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 !exp_connect_overstriping(info->mti_exp)) {
686 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
687 if (rc < 0 && rc != -ENODATA)
690 if (!exp_connect_flr(info->mti_exp)) {
692 mdt_lmm_is_flr(info->mti_big_lmm))
693 GOTO(out_put, rc = -EOPNOTSUPP);
696 if (!exp_connect_overstriping(info->mti_exp)) {
698 mdt_lmm_is_overstriping(info->mti_big_lmm))
699 GOTO(out_put, rc = -EOPNOTSUPP);
703 /* For truncate, the file size sent from client
704 * is believable, but the blocks are incorrect,
705 * which makes the block size in LSOM attribute
706 * inconsisent with the real block size.
708 rc = mdt_lsom_update(info, mo, true);
713 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
714 if (ma->ma_valid & MA_LOV)
715 GOTO(out_put, rc = -EPROTO);
717 /* MDT supports FMD for regular files due to Data-on-MDT */
718 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
719 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
720 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
723 rc = mdt_attr_set(info, mo, ma);
726 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
727 (ma->ma_valid & MA_INODE)) {
728 struct lu_buf *buf = &info->mti_buf;
729 struct lu_ucred *uc = mdt_ucred(info);
730 struct mdt_lock_handle *lh;
732 /* reject if either remote or striped dir is disabled */
733 if (ma->ma_valid & MA_LMV) {
734 if (!mdt->mdt_enable_remote_dir ||
735 !mdt->mdt_enable_striped_dir)
736 GOTO(out_put, rc = -EPERM);
738 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
739 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
740 mdt->mdt_enable_remote_dir_gid != -1)
741 GOTO(out_put, rc = -EPERM);
744 if (ma->ma_attr.la_valid != 0)
745 GOTO(out_put, rc = -EPROTO);
747 lh = &info->mti_lh[MDT_LH_PARENT];
748 mdt_lock_reg_init(lh, LCK_PW);
750 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
754 if (ma->ma_valid & MA_LOV) {
755 buf->lb_buf = ma->ma_lmm;
756 buf->lb_len = ma->ma_lmm_size;
758 buf->lb_buf = ma->ma_lmv;
759 buf->lb_len = ma->ma_lmv_size;
761 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
762 (ma->ma_valid & MA_LOV) ?
763 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
766 mdt_object_unlock(info, mo, lh, rc);
770 GOTO(out_put, rc = -EPROTO);
773 /* If file data is modified, add the dirty flag */
774 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
775 rc = mdt_add_dirty_flag(info, mo, ma);
777 ma->ma_need = MA_INODE;
779 rc = mdt_attr_get_complex(info, mo, ma);
783 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
787 mdt_object_put(info->mti_env, mo);
790 mdt_counter_incr(req, LPROC_MDT_SETATTR);
792 mdt_client_compatibility(info);
793 rc2 = mdt_fix_reply(info);
799 static int mdt_reint_create(struct mdt_thread_info *info,
800 struct mdt_lock_handle *lhc)
802 struct ptlrpc_request *req = mdt_info_req(info);
806 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
807 RETURN(err_serious(-ESTALE));
809 if (info->mti_dlm_req)
810 ldlm_request_cancel(mdt_info_req(info),
811 info->mti_dlm_req, 0, LATF_SKIP);
813 if (!lu_name_is_valid(&info->mti_rr.rr_name))
816 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
818 mdt_counter_incr(req, LPROC_MDT_MKDIR);
826 /* Special file should stay on the same node as parent. */
827 mdt_counter_incr(req, LPROC_MDT_MKNOD);
830 CERROR("%s: Unsupported mode %o\n",
831 mdt_obd_name(info->mti_mdt),
832 info->mti_attr.ma_attr.la_mode);
833 RETURN(err_serious(-EOPNOTSUPP));
836 rc = mdt_create(info);
841 * VBR: save parent version in reply and child version getting by its name.
842 * Version of child is getting and checking during its lookup. If
844 static int mdt_reint_unlink(struct mdt_thread_info *info,
845 struct mdt_lock_handle *lhc)
847 struct mdt_reint_record *rr = &info->mti_rr;
848 struct ptlrpc_request *req = mdt_info_req(info);
849 struct md_attr *ma = &info->mti_attr;
850 struct lu_fid *child_fid = &info->mti_tmp_fid1;
851 struct mdt_object *mp;
852 struct mdt_object *mc;
853 struct mdt_lock_handle *parent_lh;
854 struct mdt_lock_handle *child_lh;
855 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
857 bool cos_incompat = false;
863 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
864 PNAME(&rr->rr_name));
866 if (info->mti_dlm_req)
867 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
869 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
870 RETURN(err_serious(-ENOENT));
872 if (!fid_is_md_operative(rr->rr_fid1))
875 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
879 if (mdt_object_remote(mp)) {
882 rc = mdt_version_get_check_save(info, mp, 0);
884 GOTO(put_parent, rc);
888 parent_lh = &info->mti_lh[MDT_LH_PARENT];
889 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
890 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
893 GOTO(put_parent, rc);
895 /* lookup child object along with version checking */
897 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
899 /* Name might not be able to find during resend of
900 * remote unlink, considering following case.
901 * dir_A is a remote directory, the name entry of
902 * dir_A is on MDT0, the directory is on MDT1,
904 * 1. client sends unlink req to MDT1.
905 * 2. MDT1 sends name delete update to MDT0.
906 * 3. name entry is being deleted in MDT0 synchronously.
907 * 4. MDT1 is restarted.
908 * 5. client resends unlink req to MDT1. So it can not
909 * find the name entry on MDT0 anymore.
