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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_reint.c
34 * Lustre Metadata Target (mdt) reintegration routines
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Huang Hua <huanghua@clusterfs.com>
40 * Author: Yury Umanets <umka@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lprocfs_status.h>
46 #include "mdt_internal.h"
47 #include <lustre_lmv.h>
49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
52 ma->ma_need = MA_INODE;
57 * Get version of object by fid.
59 * Return real version or ENOENT_VERSION if object doesn't exist
61 static void mdt_obj_version_get(struct mdt_thread_info *info,
62 struct mdt_object *o, __u64 *version)
66 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
67 !fid_is_obf(mdt_object_fid(o)))
68 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
70 *version = ENOENT_VERSION;
71 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
72 PFID(mdt_object_fid(o)), *version);
76 * Check version is correct.
78 * Should be called only during replay.
80 static int mdt_version_check(struct ptlrpc_request *req,
81 __u64 version, int idx)
83 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
86 if (!exp_connect_vbr(req->rq_export))
89 LASSERT(req_is_replay(req));
90 /** VBR: version is checked always because costs nothing */
91 LASSERT(idx < PTLRPC_NUM_VERSIONS);
92 /** Sanity check for malformed buffers */
93 if (pre_ver == NULL) {
94 CERROR("No versions in request buffer\n");
95 spin_lock(&req->rq_export->exp_lock);
96 req->rq_export->exp_vbr_failed = 1;
97 spin_unlock(&req->rq_export->exp_lock);
99 } else if (pre_ver[idx] != version) {
100 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
101 pre_ver[idx], version);
102 spin_lock(&req->rq_export->exp_lock);
103 req->rq_export->exp_vbr_failed = 1;
104 spin_unlock(&req->rq_export->exp_lock);
111 * Save pre-versions in reply.
113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
118 if (!exp_connect_vbr(req->rq_export))
121 LASSERT(!req_is_replay(req));
122 LASSERT(req->rq_repmsg != NULL);
123 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
125 reply_ver[idx] = version;
129 * Save enoent version, it is needed when it is obvious that object doesn't
130 * exist, e.g. child during create.
132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
134 /* save version of file name for replay, it must be ENOENT here */
135 if (!req_is_replay(mdt_info_req(info))) {
136 info->mti_ver[idx] = ENOENT_VERSION;
137 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
142 * Get version from disk and save in reply buffer.
144 * Versions are saved in reply only during normal operations not replays.
146 void mdt_version_get_save(struct mdt_thread_info *info,
147 struct mdt_object *mto, int idx)
149 /* don't save versions during replay */
150 if (!req_is_replay(mdt_info_req(info))) {
151 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
152 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
157 * Get version from disk and check it, no save in reply.
159 int mdt_version_get_check(struct mdt_thread_info *info,
160 struct mdt_object *mto, int idx)
162 /* only check versions during replay */
163 if (!req_is_replay(mdt_info_req(info)))
166 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
167 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and check if recovery or just save.
173 int mdt_version_get_check_save(struct mdt_thread_info *info,
174 struct mdt_object *mto, int idx)
178 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
179 if (req_is_replay(mdt_info_req(info)))
180 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
183 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
188 * Lookup with version checking.
190 * This checks version of 'name'. Many reint functions uses 'name' for child not
191 * FID, therefore we need to get object by name and check its version.
193 static int mdt_lookup_version_check(struct mdt_thread_info *info,
194 struct mdt_object *p,
195 const struct lu_name *lname,
196 struct lu_fid *fid, int idx)
200 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
202 /* Check version only during replay */
203 if (!req_is_replay(mdt_info_req(info)))
206 info->mti_ver[idx] = ENOENT_VERSION;
208 struct mdt_object *child;
209 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
210 if (likely(!IS_ERR(child))) {
211 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
212 mdt_object_put(info->mti_env, child);
215 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
216 return vbrc ? vbrc : rc;
220 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
222 struct lu_ucred *uc = mdt_ucred(info);
223 struct mdt_device *mdt = info->mti_mdt;
225 if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
226 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
227 mdt->mdt_enable_remote_dir_gid != -1)
235 * mdt_remote_permission: Check whether the remote operation is permitted,
237 * Only sysadmin can create remote directory / striped directory,
238 * migrate directory and set default stripedEA on directory, unless
240 * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
242 * param[in] info: mdt_thread_info.
244 * retval = 0 remote operation is allowed.
245 * < 0 remote operation is denied.
247 static int mdt_remote_permission(struct mdt_thread_info *info)
249 struct md_op_spec *spec = &info->mti_spec;
250 struct lu_attr *attr = &info->mti_attr.ma_attr;
251 struct obd_export *exp = mdt_info_req(info)->rq_export;
254 if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
255 rc = mdt_remote_permission_check(info);
260 if (info->mti_rr.rr_opcode == REINT_CREATE &&
261 (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
262 spec->u.sp_ea.eadatalen != 0)) {
263 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
265 /* Only new clients can create remote dir( >= 2.4) and
266 * striped dir(>= 2.6), old client will return -ENOTSUPP */
267 if (!mdt_is_dne_client(exp))
270 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
271 !mdt_is_striped_client(exp))
274 rc = mdt_remote_permission_check(info);
279 if (info->mti_rr.rr_opcode == REINT_SETATTR) {
280 struct md_attr *ma = &info->mti_attr;
282 if ((ma->ma_valid & MA_LMV)) {
283 rc = mdt_remote_permission_check(info);
292 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
293 struct mdt_object *obj,
294 struct ldlm_enqueue_info *einfo,
297 union ldlm_policy_data *policy = &mti->mti_policy;
298 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
299 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
302 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
303 LASSERT(slave_locks);
305 memset(policy, 0, sizeof(*policy));
306 policy->l_inodebits.bits = einfo->ei_inodebits;
307 mdt_lock_handle_init(lh);
308 mdt_lock_reg_init(lh, einfo->ei_mode);
309 for (i = 0; i < slave_locks->ha_count; i++) {
310 if (test_bit(i, (void *)slave_locks->ha_map))
311 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
313 lh->mlh_reg_lh = slave_locks->ha_handles[i];
314 mdt_object_unlock(mti, NULL, lh, decref);
315 slave_locks->ha_handles[i].cookie = 0ull;
318 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
322 static inline int mdt_object_striped(struct mdt_thread_info *mti,
323 struct mdt_object *obj)
327 if (!S_ISDIR(obj->mot_header.loh_attr))
330 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
333 return rc == -ENODATA ? 0 : rc;
339 * Lock slave stripes if necessary, the lock handles of slave stripes
340 * will be stored in einfo->ei_cbdata.
