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 int mdt_unlock_slaves(struct mdt_thread_info *mti,
221 struct mdt_object *obj,
222 struct ldlm_enqueue_info *einfo,
225 union ldlm_policy_data *policy = &mti->mti_policy;
226 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
227 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
230 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
231 LASSERT(slave_locks);
233 memset(policy, 0, sizeof(*policy));
234 policy->l_inodebits.bits = einfo->ei_inodebits;
235 mdt_lock_handle_init(lh);
236 mdt_lock_reg_init(lh, einfo->ei_mode);
237 for (i = 0; i < slave_locks->ha_count; i++) {
238 if (test_bit(i, (void *)slave_locks->ha_map))
239 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
241 lh->mlh_reg_lh = slave_locks->ha_handles[i];
242 mdt_object_unlock(mti, NULL, lh, decref);
243 slave_locks->ha_handles[i].cookie = 0ull;
246 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
251 struct mdt_object *obj)
255 if (!S_ISDIR(obj->mot_header.loh_attr))
258 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
261 return rc == -ENODATA ? 0 : rc;
267 * Lock slave stripes if necessary, the lock handles of slave stripes
268 * will be stored in einfo->ei_cbdata.
270 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
271 enum ldlm_mode mode, __u64 ibits,
272 struct ldlm_enqueue_info *einfo)
274 union ldlm_policy_data *policy = &mti->mti_policy;
276 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
278 einfo->ei_type = LDLM_IBITS;
279 einfo->ei_mode = mode;
280 einfo->ei_cb_bl = mdt_remote_blocking_ast;
281 einfo->ei_cb_local_bl = mdt_blocking_ast;
282 einfo->ei_cb_cp = ldlm_completion_ast;
283 einfo->ei_enq_slave = 1;
284 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
285 einfo->ei_inodebits = ibits;
286 memset(policy, 0, sizeof(*policy));
287 policy->l_inodebits.bits = ibits;
289 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
293 int mdt_reint_striped_lock(struct mdt_thread_info *info,
294 struct mdt_object *o,
295 struct mdt_lock_handle *lh,
297 struct ldlm_enqueue_info *einfo,
302 LASSERT(!mdt_object_remote(o));
304 memset(einfo, 0, sizeof(*einfo));
306 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
310 rc = mdt_object_striped(info, o);
313 mdt_object_unlock(info, o, lh, rc);
317 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
319 mdt_object_unlock(info, o, lh, rc);
320 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
327 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
328 struct mdt_object *o,
329 struct mdt_lock_handle *lh,
330 struct ldlm_enqueue_info *einfo, int decref)
332 if (einfo->ei_cbdata)
333 mdt_unlock_slaves(info, o, einfo, decref);
334 mdt_object_unlock(info, o, lh, decref);
338 * VBR: we save three versions in reply:
339 * 0 - parent. Check that parent version is the same during replay.
340 * 1 - name. Version of 'name' if file exists with the same name or
341 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
342 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
345 static int mdt_create(struct mdt_thread_info *info)
347 struct mdt_device *mdt = info->mti_mdt;
348 struct mdt_object *parent;
349 struct mdt_object *child;
350 struct mdt_lock_handle *lh;
351 struct mdt_body *repbody;
352 struct md_attr *ma = &info->mti_attr;
353 struct mdt_reint_record *rr = &info->mti_rr;
354 struct md_op_spec *spec = &info->mti_spec;
358 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
360 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
362 if (!fid_is_md_operative(rr->rr_fid1))
365 if (S_ISDIR(ma->ma_attr.la_mode) &&
366 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
367 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
368 struct lu_ucred *uc = mdt_ucred(info);
369 struct obd_export *exp = mdt_info_req(info)->rq_export;
371 /* Only new clients can create remote dir( >= 2.4) and
372 * striped dir(>= 2.6), old client will return -ENOTSUPP */
373 if (!mdt_is_dne_client(exp))
376 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
377 if (!mdt_is_striped_client(exp))
380 if (!mdt->mdt_enable_striped_dir)
382 } else if (!mdt->mdt_enable_remote_dir) {
386 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
387 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
388 mdt->mdt_enable_remote_dir_gid != -1)
392 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
394 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
396 RETURN(PTR_ERR(parent));
398 if (!mdt_object_exists(parent))
399 GOTO(put_parent, rc = -ENOENT);
401 lh = &info->mti_lh[MDT_LH_PARENT];
402 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
403 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
405 GOTO(put_parent, rc);
407 if (!mdt_object_remote(parent)) {
408 rc = mdt_version_get_check_save(info, parent, 0);
410 GOTO(unlock_parent, rc);
414 * Check child name version during replay.
415 * During create replay a file may exist with same name.
417 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
418 &info->mti_tmp_fid1, 1);
420 GOTO(unlock_parent, rc = -EEXIST);
422 /* -ENOENT is expected here */
424 GOTO(unlock_parent, rc);
426 /* save version of file name for replay, it must be ENOENT here */
427 mdt_enoent_version_save(info, 1);
429 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
430 if (unlikely(IS_ERR(child)))
431 GOTO(unlock_parent, rc = PTR_ERR(child));
433 ma->ma_need = MA_INODE;
436 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
437 OBD_FAIL_MDS_REINT_CREATE_WRITE);
439 /* Version of child will be updated on disk. */
440 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
441 rc = mdt_version_get_check_save(info, child, 2);
445 /* Let lower layer know current lock mode. */
446 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
449 * Do not perform lookup sanity check. We know that name does
452 info->mti_spec.sp_cr_lookup = 0;
453 info->mti_spec.sp_feat = &dt_directory_features;
455 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
456 mdt_object_child(child), &info->mti_spec, ma);
458 rc = mdt_attr_get_complex(info, child, ma);
464 * On DNE, we need to eliminate dependey between 'mkdir a' and
465 * 'mkdir a/b' if b is a striped directory, to achieve this, two
466 * things are done below:
467 * 1. save child and slaves lock.
468 * 2. if the child is a striped directory, relock parent so to
469 * compare against with COS locks to ensure parent was
472 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
473 struct mdt_lock_handle *lhc;
474 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
477 rc = mdt_object_striped(info, child);
483 if (!mdt_object_remote(parent)) {
484 mdt_object_unlock(info, parent, lh, 1);
485 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
486 rc = mdt_reint_object_lock(info, parent, lh,
487 MDS_INODELOCK_UPDATE,
494 lhc = &info->mti_lh[MDT_LH_CHILD];
495 mdt_lock_handle_init(lhc);
496 mdt_lock_reg_init(lhc, LCK_PW);
497 rc = mdt_reint_striped_lock(info, child, lhc,
498 MDS_INODELOCK_UPDATE, einfo,
503 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
506 /* Return fid & attr to client. */
507 if (ma->ma_valid & MA_INODE)
508 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
509 mdt_object_fid(child));
511 mdt_object_put(info->mti_env, child);
513 mdt_object_unlock(info, parent, lh, rc);
515 mdt_object_put(info->mti_env, parent);
519 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
522 struct mdt_lock_handle *lh;
523 int do_vbr = ma->ma_attr.la_valid &
524 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
525 __u64 lockpart = MDS_INODELOCK_UPDATE;
526 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
531 rc = mdt_object_striped(info, mo);
537 lh = &info->mti_lh[MDT_LH_PARENT];
538 mdt_lock_reg_init(lh, LCK_PW);
540 /* Even though the new MDT will grant PERM lock to the old
541 * client, but the old client will almost ignore that during
542 * So it needs to revoke both LOOKUP and PERM lock here, so
543 * both new and old client can cancel the dcache */
544 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
545 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
547 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
552 /* all attrs are packed into mti_attr in unpack_setattr */
553 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
554 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
556 /* VBR: update version if attr changed are important for recovery */
558 /* update on-disk version of changed object */
559 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
560 rc = mdt_version_get_check_save(info, mo, 0);
562 GOTO(out_unlock, rc);
565 /* Ensure constant striping during chown(). See LU-2789. */
566 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
567 mutex_lock(&mo->mot_lov_mutex);
569 /* all attrs are packed into mti_attr in unpack_setattr */
570 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
572 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
573 mutex_unlock(&mo->mot_lov_mutex);
576 GOTO(out_unlock, rc);
577 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
580 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
585 * Check HSM flags and add HS_DIRTY flag if relevant.