910 * In this case, MDT1 only needs to destory the local
913 if (mdt_object_remote(mp) && rc == -ENOENT &&
914 !fid_is_zero(rr->rr_fid2) &&
915 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
917 *child_fid = *rr->rr_fid2;
919 GOTO(unlock_parent, rc);
923 if (!fid_is_md_operative(child_fid))
924 GOTO(unlock_parent, rc = -EPERM);
926 /* We will lock the child regardless it is local or remote. No harm. */
927 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
929 GOTO(unlock_parent, rc = PTR_ERR(mc));
932 rc = mdt_object_striped(info, mc);
934 GOTO(unlock_parent, rc);
938 mdt_object_put(info->mti_env, mc);
939 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
944 child_lh = &info->mti_lh[MDT_LH_CHILD];
945 mdt_lock_reg_init(child_lh, LCK_EX);
946 if (info->mti_spec.sp_rm_entry) {
947 struct lu_ucred *uc = mdt_ucred(info);
949 if (!mdt_is_dne_client(req->rq_export))
950 /* Return -ENOTSUPP for old client */
951 GOTO(put_child, rc = -ENOTSUPP);
953 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
954 GOTO(put_child, rc = -EPERM);
956 ma->ma_need = MA_INODE;
958 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
959 NULL, &rr->rr_name, ma, no_name);
963 if (mdt_object_remote(mc)) {
964 struct mdt_body *repbody;
966 if (!fid_is_zero(rr->rr_fid2)) {
967 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
968 mdt_obd_name(info->mti_mdt),
969 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
970 GOTO(put_child, rc = -ENOENT);
972 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
973 mdt_obd_name(info->mti_mdt),
974 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
976 if (!mdt_is_dne_client(req->rq_export))
977 /* Return -ENOTSUPP for old client */
978 GOTO(put_child, rc = -ENOTSUPP);
980 /* Revoke the LOOKUP lock of the remote object granted by
981 * this MDT. Since the unlink will happen on another MDT,
982 * it will release the LOOKUP lock right away. Then What
983 * would happen if another client try to grab the LOOKUP
984 * lock at the same time with unlink XXX */
985 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
986 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
987 LASSERT(repbody != NULL);
988 repbody->mbo_fid1 = *mdt_object_fid(mc);
989 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
990 GOTO(unlock_child, rc = -EREMOTE);
992 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
993 * this now because a running HSM restore on the child (unlink
994 * victim) will hold the layout lock. See LU-4002. */
995 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
996 if (mdt_object_remote(mp)) {
997 /* Enqueue lookup lock from parent MDT */
998 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
999 &child_lh->mlh_rreg_lh,
1000 child_lh->mlh_rreg_mode,
1001 MDS_INODELOCK_LOOKUP, false);
1003 GOTO(put_child, rc);
1005 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1008 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1011 GOTO(put_child, rc);
1014 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1015 * whether need MA_LOV and MA_COOKIE.
1017 ma->ma_need = MA_INODE;
1020 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1021 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1022 /* save version when object is locked */
1023 mdt_version_get_save(info, mc, 1);
1025 mutex_lock(&mc->mot_lov_mutex);
1027 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1028 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1030 mutex_unlock(&mc->mot_lov_mutex);
1032 GOTO(unlock_child, rc);
1034 if (!lu_object_is_dying(&mc->mot_header)) {
1035 rc = mdt_attr_get_complex(info, mc, ma);
1038 } else if (mdt_dom_check_for_discard(info, mc)) {
1039 mdt_dom_discard_data(info, mc);
1041 mdt_handle_last_unlink(info, mc, ma);
1044 if (ma->ma_valid & MA_INODE) {
1045 switch (ma->ma_attr.la_mode & S_IFMT) {
1047 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1055 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1058 LASSERTF(0, "bad file type %o unlinking\n",
1059 ma->ma_attr.la_mode);
1066 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1068 mdt_object_put(info->mti_env, mc);
1070 mdt_object_unlock(info, mp, parent_lh, rc);
1072 mdt_object_put(info->mti_env, mp);
1077 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1080 static int mdt_reint_link(struct mdt_thread_info *info,
1081 struct mdt_lock_handle *lhc)
1083 struct mdt_reint_record *rr = &info->mti_rr;
1084 struct ptlrpc_request *req = mdt_info_req(info);
1085 struct md_attr *ma = &info->mti_attr;
1086 struct mdt_object *ms;
1087 struct mdt_object *mp;
1088 struct mdt_lock_handle *lhs;
1089 struct mdt_lock_handle *lhp;
1094 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1095 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1097 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1098 RETURN(err_serious(-ENOENT));
1100 if (info->mti_dlm_req)
1101 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1103 /* Invalid case so return error immediately instead of
1105 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1108 if (!fid_is_md_operative(rr->rr_fid1) ||
1109 !fid_is_md_operative(rr->rr_fid2))
1112 /* step 1: find target parent dir */
1113 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1115 RETURN(PTR_ERR(mp));
1117 rc = mdt_version_get_check_save(info, mp, 0);
1119 GOTO(put_parent, rc);
1121 /* step 2: find source */
1122 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1124 GOTO(put_parent, rc = PTR_ERR(ms));
1126 if (!mdt_object_exists(ms)) {
1127 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1128 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1129 GOTO(put_source, rc = -ENOENT);
1132 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1134 lhp = &info->mti_lh[MDT_LH_PARENT];
1135 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1136 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1139 GOTO(put_source, rc);
1141 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1143 lhs = &info->mti_lh[MDT_LH_CHILD];
1144 mdt_lock_reg_init(lhs, LCK_EX);
1145 rc = mdt_reint_object_lock(info, ms, lhs,
1146 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1149 GOTO(unlock_parent, rc);
1151 /* step 3: link it */
1152 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1153 OBD_FAIL_MDS_REINT_LINK_WRITE);
1155 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1156 rc = mdt_version_get_check_save(info, ms, 1);
1158 GOTO(unlock_source, rc);
1160 /** check target version by name during replay */
1161 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1162 &info->mti_tmp_fid1, 2);
1163 if (rc != 0 && rc != -ENOENT)
1164 GOTO(unlock_source, rc);
1165 /* save version of file name for replay, it must be ENOENT here */
1166 if (!req_is_replay(mdt_info_req(info))) {
1167 if (rc != -ENOENT) {
1168 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1169 PNAME(&rr->rr_name));
1170 GOTO(unlock_source, rc = -EEXIST);
1172 info->mti_ver[2] = ENOENT_VERSION;
1173 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1176 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1177 mdt_object_child(ms), &rr->rr_name, ma);
1180 mdt_counter_incr(req, LPROC_MDT_LINK);
1184 mdt_object_unlock(info, ms, lhs, rc);
1186 mdt_object_unlock(info, mp, lhp, rc);
1188 mdt_object_put(info->mti_env, ms);
1190 mdt_object_put(info->mti_env, mp);
1194 * lock the part of the directory according to the hash of the name
1195 * (lh->mlh_pdo_hash) in parallel directory lock.
