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/
31 * lustre/mdt/mdt_reint.c
33 * Lustre Metadata Target (mdt) reintegration routines
35 * Author: Peter Braam <braam@clusterfs.com>
36 * Author: Andreas Dilger <adilger@clusterfs.com>
37 * Author: Phil Schwan <phil@clusterfs.com>
38 * Author: Huang Hua <huanghua@clusterfs.com>
39 * Author: Yury Umanets <umka@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_MDS
44 #include <lprocfs_status.h>
45 #include "mdt_internal.h"
46 #include <lustre_lmv.h>
47 #include <lustre_crypto.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 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;
210 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
211 if (likely(!IS_ERR(child))) {
212 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
213 mdt_object_put(info->mti_env, child);
216 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
217 return vbrc ? vbrc : rc;
221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
222 struct mdt_object *obj,
223 struct ldlm_enqueue_info *einfo,
226 union ldlm_policy_data *policy = &mti->mti_policy;
227 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
228 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
231 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
232 LASSERT(slave_locks);
234 memset(policy, 0, sizeof(*policy));
235 policy->l_inodebits.bits = einfo->ei_inodebits;
236 mdt_lock_handle_init(lh);
237 mdt_lock_reg_init(lh, einfo->ei_mode);
238 for (i = 0; i < slave_locks->ha_count; i++) {
239 if (test_bit(i, (void *)slave_locks->ha_map))
240 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
242 lh->mlh_reg_lh = slave_locks->ha_handles[i];
243 mdt_object_unlock(mti, NULL, lh, decref);
244 slave_locks->ha_handles[i].cookie = 0ull;
247 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
252 struct mdt_object *obj)
254 struct lu_device *bottom_dev;
255 struct lu_object *bottom_obj;
258 if (!S_ISDIR(obj->mot_header.loh_attr))
261 /* getxattr from bottom obj to avoid reading in shard FIDs */
262 bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
263 bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
264 mdt_object_fid(obj), NULL);
265 if (IS_ERR(bottom_obj))
266 return PTR_ERR(bottom_obj);
268 rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
270 lu_object_put(mti->mti_env, bottom_obj);
272 return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
276 * Lock slave stripes if necessary, the lock handles of slave stripes
277 * will be stored in einfo->ei_cbdata.
279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
280 enum ldlm_mode mode, __u64 ibits,
281 struct ldlm_enqueue_info *einfo)
283 union ldlm_policy_data *policy = &mti->mti_policy;
285 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
287 einfo->ei_type = LDLM_IBITS;
288 einfo->ei_mode = mode;
289 einfo->ei_cb_bl = mdt_remote_blocking_ast;
290 einfo->ei_cb_local_bl = mdt_blocking_ast;
291 einfo->ei_cb_cp = ldlm_completion_ast;
292 einfo->ei_enq_slave = 1;
293 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
294 einfo->ei_inodebits = ibits;
295 einfo->ei_req_slot = 1;
296 memset(policy, 0, sizeof(*policy));
297 policy->l_inodebits.bits = ibits;
299 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
303 int mdt_reint_striped_lock(struct mdt_thread_info *info,
304 struct mdt_object *o,
305 struct mdt_lock_handle *lh,
307 struct ldlm_enqueue_info *einfo,
312 LASSERT(!mdt_object_remote(o));
314 memset(einfo, 0, sizeof(*einfo));
316 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
320 rc = mdt_object_striped(info, o);
323 mdt_object_unlock(info, o, lh, rc);
327 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
329 mdt_object_unlock(info, o, lh, rc);
330 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
337 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
338 struct mdt_object *o,
339 struct mdt_lock_handle *lh,
340 struct ldlm_enqueue_info *einfo, int decref)
342 if (einfo->ei_cbdata)
343 mdt_unlock_slaves(info, o, einfo, decref);
344 mdt_object_unlock(info, o, lh, decref);
347 static int mdt_restripe(struct mdt_thread_info *info,
348 struct mdt_object *parent,
349 const struct lu_name *lname,
350 const struct lu_fid *tfid,
351 struct md_op_spec *spec,
354 struct mdt_device *mdt = info->mti_mdt;
355 struct lu_fid *fid = &info->mti_tmp_fid2;
356 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
357 struct lmv_user_md *lum = spec->u.sp_ea.eadata;
358 struct lmv_mds_md_v1 *lmv;
359 struct mdt_object *child;
360 struct mdt_lock_handle *lhp;
361 struct mdt_lock_handle *lhc;
362 struct mdt_body *repbody;
366 if (!mdt->mdt_enable_dir_restripe)
370 lum->lum_hash_type |= cpu_to_le32(LMV_HASH_FLAG_FIXED);
372 rc = mdt_version_get_check_save(info, parent, 0);
376 lhp = &info->mti_lh[MDT_LH_PARENT];
377 mdt_lock_pdo_init(lhp, LCK_PW, lname);
378 rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
383 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
385 GOTO(unlock_parent, rc);
387 if (ma->ma_valid & MA_LMV) {
388 /* don't allow restripe if parent dir layout is changing */
389 lmv = &ma->ma_lmv->lmv_md_v1;
390 if (!lmv_is_sane2(lmv))
391 GOTO(unlock_parent, rc = -EBADF);
393 if (lmv_is_layout_changing(lmv))
394 GOTO(unlock_parent, rc = -EBUSY);
398 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
400 GOTO(unlock_parent, rc);
402 child = mdt_object_find(info->mti_env, mdt, fid);
404 GOTO(unlock_parent, rc = PTR_ERR(child));
406 if (!mdt_object_exists(child))
407 GOTO(out_child, rc = -ENOENT);
409 if (mdt_object_remote(child)) {
410 struct mdt_body *repbody;
412 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
414 GOTO(out_child, rc = -EPROTO);
416 repbody->mbo_fid1 = *fid;
417 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
418 GOTO(out_child, rc = -EREMOTE);
421 if (!S_ISDIR(lu_object_attr(&child->mot_obj)))
422 GOTO(out_child, rc = -ENOTDIR);
424 rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
428 /* race with migrate? */
429 if ((ma->ma_valid & MA_LMV) &&
430 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1))
431 GOTO(out_child, rc = -EBUSY);
434 lhc = &info->mti_lh[MDT_LH_CHILD];
435 mdt_lock_reg_init(lhc, LCK_EX);
437 /* enqueue object remote LOOKUP lock */
438 if (mdt_object_remote(parent)) {
439 rc = mdt_remote_object_lock(info, parent, fid,
442 MDS_INODELOCK_LOOKUP, false);
447 rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
450 GOTO(unlock_child, rc);
452 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
453 rc = mdt_version_get_check_save(info, child, 1);
455 GOTO(unlock_child, rc);
457 spin_lock(&mdt->mdt_restriper.mdr_lock);
458 if (child->mot_restriping) {
460 spin_unlock(&mdt->mdt_restriper.mdr_lock);
461 GOTO(unlock_child, rc = -EBUSY);
463 child->mot_restriping = 1;
464 spin_unlock(&mdt->mdt_restriper.mdr_lock);
467 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
469 GOTO(restriping_clear, rc);
471 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
473 GOTO(restriping_clear, rc = -EPROTO);
475 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
479 child->mot_restriping = 0;
481 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
483 mdt_object_put(info->mti_env, child);
485 mdt_object_unlock(info, parent, lhp, rc);
491 * VBR: we save three versions in reply:
492 * 0 - parent. Check that parent version is the same during replay.
493 * 1 - name. Version of 'name' if file exists with the same name or
494 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
495 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
498 static int mdt_create(struct mdt_thread_info *info)
500 struct mdt_device *mdt = info->mti_mdt;
501 struct mdt_object *parent;
502 struct mdt_object *child;
503 struct mdt_lock_handle *lh;
504 struct mdt_body *repbody;
505 struct md_attr *ma = &info->mti_attr;
506 struct mdt_reint_record *rr = &info->mti_rr;
507 struct md_op_spec *spec = &info->mti_spec;
508 bool restripe = false;
512 DEBUG_REQ(D_INODE, mdt_info_req(info),
513 "Create ("DNAME"->"DFID") in "DFID,
514 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
516 if (!fid_is_md_operative(rr->rr_fid1))
519 if (S_ISDIR(ma->ma_attr.la_mode) &&
520 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
521 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
522 struct lu_ucred *uc = mdt_ucred(info);
523 struct obd_export *exp = mdt_info_req(info)->rq_export;
525 /* Only new clients can create remote dir( >= 2.4) and
526 * striped dir(>= 2.6), old client will return -ENOTSUPP
528 if (!mdt_is_dne_client(exp))
531 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
532 if (!mdt_is_striped_client(exp))
535 if (!mdt->mdt_enable_striped_dir)
537 } else if (!mdt->mdt_enable_remote_dir) {
541 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
542 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) >=
546 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
547 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
548 mdt->mdt_enable_remote_dir_gid != -1)
551 /* restripe if later found dir exists, MDS_OPEN_CREAT means
552 * this is create only, don't try restripe.
