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 mdt_lock_handle *s0_lh,
403 struct mdt_object *s0_obj,
404 struct ldlm_enqueue_info *einfo)
406 ldlm_policy_data_t *policy = &mti->mti_policy;
410 if (!S_ISDIR(obj->mot_header.loh_attr))
413 /* Unlock stripe 0 */
414 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
415 LASSERT(s0_obj != NULL);
416 mdt_object_unlock_put(mti, s0_obj, s0_lh, 1);
419 memset(policy, 0, sizeof(*policy));
420 policy->l_inodebits.bits = ibits;
422 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
428 * Lock slave stripes if necessary, the lock handles of slave stripes
429 * will be stored in einfo->ei_cbdata.
431 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
432 ldlm_mode_t mode, __u64 ibits,
433 struct mdt_lock_handle *s0_lh,
434 struct mdt_object **s0_objp,
435 struct ldlm_enqueue_info *einfo)
437 ldlm_policy_data_t *policy = &mti->mti_policy;
438 struct lu_buf *buf = &mti->mti_buf;
439 struct lmv_mds_md_v1 *lmv;
440 struct lu_fid *fid = &mti->mti_tmp_fid1;
444 if (!S_ISDIR(obj->mot_header.loh_attr))
447 buf->lb_buf = mti->mti_xattr_buf;
448 buf->lb_len = sizeof(mti->mti_xattr_buf);
449 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
452 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
454 buf->lb_buf = mti->mti_big_lmm;
455 buf->lb_len = mti->mti_big_lmmsize;
459 if (rc == -ENODATA || rc == -ENOENT)
466 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
469 /* Sigh, 0_stripe and master object are different
470 * object, though they are in the same MDT, to avoid
471 * adding osd_object_lock here, so we will enqueue the
472 * stripe0 lock in MDT0 for now */
473 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
474 *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits);
475 if (IS_ERR(*s0_objp))
476 RETURN(PTR_ERR(*s0_objp));
478 memset(einfo, 0, sizeof(*einfo));
479 einfo->ei_type = LDLM_IBITS;
480 einfo->ei_mode = mode;
481 einfo->ei_cb_bl = mdt_remote_blocking_ast;
482 einfo->ei_cb_cp = ldlm_completion_ast;
483 einfo->ei_enq_slave = 1;
484 memset(policy, 0, sizeof(*policy));
485 policy->l_inodebits.bits = ibits;
487 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
492 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
493 struct md_attr *ma, int flags)
495 struct mdt_lock_handle *lh;
496 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
497 __u64 lockpart = MDS_INODELOCK_UPDATE;
498 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
499 struct mdt_lock_handle *s0_lh;
500 struct mdt_object *s0_obj = NULL;
504 /* attr shouldn't be set on remote object */
505 LASSERT(!mdt_object_remote(mo));
507 lh = &info->mti_lh[MDT_LH_PARENT];
508 mdt_lock_reg_init(lh, LCK_PW);
510 /* Even though the new MDT will grant PERM lock to the old
511 * client, but the old client will almost ignore that during
512 * So it needs to revoke both LOOKUP and PERM lock here, so
513 * both new and old client can cancel the dcache */
514 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
515 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
517 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
521 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
522 mdt_lock_reg_init(s0_lh, LCK_EX);
523 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
525 GOTO(out_unlock, rc);
527 if (mdt_object_exists(mo) == 0)
528 GOTO(out_unlock, rc = -ENOENT);
530 /* all attrs are packed into mti_attr in unpack_setattr */
531 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
532 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
534 /* This is only for set ctime when rename's source is on remote MDS. */
535 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
536 ma->ma_attr_flags |= MDS_VTX_BYPASS;
538 /* VBR: update version if attr changed are important for recovery */
540 /* update on-disk version of changed object */
541 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
542 rc = mdt_version_get_check_save(info, mo, 0);
544 GOTO(out_unlock, rc);
547 /* Ensure constant striping during chown(). See LU-2789. */
548 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
549 mutex_lock(&mo->mot_lov_mutex);
551 /* all attrs are packed into mti_attr in unpack_setattr */
552 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
554 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
555 mutex_unlock(&mo->mot_lov_mutex);
558 GOTO(out_unlock, rc);
562 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
563 mdt_object_unlock(info, mo, lh, rc);
568 * Check HSM flags and add HS_DIRTY flag if relevant.
570 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
571 * and is not RELEASED.