1197 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1198 struct mdt_lock_handle *lh,
1199 struct mdt_object *obj, __u64 ibits,
1202 struct ldlm_res_id *res = &info->mti_res_id;
1203 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1204 union ldlm_policy_data *policy = &info->mti_policy;
1205 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1209 * Finish res_id initializing by name hash marking part of
1210 * directory which is taking modification.
1212 LASSERT(lh->mlh_pdo_hash != 0);
1213 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1214 memset(policy, 0, sizeof(*policy));
1215 policy->l_inodebits.bits = ibits;
1217 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1218 dlmflags |= LDLM_FL_COS_INCOMPAT;
1220 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1221 * going to be sent to client. If it is - mdt_intent_policy() path will
1222 * fix it up and turn FL_LOCAL flag off.
1224 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1225 policy, res, dlmflags,
1226 &info->mti_exp->exp_handle.h_cookie);
1231 * Get BFL lock for rename or migrate process.
1233 static int mdt_rename_lock(struct mdt_thread_info *info,
1234 struct lustre_handle *lh)
1239 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1240 struct lu_fid *fid = &info->mti_tmp_fid1;
1241 struct mdt_object *obj;
1243 /* XXX, right now, it has to use object API to
1244 * enqueue lock cross MDT, so it will enqueue
1245 * rename lock(with LUSTRE_BFL_FID) by root object */
1247 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1249 RETURN(PTR_ERR(obj));
1251 rc = mdt_remote_object_lock(info, obj,
1252 &LUSTRE_BFL_FID, lh,
1254 MDS_INODELOCK_UPDATE, false);
1255 mdt_object_put(info->mti_env, obj);
1257 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1258 union ldlm_policy_data *policy = &info->mti_policy;
1259 struct ldlm_res_id *res_id = &info->mti_res_id;
1262 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1263 memset(policy, 0, sizeof *policy);
1264 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1265 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1266 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1267 LDLM_IBITS, policy, LCK_EX, &flags,
1269 ldlm_completion_ast, NULL, NULL, 0,
1271 &info->mti_exp->exp_handle.h_cookie,
1278 static void mdt_rename_unlock(struct lustre_handle *lh)
1281 LASSERT(lustre_handle_is_used(lh));
1282 /* Cancel the single rename lock right away */
1283 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1287 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1288 const struct lu_fid *fid,
1291 struct mdt_object *dir;
1296 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1300 /* check early, the real version will be saved after locking */
1301 rc = mdt_version_get_check(info, dir, idx);
1305 if (!mdt_object_exists(dir))
1306 GOTO(out_put, rc = -ENOENT);
1308 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1309 GOTO(out_put, rc = -ENOTDIR);
1313 mdt_object_put(info->mti_env, dir);
1318 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1319 * herein we don't really check it, just do revoke.
1321 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1322 struct mdt_object *pobj,
1323 struct mdt_object *obj)
1325 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1328 mdt_lock_handle_init(lh);
1329 mdt_lock_reg_init(lh, LCK_EX);
1331 if (mdt_object_remote(pobj)) {
1332 /* don't bother to check if pobj and obj are on the same MDT. */
1333 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1334 &lh->mlh_rreg_lh, LCK_EX,
1335 MDS_INODELOCK_LOOKUP, false);
1336 } else if (mdt_object_remote(obj)) {
1337 struct ldlm_res_id *res = &info->mti_res_id;
1338 union ldlm_policy_data *policy = &info->mti_policy;
1339 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1340 LDLM_FL_COS_INCOMPAT;
1342 fid_build_reg_res_name(mdt_object_fid(obj), res);
1343 memset(policy, 0, sizeof(*policy));
1344 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1345 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1346 &lh->mlh_reg_lh, LCK_EX, policy, res,
1349 /* do nothing if both are local */
1357 * TODO, currently we don't save this lock because there is no place to
1358 * hold this lock handle, but to avoid race we need to save this lock.
1360 mdt_object_unlock(info, NULL, lh, 1);
1366 * operation may takes locks of linkea, or directory stripes, group them in
1369 struct mdt_sub_lock {
1370 struct mdt_object *msl_obj;
1371 struct mdt_lock_handle msl_lh;
1372 struct list_head msl_linkage;
1375 static void mdt_unlock_list(struct mdt_thread_info *info,
1376 struct list_head *list, int decref)
1378 struct mdt_sub_lock *msl;
1379 struct mdt_sub_lock *tmp;
1381 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1382 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1383 list_del(&msl->msl_linkage);
1388 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1389 struct mdt_object *obj,
1390 struct mdt_lock_handle *lh,
1391 struct ldlm_enqueue_info *einfo,
1392 struct list_head *slave_locks,
1395 if (mdt_object_remote(obj)) {
1396 mdt_unlock_list(info, slave_locks, decref);
1397 mdt_object_unlock(info, obj, lh, decref);
1399 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1404 * lock parents of links, and also check whether total locks don't exceed
1407 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1408 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1409 * \retval -ev negative errno upon error
1411 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1412 struct mdt_object *pobj,
1413 const struct md_attr *ma,
1414 struct mdt_object *obj,
1415 struct mdt_lock_handle *lhp,
1416 struct ldlm_enqueue_info *peinfo,
1417 struct list_head *parent_slave_locks,
1418 struct list_head *link_locks)
1420 struct mdt_device *mdt = info->mti_mdt;
1421 struct lu_buf *buf = &info->mti_big_buf;
1422 struct lu_name *lname = &info->mti_name;
1423 struct linkea_data ldata = { NULL };
1424 bool blocked = false;
1425 int local_lnkp_cnt = 0;
1430 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1433 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1434 if (buf->lb_buf == NULL)
1438 rc = mdt_links_read(info, obj, &ldata);
1440 if (rc == -ENOENT || rc == -ENODATA)
1445 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1446 linkea_next_entry(&ldata)) {
1447 struct mdt_object *lnkp;
1448 struct mdt_sub_lock *msl;
1452 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1455 /* check if it's also linked to parent */
1456 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1457 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1458 PFID(&fid), PNAME(lname));
1459 /* in case link is remote object, revoke LOOKUP lock */
1460 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1466 /* check if it's linked to a stripe of parent */
1467 if (ma->ma_valid & MA_LMV) {
1468 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1469 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1472 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1473 fid_le_to_cpu(stripe_fid,
1474 &lmv->lmv_stripe_fids[j]);
1475 if (lu_fid_eq(stripe_fid, &fid)) {
1476 CDEBUG(D_INFO, "skip stripe "DFID
1477 ", reovke "DNAME"\n",
1478 PFID(&fid), PNAME(lname));
1479 lnkp = mdt_object_find(info->mti_env,
1482 GOTO(out, rc = PTR_ERR(lnkp));
1488 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1490 mdt_object_put(info->mti_env, lnkp);
1495 /* Check if it's already locked */
1496 list_for_each_entry(msl, link_locks, msl_linkage) {
1497 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1499 DFID" was locked, revoke "DNAME"\n",
1500 PFID(&fid), PNAME(lname));
1501 lnkp = msl->msl_obj;
1507 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1511 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1512 PFID(&fid), PNAME(lname));
1514 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1516 CWARN("%s: cannot find obj "DFID": %ld\n",
1517 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1521 if (!mdt_object_exists(lnkp)) {
1522 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1523 PFID(&fid), PNAME(lname));
1524 mdt_object_put(info->mti_env, lnkp);
1528 if (!mdt_object_remote(lnkp))
1533 GOTO(out, rc = -ENOMEM);
1536 * we can't follow parent-child lock order like other MD
1537 * operations, use lock_try here to avoid deadlock, if the lock
1538 * cannot be taken, drop all locks taken, revoke the blocked
1539 * one, and continue processing the remaining entries, and in
1540 * the end of the loop restart from beginning.
