4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_reint.c
38 * Lustre Metadata Target (mdt) reintegration routines
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Huang Hua <huanghua@clusterfs.com>
44 * Author: Yury Umanets <umka@clusterfs.com>
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include "mdt_internal.h"
50 #include <lustre_lmv.h>
52 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
55 ma->ma_need = MA_INODE;
59 static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
60 struct mdt_object *object,
61 struct mdt_body *repbody)
65 /* for cross-ref mkdir, mds capa has been fetched from remote obj, then
66 * we won't go to below*/
67 if (repbody->mbo_valid & OBD_MD_FLMDSCAPA)
70 if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa &&
71 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
72 struct lustre_capa *capa;
74 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
76 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
77 rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa,
80 repbody->mbo_valid |= OBD_MD_FLMDSCAPA;
87 * Get version of object by fid.
89 * Return real version or ENOENT_VERSION if object doesn't exist
91 static void mdt_obj_version_get(struct mdt_thread_info *info,
92 struct mdt_object *o, __u64 *version)
95 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
96 !fid_is_obf(mdt_object_fid(o)))
97 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
99 *version = ENOENT_VERSION;
100 CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
101 PFID(mdt_object_fid(o)), *version);
105 * Check version is correct.
107 * Should be called only during replay.
109 static int mdt_version_check(struct ptlrpc_request *req,
110 __u64 version, int idx)
112 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
115 if (!exp_connect_vbr(req->rq_export))
118 LASSERT(req_is_replay(req));
119 /** VBR: version is checked always because costs nothing */
120 LASSERT(idx < PTLRPC_NUM_VERSIONS);
121 /** Sanity check for malformed buffers */
122 if (pre_ver == NULL) {
123 CERROR("No versions in request buffer\n");
124 spin_lock(&req->rq_export->exp_lock);
125 req->rq_export->exp_vbr_failed = 1;
126 spin_unlock(&req->rq_export->exp_lock);
128 } else if (pre_ver[idx] != version) {
129 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
130 pre_ver[idx], version);
131 spin_lock(&req->rq_export->exp_lock);
132 req->rq_export->exp_vbr_failed = 1;
133 spin_unlock(&req->rq_export->exp_lock);
140 * Save pre-versions in reply.
142 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
147 if (!exp_connect_vbr(req->rq_export))
150 LASSERT(!req_is_replay(req));
151 LASSERT(req->rq_repmsg != NULL);
152 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
154 reply_ver[idx] = version;
158 * Save enoent version, it is needed when it is obvious that object doesn't
159 * exist, e.g. child during create.
161 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
163 /* save version of file name for replay, it must be ENOENT here */
164 if (!req_is_replay(mdt_info_req(info))) {
165 info->mti_ver[idx] = ENOENT_VERSION;
166 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and save in reply buffer.
173 * Versions are saved in reply only during normal operations not replays.
175 void mdt_version_get_save(struct mdt_thread_info *info,
176 struct mdt_object *mto, int idx)
178 /* don't save versions during replay */
179 if (!req_is_replay(mdt_info_req(info))) {
180 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
181 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
186 * Get version from disk and check it, no save in reply.
188 int mdt_version_get_check(struct mdt_thread_info *info,
189 struct mdt_object *mto, int idx)
191 /* only check versions during replay */
192 if (!req_is_replay(mdt_info_req(info)))
195 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
196 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
200 * Get version from disk and check if recovery or just save.
202 int mdt_version_get_check_save(struct mdt_thread_info *info,
203 struct mdt_object *mto, int idx)
207 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
208 if (req_is_replay(mdt_info_req(info)))
209 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
212 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
217 * Lookup with version checking.
219 * This checks version of 'name'. Many reint functions uses 'name' for child not
220 * FID, therefore we need to get object by name and check its version.
222 int mdt_lookup_version_check(struct mdt_thread_info *info,
223 struct mdt_object *p, const struct lu_name *lname,
224 struct lu_fid *fid, int idx)
228 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
230 /* Check version only during replay */
231 if (!req_is_replay(mdt_info_req(info)))
234 info->mti_ver[idx] = ENOENT_VERSION;
236 struct mdt_object *child;
237 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
238 if (likely(!IS_ERR(child))) {
239 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
240 mdt_object_put(info->mti_env, child);
243 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
244 return vbrc ? vbrc : rc;
249 * mdt_remote_permission: Check whether the remote operation is permitted,
251 * Before we implement async cross-MDT updates (DNE phase 2). There are a few
254 * 1.Only sysadmin can create remote directory and striped directory and
255 * migrate directory now, unless
256 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
257 * 2.Remote directory can only be created on MDT0, unless
258 * lctl set_param mdt.*.enable_remote_dir = 1
259 * 3.Only new clients can access remote dir( >= 2.4) and striped dir(>= 2.6),
260 * old client will return -ENOTSUPP.
262 * XXX these check are only needed for remote synchronization, once async
263 * update is supported, these check will be removed.
265 * param[in]info: execution environment.
266 * param[in]parent: the directory of this operation.
267 * param[in]child: the child of this operation.
269 * retval = 0 remote operation is allowed.
270 * < 0 remote operation is denied.
272 static int mdt_remote_permission(struct mdt_thread_info *info,
273 struct mdt_object *parent,
274 struct mdt_object *child)
276 struct mdt_device *mdt = info->mti_mdt;
277 struct lu_ucred *uc = mdt_ucred(info);
278 struct md_op_spec *spec = &info->mti_spec;
279 struct lu_attr *attr = &info->mti_attr.ma_attr;
280 struct obd_export *exp = mdt_info_req(info)->rq_export;
282 /* Only check create remote directory, striped directory and
284 if (mdt_object_remote(parent) == 0 && mdt_object_remote(child) == 0 &&
285 !(S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
286 spec->u.sp_ea.eadatalen != 0) &&
287 info->mti_rr.rr_opcode != REINT_MIGRATE)
290 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
291 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
292 mdt->mdt_enable_remote_dir_gid != -1)
296 if (mdt->mdt_enable_remote_dir == 0) {
297 struct seq_server_site *ss = mdt_seq_site(mdt);
298 struct lu_seq_range range = { 0 };
301 fld_range_set_type(&range, LU_SEQ_RANGE_MDT);
302 rc = fld_server_lookup(info->mti_env, ss->ss_server_fld,
303 fid_seq(mdt_object_fid(parent)), &range);
307 if (range.lsr_index != 0)
311 if (!mdt_is_dne_client(exp))
314 if (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
315 spec->u.sp_ea.eadatalen != 0 && !mdt_is_striped_client(exp))
322 * VBR: we save three versions in reply:
323 * 0 - parent. Check that parent version is the same during replay.