342 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
343 enum ldlm_mode mode, __u64 ibits,
344 struct ldlm_enqueue_info *einfo)
346 union ldlm_policy_data *policy = &mti->mti_policy;
348 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
350 einfo->ei_type = LDLM_IBITS;
351 einfo->ei_mode = mode;
352 einfo->ei_cb_bl = mdt_remote_blocking_ast;
353 einfo->ei_cb_local_bl = mdt_blocking_ast;
354 einfo->ei_cb_cp = ldlm_completion_ast;
355 einfo->ei_enq_slave = 1;
356 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
357 einfo->ei_inodebits = ibits;
358 memset(policy, 0, sizeof(*policy));
359 policy->l_inodebits.bits = ibits;
361 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
365 static inline int mdt_reint_striped_lock(struct mdt_thread_info *info,
366 struct mdt_object *o,
367 struct mdt_lock_handle *lh,
369 struct ldlm_enqueue_info *einfo,
374 LASSERT(!mdt_object_remote(o));
376 memset(einfo, 0, sizeof(*einfo));
378 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
382 rc = mdt_object_striped(info, o);
385 mdt_object_unlock(info, o, lh, rc);
389 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
391 mdt_object_unlock(info, o, lh, rc);
392 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
400 mdt_reint_striped_unlock(struct mdt_thread_info *info, struct mdt_object *o,
401 struct mdt_lock_handle *lh,
402 struct ldlm_enqueue_info *einfo, int decref)
404 if (einfo->ei_cbdata)
405 mdt_unlock_slaves(info, o, einfo, decref);
406 mdt_object_unlock(info, o, lh, decref);
410 * VBR: we save three versions in reply:
411 * 0 - parent. Check that parent version is the same during replay.
412 * 1 - name. Version of 'name' if file exists with the same name or
413 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
414 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
417 static int mdt_create(struct mdt_thread_info *info)
419 struct mdt_device *mdt = info->mti_mdt;
420 struct mdt_object *parent;
421 struct mdt_object *child;
422 struct mdt_lock_handle *lh;
423 struct mdt_body *repbody;
424 struct md_attr *ma = &info->mti_attr;
425 struct mdt_reint_record *rr = &info->mti_rr;
429 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
431 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
433 if (!fid_is_md_operative(rr->rr_fid1))
436 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
438 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
440 RETURN(PTR_ERR(parent));
442 if (!mdt_object_exists(parent))
443 GOTO(put_parent, rc = -ENOENT);
445 lh = &info->mti_lh[MDT_LH_PARENT];
446 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
447 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
449 GOTO(put_parent, rc);
451 if (!mdt_object_remote(parent)) {
452 rc = mdt_version_get_check_save(info, parent, 0);
454 GOTO(unlock_parent, rc);
458 * Check child name version during replay.
459 * During create replay a file may exist with same name.
461 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
462 &info->mti_tmp_fid1, 1);
464 GOTO(unlock_parent, rc = -EEXIST);
466 /* -ENOENT is expected here */
468 GOTO(unlock_parent, rc);
470 /* save version of file name for replay, it must be ENOENT here */
471 mdt_enoent_version_save(info, 1);
473 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
474 if (unlikely(IS_ERR(child)))
475 GOTO(unlock_parent, rc = PTR_ERR(child));
477 rc = mdt_remote_permission(info);
481 ma->ma_need = MA_INODE;
484 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
485 OBD_FAIL_MDS_REINT_CREATE_WRITE);
487 /* Version of child will be updated on disk. */
488 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
489 rc = mdt_version_get_check_save(info, child, 2);
493 /* Let lower layer know current lock mode. */
494 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
497 * Do not perform lookup sanity check. We know that name does
500 info->mti_spec.sp_cr_lookup = 0;
501 info->mti_spec.sp_feat = &dt_directory_features;
503 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
504 mdt_object_child(child), &info->mti_spec, ma);
506 rc = mdt_attr_get_complex(info, child, ma);
512 * On DNE, we need to eliminate dependey between 'mkdir a' and
513 * 'mkdir a/b' if b is a striped directory, to achieve this, two
514 * things are done below:
515 * 1. save child and slaves lock.
516 * 2. if the child is a striped directory, relock parent so to
517 * compare against with COS locks to ensure parent was
520 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
521 struct mdt_lock_handle *lhc;
522 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
525 rc = mdt_object_striped(info, child);
531 if (!mdt_object_remote(parent)) {
532 mdt_object_unlock(info, parent, lh, 1);
533 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
534 rc = mdt_reint_object_lock(info, parent, lh,
535 MDS_INODELOCK_UPDATE,
542 lhc = &info->mti_lh[MDT_LH_CHILD];
543 mdt_lock_handle_init(lhc);
544 mdt_lock_reg_init(lhc, LCK_PW);
545 rc = mdt_reint_striped_lock(info, child, lhc,
546 MDS_INODELOCK_UPDATE, einfo,
551 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
554 /* Return fid & attr to client. */
555 if (ma->ma_valid & MA_INODE)
556 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
557 mdt_object_fid(child));
559 mdt_object_put(info->mti_env, child);
561 mdt_object_unlock(info, parent, lh, rc);
563 mdt_object_put(info->mti_env, parent);
567 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
570 struct mdt_lock_handle *lh;
571 int do_vbr = ma->ma_attr.la_valid &
572 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
573 __u64 lockpart = MDS_INODELOCK_UPDATE;
574 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
579 rc = mdt_object_striped(info, mo);
585 lh = &info->mti_lh[MDT_LH_PARENT];
586 mdt_lock_reg_init(lh, LCK_PW);
588 /* Even though the new MDT will grant PERM lock to the old
589 * client, but the old client will almost ignore that during
590 * So it needs to revoke both LOOKUP and PERM lock here, so
591 * both new and old client can cancel the dcache */
592 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
593 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
595 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
600 /* all attrs are packed into mti_attr in unpack_setattr */
601 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
602 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
604 /* VBR: update version if attr changed are important for recovery */
606 /* update on-disk version of changed object */
607 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
608 rc = mdt_version_get_check_save(info, mo, 0);
610 GOTO(out_unlock, rc);
613 /* Ensure constant striping during chown(). See LU-2789. */
614 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
615 mutex_lock(&mo->mot_lov_mutex);
617 /* all attrs are packed into mti_attr in unpack_setattr */
618 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
620 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
621 mutex_unlock(&mo->mot_lov_mutex);
624 GOTO(out_unlock, rc);
625 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
628 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
633 * Check HSM flags and add HS_DIRTY flag if relevant.
635 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
636 * and is not RELEASED.
638 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
644 /* If the file was modified, add the dirty flag */
645 ma->ma_need = MA_HSM;
646 rc = mdt_attr_get_complex(info, mo, ma);
648 CERROR("file attribute read error for "DFID": %d.\n",
649 PFID(mdt_object_fid(mo)), rc);
653 /* If an up2date copy exists in the backend, add dirty flag */
654 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
655 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
656 ma->ma_hsm.mh_flags |= HS_DIRTY;
658 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
660 CERROR("file attribute change error for "DFID": %d\n",
661 PFID(mdt_object_fid(mo)), rc);
667 static int mdt_reint_setattr(struct mdt_thread_info *info,
668 struct mdt_lock_handle *lhc)
670 struct md_attr *ma = &info->mti_attr;
671 struct mdt_reint_record *rr = &info->mti_rr;
672 struct ptlrpc_request *req = mdt_info_req(info);
673 struct mdt_object *mo;
674 struct mdt_body *repbody;
678 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
679 (unsigned int)ma->ma_attr.la_valid);
681 if (info->mti_dlm_req)
682 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
684 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
685 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
687 GOTO(out, rc = PTR_ERR(mo));
689 if (!mdt_object_exists(mo))
690 GOTO(out_put, rc = -ENOENT);
692 if (mdt_object_remote(mo))
693 GOTO(out_put, rc = -EREMOTE);
695 /* revoke lease lock if size is going to be changed */
696 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
697 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
698 atomic_read(&mo->mot_lease_count) > 0)) {
699 down_read(&mo->mot_open_sem);
701 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
702 lhc = &info->mti_lh[MDT_LH_LOCAL];
703 mdt_lock_reg_init(lhc, LCK_CW);
705 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
707 up_read(&mo->mot_open_sem);
711 /* revoke lease lock */
712 mdt_object_unlock(info, mo, lhc, 1);
714 up_read(&mo->mot_open_sem);
717 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
718 /* Check write access for the O_TRUNC case */
719 if (mdt_write_read(mo) < 0)
720 GOTO(out_put, rc = -ETXTBSY);
722 /* LU-10286: compatibility check for FLR.