1542 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1544 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1545 MDS_INODELOCK_UPDATE, true);
1546 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1548 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1549 PFID(&fid), PNAME(lname));
1551 mdt_unlock_list(info, link_locks, 1);
1552 /* also unlock parent locks to avoid deadlock */
1554 mdt_migrate_object_unlock(info, pobj, lhp,
1561 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1562 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1563 MDS_INODELOCK_UPDATE);
1565 mdt_object_put(info->mti_env, lnkp);
1570 if (mdt_object_remote(lnkp)) {
1571 struct ldlm_lock *lock;
1574 * for remote object, set lock cb_atomic,
1575 * so lock can be released in blocking_ast()
1576 * immediately, then the next lock_try will
1577 * have better chance of success.
1579 lock = ldlm_handle2lock(
1580 &msl->msl_lh.mlh_rreg_lh);
1581 LASSERT(lock != NULL);
1582 lock_res_and_lock(lock);
1583 ldlm_set_atomic_cb(lock);
1584 unlock_res_and_lock(lock);
1585 LDLM_LOCK_PUT(lock);
1588 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1593 INIT_LIST_HEAD(&msl->msl_linkage);
1594 msl->msl_obj = lnkp;
1595 list_add_tail(&msl->msl_linkage, link_locks);
1597 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1601 GOTO(out, rc = -EBUSY);
1606 mdt_unlock_list(info, link_locks, rc);
1607 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1609 * parent may have 3 local objects: master object and 2 stripes
1610 * (if it's being migrated too); source may have 2 local
1611 * objects: master and 1 stripe; target has 1 local object.
1617 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1618 struct mdt_object *obj,
1619 const struct md_attr *ma,
1620 struct list_head *slave_locks)
1622 struct mdt_device *mdt = info->mti_mdt;
1623 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1624 struct lu_fid *fid = &info->mti_tmp_fid1;
1625 struct mdt_object *slave;
1626 struct mdt_sub_lock *msl;
1632 LASSERT(mdt_object_remote(obj));
1633 LASSERT(ma->ma_valid & MA_LMV);
1636 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1639 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1642 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1643 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1645 if (!fid_is_sane(fid))
1648 slave = mdt_object_find(info->mti_env, mdt, fid);
1650 GOTO(out, rc = PTR_ERR(slave));
1654 mdt_object_put(info->mti_env, slave);
1655 GOTO(out, rc = -ENOMEM);
1658 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1659 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1660 MDS_INODELOCK_UPDATE, true);
1663 mdt_object_put(info->mti_env, slave);
1667 INIT_LIST_HEAD(&msl->msl_linkage);
1668 msl->msl_obj = slave;
1669 list_add_tail(&msl->msl_linkage, slave_locks);
1675 mdt_unlock_list(info, slave_locks, rc);
1679 /* lock parent and its stripes */
1680 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1681 struct mdt_object *obj,
1682 const struct md_attr *ma,
1683 struct mdt_lock_handle *lh,
1684 struct ldlm_enqueue_info *einfo,
1685 struct list_head *slave_locks)
1689 if (mdt_object_remote(obj)) {
1690 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1691 &lh->mlh_rreg_lh, LCK_PW,
1692 MDS_INODELOCK_UPDATE, false);
1697 * if obj is remote and striped, lock its stripes explicitly
1698 * because it's not striped in LOD layer on this MDT.
1700 if (ma->ma_valid & MA_LMV) {
1701 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1703 mdt_object_unlock(info, obj, lh, rc);
1706 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1714 * in migration, object may be remote, and we need take full lock of it and its
1715 * stripes if it's directory, besides, object may be a remote object on its
1716 * parent, revoke its LOOKUP lock on where its parent is located.
1718 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1719 struct mdt_object *pobj,
1720 struct mdt_object *obj,
1721 struct mdt_lock_handle *lh,
1722 struct ldlm_enqueue_info *einfo,
1723 struct list_head *slave_locks)
1727 if (mdt_object_remote(obj)) {
1728 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1732 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1733 &lh->mlh_rreg_lh, LCK_EX,
1734 MDS_INODELOCK_FULL, false);
1739 * if obj is remote and striped, lock its stripes explicitly
1740 * because it's not striped in LOD layer on this MDT.
1742 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1743 struct md_attr *ma = &info->mti_attr;
1745 ma->ma_lmv = info->mti_big_lmm;
1746 ma->ma_lmv_size = info->mti_big_lmmsize;
1748 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1750 mdt_object_unlock(info, obj, lh, rc);
1754 if (ma->ma_valid & MA_LMV) {
1755 rc = mdt_lock_remote_slaves(info, obj, ma,
1758 mdt_object_unlock(info, obj, lh, rc);
1762 if (mdt_object_remote(pobj)) {
1763 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1768 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1776 * lookup source by name, if parent is striped directory, we need to find the
1777 * corresponding stripe where source is located, and then lookup there.
