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 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 memset(policy, 0, sizeof(*policy));
296 policy->l_inodebits.bits = ibits;
298 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
303 struct mdt_object *o,
304 struct mdt_lock_handle *lh,
306 struct ldlm_enqueue_info *einfo,
311 LASSERT(!mdt_object_remote(o));
313 memset(einfo, 0, sizeof(*einfo));
315 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
319 rc = mdt_object_striped(info, o);
322 mdt_object_unlock(info, o, lh, rc);
326 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
328 mdt_object_unlock(info, o, lh, rc);
329 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
337 struct mdt_object *o,
338 struct mdt_lock_handle *lh,
339 struct ldlm_enqueue_info *einfo, int decref)
341 if (einfo->ei_cbdata)
342 mdt_unlock_slaves(info, o, einfo, decref);
343 mdt_object_unlock(info, o, lh, decref);
346 static int mdt_restripe(struct mdt_thread_info *info,
347 struct mdt_object *parent,
348 const struct lu_name *lname,
349 const struct lu_fid *tfid,
350 struct md_op_spec *spec,
353 struct mdt_device *mdt = info->mti_mdt;
354 struct lu_fid *fid = &info->mti_tmp_fid2;
355 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
356 struct lmv_mds_md_v1 *lmv;
357 struct mdt_object *child;
358 struct mdt_lock_handle *lhp;
359 struct mdt_lock_handle *lhc;
360 struct mdt_body *repbody;
364 if (!mdt->mdt_enable_dir_restripe)
367 rc = mdt_version_get_check_save(info, parent, 0);
371 lhp = &info->mti_lh[MDT_LH_PARENT];
372 mdt_lock_pdo_init(lhp, LCK_PW, lname);
373 rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
378 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
380 GOTO(unlock_parent, rc);
382 if (ma->ma_valid & MA_LMV) {
383 /* don't allow restripe if parent dir layout is changing */
384 lmv = &ma->ma_lmv->lmv_md_v1;
385 if (!lmv_is_sane(lmv))
386 GOTO(unlock_parent, rc = -EBADF);
388 if (lmv_is_layout_changing(lmv))
389 GOTO(unlock_parent, rc = -EBUSY);
393 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
395 GOTO(unlock_parent, rc);
397 child = mdt_object_find(info->mti_env, mdt, fid);
399 GOTO(unlock_parent, rc = PTR_ERR(child));
401 if (!mdt_object_exists(child))
402 GOTO(out_child, rc = -ENOENT);
404 if (mdt_object_remote(child)) {
405 struct mdt_body *repbody;
407 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
409 GOTO(out_child, rc = -EPROTO);
411 repbody->mbo_fid1 = *fid;
412 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
413 GOTO(out_child, rc = -EREMOTE);
417 lhc = &info->mti_lh[MDT_LH_CHILD];
418 mdt_lock_reg_init(lhc, LCK_EX);
420 /* enqueue object remote LOOKUP lock */
421 if (mdt_object_remote(parent)) {
422 rc = mdt_remote_object_lock(info, parent, fid,
425 MDS_INODELOCK_LOOKUP, false);
430 rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
433 GOTO(unlock_child, rc);
435 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
436 rc = mdt_version_get_check_save(info, child, 1);
438 GOTO(unlock_child, rc);
440 spin_lock(&mdt->mdt_restriper.mdr_lock);
441 if (child->mot_restriping) {
443 spin_unlock(&mdt->mdt_restriper.mdr_lock);
444 GOTO(unlock_child, rc = -EBUSY);
446 child->mot_restriping = 1;
447 spin_unlock(&mdt->mdt_restriper.mdr_lock);
450 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
452 GOTO(restriping_clear, rc);
454 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
456 GOTO(restriping_clear, rc = -EPROTO);
458 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
462 child->mot_restriping = 0;
464 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
466 mdt_object_put(info->mti_env, child);
468 mdt_object_unlock(info, parent, lhp, rc);
474 * VBR: we save three versions in reply:
475 * 0 - parent. Check that parent version is the same during replay.
