * 2. src_p and tgt_p are same directory, and tgt_c does not
1105 * exists. In this case, all of modification will happen
1106 * in the MDT where ithesource parent is, only one remote
1107 * update is needed, i.e. set c_time/m_time on the child.
1108 * And tgt_c will be still in the same MDT as the original
1111 static int mdt_reint_rename(struct mdt_thread_info *info,
1112 struct mdt_lock_handle *lhc)
1114 struct mdt_reint_record *rr = &info->mti_rr;
1115 struct md_attr *ma = &info->mti_attr;
1116 struct ptlrpc_request *req = mdt_info_req(info);
1117 struct mdt_object *msrcdir;
1118 struct mdt_object *mtgtdir;
1119 struct mdt_object *mold;
1120 struct mdt_object *mnew = NULL;
1121 struct mdt_lock_handle *lh_srcdirp;
1122 struct mdt_lock_handle *lh_tgtdirp;
1123 struct mdt_lock_handle *lh_oldp;
1124 struct mdt_lock_handle *lh_newp;
1125 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1126 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1127 struct lustre_handle rename_lh = { 0 };
1128 struct lu_name slname = { 0 };
1129 struct lu_name *lname;
1133 if (info->mti_dlm_req)
1134 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1136 DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s",
1137 PFID(rr->rr_fid1), rr->rr_name,
1138 PFID(rr->rr_fid2), rr->rr_tgt);
1140 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1141 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1144 rc = mdt_rename_lock(info, &rename_lh);
1146 CERROR("Can't lock FS for rename, rc %d\n", rc);
1150 lh_newp = &info->mti_lh[MDT_LH_NEW];
1152 /* step 1: lock the source dir. */
1153 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1154 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name,
1156 msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
1157 MDS_INODELOCK_UPDATE);
1158 if (IS_ERR(msrcdir))
1159 GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
1161 rc = mdt_version_get_check_save(info, msrcdir, 0);
1163 GOTO(out_unlock_source, rc);
1165 /* step 2: find & lock the target dir. */
1166 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1167 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt,
1169 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1170 mdt_object_get(info->mti_env, msrcdir);
1172 if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
1173 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1174 MDS_INODELOCK_UPDATE);
1176 GOTO(out_unlock_source, rc);
1177 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1180 mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
1182 if (IS_ERR(mtgtdir))
1183 GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
1185 /* check early, the real version will be saved after locking */
1186 rc = mdt_version_get_check(info, mtgtdir, 1);
1188 GOTO(out_put_target, rc);
1190 if (unlikely(mdt_object_remote(mtgtdir))) {
1191 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1192 "on different MDTs\n", PFID(rr->rr_fid1),
1194 GOTO(out_put_target, rc = -EXDEV);
1196 if (likely(mdt_object_exists(mtgtdir))) {
1197 /* we lock the target dir if it is local */
1198 rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
1199 MDS_INODELOCK_UPDATE,
1202 GOTO(out_put_target, rc);
1203 /* get and save correct version after locking */
1204 mdt_version_get_save(info, mtgtdir, 1);
1206 GOTO(out_put_target, rc = -ESTALE);
1211 /* step 3: find & lock the old object. */
1212 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1213 mdt_name_copy(&slname, lname);
1215 rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2);
1217 GOTO(out_unlock_target, rc);
1219 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1220 GOTO(out_unlock_target, rc = -EINVAL);
1222 if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid))
1223 GOTO(out_unlock_target, rc = -EPERM);
1225 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1227 GOTO(out_unlock_target, rc = PTR_ERR(mold));
1229 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1230 mdt_lock_reg_init(lh_oldp, LCK_EX);
1231 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
1234 mdt_object_put(info->mti_env, mold);
1235 GOTO(out_unlock_target, rc);
1238 info->mti_mos = mold;
1239 /* save version after locking */
1240 mdt_version_get_save(info, mold, 2);
1241 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1243 /* step 4: find & lock the new object. */
1244 /* new target object may not exist now */
1245 lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
1246 /* lookup with version checking */
1248 rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3);
1250 /* the new_fid should have been filled at this moment */
1251 if (lu_fid_eq(old_fid, new_fid))
1252 GOTO(out_unlock_old, rc);
1254 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1255 lu_fid_eq(new_fid, rr->rr_fid2))
1256 GOTO(out_unlock_old, rc = -EINVAL);
1258 if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid))
1259 GOTO(out_unlock_old, rc = -EPERM);
1261 if (mdt_object_remote(mold)) {
1262 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1264 GOTO(out_unlock_old, rc = -EXDEV);
1267 mdt_lock_reg_init(lh_newp, LCK_EX);
1268 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1270 GOTO(out_unlock_old, rc = PTR_ERR(mnew));
1272 if (mdt_object_remote(mnew)) {
1273 mdt_object_put(info->mti_env, mnew);
1274 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1276 GOTO(out_unlock_old, rc = -EXDEV);
1279 rc = mdt_object_lock(info, mnew, lh_newp,
1280 MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
1282 mdt_object_put(info->mti_env, mnew);
1283 GOTO(out_unlock_old, rc);
1285 /* get and save version after locking */
1286 mdt_version_get_save(info, mnew, 3);
1287 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1288 } else if (rc != -EREMOTE && rc != -ENOENT) {
1289 GOTO(out_unlock_old, rc);
1291 mdt_enoent_version_save(info, 3);
1294 /* step 5: rename it */
1295 mdt_reint_init_ma(info, ma);
1297 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1298 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1301 /* Check if @dst is subdir of @src. */
1302 rc = mdt_rename_sanity(info, old_fid);
1304 GOTO(out_unlock_new, rc);
1306 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1307 mdt_object_child(mtgtdir), old_fid, &slname,
1308 (mnew ? mdt_object_child(mnew) : NULL),
1311 /* handle last link of tgt object */
1313 mdt_counter_incr(req, LPROC_MDT_RENAME);
1315 mdt_handle_last_unlink(info, mnew, ma);
1317 mdt_rename_counter_tally(info, info->mti_mdt, req,
1324 mdt_object_unlock_put(info, mnew, lh_newp, rc);
1326 mdt_object_unlock_put(info, mold, lh_oldp, rc);
1328 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1330 mdt_object_put(info->mti_env, mtgtdir);
1332 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1334 if (lustre_handle_is_used(&rename_lh))
1335 mdt_rename_unlock(&rename_lh);
1339 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1340 struct mdt_lock_handle *lhc);
1342 static mdt_reinter reinters[REINT_MAX] = {
1343 [REINT_SETATTR] = mdt_reint_setattr,
1344 [REINT_CREATE] = mdt_reint_create,
1345 [REINT_LINK] = mdt_reint_link,
1346 [REINT_UNLINK] = mdt_reint_unlink,
1347 [REINT_RENAME] = mdt_reint_rename,
1348 [REINT_OPEN] = mdt_reint_open,
1349 [REINT_SETXATTR] = mdt_reint_setxattr,
1350 [REINT_RMENTRY] = mdt_reint_unlink
1353 int mdt_reint_rec(struct mdt_thread_info *info,
1354 struct mdt_lock_handle *lhc)
1359 rc = reinters[info->mti_rr.rr_opcode](info, lhc);