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 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 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))
399 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
400 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[0];
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[0];
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[0];
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 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1308 const struct lu_fid *fid,
1311 struct mdt_object *dir;
1316 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1320 /* check early, the real version will be saved after locking */
1321 rc = mdt_version_get_check(info, dir, idx);
1327 mdt_object_put(info->mti_env, dir);
1332 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1333 * herein we don't really check it, just do revoke.
1335 static int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1336 struct mdt_object *pobj,
1337 struct mdt_object *obj)
1339 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1342 mdt_lock_handle_init(lh);
1343 mdt_lock_reg_init(lh, LCK_EX);
1345 if (mdt_object_remote(pobj)) {
1346 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1347 &lh->mlh_rreg_lh, LCK_EX,
1348 MDS_INODELOCK_LOOKUP, false);
1350 struct ldlm_res_id *res = &info->mti_res_id;
1351 union ldlm_policy_data *policy = &info->mti_policy;
1352 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1353 LDLM_FL_COS_INCOMPAT;
1355 fid_build_reg_res_name(mdt_object_fid(obj), res);
1356 memset(policy, 0, sizeof(*policy));
1357 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1358 rc = mdt_fid_lock(info->mti_mdt->mdt_namespace, &lh->mlh_reg_lh,
1359 LCK_EX, policy, res, dlmflags, NULL);
1366 * TODO, currently we don't save this lock because there is no place to
1367 * hold this lock handle, but to avoid race we need to save this lock.
1369 mdt_object_unlock(info, NULL, lh, 1);
1375 * operation may takes locks of linkea, or directory stripes, group them in
1378 struct mdt_sub_lock {
1379 struct mdt_object *msl_obj;
1380 struct mdt_lock_handle msl_lh;
1381 struct list_head msl_linkage;
1384 static void mdt_unlock_list(struct mdt_thread_info *info,
1385 struct list_head *list, int decref)
1387 struct mdt_sub_lock *msl;
1388 struct mdt_sub_lock *tmp;
1390 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1391 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1392 list_del(&msl->msl_linkage);
1398 * lock parents of links, and also check whether total locks don't exceed
1401 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1402 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1403 * \retval -ev negative errno upon error
1405 static int mdt_lock_links(struct mdt_thread_info *info,
1406 struct mdt_object *pobj,
1407 const struct md_attr *ma,
1408 struct mdt_object *obj,
1409 struct list_head *link_locks)
1411 struct mdt_device *mdt = info->mti_mdt;
1412 struct lu_buf *buf = &info->mti_big_buf;
1413 struct lu_name *lname = &info->mti_name;
1414 struct linkea_data ldata = { NULL };
1415 bool blocked = false;
1417 int local_lnkp_cnt = 0;
1422 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1425 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1426 if (buf->lb_buf == NULL)
1430 rc = mdt_links_read(info, obj, &ldata);
1432 if (rc == -ENOENT || rc == -ENODATA)
1438 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1439 linkea_next_entry(&ldata)) {
1440 struct mdt_object *lnkp;
1441 struct mdt_sub_lock *msl;
1445 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1448 /* check if it's also linked to parent */
1449 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1450 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1451 PFID(&fid), PNAME(lname));
1452 /* in case link is remote object, revoke LOOKUP lock */
1453 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1459 /* check if it's linked to a stripe of parent */
1460 if (ma->ma_valid & MA_LMV) {
1461 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1462 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1465 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1466 fid_le_to_cpu(stripe_fid,
1467 &lmv->lmv_stripe_fids[j]);
1468 if (lu_fid_eq(stripe_fid, &fid)) {
1469 CDEBUG(D_INFO, "skip stripe "DFID
1470 ", reovke "DNAME"\n",
1471 PFID(&fid), PNAME(lname));
1472 lnkp = mdt_object_find(info->mti_env,
1475 GOTO(out, rc = PTR_ERR(lnkp));
1481 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1483 mdt_object_put(info->mti_env, lnkp);
1488 /* Check if it's already locked */
1489 list_for_each_entry(msl, link_locks, msl_linkage) {
1490 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1492 DFID" was locked, revoke "DNAME"\n",
1493 PFID(&fid), PNAME(lname));
1494 lnkp = msl->msl_obj;
1500 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1504 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1505 PFID(&fid), PNAME(lname));
1507 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1509 CWARN("%s: cannot find obj "DFID": %ld\n",
1510 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1514 if (!mdt_object_exists(lnkp)) {
1515 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1516 PFID(&fid), PNAME(lname));
1517 mdt_object_put(info->mti_env, lnkp);
1521 if (!mdt_object_remote(lnkp))
1526 GOTO(out, rc = -ENOMEM);
1529 * we can't follow parent-child lock order like other MD
1530 * operations, use lock_try here to avoid deadlock, if the lock
1531 * cannot be taken, drop all locks taken, revoke the blocked
1532 * one, and continue processing the remaining entries, and in
1533 * the end of the loop restart from beginning.