723 * Please check the comment in mdt_finish_open() for details */
724 if (!exp_connect_flr(info->mti_exp)) {
725 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
726 if (rc < 0 && rc != -ENODATA)
729 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
730 GOTO(out_put, rc = -EOPNOTSUPP);
733 /* For truncate, the file size sent from client
734 * is believable, but the blocks are incorrect,
735 * which makes the block size in LSOM attribute
736 * inconsisent with the real block size.
738 rc = mdt_lsom_update(info, mo, true);
743 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
744 if (ma->ma_valid & MA_LOV)
745 GOTO(out_put, rc = -EPROTO);
747 rc = mdt_attr_set(info, mo, ma);
750 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
751 (ma->ma_valid & MA_INODE)) {
752 struct lu_buf *buf = &info->mti_buf;
753 struct mdt_lock_handle *lh;
755 rc = mdt_remote_permission(info);
759 if (ma->ma_attr.la_valid != 0)
760 GOTO(out_put, rc = -EPROTO);
762 lh = &info->mti_lh[MDT_LH_PARENT];
763 mdt_lock_reg_init(lh, LCK_PW);
765 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
769 if (ma->ma_valid & MA_LOV) {
770 buf->lb_buf = ma->ma_lmm;
771 buf->lb_len = ma->ma_lmm_size;
773 buf->lb_buf = ma->ma_lmv;
774 buf->lb_len = ma->ma_lmv_size;
776 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
777 (ma->ma_valid & MA_LOV) ?
778 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
781 mdt_object_unlock(info, mo, lh, rc);
785 GOTO(out_put, rc = -EPROTO);
788 /* If file data is modified, add the dirty flag */
789 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
790 rc = mdt_add_dirty_flag(info, mo, ma);
792 ma->ma_need = MA_INODE;
794 rc = mdt_attr_get_complex(info, mo, ma);
798 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
802 mdt_object_put(info->mti_env, mo);
805 mdt_counter_incr(req, LPROC_MDT_SETATTR);
807 mdt_client_compatibility(info);
808 rc2 = mdt_fix_reply(info);
814 static int mdt_reint_create(struct mdt_thread_info *info,
815 struct mdt_lock_handle *lhc)
817 struct ptlrpc_request *req = mdt_info_req(info);
821 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
822 RETURN(err_serious(-ESTALE));
824 if (info->mti_dlm_req)
825 ldlm_request_cancel(mdt_info_req(info),
826 info->mti_dlm_req, 0, LATF_SKIP);
828 if (!lu_name_is_valid(&info->mti_rr.rr_name))
831 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
833 mdt_counter_incr(req, LPROC_MDT_MKDIR);
841 /* Special file should stay on the same node as parent. */
842 mdt_counter_incr(req, LPROC_MDT_MKNOD);
845 CERROR("%s: Unsupported mode %o\n",
846 mdt_obd_name(info->mti_mdt),
847 info->mti_attr.ma_attr.la_mode);
848 RETURN(err_serious(-EOPNOTSUPP));
851 rc = mdt_create(info);
856 * VBR: save parent version in reply and child version getting by its name.
857 * Version of child is getting and checking during its lookup. If
859 static int mdt_reint_unlink(struct mdt_thread_info *info,
860 struct mdt_lock_handle *lhc)
862 struct mdt_reint_record *rr = &info->mti_rr;
863 struct ptlrpc_request *req = mdt_info_req(info);
864 struct md_attr *ma = &info->mti_attr;
865 struct lu_fid *child_fid = &info->mti_tmp_fid1;
866 struct mdt_object *mp;
867 struct mdt_object *mc;
868 struct mdt_lock_handle *parent_lh;
869 struct mdt_lock_handle *child_lh;
870 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
872 bool cos_incompat = false, discard = false;
878 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
879 PNAME(&rr->rr_name));
881 if (info->mti_dlm_req)
882 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
884 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
885 RETURN(err_serious(-ENOENT));
887 if (!fid_is_md_operative(rr->rr_fid1))
890 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
894 if (mdt_object_remote(mp)) {
897 rc = mdt_version_get_check_save(info, mp, 0);
899 GOTO(put_parent, rc);
903 parent_lh = &info->mti_lh[MDT_LH_PARENT];
904 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
905 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
908 GOTO(put_parent, rc);
910 /* lookup child object along with version checking */
912 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
914 /* Name might not be able to find during resend of
915 * remote unlink, considering following case.
916 * dir_A is a remote directory, the name entry of
917 * dir_A is on MDT0, the directory is on MDT1,
919 * 1. client sends unlink req to MDT1.
920 * 2. MDT1 sends name delete update to MDT0.
921 * 3. name entry is being deleted in MDT0 synchronously.
922 * 4. MDT1 is restarted.
923 * 5. client resends unlink req to MDT1. So it can not
924 * find the name entry on MDT0 anymore.
925 * In this case, MDT1 only needs to destory the local
928 if (mdt_object_remote(mp) && rc == -ENOENT &&
929 !fid_is_zero(rr->rr_fid2) &&
930 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
932 *child_fid = *rr->rr_fid2;
934 GOTO(unlock_parent, rc);
938 if (!fid_is_md_operative(child_fid))
939 GOTO(unlock_parent, rc = -EPERM);
941 /* We will lock the child regardless it is local or remote. No harm. */
942 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
944 GOTO(unlock_parent, rc = PTR_ERR(mc));
947 rc = mdt_object_striped(info, mc);
949 GOTO(unlock_parent, rc = PTR_ERR(mc));
953 mdt_object_put(info->mti_env, mc);
954 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
959 child_lh = &info->mti_lh[MDT_LH_CHILD];
960 mdt_lock_reg_init(child_lh, LCK_EX);
961 if (info->mti_spec.sp_rm_entry) {
962 struct lu_ucred *uc = mdt_ucred(info);
964 if (!mdt_is_dne_client(req->rq_export))
965 /* Return -ENOTSUPP for old client */
966 GOTO(put_child, rc = -ENOTSUPP);
968 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
969 GOTO(put_child, rc = -EPERM);
971 ma->ma_need = MA_INODE;
973 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
974 NULL, &rr->rr_name, ma, no_name);
978 if (mdt_object_remote(mc)) {
979 struct mdt_body *repbody;
981 if (!fid_is_zero(rr->rr_fid2)) {
982 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
983 mdt_obd_name(info->mti_mdt),
984 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
985 GOTO(put_child, rc = -ENOENT);
987 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
988 mdt_obd_name(info->mti_mdt),
989 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
991 if (!mdt_is_dne_client(req->rq_export))
992 /* Return -ENOTSUPP for old client */
993 GOTO(put_child, rc = -ENOTSUPP);
995 /* Revoke the LOOKUP lock of the remote object granted by
996 * this MDT. Since the unlink will happen on another MDT,
997 * it will release the LOOKUP lock right away. Then What
998 * would happen if another client try to grab the LOOKUP
999 * lock at the same time with unlink XXX */
1000 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1001 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1002 LASSERT(repbody != NULL);
1003 repbody->mbo_fid1 = *mdt_object_fid(mc);
1004 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1005 GOTO(unlock_child, rc = -EREMOTE);
1007 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1008 * this now because a running HSM restore on the child (unlink
1009 * victim) will hold the layout lock. See LU-4002. */
1010 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1011 if (mdt_object_remote(mp)) {
1012 /* Enqueue lookup lock from parent MDT */
1013 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1014 &child_lh->mlh_rreg_lh,
1015 child_lh->mlh_rreg_mode,
1016 MDS_INODELOCK_LOOKUP, false);
1018 GOTO(put_child, rc);
1020 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1023 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1026 GOTO(put_child, rc);
1029 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1030 * whether need MA_LOV and MA_COOKIE.