587 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
588 * and is not RELEASED.
590 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
593 struct lu_ucred *uc = mdt_ucred(info);
598 /* If the file was modified, add the dirty flag */
599 ma->ma_need = MA_HSM;
600 rc = mdt_attr_get_complex(info, mo, ma);
602 CERROR("file attribute read error for "DFID": %d.\n",
603 PFID(mdt_object_fid(mo)), rc);
607 /* If an up2date copy exists in the backend, add dirty flag */
608 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
609 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
610 ma->ma_hsm.mh_flags |= HS_DIRTY;
612 /* Bump cap so that closes from non-owner writers can
613 * set the HSM state to dirty. */
614 cap_saved = uc->uc_cap;
615 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
616 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
617 uc->uc_cap = cap_saved;
619 CERROR("file attribute change error for "DFID": %d\n",
620 PFID(mdt_object_fid(mo)), rc);
626 static int mdt_reint_setattr(struct mdt_thread_info *info,
627 struct mdt_lock_handle *lhc)
629 struct mdt_device *mdt = info->mti_mdt;
630 struct md_attr *ma = &info->mti_attr;
631 struct mdt_reint_record *rr = &info->mti_rr;
632 struct ptlrpc_request *req = mdt_info_req(info);
633 struct mdt_object *mo;
634 struct mdt_body *repbody;
638 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
639 (unsigned int)ma->ma_attr.la_valid);
641 if (info->mti_dlm_req)
642 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
644 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
645 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
647 GOTO(out, rc = PTR_ERR(mo));
649 if (!mdt_object_exists(mo))
650 GOTO(out_put, rc = -ENOENT);
652 if (mdt_object_remote(mo))
653 GOTO(out_put, rc = -EREMOTE);
655 /* revoke lease lock if size is going to be changed */
656 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
657 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
658 atomic_read(&mo->mot_lease_count) > 0)) {
659 down_read(&mo->mot_open_sem);
661 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
662 lhc = &info->mti_lh[MDT_LH_LOCAL];
663 mdt_lock_reg_init(lhc, LCK_CW);
665 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
667 up_read(&mo->mot_open_sem);
671 /* revoke lease lock */
672 mdt_object_unlock(info, mo, lhc, 1);
674 up_read(&mo->mot_open_sem);
677 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
678 /* Check write access for the O_TRUNC case */
679 if (mdt_write_read(mo) < 0)
680 GOTO(out_put, rc = -ETXTBSY);
682 /* LU-10286: compatibility check for FLR.
683 * Please check the comment in mdt_finish_open() for details */
684 if (!exp_connect_flr(info->mti_exp)) {
685 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
686 if (rc < 0 && rc != -ENODATA)
689 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
690 GOTO(out_put, rc = -EOPNOTSUPP);
693 /* For truncate, the file size sent from client
694 * is believable, but the blocks are incorrect,
695 * which makes the block size in LSOM attribute
696 * inconsisent with the real block size.
698 rc = mdt_lsom_update(info, mo, true);
703 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
704 if (ma->ma_valid & MA_LOV)
705 GOTO(out_put, rc = -EPROTO);
707 rc = mdt_attr_set(info, mo, ma);
710 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
711 (ma->ma_valid & MA_INODE)) {
712 struct lu_buf *buf = &info->mti_buf;
713 struct lu_ucred *uc = mdt_ucred(info);
714 struct mdt_lock_handle *lh;
716 /* reject if either remote or striped dir is disabled */
717 if (ma->ma_valid & MA_LMV) {
718 if (!mdt->mdt_enable_remote_dir ||
719 !mdt->mdt_enable_striped_dir)
720 GOTO(out_put, rc = -EPERM);
722 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
723 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
724 mdt->mdt_enable_remote_dir_gid != -1)
725 GOTO(out_put, rc = -EPERM);
728 if (ma->ma_attr.la_valid != 0)
729 GOTO(out_put, rc = -EPROTO);
731 lh = &info->mti_lh[MDT_LH_PARENT];
732 mdt_lock_reg_init(lh, LCK_PW);
734 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
738 if (ma->ma_valid & MA_LOV) {
739 buf->lb_buf = ma->ma_lmm;
740 buf->lb_len = ma->ma_lmm_size;
742 buf->lb_buf = ma->ma_lmv;
743 buf->lb_len = ma->ma_lmv_size;
745 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
746 (ma->ma_valid & MA_LOV) ?
747 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
750 mdt_object_unlock(info, mo, lh, rc);
754 GOTO(out_put, rc = -EPROTO);
757 /* If file data is modified, add the dirty flag */
758 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
759 rc = mdt_add_dirty_flag(info, mo, ma);
761 ma->ma_need = MA_INODE;
763 rc = mdt_attr_get_complex(info, mo, ma);
767 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
771 mdt_object_put(info->mti_env, mo);
774 mdt_counter_incr(req, LPROC_MDT_SETATTR);
776 mdt_client_compatibility(info);
777 rc2 = mdt_fix_reply(info);
783 static int mdt_reint_create(struct mdt_thread_info *info,
784 struct mdt_lock_handle *lhc)
786 struct ptlrpc_request *req = mdt_info_req(info);
790 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
791 RETURN(err_serious(-ESTALE));
793 if (info->mti_dlm_req)
794 ldlm_request_cancel(mdt_info_req(info),
795 info->mti_dlm_req, 0, LATF_SKIP);
797 if (!lu_name_is_valid(&info->mti_rr.rr_name))
800 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
802 mdt_counter_incr(req, LPROC_MDT_MKDIR);
810 /* Special file should stay on the same node as parent. */
811 mdt_counter_incr(req, LPROC_MDT_MKNOD);
814 CERROR("%s: Unsupported mode %o\n",
815 mdt_obd_name(info->mti_mdt),
816 info->mti_attr.ma_attr.la_mode);
817 RETURN(err_serious(-EOPNOTSUPP));
820 rc = mdt_create(info);
825 * VBR: save parent version in reply and child version getting by its name.
826 * Version of child is getting and checking during its lookup. If
828 static int mdt_reint_unlink(struct mdt_thread_info *info,
829 struct mdt_lock_handle *lhc)
831 struct mdt_reint_record *rr = &info->mti_rr;
832 struct ptlrpc_request *req = mdt_info_req(info);
833 struct md_attr *ma = &info->mti_attr;
834 struct lu_fid *child_fid = &info->mti_tmp_fid1;
835 struct mdt_object *mp;
836 struct mdt_object *mc;
837 struct mdt_lock_handle *parent_lh;
838 struct mdt_lock_handle *child_lh;
839 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
841 bool cos_incompat = false, discard = false;
847 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
848 PNAME(&rr->rr_name));
850 if (info->mti_dlm_req)
851 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
853 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
854 RETURN(err_serious(-ENOENT));
856 if (!fid_is_md_operative(rr->rr_fid1))
859 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
863 if (mdt_object_remote(mp)) {
866 rc = mdt_version_get_check_save(info, mp, 0);
868 GOTO(put_parent, rc);
872 parent_lh = &info->mti_lh[MDT_LH_PARENT];
873 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
874 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
877 GOTO(put_parent, rc);
879 /* lookup child object along with version checking */
881 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
883 /* Name might not be able to find during resend of
884 * remote unlink, considering following case.