476 * 1 - name. Version of 'name' if file exists with the same name or
477 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
478 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
481 static int mdt_create(struct mdt_thread_info *info)
483 struct mdt_device *mdt = info->mti_mdt;
484 struct mdt_object *parent;
485 struct mdt_object *child;
486 struct mdt_lock_handle *lh;
487 struct mdt_body *repbody;
488 struct md_attr *ma = &info->mti_attr;
489 struct mdt_reint_record *rr = &info->mti_rr;
490 struct md_op_spec *spec = &info->mti_spec;
491 bool restripe = false;
495 DEBUG_REQ(D_INODE, mdt_info_req(info),
496 "Create ("DNAME"->"DFID") in "DFID,
497 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
499 if (!fid_is_md_operative(rr->rr_fid1))
502 if (S_ISDIR(ma->ma_attr.la_mode) &&
503 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
504 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
505 struct lu_ucred *uc = mdt_ucred(info);
506 struct obd_export *exp = mdt_info_req(info)->rq_export;
508 /* Only new clients can create remote dir( >= 2.4) and
509 * striped dir(>= 2.6), old client will return -ENOTSUPP
511 if (!mdt_is_dne_client(exp))
514 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
515 if (!mdt_is_striped_client(exp))
518 if (!mdt->mdt_enable_striped_dir)
520 } else if (!mdt->mdt_enable_remote_dir) {
524 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
525 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
529 if (!md_capable(uc, CAP_SYS_ADMIN) &&
530 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
531 mdt->mdt_enable_remote_dir_gid != -1)
534 /* restripe if later found dir exists, MDS_OPEN_CREAT means
535 * this is create only, don't try restripe.
537 if (mdt->mdt_enable_dir_restripe &&
538 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
539 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
543 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
545 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
547 RETURN(PTR_ERR(parent));
549 if (!mdt_object_exists(parent))
550 GOTO(put_parent, rc = -ENOENT);
553 * LU-10235: check if name exists locklessly first to avoid massive
554 * lock recalls on existing directories.
556 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
557 &info->mti_tmp_fid1, 1);
560 GOTO(put_parent, rc = -EEXIST);
562 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
566 /* -ENOENT is expected here */
568 GOTO(put_parent, rc);
570 /* save version of file name for replay, it must be ENOENT here */
571 mdt_enoent_version_save(info, 1);
573 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
575 lh = &info->mti_lh[MDT_LH_PARENT];
576 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
577 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
579 GOTO(put_parent, rc);
581 if (!mdt_object_remote(parent)) {
582 rc = mdt_version_get_check_save(info, parent, 0);
584 GOTO(unlock_parent, rc);
587 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
588 if (unlikely(IS_ERR(child)))
589 GOTO(unlock_parent, rc = PTR_ERR(child));
591 ma->ma_need = MA_INODE;
594 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
595 OBD_FAIL_MDS_REINT_CREATE_WRITE);
597 /* Version of child will be updated on disk. */
598 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
599 rc = mdt_version_get_check_save(info, child, 2);
603 /* Let lower layer know current lock mode. */
604 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
607 * Do not perform lookup sanity check. We know that name does
610 info->mti_spec.sp_cr_lookup = 0;
611 info->mti_spec.sp_feat = &dt_directory_features;
613 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
614 mdt_object_child(child), &info->mti_spec, ma);
616 rc = mdt_attr_get_complex(info, child, ma);
622 * On DNE, we need to eliminate dependey between 'mkdir a' and
623 * 'mkdir a/b' if b is a striped directory, to achieve this, two
624 * things are done below:
625 * 1. save child and slaves lock.