554 if (mdt->mdt_enable_dir_restripe &&
555 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
556 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
560 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
562 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
564 RETURN(PTR_ERR(parent));
566 if (!mdt_object_exists(parent))
567 GOTO(put_parent, rc = -ENOENT);
569 rc = mdt_check_enc(info, parent);
571 GOTO(put_parent, rc);
574 * LU-10235: check if name exists locklessly first to avoid massive
575 * lock recalls on existing directories.
577 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
578 &info->mti_tmp_fid1, 1);
581 GOTO(put_parent, rc = -EEXIST);
583 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
587 /* -ENOENT is expected here */
589 GOTO(put_parent, rc);
591 /* save version of file name for replay, it must be ENOENT here */
592 mdt_enoent_version_save(info, 1);
594 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
596 lh = &info->mti_lh[MDT_LH_PARENT];
597 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
598 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
600 GOTO(put_parent, rc);
602 if (!mdt_object_remote(parent)) {
603 rc = mdt_version_get_check_save(info, parent, 0);
605 GOTO(unlock_parent, rc);
608 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
609 if (unlikely(IS_ERR(child)))
610 GOTO(unlock_parent, rc = PTR_ERR(child));
612 ma->ma_need = MA_INODE;
615 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
616 OBD_FAIL_MDS_REINT_CREATE_WRITE);
618 /* Version of child will be updated on disk. */
619 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
620 rc = mdt_version_get_check_save(info, child, 2);
625 * Do not perform lookup sanity check. We know that name does
628 info->mti_spec.sp_cr_lookup = 0;
629 if (mdt_object_remote(parent))
630 info->mti_spec.sp_cr_lookup = 1;
631 info->mti_spec.sp_feat = &dt_directory_features;
633 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
634 mdt_object_child(child), &info->mti_spec, ma);
636 rc = mdt_attr_get_complex(info, child, ma);
642 * On DNE, we need to eliminate dependey between 'mkdir a' and
643 * 'mkdir a/b' if b is a striped directory, to achieve this, two
644 * things are done below:
645 * 1. save child and slaves lock.
646 * 2. if the child is a striped directory, relock parent so to
647 * compare against with COS locks to ensure parent was
650 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
651 struct mdt_lock_handle *lhc;
652 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
655 rc = mdt_object_striped(info, child);
661 if (!mdt_object_remote(parent)) {
662 mdt_object_unlock(info, parent, lh, 1);
663 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
664 rc = mdt_reint_object_lock(info, parent, lh,
665 MDS_INODELOCK_UPDATE,
672 lhc = &info->mti_lh[MDT_LH_CHILD];
673 mdt_lock_handle_init(lhc);
674 mdt_lock_reg_init(lhc, LCK_PW);
675 rc = mdt_reint_striped_lock(info, child, lhc,
676 MDS_INODELOCK_UPDATE, einfo,
681 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
684 /* Return fid & attr to client. */
685 if (ma->ma_valid & MA_INODE)
686 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
687 mdt_object_fid(child));
690 mdt_object_put(info->mti_env, child);
692 mdt_object_unlock(info, parent, lh, rc);
694 mdt_object_put(info->mti_env, parent);
698 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
701 struct mdt_lock_handle *lh;
702 int do_vbr = ma->ma_attr.la_valid &
703 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
704 __u64 lockpart = MDS_INODELOCK_UPDATE;
705 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
710 rc = mdt_object_striped(info, mo);
716 lh = &info->mti_lh[MDT_LH_PARENT];
717 mdt_lock_reg_init(lh, LCK_PW);
719 /* Even though the new MDT will grant PERM lock to the old
720 * client, but the old client will almost ignore that during
721 * So it needs to revoke both LOOKUP and PERM lock here, so
722 * both new and old client can cancel the dcache
724 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
725 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
726 /* Clear xattr cache on clients, so the virtual project ID xattr
727 * can get the new project ID
729 if (ma->ma_attr.la_valid & LA_PROJID)
730 lockpart |= MDS_INODELOCK_XATTR;
732 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
737 /* all attrs are packed into mti_attr in unpack_setattr */
738 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
739 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
741 /* VBR: update version if attr changed are important for recovery */
743 /* update on-disk version of changed object */
744 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
745 rc = mdt_version_get_check_save(info, mo, 0);
747 GOTO(out_unlock, rc);
750 /* Ensure constant striping during chown(). See LU-2789. */
751 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
752 mutex_lock(&mo->mot_lov_mutex);
754 /* all attrs are packed into mti_attr in unpack_setattr */
755 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
757 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
758 mutex_unlock(&mo->mot_lov_mutex);
761 GOTO(out_unlock, rc);
762 mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
765 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
770 * Check HSM flags and add HS_DIRTY flag if relevant.
772 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
773 * and is not RELEASED.
775 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
778 struct lu_ucred *uc = mdt_ucred(info);
779 kernel_cap_t cap_saved;
783 /* If the file was modified, add the dirty flag */
784 ma->ma_need = MA_HSM;
785 rc = mdt_attr_get_complex(info, mo, ma);
787 CERROR("file attribute read error for "DFID": %d.\n",
788 PFID(mdt_object_fid(mo)), rc);
792 /* If an up2date copy exists in the backend, add dirty flag */
793 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
794 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
795 ma->ma_hsm.mh_flags |= HS_DIRTY;
797 /* Bump cap so that closes from non-owner writers can
798 * set the HSM state to dirty.
800 cap_saved = uc->uc_cap;
801 cap_raise(uc->uc_cap, CAP_FOWNER);
802 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
803 uc->uc_cap = cap_saved;
805 CERROR("file attribute change error for "DFID": %d\n",
806 PFID(mdt_object_fid(mo)), rc);
812 static int mdt_reint_setattr(struct mdt_thread_info *info,
813 struct mdt_lock_handle *lhc)
815 struct mdt_device *mdt = info->mti_mdt;
816 struct md_attr *ma = &info->mti_attr;
817 struct mdt_reint_record *rr = &info->mti_rr;
818 struct ptlrpc_request *req = mdt_info_req(info);
819 struct mdt_object *mo;
820 struct mdt_body *repbody;
821 ktime_t kstart = ktime_get();
825 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
826 (unsigned int)ma->ma_attr.la_valid);
828 if (info->mti_dlm_req)
829 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
831 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
833 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
834 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
836 GOTO(out, rc = PTR_ERR(mo));
838 if (!mdt_object_exists(mo))
839 GOTO(out_put, rc = -ENOENT);
841 if (mdt_object_remote(mo))
842 GOTO(out_put, rc = -EREMOTE);
844 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
845 /* revoke lease lock if size is going to be changed */
846 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
847 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
848 atomic_read(&mo->mot_lease_count) > 0)) {
849 down_read(&mo->mot_open_sem);
851 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
852 lhc = &info->mti_lh[MDT_LH_LOCAL];
853 mdt_lock_reg_init(lhc, LCK_CW);
855 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
857 up_read(&mo->mot_open_sem);
861 /* revoke lease lock */
862 mdt_object_unlock(info, mo, lhc, 1);
864 up_read(&mo->mot_open_sem);
867 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
868 /* Check write access for the O_TRUNC case */
869 if (mdt_write_read(mo) < 0)
870 GOTO(out_put, rc = -ETXTBSY);
872 /* LU-10286: compatibility check for FLR.
873 * Please check the comment in mdt_finish_open() for details
875 if (!exp_connect_flr(info->mti_exp) ||
876 !exp_connect_overstriping(info->mti_exp)) {
877 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
878 if (rc < 0 && rc != -ENODATA)
881 if (!exp_connect_flr(info->mti_exp)) {
883 mdt_lmm_is_flr(info->mti_big_lmm))
884 GOTO(out_put, rc = -EOPNOTSUPP);
887 if (!exp_connect_overstriping(info->mti_exp)) {
889 mdt_lmm_is_overstriping(info->mti_big_lmm))
890 GOTO(out_put, rc = -EOPNOTSUPP);
894 /* For truncate, the file size sent from client
895 * is believable, but the blocks are incorrect,
896 * which makes the block size in LSOM attribute
897 * inconsisent with the real block size.