885 * dir_A is a remote directory, the name entry of
886 * dir_A is on MDT0, the directory is on MDT1,
888 * 1. client sends unlink req to MDT1.
889 * 2. MDT1 sends name delete update to MDT0.
890 * 3. name entry is being deleted in MDT0 synchronously.
891 * 4. MDT1 is restarted.
892 * 5. client resends unlink req to MDT1. So it can not
893 * find the name entry on MDT0 anymore.
894 * In this case, MDT1 only needs to destory the local
897 if (mdt_object_remote(mp) && rc == -ENOENT &&
898 !fid_is_zero(rr->rr_fid2) &&
899 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
901 *child_fid = *rr->rr_fid2;
903 GOTO(unlock_parent, rc);
907 if (!fid_is_md_operative(child_fid))
908 GOTO(unlock_parent, rc = -EPERM);
910 /* We will lock the child regardless it is local or remote. No harm. */
911 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
913 GOTO(unlock_parent, rc = PTR_ERR(mc));
916 rc = mdt_object_striped(info, mc);
918 GOTO(unlock_parent, rc);
922 mdt_object_put(info->mti_env, mc);
923 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
928 child_lh = &info->mti_lh[MDT_LH_CHILD];
929 mdt_lock_reg_init(child_lh, LCK_EX);
930 if (info->mti_spec.sp_rm_entry) {
931 struct lu_ucred *uc = mdt_ucred(info);
933 if (!mdt_is_dne_client(req->rq_export))
934 /* Return -ENOTSUPP for old client */
935 GOTO(put_child, rc = -ENOTSUPP);
937 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
938 GOTO(put_child, rc = -EPERM);
940 ma->ma_need = MA_INODE;
942 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
943 NULL, &rr->rr_name, ma, no_name);
947 if (mdt_object_remote(mc)) {
948 struct mdt_body *repbody;
950 if (!fid_is_zero(rr->rr_fid2)) {
951 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
952 mdt_obd_name(info->mti_mdt),
953 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
954 GOTO(put_child, rc = -ENOENT);
956 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
957 mdt_obd_name(info->mti_mdt),
958 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
960 if (!mdt_is_dne_client(req->rq_export))
961 /* Return -ENOTSUPP for old client */
962 GOTO(put_child, rc = -ENOTSUPP);
964 /* Revoke the LOOKUP lock of the remote object granted by
965 * this MDT. Since the unlink will happen on another MDT,
966 * it will release the LOOKUP lock right away. Then What
967 * would happen if another client try to grab the LOOKUP
968 * lock at the same time with unlink XXX */
969 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
970 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
971 LASSERT(repbody != NULL);
972 repbody->mbo_fid1 = *mdt_object_fid(mc);
973 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
974 GOTO(unlock_child, rc = -EREMOTE);
976 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
977 * this now because a running HSM restore on the child (unlink
978 * victim) will hold the layout lock. See LU-4002. */
979 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
980 if (mdt_object_remote(mp)) {
981 /* Enqueue lookup lock from parent MDT */
982 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
983 &child_lh->mlh_rreg_lh,
984 child_lh->mlh_rreg_mode,
985 MDS_INODELOCK_LOOKUP, false);
989 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
992 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
998 * Now we can only make sure we need MA_INODE, in mdd layer, will check
999 * whether need MA_LOV and MA_COOKIE.
1001 ma->ma_need = MA_INODE;
1004 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1005 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1006 /* save version when object is locked */
1007 mdt_version_get_save(info, mc, 1);
1009 mutex_lock(&mc->mot_lov_mutex);
1011 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1012 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1014 mutex_unlock(&mc->mot_lov_mutex);
1016 GOTO(unlock_child, rc);
1018 if (!lu_object_is_dying(&mc->mot_header)) {
1019 rc = mdt_attr_get_complex(info, mc, ma);
1023 discard = mdt_dom_check_for_discard(info, mc);
1025 mdt_handle_last_unlink(info, mc, ma);
1028 if (ma->ma_valid & MA_INODE) {
1029 switch (ma->ma_attr.la_mode & S_IFMT) {
1031 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1039 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1042 LASSERTF(0, "bad file type %o unlinking\n",
1043 ma->ma_attr.la_mode);
1050 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1052 mdt_object_put(info->mti_env, mc);
1054 mdt_object_unlock(info, mp, parent_lh, rc);
1056 mdt_object_put(info->mti_env, mp);
1058 /* discard is just a PW DOM lock to drop the data on a client
1059 * no need to keep objects being get and locked, do that after all.
1062 mdt_dom_discard_data(info, child_fid);
1068 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1071 static int mdt_reint_link(struct mdt_thread_info *info,
1072 struct mdt_lock_handle *lhc)
1074 struct mdt_reint_record *rr = &info->mti_rr;
1075 struct ptlrpc_request *req = mdt_info_req(info);
1076 struct md_attr *ma = &info->mti_attr;
1077 struct mdt_object *ms;
1078 struct mdt_object *mp;
1079 struct mdt_lock_handle *lhs;
1080 struct mdt_lock_handle *lhp;
1085 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1086 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1088 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1089 RETURN(err_serious(-ENOENT));
1091 if (info->mti_dlm_req)
1092 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1094 /* Invalid case so return error immediately instead of
1096 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1099 if (!fid_is_md_operative(rr->rr_fid1) ||
1100 !fid_is_md_operative(rr->rr_fid2))
1103 /* step 1: find target parent dir */
1104 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1106 RETURN(PTR_ERR(mp));
1108 rc = mdt_version_get_check_save(info, mp, 0);
1110 GOTO(put_parent, rc);
1112 /* step 2: find source */
1113 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1115 GOTO(put_parent, rc = PTR_ERR(ms));
1117 if (!mdt_object_exists(ms)) {
1118 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1119 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1120 GOTO(put_source, rc = -ENOENT);
1123 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1125 lhp = &info->mti_lh[MDT_LH_PARENT];
1126 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1127 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1130 GOTO(put_source, rc);
1132 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1134 lhs = &info->mti_lh[MDT_LH_CHILD];
1135 mdt_lock_reg_init(lhs, LCK_EX);
1136 rc = mdt_reint_object_lock(info, ms, lhs,
1137 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1140 GOTO(unlock_parent, rc);
1142 /* step 3: link it */
1143 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1144 OBD_FAIL_MDS_REINT_LINK_WRITE);
1146 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1147 rc = mdt_version_get_check_save(info, ms, 1);
1149 GOTO(unlock_source, rc);
1151 /** check target version by name during replay */
1152 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1153 &info->mti_tmp_fid1, 2);
1154 if (rc != 0 && rc != -ENOENT)
1155 GOTO(unlock_source, rc);
1156 /* save version of file name for replay, it must be ENOENT here */
1157 if (!req_is_replay(mdt_info_req(info))) {
1158 if (rc != -ENOENT) {
1159 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1160 PNAME(&rr->rr_name));
1161 GOTO(unlock_source, rc = -EEXIST);
1163 info->mti_ver[2] = ENOENT_VERSION;
1164 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1167 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1168 mdt_object_child(ms), &rr->rr_name, ma);
1171 mdt_counter_incr(req, LPROC_MDT_LINK);
1175 mdt_object_unlock(info, ms, lhs, rc);
1177 mdt_object_unlock(info, mp, lhp, rc);
1179 mdt_object_put(info->mti_env, ms);
1181 mdt_object_put(info->mti_env, mp);
1185 * lock the part of the directory according to the hash of the name
1186 * (lh->mlh_pdo_hash) in parallel directory lock.