1779 * besides, if parent is migrating too, and file is already in target stripe,
1780 * this should be a redo of 'lfs migrate' on client side.
1782 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1783 struct mdt_object *pobj,
1784 const struct md_attr *ma,
1785 const struct lu_name *lname,
1786 struct mdt_object **spobj,
1787 struct mdt_object **sobj)
1789 const struct lu_env *env = info->mti_env;
1790 struct lu_fid *fid = &info->mti_tmp_fid1;
1791 struct mdt_object *stripe;
1794 if (ma->ma_valid & MA_LMV) {
1795 /* if parent is striped, lookup on corresponding stripe */
1796 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1797 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1798 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1799 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1800 LMV_HASH_FLAG_MIGRATION;
1803 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1804 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1807 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1813 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1814 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1816 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1818 stripe = mdt_object_find(env, info->mti_mdt, fid);
1820 return PTR_ERR(stripe);
1823 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1825 if (rc == -ENOENT && is_migrating) {
1827 * if parent is migrating, and lookup child failed on
1828 * source stripe, lookup again on target stripe, if it
1829 * exists, it means previous migration was interrupted,
1830 * and current file was migrated already.
1832 mdt_object_put(env, stripe);
1834 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1835 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1837 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1843 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1845 stripe = mdt_object_find(env, info->mti_mdt, fid);
1847 return PTR_ERR(stripe);
1850 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1851 fid, &info->mti_spec);
1852 mdt_object_put(env, stripe);
1853 return rc ?: -EALREADY;
1855 mdt_object_put(env, stripe);
1860 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1866 mdt_object_get(env, stripe);
1871 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1872 if (IS_ERR(*sobj)) {
1873 mdt_object_put(env, stripe);
1874 rc = PTR_ERR(*sobj);
1882 /* end lease and close file for regular file */
1883 static int mdd_migrate_close(struct mdt_thread_info *info,
1884 struct mdt_object *obj)
1886 struct close_data *data;
1887 struct mdt_body *repbody;
1888 struct ldlm_lock *lease;
1893 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1895 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1899 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1904 lease = ldlm_handle2lock(&data->cd_handle);
1908 /* check if the lease was already canceled */
1909 lock_res_and_lock(lease);
1910 rc = ldlm_is_cancel(lease);
1911 unlock_res_and_lock(lease);
1915 LDLM_DEBUG(lease, DFID" lease broken",
1916 PFID(mdt_object_fid(obj)));
1920 * cancel server side lease, client side counterpart should have been
1921 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1923 ldlm_lock_cancel(lease);
1924 ldlm_reprocess_all(lease->l_resource);
1925 LDLM_LOCK_PUT(lease);
1928 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1929 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1930 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1936 * migrate file in below steps:
1937 * 1. lock parent and its stripes
1938 * 2. lookup source by name
1939 * 3. lock parents of source links if source is not directory
1940 * 4. reject if source is in HSM
1941 * 5. take source open_sem and close file if source is regular file
1942 * 6. lock source and its stripes if it's directory
1943 * 7. lock target so subsequent change to it can trigger COS
1945 * 9. unlock above locks
1946 * 10. sync device if source has links
1948 static int mdt_reint_migrate(struct mdt_thread_info *info,
1949 struct mdt_lock_handle *unused)
1951 const struct lu_env *env = info->mti_env;
1952 struct mdt_device *mdt = info->mti_mdt;
1953 struct ptlrpc_request *req = mdt_info_req(info);
1954 struct mdt_reint_record *rr = &info->mti_rr;
1955 struct lu_ucred *uc = mdt_ucred(info);
1956 struct md_attr *ma = &info->mti_attr;
1957 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1958 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1959 struct mdt_object *pobj;
1960 struct mdt_object *spobj = NULL;
1961 struct mdt_object *sobj = NULL;
1962 struct mdt_object *tobj;
1963 struct lustre_handle rename_lh = { 0 };
1964 struct mdt_lock_handle *lhp;
1965 struct mdt_lock_handle *lhs;
1966 struct mdt_lock_handle *lht;
1967 LIST_HEAD(parent_slave_locks);
1968 LIST_HEAD(child_slave_locks);
1969 LIST_HEAD(link_locks);
1970 int lock_retries = 5;
1971 bool open_sem_locked = false;
1972 bool do_sync = false;
1976 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1977 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1979 if (info->mti_dlm_req)
1980 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1982 if (!fid_is_md_operative(rr->rr_fid1) ||
1983 !fid_is_md_operative(rr->rr_fid2))
1986 /* don't allow migrate . or .. */
1987 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1990 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1993 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1994 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1995 mdt->mdt_enable_remote_dir_gid != -1)
1999 * Note: do not enqueue rename lock for replay request, because
2000 * if other MDT holds rename lock, but being blocked to wait for
2001 * this MDT to finish its recovery, and the failover MDT can not
2002 * get rename lock, which will cause deadlock.
2004 if (!req_is_replay(req)) {
2005 rc = mdt_rename_lock(info, &rename_lh);
2007 CERROR("%s: can't lock FS for rename: rc = %d\n",
2008 mdt_obd_name(info->mti_mdt), rc);
2013 /* pobj is master object of parent */
2014 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2016 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2018 if (unlikely(!info->mti_big_lmm)) {
2019 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2020 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2021 if (!info->mti_big_lmm)
2022 GOTO(put_parent, rc = -ENOMEM);
2025 ma->ma_lmv = info->mti_big_lmm;
2026 ma->ma_lmv_size = info->mti_big_lmmsize;
2028 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2030 GOTO(put_parent, rc);
2033 /* lock parent object */
2034 lhp = &info->mti_lh[MDT_LH_PARENT];
2035 mdt_lock_reg_init(lhp, LCK_PW);
2036 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2037 &parent_slave_locks);
2039 GOTO(put_parent, rc);
2042 * spobj is the corresponding stripe against name if pobj is striped
2043 * directory, which is the real parent, and no need to lock, because
2044 * we've taken full lock of pobj.
2046 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2048 GOTO(unlock_parent, rc);
2050 /* lock parents of source links, and revoke LOOKUP lock of links */
2051 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2052 &parent_slave_locks, &link_locks);
2053 if (rc == -EBUSY && lock_retries-- > 0) {
2054 mdt_object_put(env, sobj);
2055 mdt_object_put(env, spobj);
2060 GOTO(put_source, rc);
2063 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2064 * with one request, if total lock count exceeds this limit, we will
2065 * drop all locks after migration, and synchronous device in the end.