899 rc = mdt_lsom_update(info, mo, true);
904 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
905 if (ma->ma_valid & MA_LOV)
906 GOTO(out_put, rc = -EPROTO);
908 /* MDT supports FMD for regular files due to Data-on-MDT */
909 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
910 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
911 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
914 if (ma->ma_attr.la_valid & LA_MTIME) {
915 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
917 ma->ma_valid |= MA_PFID;
921 rc = mdt_attr_set(info, mo, ma);
924 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
925 (ma->ma_valid & MA_INODE)) {
926 struct lu_buf *buf = &info->mti_buf;
927 struct lu_ucred *uc = mdt_ucred(info);
928 struct mdt_lock_handle *lh;
930 __u64 lockpart = MDS_INODELOCK_XATTR;
932 /* reject if either remote or striped dir is disabled */
933 if (ma->ma_valid & MA_LMV) {
934 if (!mdt->mdt_enable_remote_dir ||
935 !mdt->mdt_enable_striped_dir)
936 GOTO(out_put, rc = -EPERM);
938 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
939 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
940 mdt->mdt_enable_remote_dir_gid != -1)
941 GOTO(out_put, rc = -EPERM);
944 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
945 GOTO(out_put, rc = -ENOTDIR);
947 if (ma->ma_attr.la_valid != 0)
948 GOTO(out_put, rc = -EPROTO);
950 lh = &info->mti_lh[MDT_LH_PARENT];
951 mdt_lock_reg_init(lh, LCK_PW);
953 if (ma->ma_valid & MA_LOV) {
954 buf->lb_buf = ma->ma_lmm;
955 buf->lb_len = ma->ma_lmm_size;
956 name = XATTR_NAME_LOV;
958 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
959 struct lu_fid *pfid = &info->mti_tmp_fid1;
960 struct lu_name *pname = &info->mti_name;
961 const char dotdot[] = "..";
962 struct mdt_object *pobj;
965 buf->lb_len = ma->ma_lmv_size;
966 name = XATTR_NAME_DEFAULT_LMV;
968 if (fid_is_root(rr->rr_fid1)) {
969 lockpart |= MDS_INODELOCK_LOOKUP;
971 /* force client to update dir default layout */
973 pname->ln_name = dotdot;
974 pname->ln_namelen = sizeof(dotdot);
975 rc = mdo_lookup(info->mti_env,
976 mdt_object_child(mo), pname,
981 pobj = mdt_object_find(info->mti_env, mdt,
984 GOTO(out_put, rc = PTR_ERR(pobj));
986 if (mdt_object_remote(pobj))
987 rc = mdt_remote_object_lock(info, pobj,
989 &lh->mlh_rreg_lh, LCK_EX,
990 MDS_INODELOCK_LOOKUP, false);
992 lockpart |= MDS_INODELOCK_LOOKUP;
994 mdt_object_put(info->mti_env, pobj);
1001 rc = mdt_object_lock(info, mo, lh, lockpart);
1005 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
1008 mdt_object_unlock(info, mo, lh, rc);
1012 GOTO(out_put, rc = -EPROTO);
1015 /* If file data is modified, add the dirty flag */
1016 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1017 rc = mdt_add_dirty_flag(info, mo, ma);
1019 ma->ma_need = MA_INODE;
1021 rc = mdt_attr_get_complex(info, mo, ma);
1025 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1029 mdt_object_put(info->mti_env, mo);
1032 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1033 ktime_us_delta(ktime_get(), kstart));
1035 mdt_client_compatibility(info);
1039 static int mdt_reint_create(struct mdt_thread_info *info,
1040 struct mdt_lock_handle *lhc)
1042 struct ptlrpc_request *req = mdt_info_req(info);
1043 ktime_t kstart = ktime_get();
1047 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1048 RETURN(err_serious(-ESTALE));
1050 if (info->mti_dlm_req)
1051 ldlm_request_cancel(mdt_info_req(info),
1052 info->mti_dlm_req, 0, LATF_SKIP);
1054 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1057 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1067 CERROR("%s: Unsupported mode %o\n",
1068 mdt_obd_name(info->mti_mdt),
1069 info->mti_attr.ma_attr.la_mode);
1070 RETURN(err_serious(-EOPNOTSUPP));
1073 rc = mdt_create(info);
1075 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1076 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1077 ktime_us_delta(ktime_get(), kstart));
1079 /* Special file should stay on the same node as parent*/
1080 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1081 ktime_us_delta(ktime_get(), kstart));
1088 * VBR: save parent version in reply and child version getting by its name.
1089 * Version of child is getting and checking during its lookup. If
1091 static int mdt_reint_unlink(struct mdt_thread_info *info,
1092 struct mdt_lock_handle *lhc)
1094 struct mdt_reint_record *rr = &info->mti_rr;
1095 struct ptlrpc_request *req = mdt_info_req(info);
1096 struct md_attr *ma = &info->mti_attr;
1097 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1098 struct mdt_object *mp;
1099 struct mdt_object *mc;
1100 struct mdt_lock_handle *parent_lh;
1101 struct mdt_lock_handle *child_lh;
1102 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1104 bool cos_incompat = false;
1106 ktime_t kstart = ktime_get();
1110 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1111 PNAME(&rr->rr_name));
1113 if (info->mti_dlm_req)
1114 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1116 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1117 RETURN(err_serious(-ENOENT));
1119 if (!fid_is_md_operative(rr->rr_fid1))
1122 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1124 RETURN(PTR_ERR(mp));
1126 if (mdt_object_remote(mp)) {
1127 cos_incompat = true;
1129 rc = mdt_version_get_check_save(info, mp, 0);
1131 GOTO(put_parent, rc);
1134 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1135 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1137 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1138 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1139 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1142 GOTO(put_parent, rc);
1144 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1145 *child_fid = *rr->rr_fid2;
1147 /* lookup child object along with version checking */
1148 fid_zero(child_fid);
1149 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1152 /* Name might not be able to find during resend of
1153 * remote unlink, considering following case.
1154 * dir_A is a remote directory, the name entry of
1155 * dir_A is on MDT0, the directory is on MDT1,
1157 * 1. client sends unlink req to MDT1.
1158 * 2. MDT1 sends name delete update to MDT0.
1159 * 3. name entry is being deleted in MDT0 synchronously.
1160 * 4. MDT1 is restarted.
1161 * 5. client resends unlink req to MDT1. So it can not
1162 * find the name entry on MDT0 anymore.
1163 * In this case, MDT1 only needs to destory the local
1166 if (mdt_object_remote(mp) && rc == -ENOENT &&
1167 !fid_is_zero(rr->rr_fid2) &&
1168 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1170 *child_fid = *rr->rr_fid2;
1172 GOTO(unlock_parent, rc);
1177 if (!fid_is_md_operative(child_fid))
1178 GOTO(unlock_parent, rc = -EPERM);
1180 /* We will lock the child regardless it is local or remote. No harm. */
1181 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1183 GOTO(unlock_parent, rc = PTR_ERR(mc));
1185 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1186 /* In this case, child fid is embedded in the request, and we do
1187 * not have a proper name as rr_name contains an encoded
1188 * hash. So find name that matches provided hash.
1190 if (!find_name_matching_hash(info, &rr->rr_name,
1192 GOTO(put_child, rc = -ENOENT);
1195 if (!cos_incompat) {
1196 rc = mdt_object_striped(info, mc);
1198 GOTO(put_child, rc);
1202 mdt_object_put(info->mti_env, mc);
1203 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1208 child_lh = &info->mti_lh[MDT_LH_CHILD];
1209 mdt_lock_reg_init(child_lh, LCK_EX);
1210 if (info->mti_spec.sp_rm_entry) {
1211 struct lu_ucred *uc = mdt_ucred(info);
1213 if (!mdt_is_dne_client(req->rq_export))
1214 /* Return -ENOTSUPP for old client */
1215 GOTO(put_child, rc = -ENOTSUPP);
1217 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1218 GOTO(put_child, rc = -EPERM);
1220 ma->ma_need = MA_INODE;
1222 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1223 NULL, &rr->rr_name, ma, no_name);
1224 GOTO(put_child, rc);
1227 if (mdt_object_remote(mc)) {
1228 struct mdt_body *repbody;
1230 if (!fid_is_zero(rr->rr_fid2)) {
1231 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1232 mdt_obd_name(info->mti_mdt),
1233 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1234 GOTO(put_child, rc = -ENOENT);
1236 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1237 mdt_obd_name(info->mti_mdt),
1238 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1240 if (!mdt_is_dne_client(req->rq_export))
1241 /* Return -ENOTSUPP for old client */
1242 GOTO(put_child, rc = -ENOTSUPP);
1244 /* Revoke the LOOKUP lock of the remote object granted by
1245 * this MDT. Since the unlink will happen on another MDT,
1246 * it will release the LOOKUP lock right away. Then What
1247 * would happen if another client try to grab the LOOKUP
1248 * lock at the same time with unlink XXX
1250 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1251 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1252 LASSERT(repbody != NULL);
1253 repbody->mbo_fid1 = *mdt_object_fid(mc);
1254 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1255 GOTO(unlock_child, rc = -EREMOTE);
1257 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1258 * this now because a running HSM restore on the child (unlink
1259 * victim) will hold the layout lock. See LU-4002.
1261 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1262 if (mdt_object_remote(mp)) {
1263 /* Enqueue lookup lock from parent MDT */
1264 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1265 &child_lh->mlh_rreg_lh,
1266 child_lh->mlh_rreg_mode,
1267 MDS_INODELOCK_LOOKUP, false);
1269 GOTO(put_child, rc);
1271 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1274 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1277 GOTO(put_child, rc);
1280 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1281 * whether need MA_LOV and MA_COOKIE.