324 * 1 - name. Version of 'name' if file exists with the same name or
325 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
326 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
329 static int mdt_md_create(struct mdt_thread_info *info)
331 struct mdt_device *mdt = info->mti_mdt;
332 struct mdt_object *parent;
333 struct mdt_object *child;
334 struct mdt_lock_handle *lh;
335 struct mdt_body *repbody;
336 struct md_attr *ma = &info->mti_attr;
337 struct mdt_reint_record *rr = &info->mti_rr;
341 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
343 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
345 if (!fid_is_md_operative(rr->rr_fid1))
348 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
350 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
352 RETURN(PTR_ERR(parent));
354 if (!mdt_object_exists(parent))
355 GOTO(put_parent, rc = -ENOENT);
357 lh = &info->mti_lh[MDT_LH_PARENT];
358 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
359 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE,
362 GOTO(put_parent, rc);
364 if (!mdt_object_remote(parent)) {
365 rc = mdt_version_get_check_save(info, parent, 0);
367 GOTO(unlock_parent, rc);
371 * Check child name version during replay.
372 * During create replay a file may exist with same name.
374 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
375 &info->mti_tmp_fid1, 1);
377 GOTO(unlock_parent, rc = -EEXIST);
379 /* -ENOENT is expected here */
381 GOTO(unlock_parent, rc);
383 /* save version of file name for replay, it must be ENOENT here */
384 mdt_enoent_version_save(info, 1);
386 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
387 if (likely(!IS_ERR(child))) {
388 struct md_object *next = mdt_object_child(parent);
390 rc = mdt_remote_permission(info, parent, child);
392 GOTO(out_put_child, rc);
394 ma->ma_need = MA_INODE;
396 /* capa for cross-ref will be stored here */
397 ma->ma_capa = req_capsule_server_get(info->mti_pill,
399 LASSERT(ma->ma_capa);
401 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
402 OBD_FAIL_MDS_REINT_CREATE_WRITE);
404 /* Version of child will be updated on disk. */
405 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
406 rc = mdt_version_get_check_save(info, child, 2);
408 GOTO(out_put_child, rc);
410 /* Let lower layer know current lock mode. */
411 info->mti_spec.sp_cr_mode =
412 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
415 * Do not perform lookup sanity check. We know that name does
418 info->mti_spec.sp_cr_lookup = 0;
419 info->mti_spec.sp_feat = &dt_directory_features;
421 rc = mdo_create(info->mti_env, next, &rr->rr_name,
422 mdt_object_child(child), &info->mti_spec, ma);
424 rc = mdt_attr_get_complex(info, child, ma);
427 /* Return fid & attr to client. */
428 if (ma->ma_valid & MA_INODE)
429 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
430 mdt_object_fid(child));
433 mdt_object_put(info->mti_env, child);
437 mdt_create_pack_capa(info, rc, child, repbody);
439 mdt_object_unlock(info, parent, lh, rc);
441 mdt_object_put(info->mti_env, parent);
445 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
446 struct mdt_object *obj, __u64 ibits,
447 struct mdt_lock_handle *s0_lh,
448 struct mdt_object *s0_obj,
449 struct ldlm_enqueue_info *einfo)
451 ldlm_policy_data_t *policy = &mti->mti_policy;
455 if (!S_ISDIR(obj->mot_header.loh_attr))
458 /* Unlock stripe 0 */
459 if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
460 LASSERT(s0_obj != NULL);
461 mdt_object_unlock_put(mti, s0_obj, s0_lh, 1);
464 memset(policy, 0, sizeof(*policy));
465 policy->l_inodebits.bits = ibits;
467 rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
473 * Lock slave stripes if necessary, the lock handles of slave stripes
474 * will be stored in einfo->ei_cbdata.
476 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
477 ldlm_mode_t mode, __u64 ibits,
478 struct mdt_lock_handle *s0_lh,
479 struct mdt_object **s0_objp,
480 struct ldlm_enqueue_info *einfo)
482 ldlm_policy_data_t *policy = &mti->mti_policy;
483 struct lu_buf *buf = &mti->mti_buf;
484 struct lmv_mds_md_v1 *lmv;
485 struct lu_fid *fid = &mti->mti_tmp_fid1;
489 if (!S_ISDIR(obj->mot_header.loh_attr))
492 buf->lb_buf = mti->mti_xattr_buf;
493 buf->lb_len = sizeof(mti->mti_xattr_buf);
494 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
497 rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
499 buf->lb_buf = mti->mti_big_lmm;
500 buf->lb_len = mti->mti_big_lmmsize;
504 if (rc == -ENODATA || rc == -ENOENT)
511 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
514 /* Sigh, 0_stripe and master object are different
515 * object, though they are in the same MDT, to avoid
516 * adding osd_object_lock here, so we will enqueue the
517 * stripe0 lock in MDT0 for now */
518 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
519 *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits);
520 if (IS_ERR(*s0_objp))
521 RETURN(PTR_ERR(*s0_objp));
523 memset(einfo, 0, sizeof(*einfo));
524 einfo->ei_type = LDLM_IBITS;
525 einfo->ei_mode = mode;
526 einfo->ei_cb_bl = mdt_remote_blocking_ast;
527 einfo->ei_cb_cp = ldlm_completion_ast;
528 einfo->ei_enq_slave = 1;
529 memset(policy, 0, sizeof(*policy));
530 policy->l_inodebits.bits = ibits;
532 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
537 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
538 struct md_attr *ma, int flags)
540 struct mdt_lock_handle *lh;
541 int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
542 __u64 lockpart = MDS_INODELOCK_UPDATE;
543 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
544 struct mdt_lock_handle *s0_lh;
545 struct mdt_object *s0_obj = NULL;
549 /* attr shouldn't be set on remote object */
550 LASSERT(!mdt_object_remote(mo));
552 lh = &info->mti_lh[MDT_LH_PARENT];
553 mdt_lock_reg_init(lh, LCK_PW);
555 /* Even though the new MDT will grant PERM lock to the old
556 * client, but the old client will almost ignore that during
557 * So it needs to revoke both LOOKUP and PERM lock here, so
558 * both new and old client can cancel the dcache */
559 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
560 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
562 rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
566 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
567 mdt_lock_reg_init(s0_lh, LCK_EX);
568 rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
570 GOTO(out_unlock, rc);
572 if (mdt_object_exists(mo) == 0)
573 GOTO(out_unlock, rc = -ENOENT);
575 /* all attrs are packed into mti_attr in unpack_setattr */
576 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
577 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
579 /* This is only for set ctime when rename's source is on remote MDS. */
580 if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
581 ma->ma_attr_flags |= MDS_VTX_BYPASS;
583 /* VBR: update version if attr changed are important for recovery */
585 /* update on-disk version of changed object */
586 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
587 rc = mdt_version_get_check_save(info, mo, 0);
589 GOTO(out_unlock, rc);
592 /* Ensure constant striping during chown(). See LU-2789. */
593 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
594 mutex_lock(&mo->mot_lov_mutex);
596 /* all attrs are packed into mti_attr in unpack_setattr */
597 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
599 if (ma->ma_attr.la_valid & (LA_UID|LA_GID))
600 mutex_unlock(&mo->mot_lov_mutex);
603 GOTO(out_unlock, rc);
607 mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
608 mdt_object_unlock(info, mo, lh, rc);
613 * Check HSM flags and add HS_DIRTY flag if relevant.