1188 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1189 struct mdt_lock_handle *lh,
1190 struct mdt_object *obj, __u64 ibits,
1193 struct ldlm_res_id *res = &info->mti_res_id;
1194 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1195 union ldlm_policy_data *policy = &info->mti_policy;
1196 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1200 * Finish res_id initializing by name hash marking part of
1201 * directory which is taking modification.
1203 LASSERT(lh->mlh_pdo_hash != 0);
1204 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1205 memset(policy, 0, sizeof(*policy));
1206 policy->l_inodebits.bits = ibits;
1208 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1209 dlmflags |= LDLM_FL_COS_INCOMPAT;
1211 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1212 * going to be sent to client. If it is - mdt_intent_policy() path will
1213 * fix it up and turn FL_LOCAL flag off.
1215 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1216 policy, res, dlmflags,
1217 &info->mti_exp->exp_handle.h_cookie);
1222 * Get BFL lock for rename or migrate process.
1224 static int mdt_rename_lock(struct mdt_thread_info *info,
1225 struct lustre_handle *lh)
1230 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1231 struct lu_fid *fid = &info->mti_tmp_fid1;
1232 struct mdt_object *obj;
1234 /* XXX, right now, it has to use object API to
1235 * enqueue lock cross MDT, so it will enqueue
1236 * rename lock(with LUSTRE_BFL_FID) by root object */
1238 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1240 RETURN(PTR_ERR(obj));
1242 rc = mdt_remote_object_lock(info, obj,
1243 &LUSTRE_BFL_FID, lh,
1245 MDS_INODELOCK_UPDATE, false);
1246 mdt_object_put(info->mti_env, obj);
1248 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1249 union ldlm_policy_data *policy = &info->mti_policy;
1250 struct ldlm_res_id *res_id = &info->mti_res_id;
1253 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1254 memset(policy, 0, sizeof *policy);
1255 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1256 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1257 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1258 LDLM_IBITS, policy, LCK_EX, &flags,
1260 ldlm_completion_ast, NULL, NULL, 0,
1262 &info->mti_exp->exp_handle.h_cookie,
1269 static void mdt_rename_unlock(struct lustre_handle *lh)
1272 LASSERT(lustre_handle_is_used(lh));
1273 /* Cancel the single rename lock right away */
1274 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1278 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1279 const struct lu_fid *fid,
1282 struct mdt_object *dir;
1287 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1291 /* check early, the real version will be saved after locking */
1292 rc = mdt_version_get_check(info, dir, idx);
1298 mdt_object_put(info->mti_env, dir);
1303 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1304 * herein we don't really check it, just do revoke.
1306 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1307 struct mdt_object *pobj,
1308 struct mdt_object *obj)
1310 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1313 mdt_lock_handle_init(lh);
1314 mdt_lock_reg_init(lh, LCK_EX);
1316 if (mdt_object_remote(pobj)) {
1317 /* don't bother to check if pobj and obj are on the same MDT. */
1318 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1319 &lh->mlh_rreg_lh, LCK_EX,
1320 MDS_INODELOCK_LOOKUP, false);
1321 } else if (mdt_object_remote(obj)) {
1322 struct ldlm_res_id *res = &info->mti_res_id;
1323 union ldlm_policy_data *policy = &info->mti_policy;
1324 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1325 LDLM_FL_COS_INCOMPAT;
1327 fid_build_reg_res_name(mdt_object_fid(obj), res);
1328 memset(policy, 0, sizeof(*policy));
1329 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1330 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1331 &lh->mlh_reg_lh, LCK_EX, policy, res,
1334 /* do nothing if both are local */
1342 * TODO, currently we don't save this lock because there is no place to
1343 * hold this lock handle, but to avoid race we need to save this lock.
1345 mdt_object_unlock(info, NULL, lh, 1);
1351 * operation may takes locks of linkea, or directory stripes, group them in
1354 struct mdt_sub_lock {
1355 struct mdt_object *msl_obj;
1356 struct mdt_lock_handle msl_lh;
1357 struct list_head msl_linkage;
1360 static void mdt_unlock_list(struct mdt_thread_info *info,
1361 struct list_head *list, int decref)
1363 struct mdt_sub_lock *msl;
1364 struct mdt_sub_lock *tmp;
1366 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1367 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1368 list_del(&msl->msl_linkage);
1374 * lock parents of links, and also check whether total locks don't exceed
1377 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1378 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1379 * \retval -ev negative errno upon error
1381 static int mdt_lock_links(struct mdt_thread_info *info,
1382 struct mdt_object *pobj,
1383 const struct md_attr *ma,
1384 struct mdt_object *obj,
1385 struct list_head *link_locks)
1387 struct mdt_device *mdt = info->mti_mdt;
1388 struct lu_buf *buf = &info->mti_big_buf;
1389 struct lu_name *lname = &info->mti_name;
1390 struct linkea_data ldata = { NULL };
1391 bool blocked = false;
1393 int local_lnkp_cnt = 0;
1398 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1401 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1402 if (buf->lb_buf == NULL)
1406 rc = mdt_links_read(info, obj, &ldata);
1408 if (rc == -ENOENT || rc == -ENODATA)
1414 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1415 linkea_next_entry(&ldata)) {
1416 struct mdt_object *lnkp;
1417 struct mdt_sub_lock *msl;
1421 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1424 /* check if it's also linked to parent */
1425 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1426 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1427 PFID(&fid), PNAME(lname));
1428 /* in case link is remote object, revoke LOOKUP lock */
1429 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1435 /* check if it's linked to a stripe of parent */
1436 if (ma->ma_valid & MA_LMV) {
1437 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1438 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1441 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1442 fid_le_to_cpu(stripe_fid,
1443 &lmv->lmv_stripe_fids[j]);
1444 if (lu_fid_eq(stripe_fid, &fid)) {
1445 CDEBUG(D_INFO, "skip stripe "DFID
1446 ", reovke "DNAME"\n",
1447 PFID(&fid), PNAME(lname));
1448 lnkp = mdt_object_find(info->mti_env,
1451 GOTO(out, rc = PTR_ERR(lnkp));
1457 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1459 mdt_object_put(info->mti_env, lnkp);
1464 /* Check if it's already locked */
1465 list_for_each_entry(msl, link_locks, msl_linkage) {
1466 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1468 DFID" was locked, revoke "DNAME"\n",
1469 PFID(&fid), PNAME(lname));
1470 lnkp = msl->msl_obj;
1476 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1480 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1481 PFID(&fid), PNAME(lname));
1483 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1485 CWARN("%s: cannot find obj "DFID": %ld\n",
1486 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1490 if (!mdt_object_exists(lnkp)) {
1491 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1492 PFID(&fid), PNAME(lname));
1493 mdt_object_put(info->mti_env, lnkp);
1497 if (!mdt_object_remote(lnkp))
1502 GOTO(out, rc = -ENOMEM);
1505 * we can't follow parent-child lock order like other MD
1506 * operations, use lock_try here to avoid deadlock, if the lock
1507 * cannot be taken, drop all locks taken, revoke the blocked
1508 * one, and continue processing the remaining entries, and in
1509 * the end of the loop restart from beginning.