1032 ma->ma_need = MA_INODE;
1035 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1036 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1037 /* save version when object is locked */
1038 mdt_version_get_save(info, mc, 1);
1040 mutex_lock(&mc->mot_lov_mutex);
1042 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1043 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1045 mutex_unlock(&mc->mot_lov_mutex);
1047 GOTO(unlock_child, rc);
1049 if (!lu_object_is_dying(&mc->mot_header)) {
1050 rc = mdt_attr_get_complex(info, mc, ma);
1054 discard = mdt_dom_check_for_discard(info, mc);
1056 mdt_handle_last_unlink(info, mc, ma);
1059 if (ma->ma_valid & MA_INODE) {
1060 switch (ma->ma_attr.la_mode & S_IFMT) {
1062 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1070 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1073 LASSERTF(0, "bad file type %o unlinking\n",
1074 ma->ma_attr.la_mode);
1081 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1083 mdt_object_put(info->mti_env, mc);
1085 mdt_object_unlock(info, mp, parent_lh, rc);
1087 mdt_object_put(info->mti_env, mp);
1089 /* discard is just a PW DOM lock to drop the data on a client
1090 * no need to keep objects being get and locked, do that after all.
1093 mdt_dom_discard_data(info, child_fid);
1099 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1102 static int mdt_reint_link(struct mdt_thread_info *info,
1103 struct mdt_lock_handle *lhc)
1105 struct mdt_reint_record *rr = &info->mti_rr;
1106 struct ptlrpc_request *req = mdt_info_req(info);
1107 struct md_attr *ma = &info->mti_attr;
1108 struct mdt_object *ms;
1109 struct mdt_object *mp;
1110 struct mdt_lock_handle *lhs;
1111 struct mdt_lock_handle *lhp;
1116 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1117 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1119 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1120 RETURN(err_serious(-ENOENT));
1122 if (info->mti_dlm_req)
1123 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1125 /* Invalid case so return error immediately instead of
1127 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1130 if (!fid_is_md_operative(rr->rr_fid1) ||
1131 !fid_is_md_operative(rr->rr_fid2))
1134 /* step 1: find target parent dir */
1135 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1137 RETURN(PTR_ERR(mp));
1139 rc = mdt_version_get_check_save(info, mp, 0);
1141 GOTO(put_parent, rc);
1143 /* step 2: find source */
1144 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1146 GOTO(put_parent, rc = PTR_ERR(ms));
1148 if (!mdt_object_exists(ms)) {
1149 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1150 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1151 GOTO(put_source, rc = -ENOENT);
1154 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1156 lhp = &info->mti_lh[MDT_LH_PARENT];
1157 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1158 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1161 GOTO(put_source, rc);
1163 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1165 lhs = &info->mti_lh[MDT_LH_CHILD];
1166 mdt_lock_reg_init(lhs, LCK_EX);
1167 rc = mdt_reint_object_lock(info, ms, lhs,
1168 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1171 GOTO(unlock_parent, rc);
1173 /* step 3: link it */
1174 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1175 OBD_FAIL_MDS_REINT_LINK_WRITE);
1177 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1178 rc = mdt_version_get_check_save(info, ms, 1);
1180 GOTO(unlock_source, rc);
1182 /** check target version by name during replay */
1183 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1184 &info->mti_tmp_fid1, 2);
1185 if (rc != 0 && rc != -ENOENT)
1186 GOTO(unlock_source, rc);
1187 /* save version of file name for replay, it must be ENOENT here */
1188 if (!req_is_replay(mdt_info_req(info))) {
1189 if (rc != -ENOENT) {
1190 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1191 PNAME(&rr->rr_name));
1192 GOTO(unlock_source, rc = -EEXIST);
1194 info->mti_ver[2] = ENOENT_VERSION;
1195 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1198 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1199 mdt_object_child(ms), &rr->rr_name, ma);
1202 mdt_counter_incr(req, LPROC_MDT_LINK);
1206 mdt_object_unlock(info, ms, lhs, rc);
1208 mdt_object_unlock(info, mp, lhp, rc);
1210 mdt_object_put(info->mti_env, ms);
1212 mdt_object_put(info->mti_env, mp);
1216 * lock the part of the directory according to the hash of the name
1217 * (lh->mlh_pdo_hash) in parallel directory lock.
1219 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1220 struct mdt_lock_handle *lh,
1221 struct mdt_object *obj, __u64 ibits,
1224 struct ldlm_res_id *res = &info->mti_res_id;
1225 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1226 union ldlm_policy_data *policy = &info->mti_policy;
1227 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1231 * Finish res_id initializing by name hash marking part of
1232 * directory which is taking modification.
1234 LASSERT(lh->mlh_pdo_hash != 0);
1235 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1236 memset(policy, 0, sizeof(*policy));
1237 policy->l_inodebits.bits = ibits;
1239 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1240 dlmflags |= LDLM_FL_COS_INCOMPAT;
1242 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1243 * going to be sent to client. If it is - mdt_intent_policy() path will
1244 * fix it up and turn FL_LOCAL flag off.
1246 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1247 res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1252 * Get BFL lock for rename or migrate process.