1535 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1537 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1538 MDS_INODELOCK_UPDATE, true);
1539 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1542 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1543 PFID(&fid), PNAME(lname), retries);
1545 mdt_unlock_list(info, link_locks, 1);
1547 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1548 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1549 MDS_INODELOCK_UPDATE);
1551 mdt_object_put(info->mti_env, lnkp);
1556 if (mdt_object_remote(lnkp)) {
1557 struct ldlm_lock *lock;
1560 * for remote object, set lock cb_atomic,
1561 * so lock can be released in blocking_ast()
1562 * immediately, then the next lock_try will
1563 * have better chance of success.
1565 lock = ldlm_handle2lock(
1566 &msl->msl_lh.mlh_rreg_lh);
1567 LASSERT(lock != NULL);
1568 lock_res_and_lock(lock);
1569 ldlm_set_atomic_cb(lock);
1570 unlock_res_and_lock(lock);
1571 LDLM_LOCK_PUT(lock);
1574 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1579 INIT_LIST_HEAD(&msl->msl_linkage);
1580 msl->msl_obj = lnkp;
1581 list_add_tail(&msl->msl_linkage, link_locks);
1583 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1588 if (--retries > 0) {
1589 mdt_unlock_list(info, link_locks, rc);
1599 mdt_unlock_list(info, link_locks, rc);
1600 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1602 * parent may have 3 local objects: master object and 2 stripes
1603 * (if it's being migrated too); source may have 2 local
1604 * objects: master and 1 stripe; target has 1 local object.
1610 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1611 struct mdt_object *obj,
1612 const struct md_attr *ma,
1613 struct list_head *slave_locks)
1615 struct mdt_device *mdt = info->mti_mdt;
1616 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1617 struct lu_fid *fid = &info->mti_tmp_fid1;
1618 struct mdt_object *slave;
1619 struct mdt_sub_lock *msl;
1625 LASSERT(mdt_object_remote(obj));
1626 LASSERT(ma->ma_valid & MA_LMV);
1629 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1632 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1635 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1636 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1638 slave = mdt_object_find(info->mti_env, mdt, fid);
1640 GOTO(out, rc = PTR_ERR(slave));
1644 mdt_object_put(info->mti_env, slave);
1645 GOTO(out, rc = -ENOMEM);
1648 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1649 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1650 MDS_INODELOCK_UPDATE, true);
1653 mdt_object_put(info->mti_env, slave);
1657 INIT_LIST_HEAD(&msl->msl_linkage);
1658 msl->msl_obj = slave;
1659 list_add_tail(&msl->msl_linkage, slave_locks);
1666 mdt_unlock_list(info, slave_locks, rc);
1670 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1671 struct mdt_object *obj,
1672 struct mdt_lock_handle *lh,
1673 struct ldlm_enqueue_info *einfo,
1674 struct list_head *slave_locks,
1677 if (mdt_object_remote(obj)) {
1678 mdt_unlock_list(info, slave_locks, decref);
1679 mdt_object_unlock(info, obj, lh, decref);
1681 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1685 /* lock parent and its stripes */
1686 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1687 struct mdt_object *obj,
1688 const struct md_attr *ma,
1689 struct mdt_lock_handle *lh,
1690 struct ldlm_enqueue_info *einfo,
1691 struct list_head *slave_locks)
1695 if (mdt_object_remote(obj)) {
1696 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1697 &lh->mlh_rreg_lh, LCK_PW,
1698 MDS_INODELOCK_UPDATE, false);
1703 * if obj is remote and striped, lock its stripes explicitly
1704 * because it's not striped in LOD layer on this MDT.