1283 ma->ma_need = MA_INODE;
1286 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1287 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1288 /* save version when object is locked */
1289 mdt_version_get_save(info, mc, 1);
1291 mutex_lock(&mc->mot_lov_mutex);
1293 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1294 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1296 mutex_unlock(&mc->mot_lov_mutex);
1298 GOTO(unlock_child, rc);
1300 if (!lu_object_is_dying(&mc->mot_header)) {
1301 rc = mdt_attr_get_complex(info, mc, ma);
1304 } else if (mdt_dom_check_for_discard(info, mc)) {
1305 mdt_dom_discard_data(info, mc);
1307 mdt_handle_last_unlink(info, mc, ma);
1310 if (ma->ma_valid & MA_INODE) {
1311 switch (ma->ma_attr.la_mode & S_IFMT) {
1313 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1314 ktime_us_delta(ktime_get(), kstart));
1322 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1323 ktime_us_delta(ktime_get(), kstart));
1326 LASSERTF(0, "bad file type %o unlinking\n",
1327 ma->ma_attr.la_mode);
1334 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1336 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1337 info->mti_big_buf.lb_buf)
1338 lu_buf_free(&info->mti_big_buf);
1339 mdt_object_put(info->mti_env, mc);
1341 mdt_object_unlock(info, mp, parent_lh, rc);
1343 mdt_object_put(info->mti_env, mp);
1344 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1349 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1352 static int mdt_reint_link(struct mdt_thread_info *info,
1353 struct mdt_lock_handle *lhc)
1355 struct mdt_reint_record *rr = &info->mti_rr;
1356 struct ptlrpc_request *req = mdt_info_req(info);
1357 struct md_attr *ma = &info->mti_attr;
1358 struct mdt_object *ms;
1359 struct mdt_object *mp;
1360 struct mdt_lock_handle *lhs;
1361 struct mdt_lock_handle *lhp;
1362 ktime_t kstart = ktime_get();
1367 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1368 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1370 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1371 RETURN(err_serious(-ENOENT));
1373 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1374 OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1375 req->rq_no_reply = 1;
1376 RETURN(err_serious(-ENOENT));
1379 if (info->mti_dlm_req)
1380 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1382 /* Invalid case so return error immediately instead of
1385 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1388 if (!fid_is_md_operative(rr->rr_fid1) ||
1389 !fid_is_md_operative(rr->rr_fid2))
1392 /* step 1: find target parent dir */
1393 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1395 RETURN(PTR_ERR(mp));
1397 rc = mdt_version_get_check_save(info, mp, 0);
1399 GOTO(put_parent, rc);
1401 rc = mdt_check_enc(info, mp);
1403 GOTO(put_parent, rc);
1405 /* step 2: find source */
1406 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1408 GOTO(put_parent, rc = PTR_ERR(ms));
1410 if (!mdt_object_exists(ms)) {
1411 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1412 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1413 GOTO(put_source, rc = -ENOENT);
1416 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1418 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1420 lhp = &info->mti_lh[MDT_LH_PARENT];
1421 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1422 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1425 GOTO(put_source, rc);
1427 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1429 lhs = &info->mti_lh[MDT_LH_CHILD];
1430 mdt_lock_reg_init(lhs, LCK_EX);
1431 rc = mdt_reint_object_lock(info, ms, lhs,
1432 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1435 GOTO(unlock_parent, rc);
1437 /* step 3: link it */
1438 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1439 OBD_FAIL_MDS_REINT_LINK_WRITE);
1441 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1442 rc = mdt_version_get_check_save(info, ms, 1);
1444 GOTO(unlock_source, rc);
1446 /** check target version by name during replay */
1447 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1448 &info->mti_tmp_fid1, 2);
1449 if (rc != 0 && rc != -ENOENT)
1450 GOTO(unlock_source, rc);
1451 /* save version of file name for replay, it must be ENOENT here */
1452 if (!req_is_replay(mdt_info_req(info))) {
1453 if (rc != -ENOENT) {
1454 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1455 PNAME(&rr->rr_name));
1456 GOTO(unlock_source, rc = -EEXIST);
1458 info->mti_ver[2] = ENOENT_VERSION;
1459 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1462 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1463 mdt_object_child(ms), &rr->rr_name, ma);
1466 mdt_counter_incr(req, LPROC_MDT_LINK,
1467 ktime_us_delta(ktime_get(), kstart));
1471 mdt_object_unlock(info, ms, lhs, rc);
1473 mdt_object_unlock(info, mp, lhp, rc);
1475 mdt_object_put(info->mti_env, ms);
1477 mdt_object_put(info->mti_env, mp);
1481 * lock the part of the directory according to the hash of the name
1482 * (lh->mlh_pdo_hash) in parallel directory lock.
1484 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1485 struct mdt_lock_handle *lh,
1486 struct mdt_object *obj, __u64 ibits,
1489 struct ldlm_res_id *res = &info->mti_res_id;
1490 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1491 union ldlm_policy_data *policy = &info->mti_policy;
1492 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1496 * Finish res_id initializing by name hash marking part of
1497 * directory which is taking modification.
1499 LASSERT(lh->mlh_pdo_hash != 0);
1500 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1501 memset(policy, 0, sizeof(*policy));
1502 policy->l_inodebits.bits = ibits;
1504 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1505 dlmflags |= LDLM_FL_COS_INCOMPAT;
1507 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1508 * going to be sent to client. If it is - mdt_intent_policy() path will
1509 * fix it up and turn FL_LOCAL flag off.
1511 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1512 policy, res, dlmflags,
1513 &info->mti_exp->exp_handle.h_cookie);
1518 * Get BFL lock for rename or migrate process.
1520 static int mdt_rename_lock(struct mdt_thread_info *info,
1521 struct lustre_handle *lh)
1526 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1527 struct lu_fid *fid = &info->mti_tmp_fid1;
1528 struct mdt_object *obj;
1530 /* XXX, right now, it has to use object API to
1531 * enqueue lock cross MDT, so it will enqueue
1532 * rename lock(with LUSTRE_BFL_FID) by root object
1535 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1537 RETURN(PTR_ERR(obj));
1539 rc = mdt_remote_object_lock(info, obj,
1540 &LUSTRE_BFL_FID, lh,
1542 MDS_INODELOCK_UPDATE, false);
1543 mdt_object_put(info->mti_env, obj);
1545 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1546 union ldlm_policy_data *policy = &info->mti_policy;
1547 struct ldlm_res_id *res_id = &info->mti_res_id;
1550 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1551 memset(policy, 0, sizeof(*policy));
1552 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1553 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1554 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1555 LDLM_IBITS, policy, LCK_EX, &flags,
1557 ldlm_completion_ast, NULL, NULL, 0,
1559 &info->mti_exp->exp_handle.h_cookie,
1566 static void mdt_rename_unlock(struct lustre_handle *lh)
1569 LASSERT(lustre_handle_is_used(lh));
1570 /* Cancel the single rename lock right away */
1571 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1575 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1576 const struct lu_fid *fid,
1579 struct mdt_object *dir;
1583 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1587 /* check early, the real version will be saved after locking */
1588 rc = mdt_version_get_check(info, dir, idx);
1592 if (!mdt_object_exists(dir))
1593 GOTO(out_put, rc = -ENOENT);
1595 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1596 GOTO(out_put, rc = -ENOTDIR);
1600 mdt_object_put(info->mti_env, dir);
1605 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1606 * herein we don't really check it, just do revoke.
1608 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1609 struct mdt_object *pobj,
1610 struct mdt_object *obj)
1612 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1615 mdt_lock_handle_init(lh);
1616 mdt_lock_reg_init(lh, LCK_EX);
1618 if (mdt_object_remote(pobj)) {
1619 /* don't bother to check if pobj and obj are on the same MDT. */
1620 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1621 &lh->mlh_rreg_lh, LCK_EX,
1622 MDS_INODELOCK_LOOKUP, false);
1623 } else if (mdt_object_remote(obj)) {
1624 struct ldlm_res_id *res = &info->mti_res_id;
1625 union ldlm_policy_data *policy = &info->mti_policy;
1626 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1627 LDLM_FL_COS_INCOMPAT;
1629 fid_build_reg_res_name(mdt_object_fid(obj), res);
1630 memset(policy, 0, sizeof(*policy));
1631 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1632 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1633 &lh->mlh_reg_lh, LCK_EX, policy, res,
1636 /* do nothing if both are local */
1644 * TODO, currently we don't save this lock because there is no place to
1645 * hold this lock handle, but to avoid race we need to save this lock.