1511 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1513 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1514 MDS_INODELOCK_UPDATE, true);
1515 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1518 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1519 PFID(&fid), PNAME(lname), retries);
1521 mdt_unlock_list(info, link_locks, 1);
1523 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1524 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1525 MDS_INODELOCK_UPDATE);
1527 mdt_object_put(info->mti_env, lnkp);
1532 if (mdt_object_remote(lnkp)) {
1533 struct ldlm_lock *lock;
1536 * for remote object, set lock cb_atomic,
1537 * so lock can be released in blocking_ast()
1538 * immediately, then the next lock_try will
1539 * have better chance of success.
1541 lock = ldlm_handle2lock(
1542 &msl->msl_lh.mlh_rreg_lh);
1543 LASSERT(lock != NULL);
1544 lock_res_and_lock(lock);
1545 ldlm_set_atomic_cb(lock);
1546 unlock_res_and_lock(lock);
1547 LDLM_LOCK_PUT(lock);
1550 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1555 INIT_LIST_HEAD(&msl->msl_linkage);
1556 msl->msl_obj = lnkp;
1557 list_add_tail(&msl->msl_linkage, link_locks);
1559 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1564 if (--retries > 0) {
1565 mdt_unlock_list(info, link_locks, rc);
1575 mdt_unlock_list(info, link_locks, rc);
1576 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1578 * parent may have 3 local objects: master object and 2 stripes
1579 * (if it's being migrated too); source may have 2 local
1580 * objects: master and 1 stripe; target has 1 local object.
1586 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1587 struct mdt_object *obj,
1588 const struct md_attr *ma,
1589 struct list_head *slave_locks)
1591 struct mdt_device *mdt = info->mti_mdt;
1592 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1593 struct lu_fid *fid = &info->mti_tmp_fid1;
1594 struct mdt_object *slave;
1595 struct mdt_sub_lock *msl;
1601 LASSERT(mdt_object_remote(obj));
1602 LASSERT(ma->ma_valid & MA_LMV);
1605 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1608 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1611 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1612 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1614 slave = mdt_object_find(info->mti_env, mdt, fid);
1616 GOTO(out, rc = PTR_ERR(slave));
1620 mdt_object_put(info->mti_env, slave);
1621 GOTO(out, rc = -ENOMEM);
1624 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1625 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1626 MDS_INODELOCK_UPDATE, true);
1629 mdt_object_put(info->mti_env, slave);
1633 INIT_LIST_HEAD(&msl->msl_linkage);
1634 msl->msl_obj = slave;
1635 list_add_tail(&msl->msl_linkage, slave_locks);
1641 mdt_unlock_list(info, slave_locks, rc);
1645 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1646 struct mdt_object *obj,
1647 struct mdt_lock_handle *lh,
1648 struct ldlm_enqueue_info *einfo,
1649 struct list_head *slave_locks,
1652 if (mdt_object_remote(obj)) {
1653 mdt_unlock_list(info, slave_locks, decref);
1654 mdt_object_unlock(info, obj, lh, decref);
1656 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1660 /* lock parent and its stripes */
1661 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1662 struct mdt_object *obj,
1663 const struct md_attr *ma,
1664 struct mdt_lock_handle *lh,
1665 struct ldlm_enqueue_info *einfo,
1666 struct list_head *slave_locks)
1670 if (mdt_object_remote(obj)) {
1671 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1672 &lh->mlh_rreg_lh, LCK_PW,
1673 MDS_INODELOCK_UPDATE, false);
1678 * if obj is remote and striped, lock its stripes explicitly
1679 * because it's not striped in LOD layer on this MDT.
1681 if (ma->ma_valid & MA_LMV) {
1682 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1684 mdt_object_unlock(info, obj, lh, rc);
1687 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1695 * in migration, object may be remote, and we need take full lock of it and its
1696 * stripes if it's directory, besides, object may be a remote object on its
1697 * parent, revoke its LOOKUP lock on where its parent is located.
1699 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1700 struct mdt_object *pobj,
1701 struct mdt_object *obj,
1702 struct mdt_lock_handle *lh,
1703 struct ldlm_enqueue_info *einfo,
1704 struct list_head *slave_locks)
1708 if (mdt_object_remote(obj)) {
1709 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1713 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1714 &lh->mlh_rreg_lh, LCK_EX,
1715 MDS_INODELOCK_FULL, false);
1720 * if obj is remote and striped, lock its stripes explicitly
1721 * because it's not striped in LOD layer on this MDT.
1723 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1724 struct md_attr *ma = &info->mti_attr;
1726 ma->ma_lmv = info->mti_big_lmm;
1727 ma->ma_lmv_size = info->mti_big_lmmsize;
1729 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1731 mdt_object_unlock(info, obj, lh, rc);
1735 if (ma->ma_valid & MA_LMV) {
1736 rc = mdt_lock_remote_slaves(info, obj, ma,
1739 mdt_object_unlock(info, obj, lh, rc);
1743 if (mdt_object_remote(pobj)) {
1744 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1749 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1757 * lookup source by name, if parent is striped directory, we need to find the
1758 * corresponding stripe where source is located, and then lookup there.
1760 * besides, if parent is migrating too, and file is already in target stripe,
1761 * this should be a redo of 'lfs migrate' on client side.
1763 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1764 struct mdt_object *pobj,
1765 const struct md_attr *ma,
1766 const struct lu_name *lname,
1767 struct mdt_object **spobj,
1768 struct mdt_object **sobj)
1770 const struct lu_env *env = info->mti_env;
1771 struct lu_fid *fid = &info->mti_tmp_fid1;
1772 struct mdt_object *stripe;
1775 if (ma->ma_valid & MA_LMV) {
1776 /* if parent is striped, lookup on corresponding stripe */
1777 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1778 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1779 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1780 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1781 LMV_HASH_FLAG_MIGRATION;
1784 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1785 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1788 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1794 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1795 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1797 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1799 stripe = mdt_object_find(env, info->mti_mdt, fid);
1801 return PTR_ERR(stripe);
1804 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1806 if (rc == -ENOENT && is_migrating) {
1808 * if parent is migrating, and lookup child failed on
1809 * source stripe, lookup again on target stripe, if it
1810 * exists, it means previous migration was interrupted,
1811 * and current file was migrated already.