626 * 2. if the child is a striped directory, relock parent so to
627 * compare against with COS locks to ensure parent was
630 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
631 struct mdt_lock_handle *lhc;
632 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
635 rc = mdt_object_striped(info, child);
641 if (!mdt_object_remote(parent)) {
642 mdt_object_unlock(info, parent, lh, 1);
643 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
644 rc = mdt_reint_object_lock(info, parent, lh,
645 MDS_INODELOCK_UPDATE,
652 lhc = &info->mti_lh[MDT_LH_CHILD];
653 mdt_lock_handle_init(lhc);
654 mdt_lock_reg_init(lhc, LCK_PW);
655 rc = mdt_reint_striped_lock(info, child, lhc,
656 MDS_INODELOCK_UPDATE, einfo,
661 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
664 /* Return fid & attr to client. */
665 if (ma->ma_valid & MA_INODE)
666 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
667 mdt_object_fid(child));
670 mdt_object_put(info->mti_env, child);
672 mdt_object_unlock(info, parent, lh, rc);
674 mdt_object_put(info->mti_env, parent);
678 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
681 struct mdt_lock_handle *lh;
682 int do_vbr = ma->ma_attr.la_valid &
683 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
684 __u64 lockpart = MDS_INODELOCK_UPDATE;
685 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
690 rc = mdt_object_striped(info, mo);
696 lh = &info->mti_lh[MDT_LH_PARENT];
697 mdt_lock_reg_init(lh, LCK_PW);
699 /* Even though the new MDT will grant PERM lock to the old
700 * client, but the old client will almost ignore that during
701 * So it needs to revoke both LOOKUP and PERM lock here, so
702 * both new and old client can cancel the dcache
704 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
705 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
707 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
712 /* all attrs are packed into mti_attr in unpack_setattr */
713 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
714 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
716 /* VBR: update version if attr changed are important for recovery */
718 /* update on-disk version of changed object */
719 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
720 rc = mdt_version_get_check_save(info, mo, 0);
722 GOTO(out_unlock, rc);
725 /* Ensure constant striping during chown(). See LU-2789. */
726 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
727 mutex_lock(&mo->mot_lov_mutex);
729 /* all attrs are packed into mti_attr in unpack_setattr */
730 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
732 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
733 mutex_unlock(&mo->mot_lov_mutex);
736 GOTO(out_unlock, rc);
737 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
740 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
745 * Check HSM flags and add HS_DIRTY flag if relevant.
747 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
748 * and is not RELEASED.
750 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
753 struct lu_ucred *uc = mdt_ucred(info);
758 /* If the file was modified, add the dirty flag */
759 ma->ma_need = MA_HSM;
760 rc = mdt_attr_get_complex(info, mo, ma);
762 CERROR("file attribute read error for "DFID": %d.\n",
763 PFID(mdt_object_fid(mo)), rc);
767 /* If an up2date copy exists in the backend, add dirty flag */
768 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
769 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
770 ma->ma_hsm.mh_flags |= HS_DIRTY;
772 /* Bump cap so that closes from non-owner writers can
773 * set the HSM state to dirty.
775 cap_saved = uc->uc_cap;
776 uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
777 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
778 uc->uc_cap = cap_saved;
780 CERROR("file attribute change error for "DFID": %d\n",
781 PFID(mdt_object_fid(mo)), rc);
787 static int mdt_reint_setattr(struct mdt_thread_info *info,
788 struct mdt_lock_handle *lhc)
790 struct mdt_device *mdt = info->mti_mdt;
791 struct md_attr *ma = &info->mti_attr;
792 struct mdt_reint_record *rr = &info->mti_rr;
793 struct ptlrpc_request *req = mdt_info_req(info);
794 struct mdt_object *mo;
795 struct mdt_body *repbody;
796 ktime_t kstart = ktime_get();
800 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
801 (unsigned int)ma->ma_attr.la_valid);
803 if (info->mti_dlm_req)
804 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
806 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
808 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
809 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
811 GOTO(out, rc = PTR_ERR(mo));
813 if (!mdt_object_exists(mo))
814 GOTO(out_put, rc = -ENOENT);
816 if (mdt_object_remote(mo))
817 GOTO(out_put, rc = -EREMOTE);
819 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
820 /* revoke lease lock if size is going to be changed */
821 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
822 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
823 atomic_read(&mo->mot_lease_count) > 0)) {
824 down_read(&mo->mot_open_sem);
826 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
827 lhc = &info->mti_lh[MDT_LH_LOCAL];
828 mdt_lock_reg_init(lhc, LCK_CW);
830 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
832 up_read(&mo->mot_open_sem);
836 /* revoke lease lock */
837 mdt_object_unlock(info, mo, lhc, 1);
839 up_read(&mo->mot_open_sem);
842 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
843 /* Check write access for the O_TRUNC case */
844 if (mdt_write_read(mo) < 0)
845 GOTO(out_put, rc = -ETXTBSY);
847 /* LU-10286: compatibility check for FLR.