573 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
579 /* If the file was modified, add the dirty flag */
580 ma->ma_need = MA_HSM;
581 rc = mdt_attr_get_complex(info, mo, ma);
583 CERROR("file attribute read error for "DFID": %d.\n",
584 PFID(mdt_object_fid(mo)), rc);
588 /* If an up2date copy exists in the backend, add dirty flag */
589 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
590 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
591 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
593 ma->ma_hsm.mh_flags |= HS_DIRTY;
595 mdt_lock_reg_init(lh, LCK_PW);
596 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
601 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
603 CERROR("file attribute change error for "DFID": %d\n",
604 PFID(mdt_object_fid(mo)), rc);
605 mdt_object_unlock(info, mo, lh, rc);
611 static int mdt_reint_setattr(struct mdt_thread_info *info,
612 struct mdt_lock_handle *lhc)
614 struct md_attr *ma = &info->mti_attr;
615 struct mdt_reint_record *rr = &info->mti_rr;
616 struct ptlrpc_request *req = mdt_info_req(info);
617 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
618 struct mdt_file_data *mfd;
619 struct mdt_object *mo;
620 struct mdt_body *repbody;
624 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
625 (unsigned int)ma->ma_attr.la_valid);
627 if (info->mti_dlm_req)
628 ldlm_request_cancel(req, info->mti_dlm_req, 0);
630 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
631 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
633 GOTO(out, rc = PTR_ERR(mo));
635 /* start a log jounal handle if needed */
636 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
637 if ((ma->ma_attr.la_valid & LA_SIZE) ||
638 (rr->rr_flags & MRF_OPEN_TRUNC)) {
639 /* Check write access for the O_TRUNC case */
640 if (mdt_write_read(mo) < 0)
641 GOTO(out_put, rc = -ETXTBSY);
643 } else if (info->mti_ioepoch &&
644 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
645 /* Truncate case. IOEpoch is opened. */
646 rc = mdt_write_get(mo);
650 mfd = mdt_mfd_new(med);
653 GOTO(out_put, rc = -ENOMEM);
656 mdt_ioepoch_open(info, mo, 0);
657 repbody->ioepoch = mo->mot_ioepoch;
659 mdt_object_get(info->mti_env, mo);
660 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
661 mfd->mfd_object = mo;
662 mfd->mfd_xid = req->rq_xid;
664 spin_lock(&med->med_open_lock);
665 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
666 spin_unlock(&med->med_open_lock);
667 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
670 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
672 /* SOM Attribute update case. Find the proper mfd and update
673 * SOM attributes on the proper object. */
674 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
675 LASSERT(info->mti_ioepoch);
677 spin_lock(&med->med_open_lock);
678 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
681 spin_unlock(&med->med_open_lock);
682 CDEBUG(D_INODE, "no handle for file close: "
683 "fid = "DFID": cookie = "LPX64"\n",
684 PFID(info->mti_rr.rr_fid1),
685 info->mti_ioepoch->handle.cookie);
686 GOTO(out_put, rc = -ESTALE);
688 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
689 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
691 class_handle_unhash(&mfd->mfd_handle);
692 cfs_list_del_init(&mfd->mfd_list);
693 spin_unlock(&med->med_open_lock);
695 mdt_mfd_close(info, mfd);
696 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
697 LASSERT((ma->ma_valid & MA_LOV) == 0);
698 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
701 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
702 struct lu_buf *buf = &info->mti_buf;
703 LASSERT(ma->ma_attr.la_valid == 0);
704 buf->lb_buf = ma->ma_lmm;
705 buf->lb_len = ma->ma_lmm_size;
706 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
707 buf, XATTR_NAME_LOV, 0);
710 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
711 struct lu_buf *buf = &info->mti_buf;
713 LASSERT(ma->ma_attr.la_valid == 0);
714 buf->lb_buf = ma->ma_lmv;
715 buf->lb_len = ma->ma_lmv_size;
716 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
717 buf, XATTR_NAME_DEFAULT_LMV, 0);
723 /* If file data is modified, add the dirty flag */
724 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
725 rc = mdt_add_dirty_flag(info, mo, ma);
727 ma->ma_need = MA_INODE;
729 rc = mdt_attr_get_complex(info, mo, ma);
733 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
735 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
736 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
737 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
738 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
739 struct lustre_capa *capa;
741 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
743 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
744 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
747 repbody->valid |= OBD_MD_FLOSSCAPA;
752 mdt_object_put(info->mti_env, mo);
755 mdt_counter_incr(req, LPROC_MDT_SETATTR);
757 mdt_client_compatibility(info);
758 rc2 = mdt_fix_reply(info);
764 static int mdt_reint_create(struct mdt_thread_info *info,
765 struct mdt_lock_handle *lhc)
767 struct ptlrpc_request *req = mdt_info_req(info);
771 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
772 RETURN(err_serious(-ESTALE));
774 if (info->mti_dlm_req)
775 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
777 if (!lu_name_is_valid(&info->mti_rr.rr_name))
780 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
782 mdt_counter_incr(req, LPROC_MDT_MKDIR);
790 /* Special file should stay on the same node as parent. */
791 mdt_counter_incr(req, LPROC_MDT_MKNOD);
794 CERROR("%s: Unsupported mode %o\n",
795 mdt_obd_name(info->mti_mdt),
796 info->mti_attr.ma_attr.la_mode);
797 RETURN(err_serious(-EOPNOTSUPP));
800 rc = mdt_md_create(info);
805 * VBR: save parent version in reply and child version getting by its name.