1813 mdt_object_put(env, stripe);
1815 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1816 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1818 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1824 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1826 stripe = mdt_object_find(env, info->mti_mdt, fid);
1828 return PTR_ERR(stripe);
1831 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1832 fid, &info->mti_spec);
1833 mdt_object_put(env, stripe);
1834 return rc ?: -EALREADY;
1836 mdt_object_put(env, stripe);
1841 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1847 mdt_object_get(env, stripe);
1852 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1853 if (IS_ERR(*sobj)) {
1854 mdt_object_put(env, stripe);
1855 rc = PTR_ERR(*sobj);
1863 /* end lease and close file for regular file */
1864 static int mdd_migrate_close(struct mdt_thread_info *info,
1865 struct mdt_object *obj)
1867 struct close_data *data;
1868 struct mdt_body *repbody;
1869 struct ldlm_lock *lease;
1874 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1876 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1880 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1885 lease = ldlm_handle2lock(&data->cd_handle);
1889 /* check if the lease was already canceled */
1890 lock_res_and_lock(lease);
1891 rc = ldlm_is_cancel(lease);
1892 unlock_res_and_lock(lease);
1896 LDLM_DEBUG(lease, DFID" lease broken",
1897 PFID(mdt_object_fid(obj)));
1901 * cancel server side lease, client side counterpart should have been
1902 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1904 ldlm_lock_cancel(lease);
1905 ldlm_reprocess_all(lease->l_resource);
1906 LDLM_LOCK_PUT(lease);
1909 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1910 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1911 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1917 * migrate file in below steps:
1918 * 1. lock parent and its stripes
1919 * 2. lookup source by name
1920 * 3. lock parents of source links if source is not directory
1921 * 4. reject if source is in HSM
1922 * 5. take source open_sem and close file if source is regular file
1923 * 6. lock source and its stripes if it's directory
1924 * 7. lock target so subsequent change to it can trigger COS
1926 * 9. unlock above locks
1927 * 10. sync device if source has links
1929 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1931 const struct lu_env *env = info->mti_env;
1932 struct mdt_device *mdt = info->mti_mdt;
1933 struct mdt_reint_record *rr = &info->mti_rr;
1934 struct lu_ucred *uc = mdt_ucred(info);
1935 struct md_attr *ma = &info->mti_attr;
1936 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1937 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1938 struct mdt_object *pobj;
1939 struct mdt_object *spobj = NULL;
1940 struct mdt_object *sobj = NULL;
1941 struct mdt_object *tobj;
1942 struct mdt_lock_handle *lhp;
1943 struct mdt_lock_handle *lhs;
1944 struct mdt_lock_handle *lht;
1945 LIST_HEAD(parent_slave_locks);
1946 LIST_HEAD(child_slave_locks);
1947 LIST_HEAD(link_locks);
1948 bool open_sem_locked = false;
1949 bool do_sync = false;
1953 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1954 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1956 /* don't allow migrate . or .. */
1957 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1960 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1963 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1964 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1965 mdt->mdt_enable_remote_dir_gid != -1)
1968 /* pobj is master object of parent */
1969 pobj = mdt_object_find_check(info, rr->rr_fid1, 0);
1971 RETURN(PTR_ERR(pobj));
1973 if (unlikely(!info->mti_big_lmm)) {
1974 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
1975 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
1976 if (!info->mti_big_lmm)
1977 GOTO(put_parent, rc = -ENOMEM);
1980 ma->ma_lmv = info->mti_big_lmm;
1981 ma->ma_lmv_size = info->mti_big_lmmsize;
1983 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
1985 GOTO(put_parent, rc);
1987 /* lock parent object */
1988 lhp = &info->mti_lh[MDT_LH_PARENT];
1989 mdt_lock_reg_init(lhp, LCK_PW);
1990 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
1991 &parent_slave_locks);
1993 GOTO(put_parent, rc);
1996 * spobj is the corresponding stripe against name if pobj is striped
1997 * directory, which is the real parent, and no need to lock, because
1998 * we've taken full lock of pobj.
2000 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2002 GOTO(unlock_parent, rc);
2004 /* lock parents of source links, and revoke LOOKUP lock of links */
2005 rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2007 GOTO(put_source, rc);
2010 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2011 * with one request, if total lock count exceeds this limit, we will
2012 * drop all locks after migration, and synchronous device in the end.
2016 /* TODO: DoM migration is not supported yet */
2017 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2018 ma->ma_lmm = info->mti_big_lmm;
2019 ma->ma_lmm_size = info->mti_big_lmmsize;
2021 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2023 GOTO(put_source, rc);
2025 if (ma->ma_valid & MA_LOV &&
2026 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2027 GOTO(put_source, rc = -EOPNOTSUPP);
2030 /* if migration HSM is allowed */
2031 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2032 ma->ma_need = MA_HSM;
2034 rc = mdt_attr_get_complex(info, sobj, ma);
2036 GOTO(unlock_links, rc);
2038 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2039 GOTO(unlock_links, rc = -EOPNOTSUPP);
2042 /* end lease and close file for regular file */
2043 if (info->mti_spec.sp_migrate_close) {
2044 /* try to hold open_sem so that nobody else can open the file */
2045 if (!down_write_trylock(&sobj->mot_open_sem)) {
2047 mdd_migrate_close(info, sobj);
2048 GOTO(unlock_links, rc = -EBUSY);
2050 open_sem_locked = true;
2051 rc = mdd_migrate_close(info, sobj);
2053 GOTO(unlock_open_sem, rc);
2058 lhs = &info->mti_lh[MDT_LH_OLD];
2059 mdt_lock_reg_init(lhs, LCK_EX);
2060 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2061 &child_slave_locks);
2063 GOTO(unlock_open_sem, rc);
2066 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2068 GOTO(unlock_source, rc = PTR_ERR(tobj));
2070 lht = &info->mti_lh[MDT_LH_NEW];
2071 mdt_lock_reg_init(lht, LCK_EX);
2072 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2074 GOTO(put_target, rc);
2076 /* Don't do lookup sanity check. We know name doesn't exist. */
2077 info->mti_spec.sp_cr_lookup = 0;
2078 info->mti_spec.sp_feat = &dt_directory_features;
2080 rc = mdo_migrate(env, mdt_object_child(pobj),
2081 mdt_object_child(sobj), &rr->rr_name,
2082 mdt_object_child(tobj), &info->mti_spec, ma);
2085 mdt_object_unlock(info, tobj, lht, rc);
2087 mdt_object_put(env, tobj);
2089 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2090 &child_slave_locks, rc);
2092 if (open_sem_locked)
2093 up_write(&sobj->mot_open_sem);
2095 mdt_unlock_list(info, &link_locks, rc);
2097 mdt_object_put(env, sobj);
2098 mdt_object_put(env, spobj);
2100 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2101 &parent_slave_locks, rc);
2103 mdt_object_put(env, pobj);
2106 mdt_device_sync(env, mdt);
2111 static int mdt_object_lock_save(struct mdt_thread_info *info,
2112 struct mdt_object *dir,
2113 struct mdt_lock_handle *lh,
2114 int idx, bool cos_incompat)
2118 /* we lock the target dir if it is local */
2119 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2124 /* get and save correct version after locking */
2125 mdt_version_get_save(info, dir, idx);
2130 * determine lock order of sobj and tobj
2132 * there are two situations we need to lock tobj before sobj:
2133 * 1. sobj is child of tobj
2134 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2135 * larger than that of tobj
2137 * \retval 1 lock tobj before sobj
2138 * \retval 0 lock sobj before tobj
2139 * \retval -ev negative errno upon error
2141 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2142 struct mdt_object *sobj,
2143 struct mdt_object *tobj)
2145 struct md_attr *ma = &info->mti_attr;
2146 struct lu_fid *spfid = &info->mti_tmp_fid1;
2147 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2148 struct lmv_mds_md_v1 *lmv;
2153 /* sobj and tobj are the same */
2157 if (fid_is_root(mdt_object_fid(sobj)))
2160 if (fid_is_root(mdt_object_fid(tobj)))
2163 /* check whether sobj is child of tobj */
2164 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2165 mdt_object_fid(tobj));
2172 /* check whether sobj and tobj are children of the same parent */
2173 rc = mdt_attr_get_pfid(info, sobj, spfid);
2177 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2181 if (!lu_fid_eq(spfid, tpfid))
2184 /* check whether sobj and tobj are sibling stripes */
2185 ma->ma_need = MA_LMV;
2187 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2188 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2189 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2193 if (!(ma->ma_valid & MA_LMV))
2196 lmv = &ma->ma_lmv->lmv_md_v1;
2197 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2199 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2202 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2206 if (!(ma->ma_valid & MA_LMV))
2209 lmv = &ma->ma_lmv->lmv_md_v1;
2210 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2212 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2214 /* check stripe index of sobj and tobj */
2215 if (sindex == tindex)
2218 return sindex < tindex ? 0 : 1;
2222 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2223 * 2 - srcdir child; 3 - tgtdir child.