848 * Please check the comment in mdt_finish_open() for details
850 if (!exp_connect_flr(info->mti_exp) ||
851 !exp_connect_overstriping(info->mti_exp)) {
852 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
853 if (rc < 0 && rc != -ENODATA)
856 if (!exp_connect_flr(info->mti_exp)) {
858 mdt_lmm_is_flr(info->mti_big_lmm))
859 GOTO(out_put, rc = -EOPNOTSUPP);
862 if (!exp_connect_overstriping(info->mti_exp)) {
864 mdt_lmm_is_overstriping(info->mti_big_lmm))
865 GOTO(out_put, rc = -EOPNOTSUPP);
869 /* For truncate, the file size sent from client
870 * is believable, but the blocks are incorrect,
871 * which makes the block size in LSOM attribute
872 * inconsisent with the real block size.
874 rc = mdt_lsom_update(info, mo, true);
879 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
880 if (ma->ma_valid & MA_LOV)
881 GOTO(out_put, rc = -EPROTO);
883 /* MDT supports FMD for regular files due to Data-on-MDT */
884 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
885 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
886 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
889 if (ma->ma_attr.la_valid & LA_MTIME) {
890 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
892 ma->ma_valid |= MA_PFID;
896 rc = mdt_attr_set(info, mo, ma);
899 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
900 (ma->ma_valid & MA_INODE)) {
901 struct lu_buf *buf = &info->mti_buf;
902 struct lu_ucred *uc = mdt_ucred(info);
903 struct mdt_lock_handle *lh;
905 __u64 lockpart = MDS_INODELOCK_XATTR;
907 /* reject if either remote or striped dir is disabled */
908 if (ma->ma_valid & MA_LMV) {
909 if (!mdt->mdt_enable_remote_dir ||
910 !mdt->mdt_enable_striped_dir)
911 GOTO(out_put, rc = -EPERM);
913 if (!md_capable(uc, CAP_SYS_ADMIN) &&
914 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
915 mdt->mdt_enable_remote_dir_gid != -1)
916 GOTO(out_put, rc = -EPERM);
919 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
920 GOTO(out_put, rc = -ENOTDIR);
922 if (ma->ma_attr.la_valid != 0)
923 GOTO(out_put, rc = -EPROTO);
925 if (ma->ma_valid & MA_LOV) {
926 buf->lb_buf = ma->ma_lmm;
927 buf->lb_len = ma->ma_lmm_size;
928 name = XATTR_NAME_LOV;
930 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
933 buf->lb_len = ma->ma_lmv_size;
934 name = XATTR_NAME_DEFAULT_LMV;
935 /* force client to update dir default layout */
936 lockpart |= MDS_INODELOCK_LOOKUP;
939 lh = &info->mti_lh[MDT_LH_PARENT];
940 mdt_lock_reg_init(lh, LCK_PW);
942 rc = mdt_object_lock(info, mo, lh, lockpart);
946 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
949 mdt_object_unlock(info, mo, lh, rc);
953 GOTO(out_put, rc = -EPROTO);
956 /* If file data is modified, add the dirty flag */
957 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
958 rc = mdt_add_dirty_flag(info, mo, ma);
960 ma->ma_need = MA_INODE;
962 rc = mdt_attr_get_complex(info, mo, ma);
966 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
970 mdt_object_put(info->mti_env, mo);
973 mdt_counter_incr(req, LPROC_MDT_SETATTR,
974 ktime_us_delta(ktime_get(), kstart));
976 mdt_client_compatibility(info);
977 rc2 = mdt_fix_reply(info);
983 static int mdt_reint_create(struct mdt_thread_info *info,
984 struct mdt_lock_handle *lhc)
986 struct ptlrpc_request *req = mdt_info_req(info);
987 ktime_t kstart = ktime_get();
991 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
992 RETURN(err_serious(-ESTALE));
994 if (info->mti_dlm_req)
995 ldlm_request_cancel(mdt_info_req(info),
996 info->mti_dlm_req, 0, LATF_SKIP);
998 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1001 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1011 CERROR("%s: Unsupported mode %o\n",
1012 mdt_obd_name(info->mti_mdt),
1013 info->mti_attr.ma_attr.la_mode);
1014 RETURN(err_serious(-EOPNOTSUPP));
1017 rc = mdt_create(info);
1019 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1020 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1021 ktime_us_delta(ktime_get(), kstart));
1023 /* Special file should stay on the same node as parent*/
1024 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1025 ktime_us_delta(ktime_get(), kstart));
1032 * VBR: save parent version in reply and child version getting by its name.