1254 static int mdt_rename_lock(struct mdt_thread_info *info,
1255 struct lustre_handle *lh)
1260 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1261 struct lu_fid *fid = &info->mti_tmp_fid1;
1262 struct mdt_object *obj;
1264 /* XXX, right now, it has to use object API to
1265 * enqueue lock cross MDT, so it will enqueue
1266 * rename lock(with LUSTRE_BFL_FID) by root object */
1268 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1270 RETURN(PTR_ERR(obj));
1272 rc = mdt_remote_object_lock(info, obj,
1273 &LUSTRE_BFL_FID, lh,
1275 MDS_INODELOCK_UPDATE, false);
1276 mdt_object_put(info->mti_env, obj);
1278 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1279 union ldlm_policy_data *policy = &info->mti_policy;
1280 struct ldlm_res_id *res_id = &info->mti_res_id;
1283 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1284 memset(policy, 0, sizeof *policy);
1285 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1286 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1287 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1288 LCK_EX, &flags, ldlm_blocking_ast,
1289 ldlm_completion_ast, NULL, NULL, 0,
1291 &info->mti_exp->exp_handle.h_cookie,
1298 static void mdt_rename_unlock(struct lustre_handle *lh)
1301 LASSERT(lustre_handle_is_used(lh));
1302 /* Cancel the single rename lock right away */
1303 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1307 /* Update object linkEA */
1308 struct mdt_lock_list {
1309 struct mdt_object *mll_obj;
1310 struct mdt_lock_handle mll_lh;
1311 struct list_head mll_list;
1314 static void mdt_unlock_list(struct mdt_thread_info *info,
1315 struct list_head *list, int rc)
1317 struct mdt_lock_list *mll;
1318 struct mdt_lock_list *mll2;
1320 list_for_each_entry_safe(mll, mll2, list, mll_list) {
1321 mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc);
1322 list_del(&mll->mll_list);
1327 static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info,
1328 struct mdt_object *obj,
1329 struct mdt_object *pobj,
1330 struct list_head *lock_list)
1332 struct lu_buf *buf = &info->mti_big_buf;
1333 struct linkea_data ldata = { NULL };
1339 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1342 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1343 if (buf->lb_buf == NULL)
1347 rc = mdt_links_read(info, obj, &ldata);
1349 if (rc == -ENOENT || rc == -ENODATA)
1354 /* ignore the migrating parent(@pobj) */
1355 retry_count = ldata.ld_leh->leh_reccount - 1;
1358 LASSERT(ldata.ld_leh != NULL);
1359 ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1);
1360 for (count = 0; count < ldata.ld_leh->leh_reccount; count++) {
1361 struct mdt_device *mdt = info->mti_mdt;
1362 struct mdt_object *mdt_pobj;
1363 struct mdt_lock_list *mll;
1364 struct lu_name name;
1368 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
1370 mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid);
1371 if (IS_ERR(mdt_pobj)) {
1372 CWARN("%s: cannot find obj "DFID": rc = %ld\n",
1373 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj));
1377 if (!mdt_object_exists(mdt_pobj)) {
1378 CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n",
1379 mdt_obd_name(mdt), PFID(&fid));
1380 mdt_object_put(info->mti_env, mdt_pobj);
1384 /* Check if the object already exists in the list */
1385 list_for_each_entry(mll, lock_list, mll_list) {
1386 if (mll->mll_obj == mdt_pobj) {
1387 mdt_object_put(info->mti_env, mdt_pobj);
1392 if (mdt_pobj == pobj) {
1393 CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n",
1394 mdt_obd_name(mdt), PFID(&fid));
1395 mdt_object_put(info->mti_env, mdt_pobj);
1401 mdt_object_put(info->mti_env, mdt_pobj);
1402 GOTO(out, rc = -ENOMEM);
1405 /* Since this needs to lock all of objects in linkea, to avoid
1406 * deadlocks, because it does not follow parent-child order as
1407 * other MDT operation, let's use try_lock here and if the lock
1408 * cannot be gotten because of conflicting locks, then drop all
1409 * current locks, send an AST to the client, and start again. */
1410 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1412 rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits,
1413 MDS_INODELOCK_UPDATE, true);
1414 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1415 mdt_unlock_list(info, lock_list, 0);
1417 CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n",
1418 mdt_obd_name(mdt), PFID(&fid), name.ln_name,
1421 if (retry_count == 0) {
1422 mdt_object_put(info->mti_env, mdt_pobj);
1424 GOTO(out, rc = -EBUSY);
1427 mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name);
1428 rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh,
1429 MDS_INODELOCK_UPDATE);
1431 mdt_object_put(info->mti_env, mdt_pobj);
1436 if (mdt_object_remote(mdt_pobj)) {
1437 struct ldlm_lock *lock;
1439 /* For remote object, Set lock to cb_atomic,
1440 * so lock can be released in blocking_ast()
1441 * immediately, then the next try_lock will
1442 * have better chance to succeds */
1444 ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh);
1445 LASSERT(lock != NULL);
1446 lock_res_and_lock(lock);
1447 ldlm_set_atomic_cb(lock);
1448 unlock_res_and_lock(lock);
1449 LDLM_LOCK_PUT(lock);
1451 mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc);
1457 INIT_LIST_HEAD(&mll->mll_list);
1458 mll->mll_obj = mdt_pobj;
1459 list_add_tail(&mll->mll_list, lock_list);
1461 ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee +
1466 mdt_unlock_list(info, lock_list, rc);
1470 /* migrate files from one MDT to another MDT */
1471 static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
1472 struct mdt_lock_handle *lhc)
1474 struct mdt_reint_record *rr = &info->mti_rr;
1475 struct md_attr *ma = &info->mti_attr;
1476 struct mdt_object *msrcdir;
1477 struct mdt_object *mold;
1478 struct mdt_object *mnew = NULL;
1479 struct mdt_lock_handle *lh_dirp;
1480 struct mdt_lock_handle *lh_childp;
1481 struct mdt_lock_handle *lh_tgtp = NULL;
1482 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1483 struct list_head lock_list;
1485 struct ldlm_lock *lease = NULL;
1486 bool lock_open_sem = false;
1490 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1491 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1493 /* 1: lock the source dir. */
1494 msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1495 if (IS_ERR(msrcdir)) {
1496 CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n",
1497 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
1498 (int)PTR_ERR(msrcdir));
1499 RETURN(PTR_ERR(msrcdir));
1502 lh_dirp = &info->mti_lh[MDT_LH_PARENT];
1503 mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name);
1504 rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE,
1507 GOTO(out_put_parent, rc);
1509 if (!mdt_object_remote(msrcdir)) {
1510 rc = mdt_version_get_check_save(info, msrcdir, 0);
1512 GOTO(out_unlock_parent, rc);
1515 /* 2: sanity check and find the object to be migrated. */
1517 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
1519 GOTO(out_unlock_parent, rc);
1521 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
1522 GOTO(out_unlock_parent, rc = -EINVAL);
1524 if (!fid_is_md_operative(old_fid))
1525 GOTO(out_unlock_parent, rc = -EPERM);
1527 if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid))
1528 GOTO(out_unlock_parent, rc = -EPERM);
1530 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
1532 GOTO(out_unlock_parent, rc = PTR_ERR(mold));
1534 if (!mdt_object_exists(mold)) {
1535 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1537 "object does not exist");
1538 GOTO(out_put_child, rc = -ENOENT);
1541 if (mdt_object_remote(mold)) {
1542 CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n",
1543 mdt_obd_name(info->mti_mdt), PFID(old_fid));
1544 GOTO(out_put_child, rc = -EREMOTE);
1547 if (S_ISREG(lu_object_attr(&mold->mot_obj)) &&
1548 !mdt_object_remote(msrcdir)) {
1549 CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same"
1550 " MDT, which should be migrated first:"
1551 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1552 PFID(mdt_object_fid(msrcdir)), -EPERM);
1553 GOTO(out_put_child, rc = -EPERM);