2069 /* TODO: DoM migration is not supported yet */
2070 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2071 ma->ma_lmm = info->mti_big_lmm;
2072 ma->ma_lmm_size = info->mti_big_lmmsize;
2074 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2076 GOTO(put_source, rc);
2078 if (ma->ma_valid & MA_LOV &&
2079 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2080 GOTO(put_source, rc = -EOPNOTSUPP);
2083 /* if migration HSM is allowed */
2084 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2085 ma->ma_need = MA_HSM;
2087 rc = mdt_attr_get_complex(info, sobj, ma);
2089 GOTO(unlock_links, rc);
2091 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2092 GOTO(unlock_links, rc = -EOPNOTSUPP);
2095 /* end lease and close file for regular file */
2096 if (info->mti_spec.sp_migrate_close) {
2097 /* try to hold open_sem so that nobody else can open the file */
2098 if (!down_write_trylock(&sobj->mot_open_sem)) {
2100 mdd_migrate_close(info, sobj);
2101 GOTO(unlock_links, rc = -EBUSY);
2103 open_sem_locked = true;
2104 rc = mdd_migrate_close(info, sobj);
2106 GOTO(unlock_open_sem, rc);
2111 lhs = &info->mti_lh[MDT_LH_OLD];
2112 mdt_lock_reg_init(lhs, LCK_EX);
2113 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2114 &child_slave_locks);
2116 GOTO(unlock_open_sem, rc);
2119 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2121 GOTO(unlock_source, rc = PTR_ERR(tobj));
2123 lht = &info->mti_lh[MDT_LH_NEW];
2124 mdt_lock_reg_init(lht, LCK_EX);
2125 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2127 GOTO(put_target, rc);
2129 /* Don't do lookup sanity check. We know name doesn't exist. */
2130 info->mti_spec.sp_cr_lookup = 0;
2131 info->mti_spec.sp_feat = &dt_directory_features;
2133 rc = mdo_migrate(env, mdt_object_child(pobj),
2134 mdt_object_child(sobj), &rr->rr_name,
2135 mdt_object_child(tobj), &info->mti_spec, ma);
2138 mdt_object_unlock(info, tobj, lht, rc);
2140 mdt_object_put(env, tobj);
2142 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2143 &child_slave_locks, rc);
2145 if (open_sem_locked)
2146 up_write(&sobj->mot_open_sem);
2148 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2150 mdt_object_put(env, sobj);
2151 mdt_object_put(env, spobj);
2153 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2154 &parent_slave_locks, rc);
2156 mdt_object_put(env, pobj);
2158 if (lustre_handle_is_used(&rename_lh))
2159 mdt_rename_unlock(&rename_lh);
2162 mdt_device_sync(env, mdt);
2167 static int mdt_object_lock_save(struct mdt_thread_info *info,
2168 struct mdt_object *dir,
2169 struct mdt_lock_handle *lh,
2170 int idx, bool cos_incompat)
2174 /* we lock the target dir if it is local */
2175 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2180 /* get and save correct version after locking */
2181 mdt_version_get_save(info, dir, idx);
2186 * determine lock order of sobj and tobj
2188 * there are two situations we need to lock tobj before sobj:
2189 * 1. sobj is child of tobj
2190 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2191 * larger than that of tobj
2193 * \retval 1 lock tobj before sobj
2194 * \retval 0 lock sobj before tobj
2195 * \retval -ev negative errno upon error
2197 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2198 struct mdt_object *sobj,
2199 struct mdt_object *tobj)
2201 struct md_attr *ma = &info->mti_attr;
2202 struct lu_fid *spfid = &info->mti_tmp_fid1;
2203 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2204 struct lmv_mds_md_v1 *lmv;
2209 /* sobj and tobj are the same */
2213 if (fid_is_root(mdt_object_fid(sobj)))
2216 if (fid_is_root(mdt_object_fid(tobj)))
2219 /* check whether sobj is child of tobj */
2220 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2221 mdt_object_fid(tobj));
2228 /* check whether sobj and tobj are children of the same parent */
2229 rc = mdt_attr_get_pfid(info, sobj, spfid);
2233 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2237 if (!lu_fid_eq(spfid, tpfid))
2240 /* check whether sobj and tobj are sibling stripes */
2241 ma->ma_need = MA_LMV;
2243 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2244 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2245 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2249 if (!(ma->ma_valid & MA_LMV))
2252 lmv = &ma->ma_lmv->lmv_md_v1;
2253 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2255 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2258 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2262 if (!(ma->ma_valid & MA_LMV))
2265 lmv = &ma->ma_lmv->lmv_md_v1;
2266 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2268 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2270 /* check stripe index of sobj and tobj */
2271 if (sindex == tindex)
2274 return sindex < tindex ? 0 : 1;
2278 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2279 * 2 - srcdir child; 3 - tgtdir child.
2280 * Update on disk version of srcdir child.
2282 static int mdt_reint_rename(struct mdt_thread_info *info,
2283 struct mdt_lock_handle *unused)
2285 struct mdt_device *mdt = info->mti_mdt;
2286 struct mdt_reint_record *rr = &info->mti_rr;
2287 struct md_attr *ma = &info->mti_attr;
2288 struct ptlrpc_request *req = mdt_info_req(info);
2289 struct mdt_object *msrcdir = NULL;
2290 struct mdt_object *mtgtdir = NULL;
2291 struct mdt_object *mold;
2292 struct mdt_object *mnew = NULL;
2293 struct lustre_handle rename_lh = { 0 };
2294 struct mdt_lock_handle *lh_srcdirp;
2295 struct mdt_lock_handle *lh_tgtdirp;
2296 struct mdt_lock_handle *lh_oldp = NULL;
2297 struct mdt_lock_handle *lh_newp = NULL;
2298 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2299 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2301 bool reverse = false, discard = false;
2306 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2307 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2308 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2310 if (info->mti_dlm_req)
2311 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2313 if (!fid_is_md_operative(rr->rr_fid1) ||
2314 !fid_is_md_operative(rr->rr_fid2))
2317 /* find both parents. */
2318 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2319 if (IS_ERR(msrcdir))
2320 RETURN(PTR_ERR(msrcdir));
2322 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2324 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2326 mdt_object_get(info->mti_env, mtgtdir);
2328 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2329 if (IS_ERR(mtgtdir))
2330 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2334 * Note: do not enqueue rename lock for replay request, because
2335 * if other MDT holds rename lock, but being blocked to wait for
2336 * this MDT to finish its recovery, and the failover MDT can not
2337 * get rename lock, which will cause deadlock.