1033 * Version of child is getting and checking during its lookup. If
1035 static int mdt_reint_unlink(struct mdt_thread_info *info,
1036 struct mdt_lock_handle *lhc)
1038 struct mdt_reint_record *rr = &info->mti_rr;
1039 struct ptlrpc_request *req = mdt_info_req(info);
1040 struct md_attr *ma = &info->mti_attr;
1041 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1042 struct mdt_object *mp;
1043 struct mdt_object *mc;
1044 struct mdt_lock_handle *parent_lh;
1045 struct mdt_lock_handle *child_lh;
1046 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1048 bool cos_incompat = false;
1050 ktime_t kstart = ktime_get();
1054 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1055 PNAME(&rr->rr_name));
1057 if (info->mti_dlm_req)
1058 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1060 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1061 RETURN(err_serious(-ENOENT));
1063 if (!fid_is_md_operative(rr->rr_fid1))
1066 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1068 RETURN(PTR_ERR(mp));
1070 if (mdt_object_remote(mp)) {
1071 cos_incompat = true;
1073 rc = mdt_version_get_check_save(info, mp, 0);
1075 GOTO(put_parent, rc);
1078 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1079 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1081 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1082 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1083 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1086 GOTO(put_parent, rc);
1088 /* lookup child object along with version checking */
1089 fid_zero(child_fid);
1090 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1092 /* Name might not be able to find during resend of
1093 * remote unlink, considering following case.
1094 * dir_A is a remote directory, the name entry of
1095 * dir_A is on MDT0, the directory is on MDT1,
1097 * 1. client sends unlink req to MDT1.
1098 * 2. MDT1 sends name delete update to MDT0.
1099 * 3. name entry is being deleted in MDT0 synchronously.
1100 * 4. MDT1 is restarted.
1101 * 5. client resends unlink req to MDT1. So it can not
1102 * find the name entry on MDT0 anymore.
1103 * In this case, MDT1 only needs to destory the local
1106 if (mdt_object_remote(mp) && rc == -ENOENT &&
1107 !fid_is_zero(rr->rr_fid2) &&
1108 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1110 *child_fid = *rr->rr_fid2;
1112 GOTO(unlock_parent, rc);
1116 if (!fid_is_md_operative(child_fid))
1117 GOTO(unlock_parent, rc = -EPERM);
1119 /* We will lock the child regardless it is local or remote. No harm. */
1120 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1122 GOTO(unlock_parent, rc = PTR_ERR(mc));
1124 if (!cos_incompat) {
1125 rc = mdt_object_striped(info, mc);
1127 GOTO(put_child, rc);
1131 mdt_object_put(info->mti_env, mc);
1132 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1137 child_lh = &info->mti_lh[MDT_LH_CHILD];
1138 mdt_lock_reg_init(child_lh, LCK_EX);
1139 if (info->mti_spec.sp_rm_entry) {
1140 struct lu_ucred *uc = mdt_ucred(info);
1142 if (!mdt_is_dne_client(req->rq_export))
1143 /* Return -ENOTSUPP for old client */
1144 GOTO(put_child, rc = -ENOTSUPP);
1146 if (!md_capable(uc, CAP_SYS_ADMIN))
1147 GOTO(put_child, rc = -EPERM);
1149 ma->ma_need = MA_INODE;
1151 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1152 NULL, &rr->rr_name, ma, no_name);
1153 GOTO(put_child, rc);
1156 if (mdt_object_remote(mc)) {
1157 struct mdt_body *repbody;
1159 if (!fid_is_zero(rr->rr_fid2)) {
1160 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1161 mdt_obd_name(info->mti_mdt),
1162 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1163 GOTO(put_child, rc = -ENOENT);
1165 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1166 mdt_obd_name(info->mti_mdt),
1167 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1169 if (!mdt_is_dne_client(req->rq_export))
1170 /* Return -ENOTSUPP for old client */
1171 GOTO(put_child, rc = -ENOTSUPP);
1173 /* Revoke the LOOKUP lock of the remote object granted by
1174 * this MDT. Since the unlink will happen on another MDT,
1175 * it will release the LOOKUP lock right away. Then What
1176 * would happen if another client try to grab the LOOKUP
1177 * lock at the same time with unlink XXX
1179 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1180 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1181 LASSERT(repbody != NULL);
1182 repbody->mbo_fid1 = *mdt_object_fid(mc);
1183 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1184 GOTO(unlock_child, rc = -EREMOTE);
1186 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1187 * this now because a running HSM restore on the child (unlink
1188 * victim) will hold the layout lock. See LU-4002.