1556 rc = mdt_remote_permission(info);
1558 GOTO(out_put_child, rc);
1560 /* 3: iterate the linkea of the object and lock all of the objects */
1561 INIT_LIST_HEAD(&lock_list);
1562 rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
1564 GOTO(out_put_child, rc);
1566 if (info->mti_spec.sp_migrate_close) {
1567 struct close_data *data;
1568 struct mdt_body *repbody;
1569 bool lease_broken = false;
1571 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1573 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1575 GOTO(out_lease, rc = -EPROTO);
1577 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1579 GOTO(out_lease, rc = -EPROTO);
1581 lease = ldlm_handle2lock(&data->cd_handle);
1583 GOTO(out_lease, rc = -ESTALE);
1585 /* try to hold open_sem so that nobody else can open the file */
1586 if (!down_write_trylock(&mold->mot_open_sem)) {
1587 ldlm_lock_cancel(lease);
1588 GOTO(out_lease, rc = -EBUSY);
1591 lock_open_sem = true;
1592 /* Check if the lease open lease has already canceled */
1593 lock_res_and_lock(lease);
1594 lease_broken = ldlm_is_cancel(lease);
1595 unlock_res_and_lock(lease);
1597 LDLM_DEBUG(lease, DFID " lease broken? %d",
1598 PFID(mdt_object_fid(mold)), lease_broken);
1600 /* Cancel server side lease. Client side counterpart should
1601 * have been cancelled. It's okay to cancel it now as we've
1602 * held mot_open_sem. */
1603 ldlm_lock_cancel(lease);
1606 GOTO(out_lease, rc = -EAGAIN);
1608 rc = mdt_close_internal(info, mdt_info_req(info), NULL);
1609 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1610 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1612 GOTO(out_unlock_list, rc);
1615 /* 4: lock of the object migrated object */
1616 lh_childp = &info->mti_lh[MDT_LH_OLD];
1617 mdt_lock_reg_init(lh_childp, LCK_EX);
1618 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1619 MDS_INODELOCK_LAYOUT;
1620 if (mdt_object_remote(msrcdir)) {
1621 /* Enqueue lookup lock from the parent MDT */
1622 rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold),
1623 &lh_childp->mlh_rreg_lh,
1624 lh_childp->mlh_rreg_mode,
1625 MDS_INODELOCK_LOOKUP, false);
1627 GOTO(out_unlock_list, rc);
1629 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1632 rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true);
1634 GOTO(out_unlock_child, rc);
1636 /* Migration is incompatible with HSM. */
1637 ma->ma_need = MA_HSM;
1639 rc = mdt_attr_get_complex(info, mold, ma);
1641 GOTO(out_unlock_child, rc);
1643 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) {
1646 "%s: cannot migrate HSM archived file "DFID": rc = %d\n",
1647 mdt_obd_name(info->mti_mdt), PFID(old_fid), rc);
1648 GOTO(out_unlock_child, rc);
1651 ma->ma_need = MA_LMV;
1653 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1654 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1655 rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV);
1657 GOTO(out_unlock_child, rc);
1659 if ((ma->ma_valid & MA_LMV)) {
1660 struct lmv_mds_md_v1 *lmm1;
1662 lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
1663 lmm1 = &ma->ma_lmv->lmv_md_v1;
1664 if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
1665 CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID
1666 ": rc = %d\n", mdt_obd_name(info->mti_mdt),
1667 PFID(mdt_object_fid(mold)), -EPERM);
1668 GOTO(out_unlock_child, rc = -EPERM);
1671 if (!fid_is_sane(&lmm1->lmv_stripe_fids[1]))
1672 GOTO(out_unlock_child, rc = -EINVAL);
1674 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1675 &lmm1->lmv_stripe_fids[1]);
1677 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1679 if (!mdt_object_remote(mnew)) {
1681 "%s: "DFID" being migrated is on this MDT:"
1682 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1683 PFID(rr->rr_fid2), -EPERM);
1684 GOTO(out_put_new, rc = -EPERM);
1687 lh_tgtp = &info->mti_lh[MDT_LH_CHILD];
1688 mdt_lock_reg_init(lh_tgtp, LCK_EX);
1689 rc = mdt_remote_object_lock(info, mnew,
1690 mdt_object_fid(mnew),
1691 &lh_tgtp->mlh_rreg_lh,
1692 lh_tgtp->mlh_rreg_mode,
1693 MDS_INODELOCK_UPDATE, false);
1696 GOTO(out_put_new, rc);
1699 mnew = mdt_object_find(info->mti_env, info->mti_mdt,
1702 GOTO(out_unlock_child, rc = PTR_ERR(mnew));
1703 if (!mdt_object_remote(mnew)) {
1704 CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:"
1705 " rc = %d\n", mdt_obd_name(info->mti_mdt),
1706 PFID(rr->rr_fid2), -EXDEV);
1707 GOTO(out_put_new, rc = -EXDEV);
1712 mdt_reint_init_ma(info, ma);
1714 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1715 OBD_FAIL_MDS_REINT_RENAME_WRITE);
1717 rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir),
1718 mdt_object_child(mold), &rr->rr_name,
1719 mdt_object_child(mnew), ma);
1721 GOTO(out_unlock_new, rc);
1724 if (lh_tgtp != NULL)
1725 mdt_object_unlock(info, mnew, lh_tgtp, rc);
1728 mdt_object_put(info->mti_env, mnew);
1730 mdt_object_unlock(info, mold, lh_childp, rc);
1732 /* we don't really modify linkea objects, so we can safely decref these
1733 * locks, and this can avoid saving them as COS locks, which may prevent
1734 * subsequent migrate. */
1735 mdt_unlock_list(info, &lock_list, 1);
1736 if (lease != NULL) {
1737 ldlm_reprocess_all(lease->l_resource);
1738 LDLM_LOCK_PUT(lease);
1742 up_write(&mold->mot_open_sem);
1744 mdt_object_put(info->mti_env, mold);
1746 mdt_object_unlock(info, msrcdir, lh_dirp, rc);
1748 mdt_object_put(info->mti_env, msrcdir);
1753 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1754 const struct lu_fid *fid,
1757 struct mdt_object *dir;
1761 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1765 /* check early, the real version will be saved after locking */
1766 rc = mdt_version_get_check(info, dir, idx);
1772 mdt_object_put(info->mti_env, dir);
1776 static int mdt_object_lock_save(struct mdt_thread_info *info,
1777 struct mdt_object *dir,
1778 struct mdt_lock_handle *lh,
1779 int idx, bool cos_incompat)
1783 /* we lock the target dir if it is local */
1784 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
1789 /* get and save correct version after locking */
1790 mdt_version_get_save(info, dir, idx);
1795 * determine lock order of sobj and tobj
1797 * there are two situations we need to lock tobj before sobj:
1798 * 1. sobj is child of tobj
1799 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
1800 * larger than that of tobj
1802 * \retval 1 lock tobj before sobj
1803 * \retval 0 lock sobj before tobj
1804 * \retval -ev negative errno upon error
1806 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
1807 struct mdt_object *sobj,
1808 struct mdt_object *tobj)
1810 struct md_attr *ma = &info->mti_attr;
1811 struct lu_fid *spfid = &info->mti_tmp_fid1;
1812 struct lu_fid *tpfid = &info->mti_tmp_fid2;
1813 struct lmv_mds_md_v1 *lmv;
1818 /* sobj and tobj are the same */
1822 if (fid_is_root(mdt_object_fid(sobj)))
1825 if (fid_is_root(mdt_object_fid(tobj)))
1828 /* check whether sobj is child of tobj */
1829 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
1830 mdt_object_fid(tobj));
1837 /* check whether sobj and tobj are children of the same parent */
1838 rc = mdt_attr_get_pfid(info, sobj, spfid);
1842 rc = mdt_attr_get_pfid(info, tobj, tpfid);
1846 if (!lu_fid_eq(spfid, tpfid))
1849 /* check whether sobj and tobj are sibling stripes */
1850 ma->ma_need = MA_LMV;
1852 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
1853 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
1854 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
1858 if (!(ma->ma_valid & MA_LMV))
1861 lmv = &ma->ma_lmv->lmv_md_v1;
1862 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
1864 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
1867 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
1871 if (!(ma->ma_valid & MA_LMV))
1874 lmv = &ma->ma_lmv->lmv_md_v1;
1875 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
1877 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
1879 /* check stripe index of sobj and tobj */
1880 if (sindex == tindex)
1883 return sindex < tindex ? 0 : 1;
1887 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
1888 * 2 - srcdir child; 3 - tgtdir child.