1706 if (ma->ma_valid & MA_LMV) {
1707 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1709 mdt_object_unlock(info, obj, lh, rc);
1712 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1720 * in migration, object may be remote, and we need take full lock of it and its
1721 * stripes if it's directory, besides, object may be a remote object on its
1722 * parent, revoke its LOOKUP lock on where its parent is located.
1724 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1725 struct mdt_object *pobj,
1726 struct mdt_object *obj,
1727 struct mdt_lock_handle *lh,
1728 struct ldlm_enqueue_info *einfo,
1729 struct list_head *slave_locks)
1733 if (mdt_object_remote(obj)) {
1734 /* don't bother to check if pobj and obj are on the same MDT. */
1735 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1739 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1740 &lh->mlh_rreg_lh, LCK_EX,
1741 MDS_INODELOCK_FULL, false);
1746 * if obj is remote and striped, lock its stripes explicitly
1747 * because it's not striped in LOD layer on this MDT.
1749 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1750 struct md_attr *ma = &info->mti_attr;
1752 ma->ma_lmv = info->mti_big_lmm;
1753 ma->ma_lmv_size = info->mti_big_lmmsize;
1755 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1757 mdt_object_unlock(info, obj, lh, rc);
1761 if (ma->ma_valid & MA_LMV) {
1762 rc = mdt_lock_remote_slaves(info, obj, ma,
1765 mdt_object_unlock(info, obj, lh, rc);
1769 if (mdt_object_remote(pobj)) {
1770 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1775 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1783 * lookup source by name, if parent is striped directory, we need to find the
1784 * corresponding stripe where source is located, and then lookup there.
1786 * besides, if parent is migrating too, and file is already in target stripe,
1787 * this should be a redo of 'lfs migrate' on client side.
1789 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1790 struct mdt_object *pobj,
1791 const struct md_attr *ma,
1792 const struct lu_name *lname,
1793 struct mdt_object **spobj,
1794 struct mdt_object **sobj)
1796 const struct lu_env *env = info->mti_env;
1797 struct lu_fid *fid = &info->mti_tmp_fid1;
1798 struct mdt_object *stripe;
1801 if (ma->ma_valid & MA_LMV) {
1802 /* if parent is striped, lookup on corresponding stripe */
1803 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1804 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1805 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1806 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1807 LMV_HASH_FLAG_MIGRATION;
1810 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1811 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1814 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1820 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1821 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1823 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1825 stripe = mdt_object_find(env, info->mti_mdt, fid);
1827 return PTR_ERR(stripe);
1830 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1832 if (rc == -ENOENT && is_migrating) {
1834 * if parent is migrating, and lookup child failed on
1835 * source stripe, lookup again on target stripe, if it
1836 * exists, it means previous migration was interrupted,
1837 * and current file was migrated already.
1839 mdt_object_put(env, stripe);
1841 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1842 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1844 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1850 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1852 stripe = mdt_object_find(env, info->mti_mdt, fid);
1854 return PTR_ERR(stripe);
1857 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1858 fid, &info->mti_spec);
1859 mdt_object_put(env, stripe);
1860 return rc ?: -EALREADY;
1862 mdt_object_put(env, stripe);
1867 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1873 mdt_object_get(env, stripe);
1878 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1879 if (IS_ERR(*sobj)) {
1880 mdt_object_put(env, stripe);
1881 rc = PTR_ERR(*sobj);
1889 /* end lease and close file for regular file */
1890 static int mdd_migrate_close(struct mdt_thread_info *info,
1891 struct mdt_object *obj)