806 * Version of child is getting and checking during its lookup. If
808 static int mdt_reint_unlink(struct mdt_thread_info *info,
809 struct mdt_lock_handle *lhc)
811 struct mdt_reint_record *rr = &info->mti_rr;
812 struct ptlrpc_request *req = mdt_info_req(info);
813 struct md_attr *ma = &info->mti_attr;
814 struct lu_fid *child_fid = &info->mti_tmp_fid1;
815 struct mdt_object *mp;
816 struct mdt_object *mc;
817 struct mdt_lock_handle *parent_lh;
818 struct mdt_lock_handle *child_lh;
819 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
820 struct mdt_lock_handle *s0_lh = NULL;
821 struct mdt_object *s0_obj = NULL;
826 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
827 PNAME(&rr->rr_name));
829 if (info->mti_dlm_req)
830 ldlm_request_cancel(req, info->mti_dlm_req, 0);
832 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
833 RETURN(err_serious(-ENOENT));
835 if (!fid_is_md_operative(rr->rr_fid1))
839 * step 1: Found the parent.
841 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
847 parent_lh = &info->mti_lh[MDT_LH_PARENT];
848 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
849 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
852 GOTO(put_parent, rc);
854 if (!mdt_object_remote(mp)) {
855 rc = mdt_version_get_check_save(info, mp, 0);
857 GOTO(unlock_parent, rc);
860 /* step 2: find & lock the child */
861 /* lookup child object along with version checking */
863 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
865 /* Name might not be able to find during resend of
866 * remote unlink, considering following case.
867 * dir_A is a remote directory, the name entry of
868 * dir_A is on MDT0, the directory is on MDT1,
870 * 1. client sends unlink req to MDT1.
871 * 2. MDT1 sends name delete update to MDT0.
872 * 3. name entry is being deleted in MDT0 synchronously.
873 * 4. MDT1 is restarted.
874 * 5. client resends unlink req to MDT1. So it can not
875 * find the name entry on MDT0 anymore.
876 * In this case, MDT1 only needs to destory the local
879 if (mdt_object_remote(mp) && rc == -ENOENT &&
880 !fid_is_zero(rr->rr_fid2) &&
881 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
883 *child_fid = *rr->rr_fid2;
885 GOTO(unlock_parent, rc);
889 if (!fid_is_md_operative(child_fid))
890 GOTO(unlock_parent, rc = -EPERM);
892 /* We will lock the child regardless it is local or remote. No harm. */
893 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
895 GOTO(unlock_parent, rc = PTR_ERR(mc));
897 child_lh = &info->mti_lh[MDT_LH_CHILD];
898 mdt_lock_reg_init(child_lh, LCK_EX);
899 if (mdt_object_remote(mc)) {
900 struct mdt_body *repbody;
902 if (!fid_is_zero(rr->rr_fid2)) {
903 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
904 mdt_obd_name(info->mti_mdt),
905 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
906 GOTO(put_child, rc = -ENOENT);
908 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
909 mdt_obd_name(info->mti_mdt),
910 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
912 if (!mdt_is_dne_client(req->rq_export))
913 /* Return -EIO for old client */
914 GOTO(put_child, rc = -EIO);
916 if (info->mti_spec.sp_rm_entry) {
917 struct lu_ucred *uc = mdt_ucred(info);
919 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
920 CERROR("%s: unlink remote entry is only "
921 "permitted for administrator: rc = %d\n",
922 mdt_obd_name(info->mti_mdt),
924 GOTO(put_child, rc = -EPERM);
927 ma->ma_need = MA_INODE;
929 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
930 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
931 NULL, &rr->rr_name, ma, no_name);
934 /* Revoke the LOOKUP lock of the remote object granted by
935 * this MDT. Since the unlink will happen on another MDT,
936 * it will release the LOOKUP lock right away. Then What
937 * would happen if another client try to grab the LOOKUP
938 * lock at the same time with unlink XXX */
939 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
941 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
942 LASSERT(repbody != NULL);
943 repbody->fid1 = *mdt_object_fid(mc);
944 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
945 GOTO(unlock_child, rc = -EREMOTE);
946 } else if (info->mti_spec.sp_rm_entry) {
948 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
950 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
954 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
955 * this now because a running HSM restore on the child (unlink
956 * victim) will hold the layout lock. See LU-4002. */
957 rc = mdt_object_lock(info, mc, child_lh,
958 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
963 * Now we can only make sure we need MA_INODE, in mdd layer, will check
964 * whether need MA_LOV and MA_COOKIE.