1647 mdt_object_unlock(info, NULL, lh, 1);
1653 * operation may takes locks of linkea, or directory stripes, group them in
1656 struct mdt_sub_lock {
1657 struct mdt_object *msl_obj;
1658 struct mdt_lock_handle msl_lh;
1659 struct list_head msl_linkage;
1662 static void mdt_unlock_list(struct mdt_thread_info *info,
1663 struct list_head *list, int decref)
1665 struct mdt_sub_lock *msl;
1666 struct mdt_sub_lock *tmp;
1668 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1669 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1670 list_del(&msl->msl_linkage);
1675 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1676 struct mdt_object *obj,
1677 struct mdt_lock_handle *lh,
1678 struct ldlm_enqueue_info *einfo,
1679 struct list_head *slave_locks,
1682 if (mdt_object_remote(obj)) {
1683 mdt_unlock_list(info, slave_locks, decref);
1684 mdt_object_unlock(info, obj, lh, decref);
1686 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1691 * lock parents of links, and also check whether total locks don't exceed
1694 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1695 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1696 * \retval -ev negative errno upon error
1698 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1699 struct mdt_object *pobj,
1700 const struct md_attr *ma,
1701 struct mdt_object *obj,
1702 struct mdt_lock_handle *lhp,
1703 struct ldlm_enqueue_info *peinfo,
1704 struct list_head *parent_slave_locks,
1705 struct list_head *link_locks)
1707 struct mdt_device *mdt = info->mti_mdt;
1708 struct lu_buf *buf = &info->mti_big_buf;
1709 struct lu_name *lname = &info->mti_name;
1710 struct linkea_data ldata = { NULL };
1711 bool blocked = false;
1712 int local_lnkp_cnt = 0;
1716 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1719 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1720 if (buf->lb_buf == NULL)
1724 rc = mdt_links_read(info, obj, &ldata);
1726 if (rc == -ENOENT || rc == -ENODATA)
1731 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1732 linkea_next_entry(&ldata)) {
1733 struct mdt_object *lnkp;
1734 struct mdt_sub_lock *msl;
1738 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1741 /* check if it's also linked to parent */
1742 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1743 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1744 PFID(&fid), PNAME(lname));
1745 /* in case link is remote object, revoke LOOKUP lock */
1746 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1752 /* check if it's linked to a stripe of parent */
1753 if (ma->ma_valid & MA_LMV) {
1754 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1755 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1758 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1759 fid_le_to_cpu(stripe_fid,
1760 &lmv->lmv_stripe_fids[j]);
1761 if (lu_fid_eq(stripe_fid, &fid)) {
1762 CDEBUG(D_INFO, "skip stripe "DFID
1763 ", reovke "DNAME"\n",
1764 PFID(&fid), PNAME(lname));
1765 lnkp = mdt_object_find(info->mti_env,
1768 GOTO(out, rc = PTR_ERR(lnkp));
1774 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1776 mdt_object_put(info->mti_env, lnkp);
1781 /* Check if it's already locked */
1782 list_for_each_entry(msl, link_locks, msl_linkage) {
1783 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1785 DFID" was locked, revoke "DNAME"\n",
1786 PFID(&fid), PNAME(lname));
1787 lnkp = msl->msl_obj;
1793 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1797 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1798 PFID(&fid), PNAME(lname));
1800 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1802 CWARN("%s: cannot find obj "DFID": %ld\n",
1803 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1807 if (!mdt_object_exists(lnkp)) {
1808 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1809 PFID(&fid), PNAME(lname));
1810 mdt_object_put(info->mti_env, lnkp);
1814 if (!mdt_object_remote(lnkp))
1819 GOTO(out, rc = -ENOMEM);
1822 * we can't follow parent-child lock order like other MD
1823 * operations, use lock_try here to avoid deadlock, if the lock
1824 * cannot be taken, drop all locks taken, revoke the blocked
1825 * one, and continue processing the remaining entries, and in
1826 * the end of the loop restart from beginning.
1828 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1830 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1831 MDS_INODELOCK_UPDATE, true);
1832 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1834 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1835 PFID(&fid), PNAME(lname));
1837 mdt_unlock_list(info, link_locks, 1);
1838 /* also unlock parent locks to avoid deadlock */
1840 mdt_migrate_object_unlock(info, pobj, lhp,
1847 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1848 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1849 MDS_INODELOCK_UPDATE);
1851 mdt_object_put(info->mti_env, lnkp);
1856 if (mdt_object_remote(lnkp)) {
1857 struct ldlm_lock *lock;
1860 * for remote object, set lock cb_atomic,
1861 * so lock can be released in blocking_ast()
1862 * immediately, then the next lock_try will
1863 * have better chance of success.
1865 lock = ldlm_handle2lock(
1866 &msl->msl_lh.mlh_rreg_lh);
1867 LASSERT(lock != NULL);
1868 lock_res_and_lock(lock);
1869 ldlm_set_atomic_cb(lock);
1870 unlock_res_and_lock(lock);
1871 LDLM_LOCK_PUT(lock);
1874 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1879 INIT_LIST_HEAD(&msl->msl_linkage);
1880 msl->msl_obj = lnkp;
1881 list_add_tail(&msl->msl_linkage, link_locks);
1883 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1887 GOTO(out, rc = -EBUSY);
1892 mdt_unlock_list(info, link_locks, rc);
1893 } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1894 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1897 * parent may have 3 local objects: master object and 2 stripes
1898 * (if it's being migrated too); source may have 1 local objects
1899 * as regular file; target has 1 local object.
1900 * Note, source may have 2 local locks if it is directory but it
1901 * can't have hardlinks, so it is not considered here.
1908 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1909 struct mdt_object *obj,
1910 const struct md_attr *ma,
1911 struct list_head *slave_locks)
1913 struct mdt_device *mdt = info->mti_mdt;
1914 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1915 struct lu_fid *fid = &info->mti_tmp_fid1;
1916 struct mdt_object *slave;
1917 struct mdt_sub_lock *msl;
1922 LASSERT(mdt_object_remote(obj));
1923 LASSERT(ma->ma_valid & MA_LMV);
1926 if (!lmv_is_sane(lmv))
1929 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1930 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1932 if (!fid_is_sane(fid))
1935 slave = mdt_object_find(info->mti_env, mdt, fid);
1937 GOTO(out, rc = PTR_ERR(slave));
1941 mdt_object_put(info->mti_env, slave);
1942 GOTO(out, rc = -ENOMEM);
1945 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1946 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1947 MDS_INODELOCK_UPDATE, true);
1950 mdt_object_put(info->mti_env, slave);
1954 INIT_LIST_HEAD(&msl->msl_linkage);
1955 msl->msl_obj = slave;
1956 list_add_tail(&msl->msl_linkage, slave_locks);
1962 mdt_unlock_list(info, slave_locks, rc);
1966 /* lock parent and its stripes */
1967 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1968 struct mdt_object *obj,
1969 const struct md_attr *ma,
1970 struct mdt_lock_handle *lh,
1971 struct ldlm_enqueue_info *einfo,
1972 struct list_head *slave_locks)
1976 if (mdt_object_remote(obj)) {
1977 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1978 &lh->mlh_rreg_lh, LCK_PW,
1979 MDS_INODELOCK_UPDATE, false);
1984 * if obj is remote and striped, lock its stripes explicitly
1985 * because it's not striped in LOD layer on this MDT.
1987 if (ma->ma_valid & MA_LMV) {
1988 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1990 mdt_object_unlock(info, obj, lh, rc);
1993 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
2001 * in migration, object may be remote, and we need take full lock of it and its
2002 * stripes if it's directory, besides, object may be a remote object on its
2003 * parent, revoke its LOOKUP lock on where its parent is located.
2005 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
2006 struct mdt_object *pobj,
2007 struct mdt_object *obj,
2008 struct mdt_lock_handle *lh,
2009 struct ldlm_enqueue_info *einfo,
2010 struct list_head *slave_locks)
2014 if (mdt_object_remote(obj)) {
2015 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2019 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
2020 &lh->mlh_rreg_lh, LCK_EX,
2021 MDS_INODELOCK_FULL, false);
2026 * if obj is remote and striped, lock its stripes explicitly
2027 * because it's not striped in LOD layer on this MDT.
2029 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
2030 struct md_attr *ma = &info->mti_attr;
2032 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
2034 mdt_object_unlock(info, obj, lh, rc);
2038 if (ma->ma_valid & MA_LMV) {
2039 rc = mdt_lock_remote_slaves(info, obj, ma,
2042 mdt_object_unlock(info, obj, lh, rc);
2046 if (mdt_object_remote(pobj)) {
2047 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2052 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
2060 * lookup source by name, if parent is striped directory, we need to find the
2061 * corresponding stripe where source is located, and then lookup there.
2063 * besides, if parent is migrating too, and file is already in target stripe,
2064 * this should be a redo of 'lfs migrate' on client side.