2224 * Update on disk version of srcdir child.
2227 * For DNE phase I, only these renames are allowed
2228 * mv src_p/src_c tgt_p/tgt_c
2229 * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2230 * 2. src_p and tgt_p are same directory, and tgt_c does not
2231 * exists. In this case, all of modification will happen
2232 * in the MDT where ithesource parent is, only one remote
2233 * update is needed, i.e. set c_time/m_time on the child.
2234 * And tgt_c will be still in the same MDT as the original
2237 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2238 struct mdt_lock_handle *lhc)
2240 struct mdt_reint_record *rr = &info->mti_rr;
2241 struct md_attr *ma = &info->mti_attr;
2242 struct ptlrpc_request *req = mdt_info_req(info);
2243 struct mdt_object *msrcdir = NULL;
2244 struct mdt_object *mtgtdir = NULL;
2245 struct mdt_object *mold;
2246 struct mdt_object *mnew = NULL;
2247 struct mdt_lock_handle *lh_srcdirp;
2248 struct mdt_lock_handle *lh_tgtdirp;
2249 struct mdt_lock_handle *lh_oldp = NULL;
2250 struct mdt_lock_handle *lh_newp = NULL;
2251 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2252 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2254 bool reverse = false;
2255 bool cos_incompat, discard = false;
2259 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2260 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2261 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2263 /* find both parents. */
2264 msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
2265 if (IS_ERR(msrcdir))
2266 RETURN(PTR_ERR(msrcdir));
2268 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2270 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2272 mdt_object_get(info->mti_env, mtgtdir);
2274 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
2275 if (IS_ERR(mtgtdir))
2276 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2279 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2281 GOTO(out_put_tgtdir, rc);
2285 /* source needs to be looked up after locking source parent, otherwise
2286 * this rename may race with unlink source, and cause rename hang, see
2287 * sanityn.sh 55b, so check parents first, if later we found source is
2288 * remote, relock parents. */
2289 cos_incompat = (mdt_object_remote(msrcdir) ||
2290 mdt_object_remote(mtgtdir));
2292 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2294 /* lock parents in the proper order. */
2295 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2296 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2299 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2300 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2303 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2306 GOTO(out_put_tgtdir, rc);
2308 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2310 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2313 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2314 GOTO(out_put_tgtdir, rc);
2317 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2320 GOTO(out_put_tgtdir, rc);
2322 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2324 if (mtgtdir != msrcdir) {
2325 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2327 } else if (lh_srcdirp->mlh_pdo_hash !=
2328 lh_tgtdirp->mlh_pdo_hash) {
2329 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2330 MDS_INODELOCK_UPDATE,
2332 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2335 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2336 GOTO(out_put_tgtdir, rc);
2340 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2341 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2343 /* find mold object. */
2345 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2347 GOTO(out_unlock_parents, rc);
2349 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2350 GOTO(out_unlock_parents, rc = -EINVAL);
2352 if (!fid_is_md_operative(old_fid))
2353 GOTO(out_unlock_parents, rc = -EPERM);
2355 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2357 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2359 if (!mdt_object_exists(mold)) {
2360 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2362 "object does not exist");
2363 GOTO(out_put_old, rc = -ENOENT);
2366 /* Check if @mtgtdir is subdir of @mold, before locking child
2367 * to avoid reverse locking. */
2368 if (mtgtdir != msrcdir) {
2369 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2374 GOTO(out_put_old, rc);
2378 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2379 /* save version after locking */
2380 mdt_version_get_save(info, mold, 2);
2382 if (!cos_incompat && mdt_object_remote(mold)) {
2383 cos_incompat = true;
2384 mdt_object_put(info->mti_env, mold);
2385 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2386 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2390 /* find mnew object:
2391 * mnew target object may not exist now
2392 * lookup with version checking */
2394 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2397 /* the new_fid should have been filled at this moment */
2398 if (lu_fid_eq(old_fid, new_fid))
2399 GOTO(out_put_old, rc);
2401 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2402 lu_fid_eq(new_fid, rr->rr_fid2))
2403 GOTO(out_put_old, rc = -EINVAL);
2405 if (!fid_is_md_operative(new_fid))
2406 GOTO(out_put_old, rc = -EPERM);
2408 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2410 GOTO(out_put_old, rc = PTR_ERR(mnew));
2412 if (!mdt_object_exists(mnew)) {
2413 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2415 "object does not exist");
2416 GOTO(out_put_new, rc = -ENOENT);
2419 if (mdt_object_remote(mnew)) {
2420 struct mdt_body *repbody;
2422 /* Always send rename req to the target child MDT */
2423 repbody = req_capsule_server_get(info->mti_pill,
2425 LASSERT(repbody != NULL);
2426 repbody->mbo_fid1 = *new_fid;
2427 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2428 GOTO(out_put_new, rc = -EXDEV);
2430 /* Before locking the target dir, check we do not replace
2431 * a dir with a non-dir, otherwise it may deadlock with
2432 * link op which tries to create a link in this dir
2433 * back to this non-dir. */
2434 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2435 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2436 GOTO(out_put_new, rc = -EISDIR);
2438 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2439 mdt_lock_reg_init(lh_oldp, LCK_EX);
2440 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2441 if (mdt_object_remote(msrcdir)) {
2442 /* Enqueue lookup lock from the parent MDT */
2443 rc = mdt_remote_object_lock(info, msrcdir,
2444 mdt_object_fid(mold),
2445 &lh_oldp->mlh_rreg_lh,
2446 lh_oldp->mlh_rreg_mode,
2447 MDS_INODELOCK_LOOKUP,
2450 GOTO(out_put_new, rc);
2452 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2455 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2458 GOTO(out_unlock_old, rc);
2460 /* Check if @msrcdir is subdir of @mnew, before locking child
2461 * to avoid reverse locking. */
2462 if (mtgtdir != msrcdir) {
2463 rc = mdo_is_subdir(info->mti_env,
2464 mdt_object_child(msrcdir), new_fid);
2468 GOTO(out_unlock_old, rc);
2472 /* We used to acquire MDS_INODELOCK_FULL here but we
2473 * can't do this now because a running HSM restore on
2474 * the rename onto victim will hold the layout
2475 * lock. See LU-4002. */
2477 lh_newp = &info->mti_lh[MDT_LH_NEW];
2478 mdt_lock_reg_init(lh_newp, LCK_EX);
2479 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2480 MDS_INODELOCK_LOOKUP |
2481 MDS_INODELOCK_UPDATE,
2484 GOTO(out_unlock_old, rc);
2486 /* get and save version after locking */
2487 mdt_version_get_save(info, mnew, 3);
2488 } else if (rc != -EREMOTE && rc != -ENOENT) {
2489 GOTO(out_put_old, rc);
2491 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2492 mdt_lock_reg_init(lh_oldp, LCK_EX);
2493 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2494 if (mdt_object_remote(msrcdir)) {
2495 /* Enqueue lookup lock from the parent MDT */
2496 rc = mdt_remote_object_lock(info, msrcdir,
2497 mdt_object_fid(mold),
2498 &lh_oldp->mlh_rreg_lh,
2499 lh_oldp->mlh_rreg_mode,
2500 MDS_INODELOCK_LOOKUP,
2503 GOTO(out_put_old, rc);
2505 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2508 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2511 GOTO(out_unlock_old, rc);
2513 mdt_enoent_version_save(info, 3);
2516 /* step 5: rename it */
2517 mdt_reint_init_ma(info, ma);
2519 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2520 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2523 mutex_lock(&mnew->mot_lov_mutex);
2525 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2526 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2527 mnew != NULL ? mdt_object_child(mnew) : NULL,
2528 &rr->rr_tgt_name, ma);
2531 mutex_unlock(&mnew->mot_lov_mutex);
2533 /* handle last link of tgt object */
2535 mdt_counter_incr(req, LPROC_MDT_RENAME);
2537 mdt_handle_last_unlink(info, mnew, ma);
2538 discard = mdt_dom_check_for_discard(info, mnew);
2541 mdt_rename_counter_tally(info, info->mti_mdt, req,
2547 mdt_object_unlock(info, mnew, lh_newp, rc);
2549 mdt_object_unlock(info, mold, lh_oldp, rc);
2552 mdt_object_put(info->mti_env, mnew);
2554 mdt_object_put(info->mti_env, mold);
2556 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2557 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2559 mdt_object_put(info->mti_env, mtgtdir);
2561 mdt_object_put(info->mti_env, msrcdir);
2563 /* If 'discard' is set then new_fid must exits.