615 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
616 * and is not RELEASED.
618 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
624 /* If the file was modified, add the dirty flag */
625 ma->ma_need = MA_HSM;
626 rc = mdt_attr_get_complex(info, mo, ma);
628 CERROR("file attribute read error for "DFID": %d.\n",
629 PFID(mdt_object_fid(mo)), rc);
633 /* If an up2date copy exists in the backend, add dirty flag */
634 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
635 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
636 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
638 ma->ma_hsm.mh_flags |= HS_DIRTY;
640 mdt_lock_reg_init(lh, LCK_PW);
641 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
646 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
648 CERROR("file attribute change error for "DFID": %d\n",
649 PFID(mdt_object_fid(mo)), rc);
650 mdt_object_unlock(info, mo, lh, rc);
656 static int mdt_reint_setattr(struct mdt_thread_info *info,
657 struct mdt_lock_handle *lhc)
659 struct md_attr *ma = &info->mti_attr;
660 struct mdt_reint_record *rr = &info->mti_rr;
661 struct ptlrpc_request *req = mdt_info_req(info);
662 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
663 struct mdt_file_data *mfd;
664 struct mdt_object *mo;
665 struct mdt_body *repbody;
669 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
670 (unsigned int)ma->ma_attr.la_valid);
672 if (info->mti_dlm_req)
673 ldlm_request_cancel(req, info->mti_dlm_req, 0);
675 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
676 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
678 GOTO(out, rc = PTR_ERR(mo));
680 /* start a log jounal handle if needed */
681 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
682 if ((ma->ma_attr.la_valid & LA_SIZE) ||
683 (rr->rr_flags & MRF_OPEN_TRUNC)) {
684 /* Check write access for the O_TRUNC case */
685 if (mdt_write_read(mo) < 0)
686 GOTO(out_put, rc = -ETXTBSY);
688 } else if (info->mti_ioepoch &&
689 (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
690 /* Truncate case. IOEpoch is opened. */
691 rc = mdt_write_get(mo);
695 mfd = mdt_mfd_new(med);
698 GOTO(out_put, rc = -ENOMEM);
701 mdt_ioepoch_open(info, mo, 0);
702 repbody->mbo_ioepoch = mo->mot_ioepoch;
704 mdt_object_get(info->mti_env, mo);
705 mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
706 mfd->mfd_object = mo;
707 mfd->mfd_xid = req->rq_xid;
709 spin_lock(&med->med_open_lock);
710 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
711 spin_unlock(&med->med_open_lock);
712 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
715 som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
717 /* SOM Attribute update case. Find the proper mfd and update
718 * SOM attributes on the proper object. */
719 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
720 LASSERT(info->mti_ioepoch);
722 spin_lock(&med->med_open_lock);
723 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
726 spin_unlock(&med->med_open_lock);
727 CDEBUG(D_INODE, "no handle for file close: "
728 "fid = "DFID": cookie = "LPX64"\n",
729 PFID(info->mti_rr.rr_fid1),
730 info->mti_ioepoch->handle.cookie);
731 GOTO(out_put, rc = -ESTALE);
733 LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
734 LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
736 class_handle_unhash(&mfd->mfd_handle);
737 cfs_list_del_init(&mfd->mfd_list);
738 spin_unlock(&med->med_open_lock);
740 mdt_mfd_close(info, mfd);
741 } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
742 LASSERT((ma->ma_valid & MA_LOV) == 0);
743 rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
746 } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
747 struct lu_buf *buf = &info->mti_buf;
748 LASSERT(ma->ma_attr.la_valid == 0);
749 buf->lb_buf = ma->ma_lmm;
750 buf->lb_len = ma->ma_lmm_size;
751 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
752 buf, XATTR_NAME_LOV, 0);
755 } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) {
756 struct lu_buf *buf = &info->mti_buf;
758 LASSERT(ma->ma_attr.la_valid == 0);
759 buf->lb_buf = ma->ma_lmv;
760 buf->lb_len = ma->ma_lmv_size;
761 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
762 buf, XATTR_NAME_DEFAULT_LMV, 0);
768 /* If file data is modified, add the dirty flag */
769 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
770 rc = mdt_add_dirty_flag(info, mo, ma);
772 ma->ma_need = MA_INODE;
774 rc = mdt_attr_get_complex(info, mo, ma);
778 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
780 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
781 exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
782 S_ISREG(lu_object_attr(&mo->mot_obj)) &&
783 (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
784 struct lustre_capa *capa;
786 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
788 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
789 rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
792 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
797 mdt_object_put(info->mti_env, mo);
800 mdt_counter_incr(req, LPROC_MDT_SETATTR);
802 mdt_client_compatibility(info);
803 rc2 = mdt_fix_reply(info);
809 static int mdt_reint_create(struct mdt_thread_info *info,
810 struct mdt_lock_handle *lhc)
812 struct ptlrpc_request *req = mdt_info_req(info);
816 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
817 RETURN(err_serious(-ESTALE));
819 if (info->mti_dlm_req)
820 ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0);
822 if (!lu_name_is_valid(&info->mti_rr.rr_name))
825 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
827 mdt_counter_incr(req, LPROC_MDT_MKDIR);
835 /* Special file should stay on the same node as parent. */
836 mdt_counter_incr(req, LPROC_MDT_MKNOD);
839 CERROR("%s: Unsupported mode %o\n",
840 mdt_obd_name(info->mti_mdt),
841 info->mti_attr.ma_attr.la_mode);
842 RETURN(err_serious(-EOPNOTSUPP));
845 rc = mdt_md_create(info);
850 * VBR: save parent version in reply and child version getting by its name.