966 ma->ma_need = MA_INODE;
969 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
970 mdt_lock_reg_init(s0_lh, LCK_EX);
971 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
974 GOTO(unlock_child, rc);
976 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
977 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
978 /* save version when object is locked */
979 mdt_version_get_save(info, mc, 1);
981 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
983 mutex_lock(&mc->mot_lov_mutex);
985 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
986 mdt_object_child(mc), &rr->rr_name, ma, no_name);
988 mutex_unlock(&mc->mot_lov_mutex);
990 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
991 rc = mdt_attr_get_complex(info, mc, ma);
993 mdt_handle_last_unlink(info, mc, ma);
995 if (ma->ma_valid & MA_INODE) {
996 switch (ma->ma_attr.la_mode & S_IFMT) {
998 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1006 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1009 LASSERTF(0, "bad file type %o unlinking\n",
1010 ma->ma_attr.la_mode);
1017 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
1018 mdt_object_unlock(info, mc, child_lh, rc);
1020 /* Since we do not need reply md striped dir info to client, so
1021 * reset mti_big_lmm_used to avoid confusing mdt_fix_reply */
1022 if (info->mti_big_lmm_used)
1023 info->mti_big_lmm_used = 0;
1025 mdt_object_put(info->mti_env, mc);
1027 mdt_object_unlock(info, mp, parent_lh, rc);
1029 mdt_object_put(info->mti_env, mp);
1035 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1038 static int mdt_reint_link(struct mdt_thread_info *info,
1039 struct mdt_lock_handle *lhc)
1041 struct mdt_reint_record *rr = &info->mti_rr;
1042 struct ptlrpc_request *req = mdt_info_req(info);
1043 struct md_attr *ma = &info->mti_attr;
1044 struct mdt_object *ms;
1045 struct mdt_object *mp;
1046 struct mdt_lock_handle *lhs;
1047 struct mdt_lock_handle *lhp;
1051 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1052 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1054 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1055 RETURN(err_serious(-ENOENT));
1057 if (info->mti_dlm_req)
1058 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1060 /* Invalid case so return error immediately instead of
1062 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1065 if (!fid_is_md_operative(rr->rr_fid1) ||
1066 !fid_is_md_operative(rr->rr_fid2))
1069 /* step 1: find & lock the target parent dir */
1070 lhp = &info->mti_lh[MDT_LH_PARENT];
1071 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1072 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1073 MDS_INODELOCK_UPDATE);
1075 RETURN(PTR_ERR(mp));
1077 rc = mdt_version_get_check_save(info, mp, 0);
1079 GOTO(out_unlock_parent, rc);
1081 /* step 2: find & lock the source */
1082 lhs = &info->mti_lh[MDT_LH_CHILD];
1083 mdt_lock_reg_init(lhs, LCK_EX);
1085 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1087 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1089 if (!mdt_object_exists(ms)) {
1090 mdt_object_put(info->mti_env, ms);
1091 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1092 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1093 GOTO(out_unlock_parent, rc = -ENOENT);
1096 if (mdt_object_remote(ms)) {
1097 mdt_object_put(info->mti_env, ms);
1098 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1099 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1101 GOTO(out_unlock_parent, rc = -EXDEV);
1104 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1105 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1107 mdt_object_put(info->mti_env, ms);
1108 GOTO(out_unlock_parent, rc);
1111 /* step 3: link it */
1112 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1113 OBD_FAIL_MDS_REINT_LINK_WRITE);
1115 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1116 rc = mdt_version_get_check_save(info, ms, 1);
1118 GOTO(out_unlock_child, rc);
1120 /** check target version by name during replay */
1121 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1122 &info->mti_tmp_fid1, 2);
1123 if (rc != 0 && rc != -ENOENT)
1124 GOTO(out_unlock_child, rc);
1125 /* save version of file name for replay, it must be ENOENT here */
1126 if (!req_is_replay(mdt_info_req(info))) {
1127 if (rc != -ENOENT) {
1128 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1129 PNAME(&rr->rr_name));
1130 GOTO(out_unlock_child, rc = -EEXIST);
1132 info->mti_ver[2] = ENOENT_VERSION;
1133 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1136 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1137 mdt_object_child(ms), &rr->rr_name, ma);
1140 mdt_counter_incr(req, LPROC_MDT_LINK);
1144 mdt_object_unlock_put(info, ms, lhs, rc);
1146 mdt_object_unlock_put(info, mp, lhp, rc);
1150 * lock the part of the directory according to the hash of the name
1151 * (lh->mlh_pdo_hash) in parallel directory lock.
1153 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1154 struct mdt_lock_handle *lh,
1155 struct mdt_object *obj, __u64 ibits)
1157 struct ldlm_res_id *res = &info->mti_res_id;
1158 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1159 ldlm_policy_data_t *policy = &info->mti_policy;
1163 * Finish res_id initializing by name hash marking part of
1164 * directory which is taking modification.
1166 LASSERT(lh->mlh_pdo_hash != 0);
1167 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1168 memset(policy, 0, sizeof(*policy));
1169 policy->l_inodebits.bits = ibits;
1171 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1172 * going to be sent to client. If it is - mdt_intent_policy() path will
1173 * fix it up and turn FL_LOCAL flag off.
1175 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1176 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1177 &info->mti_exp->exp_handle.h_cookie);
1181 enum mdt_rename_lock {
1187 * Get BFL lock for rename or migrate process, right now, it does not support
1188 * cross-MDT rename, so we only need global rename lock during migration.