2066 static int mdt_migrate_lookup(struct mdt_thread_info *info,
2067 struct mdt_object *pobj,
2068 const struct md_attr *ma,
2069 const struct lu_name *lname,
2070 struct mdt_object **spobj,
2071 struct mdt_object **sobj)
2073 const struct lu_env *env = info->mti_env;
2074 struct lu_fid *fid = &info->mti_tmp_fid1;
2075 struct mdt_object *stripe;
2078 if (ma->ma_valid & MA_LMV) {
2079 /* if parent is striped, lookup on corresponding stripe */
2080 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2082 if (!lmv_is_sane(lmv))
2085 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2090 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2092 stripe = mdt_object_find(env, info->mti_mdt, fid);
2094 return PTR_ERR(stripe);
2097 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2099 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2101 * if parent layout is changeing, and lookup child
2102 * failed on source stripe, lookup again on target
2103 * stripe, if it exists, it means previous migration
2104 * was interrupted, and current file was migrated
2107 mdt_object_put(env, stripe);
2109 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2114 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2116 stripe = mdt_object_find(env, info->mti_mdt, fid);
2118 return PTR_ERR(stripe);
2121 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2122 fid, &info->mti_spec);
2123 mdt_object_put(env, stripe);
2124 return rc ?: -EALREADY;
2126 mdt_object_put(env, stripe);
2131 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2137 mdt_object_get(env, stripe);
2142 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2143 if (IS_ERR(*sobj)) {
2144 mdt_object_put(env, stripe);
2145 rc = PTR_ERR(*sobj);
2153 /* end lease and close file for regular file */
2154 static int mdd_migrate_close(struct mdt_thread_info *info,
2155 struct mdt_object *obj)
2157 struct close_data *data;
2158 struct mdt_body *repbody;
2159 struct ldlm_lock *lease;
2164 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2166 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2170 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2175 lease = ldlm_handle2lock(&data->cd_handle);
2179 /* check if the lease was already canceled */
2180 lock_res_and_lock(lease);
2181 rc = ldlm_is_cancel(lease);
2182 unlock_res_and_lock(lease);
2186 LDLM_DEBUG(lease, DFID" lease broken",
2187 PFID(mdt_object_fid(obj)));
2191 * cancel server side lease, client side counterpart should have been
2192 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2194 ldlm_lock_cancel(lease);
2195 ldlm_reprocess_all(lease->l_resource,
2196 lease->l_policy_data.l_inodebits.bits);
2197 LDLM_LOCK_PUT(lease);
2200 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2201 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2202 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2208 * migrate file in below steps:
2209 * 1. lock parent and its stripes
2210 * 2. lookup source by name
2211 * 3. lock parents of source links if source is not directory
2212 * 4. reject if source is in HSM
2213 * 5. take source open_sem and close file if source is regular file
2214 * 6. lock source and its stripes if it's directory
2215 * 7. lock target so subsequent change to it can trigger COS
2217 * 9. unlock above locks
2218 * 10. sync device if source has links
2220 int mdt_reint_migrate(struct mdt_thread_info *info,
2221 struct mdt_lock_handle *unused)
2223 const struct lu_env *env = info->mti_env;
2224 struct mdt_device *mdt = info->mti_mdt;
2225 struct ptlrpc_request *req = mdt_info_req(info);
2226 struct mdt_reint_record *rr = &info->mti_rr;
2227 struct lu_ucred *uc = mdt_ucred(info);
2228 struct md_attr *ma = &info->mti_attr;
2229 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2230 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2231 struct mdt_object *pobj;
2232 struct mdt_object *spobj = NULL;
2233 struct mdt_object *sobj = NULL;
2234 struct mdt_object *tobj;
2235 struct lustre_handle rename_lh = { 0 };
2236 struct mdt_lock_handle *lhp;
2237 struct mdt_lock_handle *lhs;
2238 struct mdt_lock_handle *lht;
2239 LIST_HEAD(parent_slave_locks);
2240 LIST_HEAD(child_slave_locks);
2241 LIST_HEAD(link_locks);
2242 int lock_retries = 5;
2243 bool open_sem_locked = false;
2244 bool do_sync = false;
2248 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2249 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2251 if (info->mti_dlm_req)
2252 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2254 if (!fid_is_md_operative(rr->rr_fid1) ||
2255 !fid_is_md_operative(rr->rr_fid2))
2258 /* don't allow migrate . or .. */
2259 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2262 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2265 if (uc && !cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2266 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2267 mdt->mdt_enable_remote_dir_gid != -1)
2271 * Note: do not enqueue rename lock for replay request, because
2272 * if other MDT holds rename lock, but being blocked to wait for
2273 * this MDT to finish its recovery, and the failover MDT can not
2274 * get rename lock, which will cause deadlock.
2276 * req is NULL if this is called by directory auto-split.
2278 if (req && !req_is_replay(req)) {
2279 rc = mdt_rename_lock(info, &rename_lh);
2281 CERROR("%s: can't lock FS for rename: rc = %d\n",
2282 mdt_obd_name(info->mti_mdt), rc);
2287 /* pobj is master object of parent */
2288 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2290 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2293 rc = mdt_version_get_check(info, pobj, 0);
2295 GOTO(put_parent, rc);
2298 if (!mdt_object_exists(pobj))
2299 GOTO(put_parent, rc = -ENOENT);
2301 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2302 GOTO(put_parent, rc = -ENOTDIR);
2304 rc = mdt_check_enc(info, pobj);
2306 GOTO(put_parent, rc);
2308 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2310 GOTO(put_parent, rc);
2313 /* lock parent object */
2314 lhp = &info->mti_lh[MDT_LH_PARENT];
2315 mdt_lock_reg_init(lhp, LCK_PW);
2316 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2317 &parent_slave_locks);
2319 GOTO(put_parent, rc);
2322 * spobj is the corresponding stripe against name if pobj is striped
2323 * directory, which is the real parent, and no need to lock, because
2324 * we've taken full lock of pobj.
2326 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2328 GOTO(unlock_parent, rc);
2330 /* lock parents of source links, and revoke LOOKUP lock of links */
2331 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2332 &parent_slave_locks, &link_locks);
2333 if (rc == -EBUSY && lock_retries-- > 0) {
2334 mdt_object_put(env, sobj);
2335 mdt_object_put(env, spobj);
2340 GOTO(put_source, rc);
2343 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2344 * with one request, if total lock count exceeds this limit, we will
2345 * drop all locks after migration, and synchronous device in the end.
2349 /* TODO: DoM migration is not supported, migrate dirent only */
2350 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2351 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2353 GOTO(unlock_links, rc);
2355 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2356 info->mti_spec.sp_migrate_nsonly = 1;
2357 } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2358 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2360 GOTO(unlock_links, rc);
2362 /* race with restripe/auto-split? */
2363 if ((ma->ma_valid & MA_LMV) &&
2364 lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2365 GOTO(unlock_links, rc = -EBUSY);
2368 /* if migration HSM is allowed */
2369 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2370 ma->ma_need = MA_HSM;
2372 rc = mdt_attr_get_complex(info, sobj, ma);
2374 GOTO(unlock_links, rc);
2376 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2377 GOTO(unlock_links, rc = -EOPNOTSUPP);
2380 /* end lease and close file for regular file */
2381 if (info->mti_spec.sp_migrate_close) {
2382 /* try to hold open_sem so that nobody else can open the file */
2383 if (!down_write_trylock(&sobj->mot_open_sem)) {
2385 mdd_migrate_close(info, sobj);
2386 GOTO(unlock_links, rc = -EBUSY);
2388 open_sem_locked = true;
2389 rc = mdd_migrate_close(info, sobj);
2391 GOTO(unlock_open_sem, rc);
2396 lhs = &info->mti_lh[MDT_LH_OLD];
2397 mdt_lock_reg_init(lhs, LCK_EX);
2398 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2399 &child_slave_locks);
2401 GOTO(unlock_open_sem, rc);
2404 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2406 GOTO(unlock_source, rc = PTR_ERR(tobj));
2408 lht = &info->mti_lh[MDT_LH_NEW];
2409 mdt_lock_reg_init(lht, LCK_EX);
2410 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2412 GOTO(put_target, rc);
2414 /* Don't do lookup sanity check. We know name doesn't exist. */
2415 info->mti_spec.sp_cr_lookup = 0;
2416 info->mti_spec.sp_feat = &dt_directory_features;
2418 rc = mdo_migrate(env, mdt_object_child(pobj),
2419 mdt_object_child(sobj), &rr->rr_name,
2420 mdt_object_child(tobj),
2421 &info->mti_spec, ma);
2423 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2424 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2427 mdt_object_unlock(info, tobj, lht, rc);
2429 mdt_object_put(env, tobj);
2431 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2432 &child_slave_locks, rc);
2434 if (open_sem_locked)
2435 up_write(&sobj->mot_open_sem);
2437 /* if we've got too many locks to save into RPC,
2438 * then just commit before the locks are released
2441 mdt_device_sync(env, mdt);
2442 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2444 mdt_object_put(env, sobj);
2445 mdt_object_put(env, spobj);
2447 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2448 &parent_slave_locks, rc);
2450 mdt_object_put(env, pobj);
2452 if (lustre_handle_is_used(&rename_lh))
2453 mdt_rename_unlock(&rename_lh);
2458 static int mdt_object_lock_save(struct mdt_thread_info *info,
2459 struct mdt_object *dir,
2460 struct mdt_lock_handle *lh,
2461 int idx, bool cos_incompat)
2465 /* we lock the target dir if it is local */
2466 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2471 /* get and save correct version after locking */
2472 mdt_version_get_save(info, dir, idx);
2477 * determine lock order of sobj and tobj
2479 * there are two situations we need to lock tobj before sobj:
2480 * 1. sobj is child of tobj
2481 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2482 * larger than that of tobj
2484 * \retval 1 lock tobj before sobj
2485 * \retval 0 lock sobj before tobj
2486 * \retval -ev negative errno upon error
2488 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2489 struct mdt_object *sobj,
2490 struct mdt_object *tobj)
2492 struct md_attr *ma = &info->mti_attr;
2493 struct lu_fid *spfid = &info->mti_tmp_fid1;
2494 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2495 struct lmv_mds_md_v1 *lmv;
2500 /* sobj and tobj are the same */
2504 if (fid_is_root(mdt_object_fid(sobj)))
2507 if (fid_is_root(mdt_object_fid(tobj)))
2510 /* check whether sobj is child of tobj */
2511 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2512 mdt_object_fid(tobj));
2519 /* check whether sobj and tobj are children of the same parent */
2520 rc = mdt_attr_get_pfid(info, sobj, spfid);
2524 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2528 if (!lu_fid_eq(spfid, tpfid))
2531 /* check whether sobj and tobj are sibling stripes */
2532 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2536 if (!(ma->ma_valid & MA_LMV))
2539 lmv = &ma->ma_lmv->lmv_md_v1;
2540 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2542 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2545 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2549 if (!(ma->ma_valid & MA_LMV))
2552 lmv = &ma->ma_lmv->lmv_md_v1;
2553 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2555 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2557 /* check stripe index of sobj and tobj */
2558 if (sindex == tindex)
2561 return sindex < tindex ? 0 : 1;
2565 * lock rename source object.