2564 * DOM data discard need neither object nor lock,
2565 * so do this at the end.
2568 mdt_dom_discard_data(info, new_fid);
2573 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2574 struct mdt_lock_handle *lhc, bool rename)
2576 struct mdt_reint_record *rr = &info->mti_rr;
2577 struct ptlrpc_request *req = mdt_info_req(info);
2578 struct lustre_handle rename_lh = { 0 };
2582 if (info->mti_dlm_req)
2583 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2585 if (!fid_is_md_operative(rr->rr_fid1) ||
2586 !fid_is_md_operative(rr->rr_fid2))
2589 /* Note: do not enqueue rename lock for replay request, because
2590 * if other MDT holds rename lock, but being blocked to wait for
2591 * this MDT to finish its recovery, and the failover MDT can not
2592 * get rename lock, which will cause deadlock. */
2593 if (!req_is_replay(req)) {
2594 rc = mdt_rename_lock(info, &rename_lh);
2596 CERROR("%s: can't lock FS for rename: rc = %d\n",
2597 mdt_obd_name(info->mti_mdt), rc);
2603 rc = mdt_reint_rename_internal(info, lhc);
2605 rc = mdt_reint_migrate_internal(info);
2607 if (lustre_handle_is_used(&rename_lh))
2608 mdt_rename_unlock(&rename_lh);
2613 static int mdt_reint_rename(struct mdt_thread_info *info,
2614 struct mdt_lock_handle *lhc)
2616 return mdt_reint_rename_or_migrate(info, lhc, true);
2619 static int mdt_reint_migrate(struct mdt_thread_info *info,
2620 struct mdt_lock_handle *lhc)
2622 return mdt_reint_rename_or_migrate(info, lhc, false);
2625 static int mdt_reint_resync(struct mdt_thread_info *info,
2626 struct mdt_lock_handle *lhc)
2628 struct mdt_reint_record *rr = &info->mti_rr;
2629 struct ptlrpc_request *req = mdt_info_req(info);
2630 struct md_attr *ma = &info->mti_attr;
2631 struct mdt_object *mo;
2632 struct ldlm_lock *lease;
2633 struct mdt_body *repbody;
2634 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2639 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2641 if (info->mti_dlm_req)
2642 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2644 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2646 GOTO(out, rc = PTR_ERR(mo));
2648 if (!mdt_object_exists(mo))
2649 GOTO(out_obj, rc = -ENOENT);
2651 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2652 GOTO(out_obj, rc = -EINVAL);
2654 if (mdt_object_remote(mo))
2655 GOTO(out_obj, rc = -EREMOTE);
2657 lease = ldlm_handle2lock(rr->rr_lease_handle);
2659 GOTO(out_obj, rc = -ESTALE);
2661 /* It's really necessary to grab open_sem and check if the lease lock
2662 * has been lost. There would exist a concurrent writer coming in and
2663 * generating some dirty data in memory cache, the writeback would fail
2664 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2665 if (!down_write_trylock(&mo->mot_open_sem))
2666 GOTO(out_put_lease, rc = -EBUSY);
2668 lock_res_and_lock(lease);
2669 lease_broken = ldlm_is_cancel(lease);
2670 unlock_res_and_lock(lease);
2672 GOTO(out_unlock, rc = -EBUSY);
2674 /* the file has yet opened by anyone else after we took the lease. */
2675 layout.mlc_opc = MD_LAYOUT_RESYNC;
2676 rc = mdt_layout_change(info, mo, &layout);
2678 GOTO(out_unlock, rc);
2680 ma->ma_need = MA_INODE;
2682 rc = mdt_attr_get_complex(info, mo, ma);
2684 GOTO(out_unlock, rc);
2686 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2687 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2691 up_write(&mo->mot_open_sem);
2693 LDLM_LOCK_PUT(lease);
2695 mdt_object_put(info->mti_env, mo);
2697 mdt_client_compatibility(info);
2698 rc2 = mdt_fix_reply(info);
2704 struct mdt_reinter {
2705 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2706 enum lprocfs_extra_opc mr_extra_opc;
2709 static const struct mdt_reinter mdt_reinters[] = {
2711 .mr_handler = &mdt_reint_setattr,
2712 .mr_extra_opc = MDS_REINT_SETATTR,
2715 .mr_handler = &mdt_reint_create,
2716 .mr_extra_opc = MDS_REINT_CREATE,
2719 .mr_handler = &mdt_reint_link,
2720 .mr_extra_opc = MDS_REINT_LINK,
2723 .mr_handler = &mdt_reint_unlink,
2724 .mr_extra_opc = MDS_REINT_UNLINK,
2727 .mr_handler = &mdt_reint_rename,
2728 .mr_extra_opc = MDS_REINT_RENAME,
2731 .mr_handler = &mdt_reint_open,
2732 .mr_extra_opc = MDS_REINT_OPEN,
2734 [REINT_SETXATTR] = {
2735 .mr_handler = &mdt_reint_setxattr,
2736 .mr_extra_opc = MDS_REINT_SETXATTR,
2739 .mr_handler = &mdt_reint_unlink,
2740 .mr_extra_opc = MDS_REINT_UNLINK,
2743 .mr_handler = &mdt_reint_migrate,
2744 .mr_extra_opc = MDS_REINT_RENAME,
2747 .mr_handler = &mdt_reint_resync,
2748 .mr_extra_opc = MDS_REINT_RESYNC,
2752 int mdt_reint_rec(struct mdt_thread_info *info,
2753 struct mdt_lock_handle *lhc)
2755 const struct mdt_reinter *mr;
2759 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2762 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2763 if (mr->mr_handler == NULL)
2766 rc = (*mr->mr_handler)(info, lhc);
2768 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2769 PTLRPC_LAST_CNTR + mr->mr_extra_opc);