1190 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1191 if (mdt_object_remote(mp)) {
1192 /* Enqueue lookup lock from parent MDT */
1193 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1194 &child_lh->mlh_rreg_lh,
1195 child_lh->mlh_rreg_mode,
1196 MDS_INODELOCK_LOOKUP, false);
1198 GOTO(put_child, rc);
1200 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1203 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1206 GOTO(put_child, rc);
1209 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1210 * whether need MA_LOV and MA_COOKIE.
1212 ma->ma_need = MA_INODE;
1215 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1216 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1217 /* save version when object is locked */
1218 mdt_version_get_save(info, mc, 1);
1220 mutex_lock(&mc->mot_lov_mutex);
1222 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1223 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1225 mutex_unlock(&mc->mot_lov_mutex);
1227 GOTO(unlock_child, rc);
1229 if (!lu_object_is_dying(&mc->mot_header)) {
1230 rc = mdt_attr_get_complex(info, mc, ma);
1233 } else if (mdt_dom_check_for_discard(info, mc)) {
1234 mdt_dom_discard_data(info, mc);
1236 mdt_handle_last_unlink(info, mc, ma);
1239 if (ma->ma_valid & MA_INODE) {
1240 switch (ma->ma_attr.la_mode & S_IFMT) {
1242 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1243 ktime_us_delta(ktime_get(), kstart));
1251 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1252 ktime_us_delta(ktime_get(), kstart));
1255 LASSERTF(0, "bad file type %o unlinking\n",
1256 ma->ma_attr.la_mode);
1263 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1265 mdt_object_put(info->mti_env, mc);
1267 mdt_object_unlock(info, mp, parent_lh, rc);
1269 mdt_object_put(info->mti_env, mp);
1270 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1275 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1278 static int mdt_reint_link(struct mdt_thread_info *info,
1279 struct mdt_lock_handle *lhc)
1281 struct mdt_reint_record *rr = &info->mti_rr;
1282 struct ptlrpc_request *req = mdt_info_req(info);
1283 struct md_attr *ma = &info->mti_attr;
1284 struct mdt_object *ms;
1285 struct mdt_object *mp;
1286 struct mdt_lock_handle *lhs;
1287 struct mdt_lock_handle *lhp;
1288 ktime_t kstart = ktime_get();
1293 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1294 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1296 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1297 RETURN(err_serious(-ENOENT));
1299 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1300 req->rq_no_reply = 1;
1301 RETURN(err_serious(-ENOENT));
1304 if (info->mti_dlm_req)
1305 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1307 /* Invalid case so return error immediately instead of
1310 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1313 if (!fid_is_md_operative(rr->rr_fid1) ||
1314 !fid_is_md_operative(rr->rr_fid2))
1317 /* step 1: find target parent dir */
1318 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1320 RETURN(PTR_ERR(mp));
1322 rc = mdt_version_get_check_save(info, mp, 0);
1324 GOTO(put_parent, rc);
1326 /* step 2: find source */
1327 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1329 GOTO(put_parent, rc = PTR_ERR(ms));
1331 if (!mdt_object_exists(ms)) {
1332 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1333 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1334 GOTO(put_source, rc = -ENOENT);
1337 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1339 lhp = &info->mti_lh[MDT_LH_PARENT];
1340 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1341 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1344 GOTO(put_source, rc);
1346 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1348 lhs = &info->mti_lh[MDT_LH_CHILD];
1349 mdt_lock_reg_init(lhs, LCK_EX);
1350 rc = mdt_reint_object_lock(info, ms, lhs,
1351 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1354 GOTO(unlock_parent, rc);
1356 /* step 3: link it */
1357 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1358 OBD_FAIL_MDS_REINT_LINK_WRITE);
1360 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1361 rc = mdt_version_get_check_save(info, ms, 1);
1363 GOTO(unlock_source, rc);