1893 struct close_data *data;
1894 struct mdt_body *repbody;
1895 struct ldlm_lock *lease;
1900 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1902 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1906 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1911 lease = ldlm_handle2lock(&data->cd_handle);
1915 /* check if the lease was already canceled */
1916 lock_res_and_lock(lease);
1917 rc = ldlm_is_cancel(lease);
1918 unlock_res_and_lock(lease);
1922 LDLM_DEBUG(lease, DFID" lease broken",
1923 PFID(mdt_object_fid(obj)));
1927 * cancel server side lease, client side counterpart should have been
1928 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1930 ldlm_lock_cancel(lease);
1931 ldlm_reprocess_all(lease->l_resource);
1932 LDLM_LOCK_PUT(lease);
1935 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1936 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1937 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1943 * migrate file in below steps:
1944 * 1. lock parent and its stripes
1945 * 2. lookup source by name
1946 * 3. lock parents of source links if source is not directory
1947 * 4. reject if source is in HSM
1948 * 5. take source open_sem and close file if source is regular file
1949 * 6. lock source and its stripes if it's directory
1950 * 7. lock target so subsequent change to it can trigger COS
1952 * 9. unlock above locks
1953 * 10. sync device if source has links
1955 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1957 const struct lu_env *env = info->mti_env;
1958 struct mdt_device *mdt = info->mti_mdt;
1959 struct mdt_reint_record *rr = &info->mti_rr;
1960 struct md_attr *ma = &info->mti_attr;
1961 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1962 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1963 struct mdt_object *pobj;
1964 struct mdt_object *spobj = NULL;
1965 struct mdt_object *sobj = NULL;
1966 struct mdt_object *tobj;
1967 struct mdt_lock_handle *lhp;
1968 struct mdt_lock_handle *lhs;
1969 struct mdt_lock_handle *lht;
1970 LIST_HEAD(parent_slave_locks);
1971 LIST_HEAD(child_slave_locks);
1972 LIST_HEAD(link_locks);
1973 bool open_sem_locked = false;
1974 bool do_sync = false;
1978 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1979 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1981 /* don't allow migrate . or .. */
1982 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1985 rc = mdt_remote_permission(info);
1989 /* pobj is master object of parent */
1990 pobj = mdt_object_find_check(info, rr->rr_fid1, 0);
1992 RETURN(PTR_ERR(pobj));
1994 if (unlikely(!info->mti_big_lmm)) {
1995 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
1996 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
1997 if (!info->mti_big_lmm)
1998 GOTO(put_parent, rc = -ENOMEM);
2001 ma->ma_lmv = info->mti_big_lmm;
2002 ma->ma_lmv_size = info->mti_big_lmmsize;
2004 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2006 GOTO(put_parent, rc);
2008 /* lock parent object */
2009 lhp = &info->mti_lh[MDT_LH_PARENT];
2010 mdt_lock_reg_init(lhp, LCK_PW);
2011 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2012 &parent_slave_locks);
2014 GOTO(put_parent, rc);
2017 * spobj is the corresponding stripe against name if pobj is striped
2018 * directory, which is the real parent, and no need to lock, because
2019 * we've taken full lock of pobj.
2021 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2023 GOTO(unlock_parent, rc);
2025 /* lock parents of source links, and revoke LOOKUP lock of links */
2026 rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2028 GOTO(put_source, rc);
2031 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2032 * with one request, if total lock count exceeds this limit, we will
2033 * drop all locks after migration, and synchronous device in the end.
2037 /* if migration HSM is allowed */
2038 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2039 ma->ma_need = MA_HSM;
2041 rc = mdt_attr_get_complex(info, sobj, ma);
2043 GOTO(unlock_links, rc);
2045 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2046 GOTO(unlock_links, rc = -EOPNOTSUPP);
2049 /* end lease and close file for regular file */
2050 if (info->mti_spec.sp_migrate_close) {
2051 /* try to hold open_sem so that nobody else can open the file */