2339 if (!req_is_replay(req)) {
2341 * Normally rename RPC is handled on the MDT with the target
2342 * directory (if target exists, it's on the MDT with the
2343 * target), if the source directory is remote, it's a hint that
2344 * source is remote too (this may not be true, but it won't
2345 * cause any issue), return -EXDEV early to avoid taking
2348 if (!mdt->mdt_enable_remote_rename &&
2349 mdt_object_remote(msrcdir))
2350 GOTO(out_put_tgtdir, rc = -EXDEV);
2352 rc = mdt_rename_lock(info, &rename_lh);
2354 CERROR("%s: can't lock FS for rename: rc = %d\n",
2355 mdt_obd_name(mdt), rc);
2356 GOTO(out_put_tgtdir, rc);
2360 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2362 GOTO(out_unlock_rename, rc);
2366 /* source needs to be looked up after locking source parent, otherwise
2367 * this rename may race with unlink source, and cause rename hang, see
2368 * sanityn.sh 55b, so check parents first, if later we found source is
2369 * remote, relock parents. */
2370 cos_incompat = (mdt_object_remote(msrcdir) ||
2371 mdt_object_remote(mtgtdir));
2373 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2375 /* lock parents in the proper order. */
2376 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2377 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2380 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2381 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2384 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2387 GOTO(out_unlock_rename, rc);
2389 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2391 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2394 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2395 GOTO(out_unlock_rename, rc);
2398 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2401 GOTO(out_unlock_rename, rc);
2403 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2405 if (mtgtdir != msrcdir) {
2406 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2408 } else if (!mdt_object_remote(mtgtdir) &&
2409 lh_srcdirp->mlh_pdo_hash !=
2410 lh_tgtdirp->mlh_pdo_hash) {
2411 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2412 MDS_INODELOCK_UPDATE,
2414 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2417 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2418 GOTO(out_unlock_rename, rc);
2422 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2423 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2425 /* find mold object. */
2427 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2429 GOTO(out_unlock_parents, rc);
2431 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2432 GOTO(out_unlock_parents, rc = -EINVAL);
2434 if (!fid_is_md_operative(old_fid))
2435 GOTO(out_unlock_parents, rc = -EPERM);
2437 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2439 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2441 if (!mdt_object_exists(mold)) {
2442 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2444 "object does not exist");
2445 GOTO(out_put_old, rc = -ENOENT);
2448 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2449 GOTO(out_put_old, rc = -EXDEV);
2451 /* Check if @mtgtdir is subdir of @mold, before locking child
2452 * to avoid reverse locking. */
2453 if (mtgtdir != msrcdir) {
2454 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2459 GOTO(out_put_old, rc);
2463 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2464 /* save version after locking */
2465 mdt_version_get_save(info, mold, 2);
2467 if (!cos_incompat && mdt_object_remote(mold)) {
2468 cos_incompat = true;
2469 mdt_object_put(info->mti_env, mold);
2470 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2471 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2475 /* find mnew object:
2476 * mnew target object may not exist now
2477 * lookup with version checking */
2479 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2482 /* the new_fid should have been filled at this moment */
2483 if (lu_fid_eq(old_fid, new_fid))
2484 GOTO(out_put_old, rc);
2486 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2487 lu_fid_eq(new_fid, rr->rr_fid2))
2488 GOTO(out_put_old, rc = -EINVAL);
2490 if (!fid_is_md_operative(new_fid))
2491 GOTO(out_put_old, rc = -EPERM);
2493 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2495 GOTO(out_put_old, rc = PTR_ERR(mnew));
2497 if (!mdt_object_exists(mnew)) {
2498 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2500 "object does not exist");
2501 GOTO(out_put_new, rc = -ENOENT);
2504 if (mdt_object_remote(mnew)) {
2505 struct mdt_body *repbody;
2507 /* Always send rename req to the target child MDT */
2508 repbody = req_capsule_server_get(info->mti_pill,
2510 LASSERT(repbody != NULL);
2511 repbody->mbo_fid1 = *new_fid;
2512 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2513 GOTO(out_put_new, rc = -EXDEV);
2515 /* Before locking the target dir, check we do not replace
2516 * a dir with a non-dir, otherwise it may deadlock with
2517 * link op which tries to create a link in this dir
2518 * back to this non-dir. */
2519 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2520 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2521 GOTO(out_put_new, rc = -EISDIR);
2523 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2524 mdt_lock_reg_init(lh_oldp, LCK_EX);
2525 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2526 if (mdt_object_remote(msrcdir)) {
2527 /* Enqueue lookup lock from the parent MDT */
2528 rc = mdt_remote_object_lock(info, msrcdir,
2529 mdt_object_fid(mold),
2530 &lh_oldp->mlh_rreg_lh,
2531 lh_oldp->mlh_rreg_mode,
2532 MDS_INODELOCK_LOOKUP,
2535 GOTO(out_put_new, rc);
2537 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2540 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2543 GOTO(out_unlock_old, rc);
2545 /* Check if @msrcdir is subdir of @mnew, before locking child
2546 * to avoid reverse locking. */
2547 if (mtgtdir != msrcdir) {
2548 rc = mdo_is_subdir(info->mti_env,
2549 mdt_object_child(msrcdir), new_fid);
2553 GOTO(out_unlock_old, rc);
2557 /* We used to acquire MDS_INODELOCK_FULL here but we
2558 * can't do this now because a running HSM restore on