1190 static int mdt_rename_lock(struct mdt_thread_info *info,
1191 struct lustre_handle *lh,
1192 enum mdt_rename_lock rename_lock)
1194 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1195 ldlm_policy_data_t *policy = &info->mti_policy;
1196 struct ldlm_res_id *res_id = &info->mti_res_id;
1201 /* XXX only do global rename lock for migration */
1202 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 &&
1203 rename_lock == MRL_MIGRATE) {
1204 struct lu_fid *fid = &info->mti_tmp_fid1;
1205 struct mdt_object *obj;
1207 /* XXX, right now, it has to use object API to
1208 * enqueue lock cross MDT, so it will enqueue
1209 * rename lock(with LUSTRE_BFL_FID) by root object */
1211 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1213 RETURN(PTR_ERR(obj));
1215 LASSERT(mdt_object_remote(obj));
1216 rc = mdt_remote_object_lock(info, obj,
1217 &LUSTRE_BFL_FID, lh,
1219 MDS_INODELOCK_UPDATE);
1220 mdt_object_put(info->mti_env, obj);
1222 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1223 memset(policy, 0, sizeof *policy);
1224 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1225 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1226 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1227 LCK_EX, &flags, ldlm_blocking_ast,
1228 ldlm_completion_ast, NULL, NULL, 0,
1230 &info->mti_exp->exp_handle.h_cookie,
1237 static void mdt_rename_unlock(struct lustre_handle *lh)
1240 LASSERT(lustre_handle_is_used(lh));
1241 /* Cancel the single rename lock right away */
1242 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1247 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1248 * target. Source should not be ancestor of target dir. May be other rename
1249 * checks can be moved here later.
1251 static int mdt_rename_sanity(struct mdt_thread_info *info,
1252 const struct lu_fid *dir_fid,
1253 const struct lu_fid *fid)
1255 struct mdt_object *dst;
1256 struct lu_fid dst_fid = *dir_fid;
1260 /* If the source and target are in the same directory, they can not
1261 * be parent/child relationship, so subdir check is not needed */
1262 if (lu_fid_eq(dir_fid, fid))
1266 LASSERT(fid_is_sane(&dst_fid));
1267 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1269 rc = mdo_is_subdir(info->mti_env,
1270 mdt_object_child(dst), fid,
1272 mdt_object_put(info->mti_env, dst);
1273 if (rc != -EREMOTE && rc < 0) {
1274 CERROR("%s: failed subdir check in "DFID" for "
1276 mdt_obd_name(info->mti_mdt),
1277 PFID(dir_fid), PFID(fid), rc);
1278 /* Return EINVAL only if a parent is the @fid */
1282 /* check the found fid */
1283 if (lu_fid_eq(&dst_fid, fid))
1289 } while (rc == -EREMOTE);
1294 /* Update object linkEA */
1295 struct mdt_lock_list {
1296 struct mdt_object *mll_obj;
1297 struct mdt_lock_handle mll_lh;
1298 struct list_head mll_list;
1301 static void mdt_unlock_list(struct mdt_thread_info *info,
1302 struct list_head *list, int rc)
1304 struct mdt_lock_list *mll;
1305 struct mdt_lock_list *mll2;
1307 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1308 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1309 list_del(&mll->mll_list);
1314 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1315 struct mdt_object *obj,
1316 struct mdt_object *pobj,
1317 struct list_head *lock_list)
1319 struct lu_buf *buf = &info->mti_big_buf;
1320 struct linkea_data ldata = { 0 };
1325 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1328 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1329 if (buf->lb_buf == NULL)
1333 rc = mdt_links_read(info, obj, &ldata);
1335 if (rc == -ENOENT || rc == -ENODATA)
1340 LASSERT(ldata.ld_leh != NULL);
1341 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1342 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1343 struct mdt_device *mdt = info->mti_mdt;
1344 struct mdt_object *mdt_pobj;
1345 struct mdt_lock_list *mll;
1346 struct lu_name name;
1349 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1351 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1353 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1354 if (IS_ERR(mdt_pobj)) {
1355 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1356 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1360 if (!mdt_object_exists(mdt_pobj)) {
1361 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1362 mdt_obd_name(mdt), PFID(&fid));
1363 mdt_object_put(info->mti_env, mdt_pobj);
1367 if (mdt_pobj == pobj) {
1368 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1369 mdt_obd_name(mdt), PFID(&fid));
1370 mdt_object_put(info->mti_env, mdt_pobj);
1376 mdt_object_put(info->mti_env, mdt_pobj);
1377 GOTO(out, rc = -ENOMEM);
1380 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1381 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1382 MDS_INODELOCK_UPDATE,
1385 CERROR("%s: cannot lock "DFID": rc =%d\n",
1386 mdt_obd_name(mdt), PFID(&fid), rc);
1387 mdt_object_put(info->mti_env, mdt_pobj);
1392 CFS_INIT_LIST_HEAD(&mll->mll_list);
1393 mll->mll_obj = mdt_pobj;
1394 list_add_tail(&mll->mll_list, lock_list);
1398 mdt_unlock_list(info, lock_list, rc);
1402 /* migrate files from one MDT to another MDT */
1403 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1404 struct mdt_lock_handle *lhc)
1406 struct mdt_reint_record *rr = &info->mti_rr;
1407 struct md_attr *ma = &info->mti_attr;
1408 struct mdt_object *msrcdir;
1409 struct mdt_object *mold;
1410 struct mdt_object *mnew = NULL;
1411 struct mdt_lock_handle *lh_dirp;
1412 struct mdt_lock_handle *lh_childp;
1413 struct mdt_lock_handle *lh_tgtp = NULL;
1414 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1415 struct list_head lock_list;
1419 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1420 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1421 /* 1: lock the source dir. */