2567 * Both source and source parent may be remote, and source may be a remote
2568 * object on source parent, to avoid overriding lock handle, store remote
2569 * LOOKUP lock separately in @lhr.
2571 * \retval 0 on success
2572 * \retval -ev negative errno upon error
2574 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2575 struct mdt_object *parent,
2576 struct mdt_object *child,
2577 struct mdt_lock_handle *lhc,
2578 struct mdt_lock_handle *lhr,
2584 rc = mdt_is_remote_object(info, parent, child);
2589 /* enqueue remote LOOKUP lock from the parent MDT */
2590 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2592 if (mdt_object_remote(parent)) {
2593 rc = mdt_remote_object_lock(info, parent,
2594 mdt_object_fid(child),
2601 LASSERT(mdt_object_remote(child));
2602 rc = mdt_object_local_lock(info, child, lhr,
2603 &rmt_ibits, 0, true);
2608 ibits &= ~MDS_INODELOCK_LOOKUP;
2611 if (mdt_object_remote(child)) {
2612 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2619 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2624 mdt_object_unlock(info, child, lhr, rc);
2629 /* Helper function for mdt_reint_rename so we don't need to opencode
2630 * two different order lockings
2632 static int mdt_lock_two_dirs(struct mdt_thread_info *info,
2633 struct mdt_object *mfirstdir,
2634 struct mdt_lock_handle *lh_firstdirp,
2635 struct mdt_object *mseconddir,
2636 struct mdt_lock_handle *lh_seconddirp,
2641 rc = mdt_object_lock_save(info, mfirstdir, lh_firstdirp, 0,
2646 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2648 if (mfirstdir != mseconddir) {
2649 rc = mdt_object_lock_save(info, mseconddir, lh_seconddirp, 1,
2651 } else if (!mdt_object_remote(mseconddir) &&
2652 lh_firstdirp->mlh_pdo_hash !=
2653 lh_seconddirp->mlh_pdo_hash) {
2654 rc = mdt_pdir_hash_lock(info, lh_seconddirp, mseconddir,
2655 MDS_INODELOCK_UPDATE,
2657 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2661 mdt_object_unlock(info, mfirstdir, lh_firstdirp, rc);
2667 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2668 * 2 - srcdir child; 3 - tgtdir child.
2669 * Update on disk version of srcdir child.
2671 static int mdt_reint_rename(struct mdt_thread_info *info,
2672 struct mdt_lock_handle *unused)
2674 struct mdt_device *mdt = info->mti_mdt;
2675 struct mdt_reint_record *rr = &info->mti_rr;
2676 struct md_attr *ma = &info->mti_attr;
2677 struct ptlrpc_request *req = mdt_info_req(info);
2678 struct mdt_object *msrcdir = NULL;
2679 struct mdt_object *mtgtdir = NULL;
2680 struct mdt_object *mold;
2681 struct mdt_object *mnew = NULL;
2682 struct lustre_handle rename_lh = { 0 };
2683 struct mdt_lock_handle *lh_srcdirp;
2684 struct mdt_lock_handle *lh_tgtdirp;
2685 struct mdt_lock_handle *lh_oldp = NULL;
2686 struct mdt_lock_handle *lh_rmt = NULL;
2687 struct mdt_lock_handle *lh_newp = NULL;
2688 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2689 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2691 bool reverse = false, discard = false;
2693 ktime_t kstart = ktime_get();
2694 enum mdt_stat_idx msi = 0;
2698 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2699 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2700 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2702 if (info->mti_dlm_req)
2703 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2705 if (!fid_is_md_operative(rr->rr_fid1) ||
2706 !fid_is_md_operative(rr->rr_fid2))
2709 /* find both parents. */
2710 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2711 if (IS_ERR(msrcdir))
2712 RETURN(PTR_ERR(msrcdir));
2714 rc = mdt_check_enc(info, msrcdir);
2716 GOTO(out_put_srcdir, rc);
2718 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2720 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2722 mdt_object_get(info->mti_env, mtgtdir);
2724 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2725 if (IS_ERR(mtgtdir))
2726 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2729 rc = mdt_check_enc(info, mtgtdir);
2731 GOTO(out_put_tgtdir, rc);
2734 * Note: do not enqueue rename lock for replay request, because
2735 * if other MDT holds rename lock, but being blocked to wait for
2736 * this MDT to finish its recovery, and the failover MDT can not
2737 * get rename lock, which will cause deadlock.
2739 if (!req_is_replay(req)) {
2740 bool remote = mdt_object_remote(msrcdir);
2743 * Normally rename RPC is handled on the MDT with the target
2744 * directory (if target exists, it's on the MDT with the
2745 * target), if the source directory is remote, it's a hint that
2746 * source is remote too (this may not be true, but it won't
2747 * cause any issue), return -EXDEV early to avoid taking
2750 if (!mdt->mdt_enable_remote_rename && remote)
2751 GOTO(out_put_tgtdir, rc = -EXDEV);
2753 /* This might be further relaxed in the future for regular file
2754 * renames in different source and target parents. Start with
2755 * only same-directory renames for simplicity and because this
2756 * is by far the most the common use case.
2758 * Striped directories should be considered "remote".
2760 if (msrcdir != mtgtdir || remote ||
2761 (S_ISDIR(ma->ma_attr.la_mode) &&
2762 !mdt->mdt_enable_parallel_rename_dir) ||
2763 (!S_ISDIR(ma->ma_attr.la_mode) &&
2764 !mdt->mdt_enable_parallel_rename_file)) {
2765 rc = mdt_rename_lock(info, &rename_lh);
2767 CERROR("%s: cannot lock for rename: rc = %d\n",
2768 mdt_obd_name(mdt), rc);
2769 GOTO(out_put_tgtdir, rc);
2772 if (S_ISDIR(ma->ma_attr.la_mode))
2773 msi = LPROC_MDT_RENAME_PAR_DIR;
2775 msi = LPROC_MDT_RENAME_PAR_FILE;
2778 "%s: samedir parallel rename "DFID"/"DNAME"\n",
2779 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2780 PNAME(&rr->rr_name));
2784 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2786 GOTO(out_unlock_rename, rc);
2789 /* source needs to be looked up after locking source parent, otherwise
2790 * this rename may race with unlink source, and cause rename hang, see
2791 * sanityn.sh 55b, so check parents first, if later we found source is
2792 * remote, relock parents.
2794 cos_incompat = (mdt_object_remote(msrcdir) ||
2795 mdt_object_remote(mtgtdir));
2797 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2799 /* lock parents in the proper order. */
2800 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2801 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2803 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2804 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2806 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2807 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2809 /* In case of same dir local rename we must sort by the hash,
2810 * otherwise a lock deadlock is possible when renaming
2811 * a to b and b to a at the same time LU-15285
2813 if (!mdt_object_remote(mtgtdir) && mtgtdir == msrcdir)
2814 reverse = lh_srcdirp->mlh_pdo_hash > lh_tgtdirp->mlh_pdo_hash;
2815 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)))
2819 rc = mdt_lock_two_dirs(info, mtgtdir, lh_tgtdirp, msrcdir,
2820 lh_srcdirp, cos_incompat);
2822 rc = mdt_lock_two_dirs(info, msrcdir, lh_srcdirp, mtgtdir,
2823 lh_tgtdirp, cos_incompat);
2826 GOTO(out_unlock_rename, rc);
2828 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2829 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2831 /* find mold object. */
2833 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2835 GOTO(out_unlock_parents, rc);
2837 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2838 GOTO(out_unlock_parents, rc = -EINVAL);
2840 if (!fid_is_md_operative(old_fid))
2841 GOTO(out_unlock_parents, rc = -EPERM);
2843 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2845 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2847 if (!mdt_object_exists(mold)) {
2848 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2850 "object does not exist");
2851 GOTO(out_put_old, rc = -ENOENT);
2854 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2855 GOTO(out_put_old, rc = -EXDEV);
2857 /* Check if @mtgtdir is subdir of @mold, before locking child
2858 * to avoid reverse locking.