2559 * the rename onto victim will hold the layout
2560 * lock. See LU-4002. */
2562 lh_newp = &info->mti_lh[MDT_LH_NEW];
2563 mdt_lock_reg_init(lh_newp, LCK_EX);
2564 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2565 MDS_INODELOCK_LOOKUP |
2566 MDS_INODELOCK_UPDATE,
2569 GOTO(out_unlock_old, rc);
2571 /* get and save version after locking */
2572 mdt_version_get_save(info, mnew, 3);
2573 } else if (rc != -EREMOTE && rc != -ENOENT) {
2574 GOTO(out_put_old, rc);
2576 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2577 mdt_lock_reg_init(lh_oldp, LCK_EX);
2578 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2579 if (mdt_object_remote(msrcdir)) {
2580 /* Enqueue lookup lock from the parent MDT */
2581 rc = mdt_remote_object_lock(info, msrcdir,
2582 mdt_object_fid(mold),
2583 &lh_oldp->mlh_rreg_lh,
2584 lh_oldp->mlh_rreg_mode,
2585 MDS_INODELOCK_LOOKUP,
2588 GOTO(out_put_old, rc);
2590 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2593 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2596 GOTO(out_unlock_old, rc);
2598 mdt_enoent_version_save(info, 3);
2601 /* step 5: rename it */
2602 mdt_reint_init_ma(info, ma);
2604 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2605 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2608 mutex_lock(&mnew->mot_lov_mutex);
2610 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2611 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2612 mnew != NULL ? mdt_object_child(mnew) : NULL,
2613 &rr->rr_tgt_name, ma);
2616 mutex_unlock(&mnew->mot_lov_mutex);
2618 /* handle last link of tgt object */
2620 mdt_counter_incr(req, LPROC_MDT_RENAME);
2622 mdt_handle_last_unlink(info, mnew, ma);
2623 discard = mdt_dom_check_for_discard(info, mnew);
2625 mdt_rename_counter_tally(info, info->mti_mdt, req,
2631 mdt_object_unlock(info, mnew, lh_newp, rc);
2633 mdt_object_unlock(info, mold, lh_oldp, rc);
2635 if (mnew && !discard)
2636 mdt_object_put(info->mti_env, mnew);
2638 mdt_object_put(info->mti_env, mold);
2640 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2641 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2643 if (lustre_handle_is_used(&rename_lh))
2644 mdt_rename_unlock(&rename_lh);
2646 mdt_object_put(info->mti_env, mtgtdir);
2648 mdt_object_put(info->mti_env, msrcdir);
2650 /* The DoM discard can be done right in the place above where it is
2651 * assigned, meanwhile it is done here after rename unlock due to
2652 * compatibility with old clients, for them the discard blocks
2653 * the main thread until completion. Check LU-11359 for details.
2656 mdt_dom_discard_data(info, mnew);
2657 mdt_object_put(info->mti_env, mnew);
2662 static int mdt_reint_resync(struct mdt_thread_info *info,
2663 struct mdt_lock_handle *lhc)
2665 struct mdt_reint_record *rr = &info->mti_rr;
2666 struct ptlrpc_request *req = mdt_info_req(info);
2667 struct md_attr *ma = &info->mti_attr;
2668 struct mdt_object *mo;
2669 struct ldlm_lock *lease;
2670 struct mdt_body *repbody;
2671 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2676 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2678 if (info->mti_dlm_req)
2679 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2681 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2683 GOTO(out, rc = PTR_ERR(mo));
2685 if (!mdt_object_exists(mo))
2686 GOTO(out_obj, rc = -ENOENT);
2688 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2689 GOTO(out_obj, rc = -EINVAL);
2691 if (mdt_object_remote(mo))
2692 GOTO(out_obj, rc = -EREMOTE);
2694 lease = ldlm_handle2lock(rr->rr_lease_handle);
2696 GOTO(out_obj, rc = -ESTALE);
2698 /* It's really necessary to grab open_sem and check if the lease lock
2699 * has been lost. There would exist a concurrent writer coming in and
2700 * generating some dirty data in memory cache, the writeback would fail
2701 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2702 if (!down_write_trylock(&mo->mot_open_sem))
2703 GOTO(out_put_lease, rc = -EBUSY);
2705 lock_res_and_lock(lease);
2706 lease_broken = ldlm_is_cancel(lease);
2707 unlock_res_and_lock(lease);
2709 GOTO(out_unlock, rc = -EBUSY);
2711 /* the file has yet opened by anyone else after we took the lease. */
2712 layout.mlc_opc = MD_LAYOUT_RESYNC;
2713 rc = mdt_layout_change(info, mo, &layout);
2715 GOTO(out_unlock, rc);
2717 ma->ma_need = MA_INODE;
2719 rc = mdt_attr_get_complex(info, mo, ma);
2721 GOTO(out_unlock, rc);
2723 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2724 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2728 up_write(&mo->mot_open_sem);
2730 LDLM_LOCK_PUT(lease);
2732 mdt_object_put(info->mti_env, mo);
2734 mdt_client_compatibility(info);
2735 rc2 = mdt_fix_reply(info);
2741 struct mdt_reinter {
2742 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2743 enum lprocfs_extra_opc mr_extra_opc;
2746 static const struct mdt_reinter mdt_reinters[] = {
2748 .mr_handler = &mdt_reint_setattr,
2749 .mr_extra_opc = MDS_REINT_SETATTR,
2752 .mr_handler = &mdt_reint_create,
2753 .mr_extra_opc = MDS_REINT_CREATE,
2756 .mr_handler = &mdt_reint_link,
2757 .mr_extra_opc = MDS_REINT_LINK,
2760 .mr_handler = &mdt_reint_unlink,
2761 .mr_extra_opc = MDS_REINT_UNLINK,
2764 .mr_handler = &mdt_reint_rename,
2765 .mr_extra_opc = MDS_REINT_RENAME,
2768 .mr_handler = &mdt_reint_open,
2769 .mr_extra_opc = MDS_REINT_OPEN,
2771 [REINT_SETXATTR] = {
2772 .mr_handler = &mdt_reint_setxattr,
2773 .mr_extra_opc = MDS_REINT_SETXATTR,
2776 .mr_handler = &mdt_reint_unlink,
2777 .mr_extra_opc = MDS_REINT_UNLINK,
2780 .mr_handler = &mdt_reint_migrate,
2781 .mr_extra_opc = MDS_REINT_RENAME,
2784 .mr_handler = &mdt_reint_resync,
2785 .mr_extra_opc = MDS_REINT_RESYNC,
2789 int mdt_reint_rec(struct mdt_thread_info *info,
2790 struct mdt_lock_handle *lhc)
2792 const struct mdt_reinter *mr;
2796 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2799 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2800 if (mr->mr_handler == NULL)
2803 rc = (*mr->mr_handler)(info, lhc);
2805 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2806 PTLRPC_LAST_CNTR + mr->mr_extra_opc);