1889 * Update on disk version of srcdir child.
1892 * For DNE phase I, only these renames are allowed
1893 * mv src_p/src_c tgt_p/tgt_c
1894 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
1895 * 2. src_p and tgt_p are same directory, and tgt_c does not
1896 * exists. In this case, all of modification will happen
1897 * in the MDT where ithesource parent is, only one remote
1898 * update is needed, i.e. set c_time/m_time on the child.
1899 * And tgt_c will be still in the same MDT as the original
1902 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
1903 struct mdt_lock_handle *lhc)
1905 struct mdt_reint_record *rr = &info->mti_rr;
1906 struct md_attr *ma = &info->mti_attr;
1907 struct ptlrpc_request *req = mdt_info_req(info);
1908 struct mdt_object *msrcdir = NULL;
1909 struct mdt_object *mtgtdir = NULL;
1910 struct mdt_object *mold;
1911 struct mdt_object *mnew = NULL;
1912 struct mdt_lock_handle *lh_srcdirp;
1913 struct mdt_lock_handle *lh_tgtdirp;
1914 struct mdt_lock_handle *lh_oldp = NULL;
1915 struct mdt_lock_handle *lh_newp = NULL;
1916 struct lu_fid *old_fid = &info->mti_tmp_fid1;
1917 struct lu_fid *new_fid = &info->mti_tmp_fid2;
1919 bool reverse = false;
1920 bool cos_incompat, discard = false;
1924 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
1925 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1926 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
1928 /* find both parents. */
1929 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
1930 if (IS_ERR(msrcdir))
1931 RETURN(PTR_ERR(msrcdir));
1933 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1935 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1937 mdt_object_get(info->mti_env, mtgtdir);
1939 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
1940 if (IS_ERR(mtgtdir))
1941 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
1944 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
1946 GOTO(out_put_tgtdir, rc);
1950 /* source needs to be looked up after locking source parent, otherwise
1951 * this rename may race with unlink source, and cause rename hang, see
1952 * sanityn.sh 55b, so check parents first, if later we found source is
1953 * remote, relock parents. */
1954 cos_incompat = (mdt_object_remote(msrcdir) ||
1955 mdt_object_remote(mtgtdir));
1957 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
1959 /* lock parents in the proper order. */
1960 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
1961 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
1964 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
1965 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
1968 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1971 GOTO(out_put_tgtdir, rc);
1973 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1975 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1978 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
1979 GOTO(out_put_tgtdir, rc);
1982 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
1985 GOTO(out_put_tgtdir, rc);
1987 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
1989 if (mtgtdir != msrcdir) {
1990 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
1992 } else if (lh_srcdirp->mlh_pdo_hash !=
1993 lh_tgtdirp->mlh_pdo_hash) {
1994 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
1995 MDS_INODELOCK_UPDATE,
1997 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2000 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2001 GOTO(out_put_tgtdir, rc);
2005 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2006 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2008 /* find mold object. */
2010 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2012 GOTO(out_unlock_parents, rc);
2014 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2015 GOTO(out_unlock_parents, rc = -EINVAL);
2017 if (!fid_is_md_operative(old_fid))
2018 GOTO(out_unlock_parents, rc = -EPERM);
2020 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2022 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2024 if (!mdt_object_exists(mold)) {
2025 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2027 "object does not exist");
2028 GOTO(out_put_old, rc = -ENOENT);
2031 /* Check if @mtgtdir is subdir of @mold, before locking child
2032 * to avoid reverse locking. */
2033 if (mtgtdir != msrcdir) {
2034 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2039 GOTO(out_put_old, rc);
2043 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2044 /* save version after locking */
2045 mdt_version_get_save(info, mold, 2);
2047 if (!cos_incompat && mdt_object_remote(mold)) {
2048 cos_incompat = true;
2049 mdt_object_put(info->mti_env, mold);
2050 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2051 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2055 /* find mnew object:
2056 * mnew target object may not exist now
2057 * lookup with version checking */
2059 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2062 /* the new_fid should have been filled at this moment */
2063 if (lu_fid_eq(old_fid, new_fid))
2064 GOTO(out_put_old, rc);
2066 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2067 lu_fid_eq(new_fid, rr->rr_fid2))
2068 GOTO(out_put_old, rc = -EINVAL);
2070 if (!fid_is_md_operative(new_fid))
2071 GOTO(out_put_old, rc = -EPERM);
2073 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2075 GOTO(out_put_old, rc = PTR_ERR(mnew));
2077 if (!mdt_object_exists(mnew)) {
2078 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2080 "object does not exist");
2081 GOTO(out_put_new, rc = -ENOENT);
2084 if (mdt_object_remote(mnew)) {
2085 struct mdt_body *repbody;
2087 /* Always send rename req to the target child MDT */
2088 repbody = req_capsule_server_get(info->mti_pill,
2090 LASSERT(repbody != NULL);
2091 repbody->mbo_fid1 = *new_fid;
2092 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2093 GOTO(out_put_new, rc = -EXDEV);
2095 /* Before locking the target dir, check we do not replace
2096 * a dir with a non-dir, otherwise it may deadlock with
2097 * link op which tries to create a link in this dir
2098 * back to this non-dir. */
2099 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2100 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2101 GOTO(out_put_new, rc = -EISDIR);
2103 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2104 mdt_lock_reg_init(lh_oldp, LCK_EX);
2105 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2106 if (mdt_object_remote(msrcdir)) {
2107 /* Enqueue lookup lock from the parent MDT */
2108 rc = mdt_remote_object_lock(info, msrcdir,
2109 mdt_object_fid(mold),
2110 &lh_oldp->mlh_rreg_lh,
2111 lh_oldp->mlh_rreg_mode,
2112 MDS_INODELOCK_LOOKUP,
2115 GOTO(out_put_new, rc);
2117 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2120 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2123 GOTO(out_unlock_old, rc);
2125 /* Check if @msrcdir is subdir of @mnew, before locking child
2126 * to avoid reverse locking. */
2127 if (mtgtdir != msrcdir) {
2128 rc = mdo_is_subdir(info->mti_env,
2129 mdt_object_child(msrcdir), new_fid);
2133 GOTO(out_unlock_old, rc);
2137 /* We used to acquire MDS_INODELOCK_FULL here but we
2138 * can't do this now because a running HSM restore on
2139 * the rename onto victim will hold the layout
2140 * lock. See LU-4002. */
2142 lh_newp = &info->mti_lh[MDT_LH_NEW];
2143 mdt_lock_reg_init(lh_newp, LCK_EX);
2144 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2145 MDS_INODELOCK_LOOKUP |
2146 MDS_INODELOCK_UPDATE,
2149 GOTO(out_unlock_old, rc);
2151 /* get and save version after locking */
2152 mdt_version_get_save(info, mnew, 3);
2153 } else if (rc != -EREMOTE && rc != -ENOENT) {
2154 GOTO(out_put_old, rc);
2156 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2157 mdt_lock_reg_init(lh_oldp, LCK_EX);
2158 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2159 if (mdt_object_remote(msrcdir)) {
2160 /* Enqueue lookup lock from the parent MDT */
2161 rc = mdt_remote_object_lock(info, msrcdir,
2162 mdt_object_fid(mold),
2163 &lh_oldp->mlh_rreg_lh,
2164 lh_oldp->mlh_rreg_mode,
2165 MDS_INODELOCK_LOOKUP,
2168 GOTO(out_put_old, rc);
2170 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2173 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2176 GOTO(out_unlock_old, rc);
2178 mdt_enoent_version_save(info, 3);
2181 /* step 5: rename it */
2182 mdt_reint_init_ma(info, ma);
2184 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2185 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2188 mutex_lock(&mnew->mot_lov_mutex);
2190 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2191 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2192 mnew != NULL ? mdt_object_child(mnew) : NULL,
2193 &rr->rr_tgt_name, ma);
2196 mutex_unlock(&mnew->mot_lov_mutex);
2198 /* handle last link of tgt object */
2200 mdt_counter_incr(req, LPROC_MDT_RENAME);
2202 mdt_handle_last_unlink(info, mnew, ma);
2203 discard = mdt_dom_check_for_discard(info, mnew);
2206 mdt_rename_counter_tally(info, info->mti_mdt, req,
2212 mdt_object_unlock(info, mnew, lh_newp, rc);
2214 mdt_object_unlock(info, mold, lh_oldp, rc);
2217 mdt_object_put(info->mti_env, mnew);
2219 mdt_object_put(info->mti_env, mold);
2221 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2222 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2224 mdt_object_put(info->mti_env, mtgtdir);
2226 mdt_object_put(info->mti_env, msrcdir);
2228 /* If 'discard' is set then new_fid must exits.