2052 if (!down_write_trylock(&sobj->mot_open_sem)) {
2054 mdd_migrate_close(info, sobj);
2055 GOTO(unlock_links, rc = -EBUSY);
2057 open_sem_locked = true;
2058 rc = mdd_migrate_close(info, sobj);
2060 GOTO(unlock_open_sem, rc);
2065 lhs = &info->mti_lh[MDT_LH_OLD];
2066 mdt_lock_reg_init(lhs, LCK_EX);
2067 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2068 &child_slave_locks);
2070 GOTO(unlock_open_sem, rc);
2073 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2075 GOTO(unlock_source, rc = PTR_ERR(tobj));
2077 lht = &info->mti_lh[MDT_LH_NEW];
2078 mdt_lock_reg_init(lht, LCK_EX);
2079 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2081 GOTO(put_target, rc);
2083 /* Don't do lookup sanity check. We know name doesn't exist. */
2084 info->mti_spec.sp_cr_lookup = 0;
2085 info->mti_spec.sp_feat = &dt_directory_features;
2087 rc = mdo_migrate(env, mdt_object_child(pobj),
2088 mdt_object_child(sobj), &rr->rr_name,
2089 mdt_object_child(tobj), &info->mti_spec, ma);
2092 mdt_object_unlock(info, tobj, lht, rc);
2094 mdt_object_put(env, tobj);
2096 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2097 &child_slave_locks, rc);
2099 if (open_sem_locked)
2100 up_write(&sobj->mot_open_sem);
2102 mdt_unlock_list(info, &link_locks, rc);
2104 mdt_object_put(env, sobj);
2105 mdt_object_put(env, spobj);
2107 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2108 &parent_slave_locks, rc);
2110 mdt_object_put(env, pobj);
2113 mdt_device_sync(env, mdt);
2118 static int mdt_object_lock_save(struct mdt_thread_info *info,
2119 struct mdt_object *dir,
2120 struct mdt_lock_handle *lh,
2121 int idx, bool cos_incompat)
2125 /* we lock the target dir if it is local */
2126 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2131 /* get and save correct version after locking */
2132 mdt_version_get_save(info, dir, idx);
2137 * determine lock order of sobj and tobj
2139 * there are two situations we need to lock tobj before sobj:
2140 * 1. sobj is child of tobj
2141 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2142 * larger than that of tobj
2144 * \retval 1 lock tobj before sobj
2145 * \retval 0 lock sobj before tobj
2146 * \retval -ev negative errno upon error
2148 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2149 struct mdt_object *sobj,
2150 struct mdt_object *tobj)
2152 struct md_attr *ma = &info->mti_attr;
2153 struct lu_fid *spfid = &info->mti_tmp_fid1;
2154 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2155 struct lmv_mds_md_v1 *lmv;
2160 /* sobj and tobj are the same */
2164 if (fid_is_root(mdt_object_fid(sobj)))
2167 if (fid_is_root(mdt_object_fid(tobj)))
2170 /* check whether sobj is child of tobj */
2171 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2172 mdt_object_fid(tobj));
2179 /* check whether sobj and tobj are children of the same parent */
2180 rc = mdt_attr_get_pfid(info, sobj, spfid);
2184 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2188 if (!lu_fid_eq(spfid, tpfid))
2191 /* check whether sobj and tobj are sibling stripes */
2192 ma->ma_need = MA_LMV;
2194 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2195 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2196 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2200 if (!(ma->ma_valid & MA_LMV))
2203 lmv = &ma->ma_lmv->lmv_md_v1;
2204 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2206 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2209 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2213 if (!(ma->ma_valid & MA_LMV))
2216 lmv = &ma->ma_lmv->lmv_md_v1;
2217 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2219 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2221 /* check stripe index of sobj and tobj */
2222 if (sindex == tindex)
2225 return sindex < tindex ? 0 : 1;
2229 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2230 * 2 - srcdir child; 3 - tgtdir child.
2231 * Update on disk version of srcdir child.
2234 * For DNE phase I, only these renames are allowed
2235 * mv src_p/src_c tgt_p/tgt_c
2236 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2237 * 2. src_p and tgt_p are same directory, and tgt_c does not
2238 * exists. In this case, all of modification will happen
2239 * in the MDT where ithesource parent is, only one remote
2240 * update is needed, i.e. set c_time/m_time on the child.