1422 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1423 if (IS_ERR(msrcdir)) {
1424 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1425 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1426 (int)PTR_ERR(msrcdir));
1427 RETURN(PTR_ERR(msrcdir));
1430 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1431 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1432 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1433 MDS_INODELOCK_UPDATE,
1436 GOTO(out_put_parent, rc);
1438 if (!mdt_object_remote(msrcdir)) {
1439 rc = mdt_version_get_check_save(info, msrcdir, 0);
1441 GOTO(out_unlock_parent, rc);
1444 /* 2: sanity check and find the object to be migrated. */
1446 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1448 GOTO(out_unlock_parent, rc);
1450 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1451 GOTO(out_unlock_parent, rc = -EINVAL);
1453 if (!fid_is_md_operative(old_fid))
1454 GOTO(out_unlock_parent, rc = -EPERM);
1456 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1458 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1460 if (mdt_object_remote(mold)) {
1461 CERROR("%s: source "DFID" is on the remote MDT\n",
1462 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1463 GOTO(out_put_child, rc = -EREMOTE);
1466 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1467 !mdt_object_remote(msrcdir)) {
1468 CERROR("%s: parent "DFID" is still on the same"
1469 " MDT, which should be migrated first:"
1470 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1471 PFID(mdt_object_fid(msrcdir)), -EPERM);
1472 GOTO(out_put_child, rc = -EPERM);
1475 /* 3: iterate the linkea of the object and lock all of the objects */
1476 CFS_INIT_LIST_HEAD(&lock_list);
1477 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1479 GOTO(out_put_child, rc);
1481 /* 4: lock of the object migrated object */
1482 lh_childp = &info->mti_lh[MDT_LH_OLD];
1483 mdt_lock_reg_init(lh_childp, LCK_EX);
1484 rc = mdt_object_lock(info, mold, lh_childp,
1485 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1486 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1488 GOTO(out_unlock_list, rc);
1490 ma->ma_need = MA_LMV;
1492 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1493 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1494 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1496 GOTO(out_unlock_list, rc);
1498 if ((ma->ma_valid & MA_LMV)) {
1499 struct lmv_mds_md_v1 *lmm1;
1501 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1502 lmm1 = &ma->ma_lmv->lmv_md_v1;
1503 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1504 CERROR("%s: can not migrate striped dir "DFID
1505 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1506 PFID(mdt_object_fid(mold)), -EPERM);
1507 GOTO(out_unlock_child, rc = -EPERM);
1510 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1511 GOTO(out_unlock_child, rc = -EINVAL);
1513 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1514 &lmm1->lmv_stripe_fids[1]);
1516 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1518 if (!mdt_object_remote(mnew)) {
1519 CERROR("%s: "DFID" being migrated is on this MDT:"
1520 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1521 PFID(rr->rr_fid2), -EPERM);
1522 GOTO(out_put_new, rc = -EPERM);
1525 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1526 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1527 rc = mdt_remote_object_lock(info, mnew,
1528 mdt_object_fid(mnew),
1529 &lh_tgtp->mlh_rreg_lh,
1530 lh_tgtp->mlh_rreg_mode,
1531 MDS_INODELOCK_UPDATE);
1534 GOTO(out_put_new, rc);
1537 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1540 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1541 if (!mdt_object_remote(mnew)) {
1542 CERROR("%s: Migration "DFID" is on this MDT:"
1543 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1544 PFID(rr->rr_fid2), -EXDEV);
1545 GOTO(out_put_new, rc = -EXDEV);
1550 mdt_reint_init_ma(info, ma);
1552 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1553 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1555 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1556 mdt_object_child(mold), &rr->rr_name,
1557 mdt_object_child(mnew), ma);
1559 GOTO(out_unlock_new, rc);
1561 if (lh_tgtp != NULL)
1562 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1565 mdt_object_put(info->mti_env, mnew);
1567 mdt_object_unlock(info, mold, lh_childp, rc);
1569 mdt_unlock_list(info, &lock_list, rc);
1571 mdt_object_put(info->mti_env, mold);
1573 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1575 mdt_object_put(info->mti_env, msrcdir);
1580 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1581 const struct lu_fid *fid,
1584 struct mdt_object *dir;
1588 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1592 /* check early, the real version will be saved after locking */
1593 rc = mdt_version_get_check(info, dir, idx);
1599 mdt_object_put(info->mti_env, dir);
1603 static int mdt_object_lock_save(struct mdt_thread_info *info,
1604 struct mdt_object *dir,
1605 struct mdt_lock_handle *lh,
1610 /* we lock the target dir if it is local */
1611 rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1616 /* get and save correct version after locking */
1617 mdt_version_get_save(info, dir, idx);
1622 static int mdt_rename_parents_lock(struct mdt_thread_info *info,
1623 struct mdt_object **srcp,
1624 struct mdt_object **tgtp)
1626 struct mdt_reint_record *rr = &info->mti_rr;
1627 const struct lu_fid *fid_src = rr->rr_fid1;
1628 const struct lu_fid *fid_tgt = rr->rr_fid2;
1629 struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT];
1630 struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
1631 struct mdt_object *src;
1632 struct mdt_object *tgt;
1637 /* Check if the @src is not a child of the @tgt, otherwise a
1638 * reverse locking must take place.