851 * Version of child is getting and checking during its lookup. If
853 static int mdt_reint_unlink(struct mdt_thread_info *info,
854 struct mdt_lock_handle *lhc)
856 struct mdt_reint_record *rr = &info->mti_rr;
857 struct ptlrpc_request *req = mdt_info_req(info);
858 struct md_attr *ma = &info->mti_attr;
859 struct lu_fid *child_fid = &info->mti_tmp_fid1;
860 struct mdt_object *mp;
861 struct mdt_object *mc;
862 struct mdt_lock_handle *parent_lh;
863 struct mdt_lock_handle *child_lh;
864 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
865 struct mdt_lock_handle *s0_lh = NULL;
866 struct mdt_object *s0_obj = NULL;
871 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
872 PNAME(&rr->rr_name));
874 if (info->mti_dlm_req)
875 ldlm_request_cancel(req, info->mti_dlm_req, 0);
877 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
878 RETURN(err_serious(-ENOENT));
880 if (!fid_is_md_operative(rr->rr_fid1))
884 * step 1: Found the parent.
886 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
892 parent_lh = &info->mti_lh[MDT_LH_PARENT];
893 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
894 rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
897 GOTO(put_parent, rc);
899 if (!mdt_object_remote(mp)) {
900 rc = mdt_version_get_check_save(info, mp, 0);
902 GOTO(unlock_parent, rc);
905 /* step 2: find & lock the child */
906 /* lookup child object along with version checking */
908 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
910 /* Name might not be able to find during resend of
911 * remote unlink, considering following case.
912 * dir_A is a remote directory, the name entry of
913 * dir_A is on MDT0, the directory is on MDT1,
915 * 1. client sends unlink req to MDT1.
916 * 2. MDT1 sends name delete update to MDT0.
917 * 3. name entry is being deleted in MDT0 synchronously.
918 * 4. MDT1 is restarted.
919 * 5. client resends unlink req to MDT1. So it can not
920 * find the name entry on MDT0 anymore.
921 * In this case, MDT1 only needs to destory the local
924 if (mdt_object_remote(mp) && rc == -ENOENT &&
925 !fid_is_zero(rr->rr_fid2) &&
926 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
928 *child_fid = *rr->rr_fid2;
930 GOTO(unlock_parent, rc);
934 if (!fid_is_md_operative(child_fid))
935 GOTO(unlock_parent, rc = -EPERM);
937 /* We will lock the child regardless it is local or remote. No harm. */
938 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
940 GOTO(unlock_parent, rc = PTR_ERR(mc));
942 child_lh = &info->mti_lh[MDT_LH_CHILD];
943 mdt_lock_reg_init(child_lh, LCK_EX);
944 if (mdt_object_remote(mc)) {
945 struct mdt_body *repbody;
947 if (!fid_is_zero(rr->rr_fid2)) {
948 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
949 mdt_obd_name(info->mti_mdt),
950 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
951 GOTO(put_child, rc = -ENOENT);
953 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
954 mdt_obd_name(info->mti_mdt),
955 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
957 if (!mdt_is_dne_client(req->rq_export))
958 /* Return -ENOTSUPP for old client */
959 GOTO(put_child, rc = -ENOTSUPP);
961 if (info->mti_spec.sp_rm_entry) {
962 struct lu_ucred *uc = mdt_ucred(info);
964 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
965 CERROR("%s: unlink remote entry is only "
966 "permitted for administrator: rc = %d\n",
967 mdt_obd_name(info->mti_mdt),
969 GOTO(put_child, rc = -EPERM);
972 ma->ma_need = MA_INODE;
974 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
975 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
976 NULL, &rr->rr_name, ma, no_name);
979 /* Revoke the LOOKUP lock of the remote object granted by
980 * this MDT. Since the unlink will happen on another MDT,
981 * it will release the LOOKUP lock right away. Then What
982 * would happen if another client try to grab the LOOKUP
983 * lock at the same time with unlink XXX */
984 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
986 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
987 LASSERT(repbody != NULL);
988 repbody->mbo_fid1 = *mdt_object_fid(mc);
989 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
990 GOTO(unlock_child, rc = -EREMOTE);
991 } else if (info->mti_spec.sp_rm_entry) {
993 CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': "
995 mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc);
999 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1000 * this now because a running HSM restore on the child (unlink
1001 * victim) will hold the layout lock. See LU-4002. */
1002 rc = mdt_object_lock(info, mc, child_lh,
1003 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
1006 GOTO(put_child, rc);
1008 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1009 * whether need MA_LOV and MA_COOKIE.