2241 * And tgt_c will be still in the same MDT as the original
2244 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2245 struct mdt_lock_handle *lhc)
2247 struct mdt_reint_record *rr = &info->mti_rr;
2248 struct md_attr *ma = &info->mti_attr;
2249 struct ptlrpc_request *req = mdt_info_req(info);
2250 struct mdt_object *msrcdir = NULL;
2251 struct mdt_object *mtgtdir = NULL;
2252 struct mdt_object *mold;
2253 struct mdt_object *mnew = NULL;
2254 struct mdt_lock_handle *lh_srcdirp;
2255 struct mdt_lock_handle *lh_tgtdirp;
2256 struct mdt_lock_handle *lh_oldp = NULL;
2257 struct mdt_lock_handle *lh_newp = NULL;
2258 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2259 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2261 bool reverse = false;
2262 bool cos_incompat, discard = false;
2266 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2267 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2268 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2270 /* find both parents. */
2271 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
2272 if (IS_ERR(msrcdir))
2273 RETURN(PTR_ERR(msrcdir));
2275 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2277 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2279 mdt_object_get(info->mti_env, mtgtdir);
2281 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
2282 if (IS_ERR(mtgtdir))
2283 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2286 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2288 GOTO(out_put_tgtdir, rc);
2292 /* source needs to be looked up after locking source parent, otherwise
2293 * this rename may race with unlink source, and cause rename hang, see
2294 * sanityn.sh 55b, so check parents first, if later we found source is
2295 * remote, relock parents. */
2296 cos_incompat = (mdt_object_remote(msrcdir) ||
2297 mdt_object_remote(mtgtdir));
2299 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2301 /* lock parents in the proper order. */
2302 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2303 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2306 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2307 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2310 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2313 GOTO(out_put_tgtdir, rc);
2315 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2317 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2320 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2321 GOTO(out_put_tgtdir, rc);
2324 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2327 GOTO(out_put_tgtdir, rc);
2329 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2331 if (mtgtdir != msrcdir) {
2332 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2334 } else if (lh_srcdirp->mlh_pdo_hash !=
2335 lh_tgtdirp->mlh_pdo_hash) {
2336 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2337 MDS_INODELOCK_UPDATE,
2339 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2342 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2343 GOTO(out_put_tgtdir, rc);
2347 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2348 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2350 /* find mold object. */
2352 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2354 GOTO(out_unlock_parents, rc);
2356 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2357 GOTO(out_unlock_parents, rc = -EINVAL);
2359 if (!fid_is_md_operative(old_fid))
2360 GOTO(out_unlock_parents, rc = -EPERM);
2362 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2364 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2366 if (!mdt_object_exists(mold)) {
2367 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2369 "object does not exist");
2370 GOTO(out_put_old, rc = -ENOENT);
2373 /* Check if @mtgtdir is subdir of @mold, before locking child
2374 * to avoid reverse locking. */
2375 if (mtgtdir != msrcdir) {
2376 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2381 GOTO(out_put_old, rc);
2385 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2386 /* save version after locking */
2387 mdt_version_get_save(info, mold, 2);
2389 if (!cos_incompat && mdt_object_remote(mold)) {
2390 cos_incompat = true;
2391 mdt_object_put(info->mti_env, mold);
2392 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2393 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2397 /* find mnew object:
2398 * mnew target object may not exist now
2399 * lookup with version checking */
2401 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2404 /* the new_fid should have been filled at this moment */
2405 if (lu_fid_eq(old_fid, new_fid))
2406 GOTO(out_put_old, rc);