1640 * Note: cannot be called after object_find, because if the object
1641 * is destroyed in between it gets stuck in lu_object_find_at(),
1642 * waiting for the last ref. */
1643 rc = mdt_rename_sanity(info, fid_src, fid_tgt);
1649 /* find both parents. */
1650 src = mdt_object_find_check(info, fid_src, 0);
1652 RETURN(PTR_ERR(src));
1654 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1656 if (lu_fid_eq(fid_src, fid_tgt)) {
1658 mdt_object_get(info->mti_env, tgt);
1660 tgt = mdt_object_find_check(info, fid_tgt, 1);
1662 GOTO(err_src_put, rc = PTR_ERR(tgt));
1664 if (unlikely(mdt_object_remote(tgt))) {
1665 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1666 "on different MDTs\n", PFID(fid_src),
1668 GOTO(err_tgt_put, rc = -EXDEV);
1672 /* lock parents in the proper order. */
1674 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1676 GOTO(err_tgt_put, rc);
1678 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1680 rc = mdt_object_lock_save(info, src, lh_src, 0);
1682 rc = mdt_object_lock_save(info, src, lh_src, 0);
1684 GOTO(err_tgt_put, rc);
1686 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1689 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1690 else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
1691 rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
1692 MDS_INODELOCK_UPDATE);
1693 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1697 GOTO(err_unlock, rc);
1699 if (lu_object_is_dying(&tgt->mot_header))
1700 GOTO(err_unlock, rc = -ENOENT);
1707 /* The order does not matter as the handle is checked inside,
1708 * as well as not used handle. */
1709 mdt_object_unlock(info, src, lh_src, rc);
1710 mdt_object_unlock(info, tgt, lh_tgt, rc);
1712 mdt_object_put(info->mti_env, tgt);
1714 mdt_object_put(info->mti_env, src);
1719 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1720 * 2 - src child; 3 - tgt child.
1721 * Update on disk version of src child.
1724 * For DNE phase I, only these renames are allowed
1725 * mv src_p/src_c tgt_p/tgt_c
1726 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1727 * 2. src_p and tgt_p are same directory, and tgt_c does not
1728 * exists. In this case, all of modification will happen
1729 * in the MDT where ithesource parent is, only one remote
1730 * update is needed, i.e. set c_time/m_time on the child.
1731 * And tgt_c will be still in the same MDT as the original
1734 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1735 struct mdt_lock_handle *lhc)
1737 struct mdt_reint_record *rr = &info->mti_rr;
1738 struct md_attr *ma = &info->mti_attr;
1739 struct ptlrpc_request *req = mdt_info_req(info);
1740 struct mdt_object *msrcdir = NULL;
1741 struct mdt_object *mtgtdir = NULL;
1742 struct mdt_object *mold;
1743 struct mdt_object *mnew = NULL;
1744 struct mdt_lock_handle *lh_srcdirp;
1745 struct mdt_lock_handle *lh_tgtdirp;
1746 struct mdt_lock_handle *lh_oldp = NULL;
1747 struct mdt_lock_handle *lh_newp = NULL;
1748 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1749 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1753 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1754 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1755 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1757 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1758 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1759 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1760 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1762 /* step 1&2: lock the source and target dirs. */
1763 rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
1767 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1769 /* step 3: find & lock the old object. */
1771 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1773 GOTO(out_unlock_parents, rc);
1775 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1776 GOTO(out_unlock_parents, rc = -EINVAL);
1778 if (!fid_is_md_operative(old_fid))
1779 GOTO(out_unlock_parents, rc = -EPERM);
1781 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1783 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1785 /* Check if @mtgtdir is subdir of @mold, before locking child
1786 * to avoid reverse locking. */
1787 rc = mdt_rename_sanity(info, rr->rr_fid2, old_fid);
1789 GOTO(out_put_old, rc);
1791 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1792 /* save version after locking */
1793 mdt_version_get_save(info, mold, 2);
1794 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1796 /* step 4: find & lock the new object. */
1797 /* new target object may not exist now */
1798 /* lookup with version checking */
1800 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1803 /* the new_fid should have been filled at this moment */
1804 if (lu_fid_eq(old_fid, new_fid))
1805 GOTO(out_put_old, rc);
1807 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1808 lu_fid_eq(new_fid, rr->rr_fid2))