1011 ma->ma_need = MA_INODE;
1014 s0_lh = &info->mti_lh[MDT_LH_LOCAL];
1015 mdt_lock_reg_init(s0_lh, LCK_EX);
1016 rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
1019 GOTO(unlock_child, rc);
1021 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1022 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1023 /* save version when object is locked */
1024 mdt_version_get_save(info, mc, 1);
1026 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1028 mutex_lock(&mc->mot_lov_mutex);
1030 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1031 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1033 mutex_unlock(&mc->mot_lov_mutex);
1035 if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
1036 rc = mdt_attr_get_complex(info, mc, ma);
1038 mdt_handle_last_unlink(info, mc, ma);
1040 if (ma->ma_valid & MA_INODE) {
1041 switch (ma->ma_attr.la_mode & S_IFMT) {
1043 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1051 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1054 LASSERTF(0, "bad file type %o unlinking\n",
1055 ma->ma_attr.la_mode);
1062 mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
1063 mdt_object_unlock(info, mc, child_lh, rc);
1065 mdt_object_put(info->mti_env, mc);
1067 mdt_object_unlock(info, mp, parent_lh, rc);
1069 mdt_object_put(info->mti_env, mp);
1075 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1078 static int mdt_reint_link(struct mdt_thread_info *info,
1079 struct mdt_lock_handle *lhc)
1081 struct mdt_reint_record *rr = &info->mti_rr;
1082 struct ptlrpc_request *req = mdt_info_req(info);
1083 struct md_attr *ma = &info->mti_attr;
1084 struct mdt_object *ms;
1085 struct mdt_object *mp;
1086 struct mdt_lock_handle *lhs;
1087 struct mdt_lock_handle *lhp;
1091 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1092 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1094 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1095 RETURN(err_serious(-ENOENT));
1097 if (info->mti_dlm_req)
1098 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1100 /* Invalid case so return error immediately instead of
1102 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1105 if (!fid_is_md_operative(rr->rr_fid1) ||
1106 !fid_is_md_operative(rr->rr_fid2))
1109 /* step 1: find & lock the target parent dir */
1110 lhp = &info->mti_lh[MDT_LH_PARENT];
1111 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1112 mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
1113 MDS_INODELOCK_UPDATE);
1115 RETURN(PTR_ERR(mp));
1117 rc = mdt_version_get_check_save(info, mp, 0);
1119 GOTO(out_unlock_parent, rc);
1121 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1123 /* step 2: find & lock the source */
1124 lhs = &info->mti_lh[MDT_LH_CHILD];
1125 mdt_lock_reg_init(lhs, LCK_EX);
1127 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1129 GOTO(out_unlock_parent, rc = PTR_ERR(ms));
1131 if (!mdt_object_exists(ms)) {
1132 mdt_object_put(info->mti_env, ms);
1133 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1134 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1135 GOTO(out_unlock_parent, rc = -ENOENT);
1138 if (mdt_object_remote(ms)) {
1139 mdt_object_put(info->mti_env, ms);
1140 CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
1141 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1143 GOTO(out_unlock_parent, rc = -EXDEV);
1146 rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
1147 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1149 mdt_object_put(info->mti_env, ms);
1150 GOTO(out_unlock_parent, rc);
1153 /* step 3: link it */
1154 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1155 OBD_FAIL_MDS_REINT_LINK_WRITE);
1157 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1158 rc = mdt_version_get_check_save(info, ms, 1);
1160 GOTO(out_unlock_child, rc);
1162 /** check target version by name during replay */
1163 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1164 &info->mti_tmp_fid1, 2);
1165 if (rc != 0 && rc != -ENOENT)
1166 GOTO(out_unlock_child, rc);
1167 /* save version of file name for replay, it must be ENOENT here */
1168 if (!req_is_replay(mdt_info_req(info))) {
1169 if (rc != -ENOENT) {
1170 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1171 PNAME(&rr->rr_name));
1172 GOTO(out_unlock_child, rc = -EEXIST);
1174 info->mti_ver[2] = ENOENT_VERSION;
1175 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1178 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1179 mdt_object_child(ms), &rr->rr_name, ma);
1182 mdt_counter_incr(req, LPROC_MDT_LINK);
1186 mdt_object_unlock_put(info, ms, lhs, rc);
1188 mdt_object_unlock_put(info, mp, lhp, rc);
1192 * lock the part of the directory according to the hash of the name
1193 * (lh->mlh_pdo_hash) in parallel directory lock.
1195 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1196 struct mdt_lock_handle *lh,
1197 struct mdt_object *obj, __u64 ibits)
1199 struct ldlm_res_id *res = &info->mti_res_id;
1200 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1201 ldlm_policy_data_t *policy = &info->mti_policy;
1205 * Finish res_id initializing by name hash marking part of
1206 * directory which is taking modification.
1208 LASSERT(lh->mlh_pdo_hash != 0);
1209 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1210 memset(policy, 0, sizeof(*policy));
1211 policy->l_inodebits.bits = ibits;
1213 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1214 * going to be sent to client. If it is - mdt_intent_policy() path will
1215 * fix it up and turn FL_LOCAL flag off.
1217 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1218 res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
1219 &info->mti_exp->exp_handle.h_cookie);
1223 enum mdt_rename_lock {
1229 * Get BFL lock for rename or migrate process, right now, it does not support
1230 * cross-MDT rename, so we only need global rename lock during migration.
1232 static int mdt_rename_lock(struct mdt_thread_info *info,
1233 struct lustre_handle *lh,
1234 enum mdt_rename_lock rename_lock)
1236 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1237 ldlm_policy_data_t *policy = &info->mti_policy;
1238 struct ldlm_res_id *res_id = &info->mti_res_id;
1243 /* XXX only do global rename lock for migration */
1244 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 &&
1245 rename_lock == MRL_MIGRATE) {
1246 struct lu_fid *fid = &info->mti_tmp_fid1;
1247 struct mdt_object *obj;
1249 /* XXX, right now, it has to use object API to
1250 * enqueue lock cross MDT, so it will enqueue
1251 * rename lock(with LUSTRE_BFL_FID) by root object */
1253 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1255 RETURN(PTR_ERR(obj));
1257 LASSERT(mdt_object_remote(obj));
1258 rc = mdt_remote_object_lock(info, obj,
1259 &LUSTRE_BFL_FID, lh,
1261 MDS_INODELOCK_UPDATE);
1262 mdt_object_put(info->mti_env, obj);
1264 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1265 memset(policy, 0, sizeof *policy);
1266 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1267 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1268 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1269 LCK_EX, &flags, ldlm_blocking_ast,
1270 ldlm_completion_ast, NULL, NULL, 0,
1272 &info->mti_exp->exp_handle.h_cookie,
1279 static void mdt_rename_unlock(struct lustre_handle *lh)
1282 LASSERT(lustre_handle_is_used(lh));
1283 /* Cancel the single rename lock right away */
1284 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1289 * This is is_subdir() variant, it is CMD if cmm forwards it to correct
1290 * target. Source should not be ancestor of target dir. May be other rename
1291 * checks can be moved here later.