2408 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2409 lu_fid_eq(new_fid, rr->rr_fid2))
2410 GOTO(out_put_old, rc = -EINVAL);
2412 if (!fid_is_md_operative(new_fid))
2413 GOTO(out_put_old, rc = -EPERM);
2415 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2417 GOTO(out_put_old, rc = PTR_ERR(mnew));
2419 if (!mdt_object_exists(mnew)) {
2420 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2422 "object does not exist");
2423 GOTO(out_put_new, rc = -ENOENT);
2426 if (mdt_object_remote(mnew)) {
2427 struct mdt_body *repbody;
2429 /* Always send rename req to the target child MDT */
2430 repbody = req_capsule_server_get(info->mti_pill,
2432 LASSERT(repbody != NULL);
2433 repbody->mbo_fid1 = *new_fid;
2434 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2435 GOTO(out_put_new, rc = -EXDEV);
2437 /* Before locking the target dir, check we do not replace
2438 * a dir with a non-dir, otherwise it may deadlock with
2439 * link op which tries to create a link in this dir
2440 * back to this non-dir. */
2441 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2442 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2443 GOTO(out_put_new, rc = -EISDIR);
2445 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2446 mdt_lock_reg_init(lh_oldp, LCK_EX);
2447 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2448 if (mdt_object_remote(msrcdir)) {
2449 /* Enqueue lookup lock from the parent MDT */
2450 rc = mdt_remote_object_lock(info, msrcdir,
2451 mdt_object_fid(mold),
2452 &lh_oldp->mlh_rreg_lh,
2453 lh_oldp->mlh_rreg_mode,
2454 MDS_INODELOCK_LOOKUP,
2457 GOTO(out_put_new, rc);
2459 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2462 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2465 GOTO(out_unlock_old, rc);
2467 /* Check if @msrcdir is subdir of @mnew, before locking child
2468 * to avoid reverse locking. */
2469 if (mtgtdir != msrcdir) {
2470 rc = mdo_is_subdir(info->mti_env,
2471 mdt_object_child(msrcdir), new_fid);
2475 GOTO(out_unlock_old, rc);
2479 /* We used to acquire MDS_INODELOCK_FULL here but we
2480 * can't do this now because a running HSM restore on
2481 * the rename onto victim will hold the layout
2482 * lock. See LU-4002. */
2484 lh_newp = &info->mti_lh[MDT_LH_NEW];
2485 mdt_lock_reg_init(lh_newp, LCK_EX);
2486 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2487 MDS_INODELOCK_LOOKUP |
2488 MDS_INODELOCK_UPDATE,
2491 GOTO(out_unlock_old, rc);
2493 /* get and save version after locking */
2494 mdt_version_get_save(info, mnew, 3);
2495 } else if (rc != -EREMOTE && rc != -ENOENT) {
2496 GOTO(out_put_old, rc);
2498 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2499 mdt_lock_reg_init(lh_oldp, LCK_EX);
2500 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2501 if (mdt_object_remote(msrcdir)) {
2502 /* Enqueue lookup lock from the parent MDT */
2503 rc = mdt_remote_object_lock(info, msrcdir,
2504 mdt_object_fid(mold),
2505 &lh_oldp->mlh_rreg_lh,
2506 lh_oldp->mlh_rreg_mode,
2507 MDS_INODELOCK_LOOKUP,
2510 GOTO(out_put_old, rc);
2512 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2515 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2518 GOTO(out_unlock_old, rc);
2520 mdt_enoent_version_save(info, 3);
2523 /* step 5: rename it */
2524 mdt_reint_init_ma(info, ma);
2526 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2527 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2530 mutex_lock(&mnew->mot_lov_mutex);
2532 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2533 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2534 mnew != NULL ? mdt_object_child(mnew) : NULL,
2535 &rr->rr_tgt_name, ma);
2538 mutex_unlock(&mnew->mot_lov_mutex);
2540 /* handle last link of tgt object */
2542 mdt_counter_incr(req, LPROC_MDT_RENAME);
2544 mdt_handle_last_unlink(info, mnew, ma);
2545 discard = mdt_dom_check_for_discard(info, mnew);
2548 mdt_rename_counter_tally(info, info->mti_mdt, req,
2554 mdt_object_unlock(info, mnew, lh_newp, rc);
2556 mdt_object_unlock(info, mold, lh_oldp, rc);
2559 mdt_object_put(info->mti_env, mnew);
2561 mdt_object_put(info->mti_env, mold);
2563 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2564 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2566 mdt_object_put(info->mti_env, mtgtdir);
2568 mdt_object_put(info->mti_env, msrcdir);
2570 /* If 'discard' is set then new_fid must exits.
2571 * DOM data discard need neither object nor lock,
2572 * so do this at the end.
2575 mdt_dom_discard_data(info, new_fid);
2580 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2581 struct mdt_lock_handle *lhc, bool rename)