2860 if (mtgtdir != msrcdir) {
2861 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2866 GOTO(out_put_old, rc);
2870 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2871 /* save version after locking */
2872 mdt_version_get_save(info, mold, 2);
2874 if (!cos_incompat && mdt_object_remote(mold)) {
2875 cos_incompat = true;
2876 mdt_object_put(info->mti_env, mold);
2877 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2878 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2882 /* find mnew object:
2883 * mnew target object may not exist now
2884 * lookup with version checking
2887 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2890 /* the new_fid should have been filled at this moment */
2891 if (lu_fid_eq(old_fid, new_fid))
2892 GOTO(out_put_old, rc);
2894 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2895 lu_fid_eq(new_fid, rr->rr_fid2))
2896 GOTO(out_put_old, rc = -EINVAL);
2898 if (!fid_is_md_operative(new_fid))
2899 GOTO(out_put_old, rc = -EPERM);
2901 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2903 GOTO(out_put_old, rc = PTR_ERR(mnew));
2905 if (!mdt_object_exists(mnew)) {
2906 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2908 "object does not exist");
2909 GOTO(out_put_new, rc = -ENOENT);
2912 if (mdt_object_remote(mnew)) {
2913 struct mdt_body *repbody;
2915 /* Always send rename req to the target child MDT */
2916 repbody = req_capsule_server_get(info->mti_pill,
2918 LASSERT(repbody != NULL);
2919 repbody->mbo_fid1 = *new_fid;
2920 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2921 GOTO(out_put_new, rc = -EXDEV);
2923 /* Before locking the target dir, check we do not replace
2924 * a dir with a non-dir, otherwise it may deadlock with
2925 * link op which tries to create a link in this dir
2926 * back to this non-dir.
2928 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2929 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2930 GOTO(out_put_new, rc = -EISDIR);
2932 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2933 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2934 mdt_lock_reg_init(lh_oldp, LCK_EX);
2935 mdt_lock_reg_init(lh_rmt, LCK_EX);
2936 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2937 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2938 lh_rmt, lock_ibits, cos_incompat);
2940 GOTO(out_put_new, rc);
2942 /* Check if @msrcdir is subdir of @mnew, before locking child
2943 * to avoid reverse locking.
2945 if (mtgtdir != msrcdir) {
2946 rc = mdo_is_subdir(info->mti_env,
2947 mdt_object_child(msrcdir), new_fid);
2951 GOTO(out_unlock_old, rc);
2955 /* We used to acquire MDS_INODELOCK_FULL here but we
2956 * can't do this now because a running HSM restore on
2957 * the rename onto victim will hold the layout
2958 * lock. See LU-4002.
2961 lh_newp = &info->mti_lh[MDT_LH_NEW];
2962 mdt_lock_reg_init(lh_newp, LCK_EX);
2963 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2964 if (mdt_object_remote(mtgtdir)) {
2965 rc = mdt_remote_object_lock(info, mtgtdir,
2966 mdt_object_fid(mnew),
2967 &lh_newp->mlh_rreg_lh,
2968 lh_newp->mlh_rreg_mode,
2969 MDS_INODELOCK_LOOKUP,
2972 GOTO(out_unlock_old, rc);
2974 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2976 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2979 GOTO(out_unlock_new, rc);
2981 /* get and save version after locking */
2982 mdt_version_get_save(info, mnew, 3);
2983 } else if (rc != -ENOENT) {
2984 GOTO(out_put_old, rc);
2986 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2987 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2988 mdt_lock_reg_init(lh_oldp, LCK_EX);
2989 mdt_lock_reg_init(lh_rmt, LCK_EX);
2990 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2991 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2992 lh_rmt, lock_ibits, cos_incompat);
2994 GOTO(out_put_old, rc);
2996 mdt_enoent_version_save(info, 3);
2999 /* step 5: rename it */
3000 mdt_reint_init_ma(info, ma);
3002 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
3003 OBD_FAIL_MDS_REINT_RENAME_WRITE);
3006 mutex_lock(&mnew->mot_lov_mutex);
3008 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
3009 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
3010 mnew != NULL ? mdt_object_child(mnew) : NULL,
3011 &rr->rr_tgt_name, ma);
3014 mutex_unlock(&mnew->mot_lov_mutex);
3016 /* handle last link of tgt object */
3019 mdt_handle_last_unlink(info, mnew, ma);
3020 discard = mdt_dom_check_for_discard(info, mnew);
3022 mdt_rename_counter_tally(info, info->mti_mdt, req,
3023 msrcdir, mtgtdir, msi,
3024 ktime_us_delta(ktime_get(), kstart));
3030 mdt_object_unlock(info, mnew, lh_newp, rc);
3032 mdt_object_unlock(info, NULL, lh_rmt, rc);
3033 mdt_object_unlock(info, mold, lh_oldp, rc);
3035 if (mnew && !discard)
3036 mdt_object_put(info->mti_env, mnew);
3038 mdt_object_put(info->mti_env, mold);
3040 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
3041 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
3043 if (lustre_handle_is_used(&rename_lh))
3044 mdt_rename_unlock(&rename_lh);
3046 mdt_object_put(info->mti_env, mtgtdir);
3048 mdt_object_put(info->mti_env, msrcdir);
3050 /* The DoM discard can be done right in the place above where it is
3051 * assigned, meanwhile it is done here after rename unlock due to
3052 * compatibility with old clients, for them the discard blocks
3053 * the main thread until completion. Check LU-11359 for details.
3056 mdt_dom_discard_data(info, mnew);
3057 mdt_object_put(info->mti_env, mnew);
3059 OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
3063 static int mdt_reint_resync(struct mdt_thread_info *info,
3064 struct mdt_lock_handle *lhc)
3066 struct mdt_reint_record *rr = &info->mti_rr;
3067 struct ptlrpc_request *req = mdt_info_req(info);
3068 struct md_attr *ma = &info->mti_attr;
3069 struct mdt_object *mo;
3070 struct ldlm_lock *lease;
3071 struct mdt_body *repbody;
3072 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
3077 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
3079 if (info->mti_dlm_req)
3080 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
3082 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
3084 GOTO(out, rc = PTR_ERR(mo));
3086 if (!mdt_object_exists(mo))
3087 GOTO(out_obj, rc = -ENOENT);
3089 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
3090 GOTO(out_obj, rc = -EINVAL);
3092 if (mdt_object_remote(mo))
3093 GOTO(out_obj, rc = -EREMOTE);
3095 lease = ldlm_handle2lock(rr->rr_lease_handle);
3097 GOTO(out_obj, rc = -ESTALE);
3099 /* It's really necessary to grab open_sem and check if the lease lock
3100 * has been lost. There would exist a concurrent writer coming in and
3101 * generating some dirty data in memory cache, the writeback would fail
3102 * after the layout version is increased by MDS_REINT_RESYNC RPC.
3104 if (!down_write_trylock(&mo->mot_open_sem))
3105 GOTO(out_put_lease, rc = -EBUSY);
3107 lock_res_and_lock(lease);
3108 lease_broken = ldlm_is_cancel(lease);
3109 unlock_res_and_lock(lease);
3111 GOTO(out_unlock, rc = -EBUSY);
3113 /* the file has yet opened by anyone else after we took the lease. */
3114 layout.mlc_opc = MD_LAYOUT_RESYNC;
3115 lhc = &info->mti_lh[MDT_LH_LOCAL];
3116 rc = mdt_layout_change(info, mo, lhc, &layout);
3118 GOTO(out_unlock, rc);
3120 mdt_object_unlock(info, mo, lhc, 0);
3122 ma->ma_need = MA_INODE;
3124 rc = mdt_attr_get_complex(info, mo, ma);
3126 GOTO(out_unlock, rc);
3128 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3129 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3133 up_write(&mo->mot_open_sem);
3135 LDLM_LOCK_PUT(lease);
3137 mdt_object_put(info->mti_env, mo);
3139 mdt_client_compatibility(info);
3143 struct mdt_reinter {
3144 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3145 enum lprocfs_extra_opc mr_extra_opc;
3148 static const struct mdt_reinter mdt_reinters[] = {
3150 .mr_handler = &mdt_reint_setattr,
3151 .mr_extra_opc = MDS_REINT_SETATTR,
3154 .mr_handler = &mdt_reint_create,
3155 .mr_extra_opc = MDS_REINT_CREATE,
3158 .mr_handler = &mdt_reint_link,
3159 .mr_extra_opc = MDS_REINT_LINK,
3162 .mr_handler = &mdt_reint_unlink,
3163 .mr_extra_opc = MDS_REINT_UNLINK,
3166 .mr_handler = &mdt_reint_rename,
3167 .mr_extra_opc = MDS_REINT_RENAME,
3170 .mr_handler = &mdt_reint_open,
3171 .mr_extra_opc = MDS_REINT_OPEN,
3173 [REINT_SETXATTR] = {
3174 .mr_handler = &mdt_reint_setxattr,
3175 .mr_extra_opc = MDS_REINT_SETXATTR,
3178 .mr_handler = &mdt_reint_unlink,
3179 .mr_extra_opc = MDS_REINT_UNLINK,
3182 .mr_handler = &mdt_reint_migrate,
3183 .mr_extra_opc = MDS_REINT_RENAME,
3186 .mr_handler = &mdt_reint_resync,
3187 .mr_extra_opc = MDS_REINT_RESYNC,
3191 int mdt_reint_rec(struct mdt_thread_info *info,
3192 struct mdt_lock_handle *lhc)
3194 const struct mdt_reinter *mr;
3198 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3201 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3202 if (mr->mr_handler == NULL)
3205 rc = (*mr->mr_handler)(info, lhc);
3207 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3208 PTLRPC_LAST_CNTR + mr->mr_extra_opc);