1293 static int mdt_rename_sanity(struct mdt_thread_info *info,
1294 const struct lu_fid *dir_fid,
1295 const struct lu_fid *fid)
1297 struct mdt_object *dst;
1298 struct lu_fid dst_fid = *dir_fid;
1302 /* If the source and target are in the same directory, they can not
1303 * be parent/child relationship, so subdir check is not needed */
1304 if (lu_fid_eq(dir_fid, fid))
1308 LASSERT(fid_is_sane(&dst_fid));
1309 dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid);
1311 /* XXX: this object might not be protected by LDLM lock
1312 * here, (see mdt_rename_parents_lock), but LOHA_EXISTS
1313 * will not change once it is being set, but LFSCK might
1314 * change this later.(LU-5069) */
1315 if (!mdt_object_exists(dst))
1318 rc = mdo_is_subdir(info->mti_env,
1319 mdt_object_child(dst), fid,
1321 mdt_object_put(info->mti_env, dst);
1322 if (rc != -EREMOTE && rc < 0) {
1323 CERROR("%s: failed subdir check in "DFID" for "
1325 mdt_obd_name(info->mti_mdt),
1326 PFID(dir_fid), PFID(fid), rc);
1327 /* Return EINVAL only if a parent is the @fid */
1331 /* check the found fid */
1332 if (lu_fid_eq(&dst_fid, fid))
1338 } while (rc == -EREMOTE);
1343 /* Update object linkEA */
1344 struct mdt_lock_list {
1345 struct mdt_object *mll_obj;
1346 struct mdt_lock_handle mll_lh;
1347 struct list_head mll_list;
1350 static void mdt_unlock_list(struct mdt_thread_info *info,
1351 struct list_head *list, int rc)
1353 struct mdt_lock_list *mll;
1354 struct mdt_lock_list *mll2;
1356 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1357 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1358 list_del(&mll->mll_list);
1363 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1364 struct mdt_object *obj,
1365 struct mdt_object *pobj,
1366 struct list_head *lock_list)
1368 struct lu_buf *buf = &info->mti_big_buf;
1369 struct linkea_data ldata = { 0 };
1374 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1377 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1378 if (buf->lb_buf == NULL)
1382 rc = mdt_links_read(info, obj, &ldata);
1384 if (rc == -ENOENT || rc == -ENODATA)
1389 LASSERT(ldata.ld_leh != NULL);
1390 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1391 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1392 struct mdt_device *mdt = info->mti_mdt;
1393 struct mdt_object *mdt_pobj;
1394 struct mdt_lock_list *mll;
1395 struct lu_name name;
1398 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1400 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1402 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1403 if (IS_ERR(mdt_pobj)) {
1404 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1405 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1409 if (!mdt_object_exists(mdt_pobj)) {
1410 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1411 mdt_obd_name(mdt), PFID(&fid));
1412 mdt_object_put(info->mti_env, mdt_pobj);
1416 if (mdt_pobj == pobj) {
1417 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1418 mdt_obd_name(mdt), PFID(&fid));
1419 mdt_object_put(info->mti_env, mdt_pobj);
1425 mdt_object_put(info->mti_env, mdt_pobj);
1426 GOTO(out, rc = -ENOMEM);
1429 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1430 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1431 MDS_INODELOCK_UPDATE,
1434 CERROR("%s: cannot lock "DFID": rc =%d\n",
1435 mdt_obd_name(mdt), PFID(&fid), rc);
1436 mdt_object_put(info->mti_env, mdt_pobj);
1441 CFS_INIT_LIST_HEAD(&mll->mll_list);
1442 mll->mll_obj = mdt_pobj;
1443 list_add_tail(&mll->mll_list, lock_list);
1447 mdt_unlock_list(info, lock_list, rc);
1451 /* migrate files from one MDT to another MDT */
1452 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1453 struct mdt_lock_handle *lhc)
1455 struct mdt_reint_record *rr = &info->mti_rr;
1456 struct md_attr *ma = &info->mti_attr;
1457 struct mdt_object *msrcdir;
1458 struct mdt_object *mold;
1459 struct mdt_object *mnew = NULL;
1460 struct mdt_lock_handle *lh_dirp;
1461 struct mdt_lock_handle *lh_childp;
1462 struct mdt_lock_handle *lh_tgtp = NULL;
1463 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1464 struct list_head lock_list;
1468 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1469 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1471 /* 1: lock the source dir. */
1472 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1473 if (IS_ERR(msrcdir)) {
1474 CERROR("%s: cannot find source dir "DFID" : rc = %d\n",
1475 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1476 (int)PTR_ERR(msrcdir));
1477 RETURN(PTR_ERR(msrcdir));
1480 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1481 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1482 rc = mdt_object_lock(info, msrcdir, lh_dirp,
1483 MDS_INODELOCK_UPDATE,
1486 GOTO(out_put_parent, rc);
1488 if (!mdt_object_remote(msrcdir)) {
1489 rc = mdt_version_get_check_save(info, msrcdir, 0);
1491 GOTO(out_unlock_parent, rc);
1494 /* 2: sanity check and find the object to be migrated. */
1496 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1498 GOTO(out_unlock_parent, rc);
1500 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1501 GOTO(out_unlock_parent, rc = -EINVAL);
1503 if (!fid_is_md_operative(old_fid))
1504 GOTO(out_unlock_parent, rc = -EPERM);
1506 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1508 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1510 if (mdt_object_remote(mold)) {
1511 CERROR("%s: source "DFID" is on the remote MDT\n",
1512 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1513 GOTO(out_put_child, rc = -EREMOTE);
1516 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1517 !mdt_object_remote(msrcdir)) {
1518 CERROR("%s: parent "DFID" is still on the same"
1519 " MDT, which should be migrated first:"
1520 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1521 PFID(mdt_object_fid(msrcdir)), -EPERM);
1522 GOTO(out_put_child, rc = -EPERM);
1525 rc = mdt_remote_permission(info, msrcdir, mold);
1527 GOTO(out_put_child, rc);
1529 /* 3: iterate the linkea of the object and lock all of the objects */
1530 CFS_INIT_LIST_HEAD(&lock_list);
1531 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1533 GOTO(out_put_child, rc);
1535 /* 4: lock of the object migrated object */
1536 lh_childp = &info->mti_lh[MDT_LH_OLD];