2583 struct mdt_reint_record *rr = &info->mti_rr;
2584 struct ptlrpc_request *req = mdt_info_req(info);
2585 struct lustre_handle rename_lh = { 0 };
2589 if (info->mti_dlm_req)
2590 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2592 if (!fid_is_md_operative(rr->rr_fid1) ||
2593 !fid_is_md_operative(rr->rr_fid2))
2596 /* Note: do not enqueue rename lock for replay request, because
2597 * if other MDT holds rename lock, but being blocked to wait for
2598 * this MDT to finish its recovery, and the failover MDT can not
2599 * get rename lock, which will cause deadlock. */
2600 if (!req_is_replay(req)) {
2601 rc = mdt_rename_lock(info, &rename_lh);
2603 CERROR("%s: can't lock FS for rename: rc = %d\n",
2604 mdt_obd_name(info->mti_mdt), rc);
2610 rc = mdt_reint_rename_internal(info, lhc);
2612 rc = mdt_reint_migrate_internal(info);
2614 if (lustre_handle_is_used(&rename_lh))
2615 mdt_rename_unlock(&rename_lh);
2620 static int mdt_reint_rename(struct mdt_thread_info *info,
2621 struct mdt_lock_handle *lhc)
2623 return mdt_reint_rename_or_migrate(info, lhc, true);
2626 static int mdt_reint_migrate(struct mdt_thread_info *info,
2627 struct mdt_lock_handle *lhc)
2629 return mdt_reint_rename_or_migrate(info, lhc, false);
2632 static int mdt_reint_resync(struct mdt_thread_info *info,
2633 struct mdt_lock_handle *lhc)
2635 struct mdt_reint_record *rr = &info->mti_rr;
2636 struct ptlrpc_request *req = mdt_info_req(info);
2637 struct md_attr *ma = &info->mti_attr;
2638 struct mdt_object *mo;
2639 struct ldlm_lock *lease;
2640 struct mdt_body *repbody;
2641 struct md_layout_change layout = { 0 };
2646 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2648 if (info->mti_dlm_req)
2649 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2651 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2653 GOTO(out, rc = PTR_ERR(mo));
2655 if (!mdt_object_exists(mo))
2656 GOTO(out_obj, rc = -ENOENT);
2658 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2659 GOTO(out_obj, rc = -EINVAL);
2661 if (mdt_object_remote(mo))
2662 GOTO(out_obj, rc = -EREMOTE);
2664 lease = ldlm_handle2lock(rr->rr_handle);
2666 GOTO(out_obj, rc = -ESTALE);
2668 /* It's really necessary to grab open_sem and check if the lease lock
2669 * has been lost. There would exist a concurrent writer coming in and
2670 * generating some dirty data in memory cache, the writeback would fail
2671 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2672 if (!down_write_trylock(&mo->mot_open_sem))
2673 GOTO(out_put_lease, rc = -EBUSY);
2675 lock_res_and_lock(lease);
2676 lease_broken = ldlm_is_cancel(lease);
2677 unlock_res_and_lock(lease);
2679 GOTO(out_unlock, rc = -EBUSY);
2681 /* the file has yet opened by anyone else after we took the lease. */
2682 layout.mlc_opc = MD_LAYOUT_RESYNC;
2683 rc = mdt_layout_change(info, mo, &layout);
2685 GOTO(out_unlock, rc);
2687 ma->ma_need = MA_INODE;
2689 rc = mdt_attr_get_complex(info, mo, ma);
2691 GOTO(out_unlock, rc);
2693 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2694 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2698 up_write(&mo->mot_open_sem);
2700 LDLM_LOCK_PUT(lease);
2702 mdt_object_put(info->mti_env, mo);
2704 mdt_client_compatibility(info);
2705 rc2 = mdt_fix_reply(info);
2711 struct mdt_reinter {
2712 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2713 enum lprocfs_extra_opc mr_extra_opc;
2716 static const struct mdt_reinter mdt_reinters[] = {
2718 .mr_handler = &mdt_reint_setattr,
2719 .mr_extra_opc = MDS_REINT_SETATTR,
2722 .mr_handler = &mdt_reint_create,
2723 .mr_extra_opc = MDS_REINT_CREATE,
2726 .mr_handler = &mdt_reint_link,
2727 .mr_extra_opc = MDS_REINT_LINK,
2730 .mr_handler = &mdt_reint_unlink,
2731 .mr_extra_opc = MDS_REINT_UNLINK,
2734 .mr_handler = &mdt_reint_rename,
2735 .mr_extra_opc = MDS_REINT_RENAME,
2738 .mr_handler = &mdt_reint_open,
2739 .mr_extra_opc = MDS_REINT_OPEN,
2741 [REINT_SETXATTR] = {
2742 .mr_handler = &mdt_reint_setxattr,
2743 .mr_extra_opc = MDS_REINT_SETXATTR,
2746 .mr_handler = &mdt_reint_unlink,
2747 .mr_extra_opc = MDS_REINT_UNLINK,
2750 .mr_handler = &mdt_reint_migrate,
2751 .mr_extra_opc = MDS_REINT_RENAME,
2754 .mr_handler = &mdt_reint_resync,
2755 .mr_extra_opc = MDS_REINT_RESYNC,
2759 int mdt_reint_rec(struct mdt_thread_info *info,
2760 struct mdt_lock_handle *lhc)
2762 const struct mdt_reinter *mr;
2766 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2769 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2770 if (mr->mr_handler == NULL)
2773 rc = (*mr->mr_handler)(info, lhc);
2775 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2776 PTLRPC_LAST_CNTR + mr->mr_extra_opc);