2229 * DOM data discard need neither object nor lock,
2230 * so do this at the end.
2233 mdt_dom_discard_data(info, new_fid);
2238 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2239 struct mdt_lock_handle *lhc, bool rename)
2241 struct mdt_reint_record *rr = &info->mti_rr;
2242 struct ptlrpc_request *req = mdt_info_req(info);
2243 struct lustre_handle rename_lh = { 0 };
2247 if (info->mti_dlm_req)
2248 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2250 if (!fid_is_md_operative(rr->rr_fid1) ||
2251 !fid_is_md_operative(rr->rr_fid2))
2254 /* Note: do not enqueue rename lock for replay request, because
2255 * if other MDT holds rename lock, but being blocked to wait for
2256 * this MDT to finish its recovery, and the failover MDT can not
2257 * get rename lock, which will cause deadlock. */
2258 if (!req_is_replay(req)) {
2259 rc = mdt_rename_lock(info, &rename_lh);
2261 CERROR("%s: can't lock FS for rename: rc = %d\n",
2262 mdt_obd_name(info->mti_mdt), rc);
2268 rc = mdt_reint_rename_internal(info, lhc);
2270 rc = mdt_reint_migrate_internal(info, lhc);
2272 if (lustre_handle_is_used(&rename_lh))
2273 mdt_rename_unlock(&rename_lh);
2278 static int mdt_reint_rename(struct mdt_thread_info *info,
2279 struct mdt_lock_handle *lhc)
2281 return mdt_reint_rename_or_migrate(info, lhc, true);
2284 static int mdt_reint_migrate(struct mdt_thread_info *info,
2285 struct mdt_lock_handle *lhc)
2287 return mdt_reint_rename_or_migrate(info, lhc, false);
2290 static int mdt_reint_resync(struct mdt_thread_info *info,
2291 struct mdt_lock_handle *lhc)
2293 struct mdt_reint_record *rr = &info->mti_rr;
2294 struct ptlrpc_request *req = mdt_info_req(info);
2295 struct md_attr *ma = &info->mti_attr;
2296 struct mdt_object *mo;
2297 struct ldlm_lock *lease;
2298 struct mdt_body *repbody;
2299 struct md_layout_change layout = { 0 };
2304 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2306 if (info->mti_dlm_req)
2307 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2309 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2311 GOTO(out, rc = PTR_ERR(mo));
2313 if (!mdt_object_exists(mo))
2314 GOTO(out_obj, rc = -ENOENT);
2316 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2317 GOTO(out_obj, rc = -EINVAL);
2319 if (mdt_object_remote(mo))
2320 GOTO(out_obj, rc = -EREMOTE);
2322 lease = ldlm_handle2lock(rr->rr_handle);
2324 GOTO(out_obj, rc = -ESTALE);
2326 /* It's really necessary to grab open_sem and check if the lease lock
2327 * has been lost. There would exist a concurrent writer coming in and
2328 * generating some dirty data in memory cache, the writeback would fail
2329 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2330 if (!down_write_trylock(&mo->mot_open_sem))
2331 GOTO(out_put_lease, rc = -EBUSY);
2333 lock_res_and_lock(lease);
2334 lease_broken = ldlm_is_cancel(lease);
2335 unlock_res_and_lock(lease);
2337 GOTO(out_unlock, rc = -EBUSY);
2339 /* the file has yet opened by anyone else after we took the lease. */
2340 layout.mlc_opc = MD_LAYOUT_RESYNC;
2341 rc = mdt_layout_change(info, mo, &layout);
2343 GOTO(out_unlock, rc);
2345 ma->ma_need = MA_INODE;
2347 rc = mdt_attr_get_complex(info, mo, ma);
2349 GOTO(out_unlock, rc);
2351 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2352 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2356 up_write(&mo->mot_open_sem);
2358 LDLM_LOCK_PUT(lease);
2360 mdt_object_put(info->mti_env, mo);
2362 mdt_client_compatibility(info);
2363 rc2 = mdt_fix_reply(info);
2369 struct mdt_reinter {
2370 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2371 enum lprocfs_extra_opc mr_extra_opc;
2374 static const struct mdt_reinter mdt_reinters[] = {
2376 .mr_handler = &mdt_reint_setattr,
2377 .mr_extra_opc = MDS_REINT_SETATTR,
2380 .mr_handler = &mdt_reint_create,
2381 .mr_extra_opc = MDS_REINT_CREATE,
2384 .mr_handler = &mdt_reint_link,
2385 .mr_extra_opc = MDS_REINT_LINK,
2388 .mr_handler = &mdt_reint_unlink,
2389 .mr_extra_opc = MDS_REINT_UNLINK,
2392 .mr_handler = &mdt_reint_rename,
2393 .mr_extra_opc = MDS_REINT_RENAME,
2396 .mr_handler = &mdt_reint_open,
2397 .mr_extra_opc = MDS_REINT_OPEN,
2399 [REINT_SETXATTR] = {
2400 .mr_handler = &mdt_reint_setxattr,
2401 .mr_extra_opc = MDS_REINT_SETXATTR,
2404 .mr_handler = &mdt_reint_unlink,
2405 .mr_extra_opc = MDS_REINT_UNLINK,
2408 .mr_handler = &mdt_reint_migrate,
2409 .mr_extra_opc = MDS_REINT_RENAME,
2412 .mr_handler = &mdt_reint_resync,
2413 .mr_extra_opc = MDS_REINT_RESYNC,
2417 int mdt_reint_rec(struct mdt_thread_info *info,
2418 struct mdt_lock_handle *lhc)
2420 const struct mdt_reinter *mr;
2424 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2427 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2428 if (mr->mr_handler == NULL)
2431 rc = (*mr->mr_handler)(info, lhc);
2433 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2434 PTLRPC_LAST_CNTR + mr->mr_extra_opc);