1537 mdt_lock_reg_init(lh_childp, LCK_EX);
1538 rc = mdt_object_lock(info, mold, lh_childp,
1539 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1540 MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK);
1542 GOTO(out_unlock_list, rc);
1544 ma->ma_need = MA_LMV;
1546 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1547 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1548 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1550 GOTO(out_unlock_list, rc);
1552 if ((ma->ma_valid & MA_LMV)) {
1553 struct lmv_mds_md_v1 *lmm1;
1555 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1556 lmm1 = &ma->ma_lmv->lmv_md_v1;
1557 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1558 CERROR("%s: can not migrate striped dir "DFID
1559 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1560 PFID(mdt_object_fid(mold)), -EPERM);
1561 GOTO(out_unlock_child, rc = -EPERM);
1564 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1565 GOTO(out_unlock_child, rc = -EINVAL);
1567 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1568 &lmm1->lmv_stripe_fids[1]);
1570 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1572 if (!mdt_object_remote(mnew)) {
1573 CERROR("%s: "DFID" being migrated is on this MDT:"
1574 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1575 PFID(rr->rr_fid2), -EPERM);
1576 GOTO(out_put_new, rc = -EPERM);
1579 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1580 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1581 rc = mdt_remote_object_lock(info, mnew,
1582 mdt_object_fid(mnew),
1583 &lh_tgtp->mlh_rreg_lh,
1584 lh_tgtp->mlh_rreg_mode,
1585 MDS_INODELOCK_UPDATE);
1588 GOTO(out_put_new, rc);
1591 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1594 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1595 if (!mdt_object_remote(mnew)) {
1596 CERROR("%s: Migration "DFID" is on this MDT:"
1597 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1598 PFID(rr->rr_fid2), -EXDEV);
1599 GOTO(out_put_new, rc = -EXDEV);
1604 mdt_reint_init_ma(info, ma);
1606 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1607 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1609 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1610 mdt_object_child(mold), &rr->rr_name,
1611 mdt_object_child(mnew), ma);
1613 GOTO(out_unlock_new, rc);
1615 if (lh_tgtp != NULL)
1616 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1619 mdt_object_put(info->mti_env, mnew);
1621 mdt_object_unlock(info, mold, lh_childp, rc);
1623 mdt_unlock_list(info, &lock_list, rc);
1625 mdt_object_put(info->mti_env, mold);
1627 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1629 mdt_object_put(info->mti_env, msrcdir);
1634 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1635 const struct lu_fid *fid,
1638 struct mdt_object *dir;
1642 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1646 /* check early, the real version will be saved after locking */
1647 rc = mdt_version_get_check(info, dir, idx);
1653 mdt_object_put(info->mti_env, dir);
1657 static int mdt_object_lock_save(struct mdt_thread_info *info,
1658 struct mdt_object *dir,
1659 struct mdt_lock_handle *lh,
1664 /* we lock the target dir if it is local */
1665 rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1670 /* get and save correct version after locking */
1671 mdt_version_get_save(info, dir, idx);
1676 static int mdt_rename_parents_lock(struct mdt_thread_info *info,
1677 struct mdt_object **srcp,
1678 struct mdt_object **tgtp)
1680 struct mdt_reint_record *rr = &info->mti_rr;
1681 const struct lu_fid *fid_src = rr->rr_fid1;
1682 const struct lu_fid *fid_tgt = rr->rr_fid2;
1683 struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT];
1684 struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
1685 struct mdt_object *src;
1686 struct mdt_object *tgt;
1691 /* Check if the @src is not a child of the @tgt, otherwise a
1692 * reverse locking must take place.
1694 * Note: cannot be called after object_find, because if the object
1695 * is destroyed in between it gets stuck in lu_object_find_at(),
1696 * waiting for the last ref. */
1697 rc = mdt_rename_sanity(info, fid_src, fid_tgt);
1703 /* find both parents. */
1704 src = mdt_object_find_check(info, fid_src, 0);
1706 RETURN(PTR_ERR(src));
1708 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1710 if (lu_fid_eq(fid_src, fid_tgt)) {
1712 mdt_object_get(info->mti_env, tgt);
1714 tgt = mdt_object_find_check(info, fid_tgt, 1);
1716 GOTO(err_src_put, rc = PTR_ERR(tgt));
1718 if (unlikely(mdt_object_remote(tgt))) {
1719 CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
1720 "on different MDTs\n", PFID(fid_src),
1722 GOTO(err_tgt_put, rc = -EXDEV);
1726 /* lock parents in the proper order. */
1728 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1730 GOTO(err_tgt_put, rc);
1732 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1734 rc = mdt_object_lock_save(info, src, lh_src, 0);
1736 rc = mdt_object_lock_save(info, src, lh_src, 0);
1738 GOTO(err_tgt_put, rc);
1740 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1743 rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
1744 else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
1745 rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
1746 MDS_INODELOCK_UPDATE);
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
1751 GOTO(err_unlock, rc);
1753 if (lu_object_is_dying(&tgt->mot_header))
1754 GOTO(err_unlock, rc = -ENOENT);
1761 /* The order does not matter as the handle is checked inside,
1762 * as well as not used handle. */
1763 mdt_object_unlock(info, src, lh_src, rc);
1764 mdt_object_unlock(info, tgt, lh_tgt, rc);
1766 mdt_object_put(info->mti_env, tgt);
1768 mdt_object_put(info->mti_env, src);
1773 * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
1774 * 2 - src child; 3 - tgt child.
1775 * Update on disk version of src child.
1778 * For DNE phase I, only these renames are allowed
1779 * mv src_p/src_c tgt_p/tgt_c
1780 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1781 * 2. src_p and tgt_p are same directory, and tgt_c does not
1782 * exists. In this case, all of modification will happen
1783 * in the MDT where ithesource parent is, only one remote
1784 * update is needed, i.e. set c_time/m_time on the child.