1809 GOTO(out_put_old, rc = -EINVAL);
1811 if (!fid_is_md_operative(new_fid))
1812 GOTO(out_put_old, rc = -EPERM);
1814 if (mdt_object_remote(mold)) {
1815 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1817 GOTO(out_put_old, rc = -EXDEV);
1820 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1822 GOTO(out_put_old, rc = PTR_ERR(mnew));
1824 if (mdt_object_remote(mnew)) {
1825 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1827 GOTO(out_put_new, rc = -EXDEV);
1830 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1831 mdt_lock_reg_init(lh_oldp, LCK_EX);
1832 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1833 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1835 GOTO(out_put_new, rc);
1837 /* Check if @msrcdir is subdir of @mnew, before locking child
1838 * to avoid reverse locking. */
1839 rc = mdt_rename_sanity(info, rr->rr_fid1, new_fid);
1841 GOTO(out_unlock_old, rc);
1843 /* We used to acquire MDS_INODELOCK_FULL here but we
1844 * can't do this now because a running HSM restore on
1845 * the rename onto victim will hold the layout
1846 * lock. See LU-4002. */
1848 lh_newp = &info->mti_lh[MDT_LH_NEW];
1849 mdt_lock_reg_init(lh_newp, LCK_EX);
1850 rc = mdt_object_lock(info, mnew, lh_newp,
1851 MDS_INODELOCK_LOOKUP |
1852 MDS_INODELOCK_UPDATE,
1855 GOTO(out_unlock_old, rc);
1857 /* get and save version after locking */
1858 mdt_version_get_save(info, mnew, 3);
1859 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1860 } else if (rc != -EREMOTE && rc != -ENOENT) {
1861 GOTO(out_put_old, rc);
1863 /* If mnew does not exist and mold are remote directory,
1864 * it only allows rename if they are under same directory */
1865 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1866 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1868 GOTO(out_put_old, rc = -EXDEV);
1871 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1872 mdt_lock_reg_init(lh_oldp, LCK_EX);
1873 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1874 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1876 GOTO(out_put_old, rc);
1878 mdt_enoent_version_save(info, 3);
1881 /* step 5: rename it */
1882 mdt_reint_init_ma(info, ma);
1884 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1885 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1888 mutex_lock(&mnew->mot_lov_mutex);
1890 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1891 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1892 mnew != NULL ? mdt_object_child(mnew) : NULL,
1893 &rr->rr_tgt_name, ma);
1896 mutex_unlock(&mnew->mot_lov_mutex);
1898 /* handle last link of tgt object */
1900 mdt_counter_incr(req, LPROC_MDT_RENAME);
1902 mdt_handle_last_unlink(info, mnew, ma);
1904 mdt_rename_counter_tally(info, info->mti_mdt, req,
1910 mdt_object_unlock(info, mnew, lh_newp, rc);
1912 mdt_object_unlock(info, mold, lh_oldp, rc);
1915 mdt_object_put(info->mti_env, mnew);
1917 mdt_object_put(info->mti_env, mold);
1919 mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc);
1920 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1924 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1925 struct mdt_lock_handle *lhc,
1926 enum mdt_rename_lock rename_lock)
1928 struct mdt_reint_record *rr = &info->mti_rr;
1929 struct ptlrpc_request *req = mdt_info_req(info);
1930 struct lustre_handle rename_lh = { 0 };
1934 if (info->mti_dlm_req)
1935 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1937 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1938 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1941 rc = mdt_rename_lock(info, &rename_lh, rename_lock);
1943 CERROR("%s: can't lock FS for rename: rc = %d\n",
1944 mdt_obd_name(info->mti_mdt), rc);
1948 if (rename_lock == MRL_RENAME)
1949 rc = mdt_reint_rename_internal(info, lhc);
1951 rc = mdt_reint_migrate_internal(info, lhc);
1953 if (lustre_handle_is_used(&rename_lh))
1954 mdt_rename_unlock(&rename_lh);
1959 static int mdt_reint_rename(struct mdt_thread_info *info,
1960 struct mdt_lock_handle *lhc)
1962 return mdt_reint_rename_or_migrate(info, lhc, MRL_RENAME);
1965 static int mdt_reint_migrate(struct mdt_thread_info *info,
1966 struct mdt_lock_handle *lhc)
1968 return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE);
1971 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
1972 struct mdt_lock_handle *lhc);
1974 static mdt_reinter reinters[REINT_MAX] = {
1975 [REINT_SETATTR] = mdt_reint_setattr,
1976 [REINT_CREATE] = mdt_reint_create,
1977 [REINT_LINK] = mdt_reint_link,
1978 [REINT_UNLINK] = mdt_reint_unlink,
1979 [REINT_RENAME] = mdt_reint_rename,
1980 [REINT_OPEN] = mdt_reint_open,
1981 [REINT_SETXATTR] = mdt_reint_setxattr,
1982 [REINT_RMENTRY] = mdt_reint_unlink,
1983 [REINT_MIGRATE] = mdt_reint_migrate,
1986 int mdt_reint_rec(struct mdt_thread_info *info,
1987 struct mdt_lock_handle *lhc)
1992 rc = reinters[info->mti_rr.rr_opcode](info, lhc);