1785 * And tgt_c will be still in the same MDT as the original
1788 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1789 struct mdt_lock_handle *lhc)
1791 struct mdt_reint_record *rr = &info->mti_rr;
1792 struct md_attr *ma = &info->mti_attr;
1793 struct ptlrpc_request *req = mdt_info_req(info);
1794 struct mdt_object *msrcdir = NULL;
1795 struct mdt_object *mtgtdir = NULL;
1796 struct mdt_object *mold;
1797 struct mdt_object *mnew = NULL;
1798 struct mdt_lock_handle *lh_srcdirp;
1799 struct mdt_lock_handle *lh_tgtdirp;
1800 struct mdt_lock_handle *lh_oldp = NULL;
1801 struct mdt_lock_handle *lh_newp = NULL;
1802 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1803 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1807 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1808 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1809 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1811 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1812 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1813 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1814 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1816 /* step 1&2: lock the source and target dirs. */
1817 rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
1821 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
1823 /* step 3: find & lock the old object. */
1825 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1827 GOTO(out_unlock_parents, rc);
1829 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1830 GOTO(out_unlock_parents, rc = -EINVAL);
1832 if (!fid_is_md_operative(old_fid))
1833 GOTO(out_unlock_parents, rc = -EPERM);
1835 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1837 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
1839 /* Check if @mtgtdir is subdir of @mold, before locking child
1840 * to avoid reverse locking. */
1841 rc = mdt_rename_sanity(info, rr->rr_fid2, old_fid);
1843 GOTO(out_put_old, rc);
1845 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
1846 /* save version after locking */
1847 mdt_version_get_save(info, mold, 2);
1848 mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
1850 /* step 4: find & lock the new object. */
1851 /* new target object may not exist now */
1852 /* lookup with version checking */
1854 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
1857 /* the new_fid should have been filled at this moment */
1858 if (lu_fid_eq(old_fid, new_fid))
1859 GOTO(out_put_old, rc);
1861 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
1862 lu_fid_eq(new_fid, rr->rr_fid2))
1863 GOTO(out_put_old, rc = -EINVAL);
1865 if (!fid_is_md_operative(new_fid))
1866 GOTO(out_put_old, rc = -EPERM);
1868 if (mdt_object_remote(mold)) {
1869 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1871 GOTO(out_put_old, rc = -EXDEV);
1874 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
1876 GOTO(out_put_old, rc = PTR_ERR(mnew));
1878 if (mdt_object_remote(mnew)) {
1879 CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
1881 GOTO(out_put_new, rc = -EXDEV);
1884 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
1885 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
1886 GOTO(out_put_new, rc = -EISDIR);
1888 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1889 mdt_lock_reg_init(lh_oldp, LCK_EX);
1890 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1891 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1893 GOTO(out_put_new, rc);
1895 /* Check if @msrcdir is subdir of @mnew, before locking child
1896 * to avoid reverse locking. */
1897 rc = mdt_rename_sanity(info, rr->rr_fid1, new_fid);
1899 GOTO(out_unlock_old, rc);
1901 /* We used to acquire MDS_INODELOCK_FULL here but we
1902 * can't do this now because a running HSM restore on
1903 * the rename onto victim will hold the layout
1904 * lock. See LU-4002. */
1906 lh_newp = &info->mti_lh[MDT_LH_NEW];
1907 mdt_lock_reg_init(lh_newp, LCK_EX);
1908 rc = mdt_object_lock(info, mnew, lh_newp,
1909 MDS_INODELOCK_LOOKUP |
1910 MDS_INODELOCK_UPDATE,
1913 GOTO(out_unlock_old, rc);
1915 /* get and save version after locking */
1916 mdt_version_get_save(info, mnew, 3);
1917 mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
1918 } else if (rc != -EREMOTE && rc != -ENOENT) {
1919 GOTO(out_put_old, rc);
1921 /* If mnew does not exist and mold are remote directory,
1922 * it only allows rename if they are under same directory */
1923 if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
1924 CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
1926 GOTO(out_put_old, rc = -EXDEV);
1929 lh_oldp = &info->mti_lh[MDT_LH_OLD];
1930 mdt_lock_reg_init(lh_oldp, LCK_EX);
1931 rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
1932 MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
1934 GOTO(out_put_old, rc);
1936 mdt_enoent_version_save(info, 3);
1939 /* step 5: rename it */
1940 mdt_reint_init_ma(info, ma);
1942 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1943 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1946 mutex_lock(&mnew->mot_lov_mutex);
1948 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
1949 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
1950 mnew != NULL ? mdt_object_child(mnew) : NULL,
1951 &rr->rr_tgt_name, ma);
1954 mutex_unlock(&mnew->mot_lov_mutex);
1956 /* handle last link of tgt object */
1958 mdt_counter_incr(req, LPROC_MDT_RENAME);
1960 mdt_handle_last_unlink(info, mnew, ma);
1962 mdt_rename_counter_tally(info, info->mti_mdt, req,
1968 mdt_object_unlock(info, mnew, lh_newp, rc);
1970 mdt_object_unlock(info, mold, lh_oldp, rc);
1973 mdt_object_put(info->mti_env, mnew);
1975 mdt_object_put(info->mti_env, mold);
1977 mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc);
1978 mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
1982 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
1983 struct mdt_lock_handle *lhc,
1984 enum mdt_rename_lock rename_lock)
1986 struct mdt_reint_record *rr = &info->mti_rr;
1987 struct ptlrpc_request *req = mdt_info_req(info);
1988 struct lustre_handle rename_lh = { 0 };
1992 if (info->mti_dlm_req)
1993 ldlm_request_cancel(req, info->mti_dlm_req, 0);
1995 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
1996 fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
1999 rc = mdt_rename_lock(info, &rename_lh, rename_lock);
2001 CERROR("%s: can't lock FS for rename: rc = %d\n",
2002 mdt_obd_name(info->mti_mdt), rc);
2006 if (rename_lock == MRL_RENAME)
2007 rc = mdt_reint_rename_internal(info, lhc);
2009 rc = mdt_reint_migrate_internal(info, lhc);
2011 if (lustre_handle_is_used(&rename_lh))
2012 mdt_rename_unlock(&rename_lh);
2017 static int mdt_reint_rename(struct mdt_thread_info *info,
2018 struct mdt_lock_handle *lhc)
2020 return mdt_reint_rename_or_migrate(info, lhc, MRL_RENAME);
2023 static int mdt_reint_migrate(struct mdt_thread_info *info,
2024 struct mdt_lock_handle *lhc)
2026 return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE);
2029 typedef int (*mdt_reinter)(struct mdt_thread_info *info,
2030 struct mdt_lock_handle *lhc);
2032 static mdt_reinter reinters[REINT_MAX] = {
2033 [REINT_SETATTR] = mdt_reint_setattr,
2034 [REINT_CREATE] = mdt_reint_create,
2035 [REINT_LINK] = mdt_reint_link,
2036 [REINT_UNLINK] = mdt_reint_unlink,
2037 [REINT_RENAME] = mdt_reint_rename,
2038 [REINT_OPEN] = mdt_reint_open,
2039 [REINT_SETXATTR] = mdt_reint_setxattr,
2040 [REINT_RMENTRY] = mdt_reint_unlink,
2041 [REINT_MIGRATE] = mdt_reint_migrate,
2044 int mdt_reint_rec(struct mdt_thread_info *info,
2045 struct mdt_lock_handle *lhc)
2050 rc = reinters[info->mti_rr.rr_opcode](info, lhc);