1365 /** check target version by name during replay */
1366 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1367 &info->mti_tmp_fid1, 2);
1368 if (rc != 0 && rc != -ENOENT)
1369 GOTO(unlock_source, rc);
1370 /* save version of file name for replay, it must be ENOENT here */
1371 if (!req_is_replay(mdt_info_req(info))) {
1372 if (rc != -ENOENT) {
1373 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1374 PNAME(&rr->rr_name));
1375 GOTO(unlock_source, rc = -EEXIST);
1377 info->mti_ver[2] = ENOENT_VERSION;
1378 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1381 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1382 mdt_object_child(ms), &rr->rr_name, ma);
1385 mdt_counter_incr(req, LPROC_MDT_LINK,
1386 ktime_us_delta(ktime_get(), kstart));
1390 mdt_object_unlock(info, ms, lhs, rc);
1392 mdt_object_unlock(info, mp, lhp, rc);
1394 mdt_object_put(info->mti_env, ms);
1396 mdt_object_put(info->mti_env, mp);
1400 * lock the part of the directory according to the hash of the name
1401 * (lh->mlh_pdo_hash) in parallel directory lock.
1403 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1404 struct mdt_lock_handle *lh,
1405 struct mdt_object *obj, __u64 ibits,
1408 struct ldlm_res_id *res = &info->mti_res_id;
1409 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1410 union ldlm_policy_data *policy = &info->mti_policy;
1411 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1415 * Finish res_id initializing by name hash marking part of
1416 * directory which is taking modification.
1418 LASSERT(lh->mlh_pdo_hash != 0);
1419 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1420 memset(policy, 0, sizeof(*policy));
1421 policy->l_inodebits.bits = ibits;
1423 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1424 dlmflags |= LDLM_FL_COS_INCOMPAT;
1426 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1427 * going to be sent to client. If it is - mdt_intent_policy() path will
1428 * fix it up and turn FL_LOCAL flag off.
1430 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1431 policy, res, dlmflags,
1432 &info->mti_exp->exp_handle.h_cookie);
1437 * Get BFL lock for rename or migrate process.
1439 static int mdt_rename_lock(struct mdt_thread_info *info,
1440 struct lustre_handle *lh)
1445 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1446 struct lu_fid *fid = &info->mti_tmp_fid1;
1447 struct mdt_object *obj;
1449 /* XXX, right now, it has to use object API to
1450 * enqueue lock cross MDT, so it will enqueue
1451 * rename lock(with LUSTRE_BFL_FID) by root object
1454 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1456 RETURN(PTR_ERR(obj));
1458 rc = mdt_remote_object_lock(info, obj,
1459 &LUSTRE_BFL_FID, lh,
1461 MDS_INODELOCK_UPDATE, false);
1462 mdt_object_put(info->mti_env, obj);
1464 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1465 union ldlm_policy_data *policy = &info->mti_policy;
1466 struct ldlm_res_id *res_id = &info->mti_res_id;
1469 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1470 memset(policy, 0, sizeof(*policy));
1471 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1472 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1473 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1474 LDLM_IBITS, policy, LCK_EX, &flags,
1476 ldlm_completion_ast, NULL, NULL, 0,
1478 &info->mti_exp->exp_handle.h_cookie,
1485 static void mdt_rename_unlock(struct lustre_handle *lh)
1488 LASSERT(lustre_handle_is_used(lh));
1489 /* Cancel the single rename lock right away */
1490 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1494 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1495 const struct lu_fid *fid,
1498 struct mdt_object *dir;
1502 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1506 /* check early, the real version will be saved after locking */
1507 rc = mdt_version_get_check(info, dir, idx);
1511 if (!mdt_object_exists(dir))
1512 GOTO(out_put, rc = -ENOENT);
1514 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1515 GOTO(out_put, rc = -ENOTDIR);
1519 mdt_object_put(info->mti_env, dir);
1524 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1525 * herein we don't really check it, just do revoke.
1527 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1528 struct mdt_object *pobj,
1529 struct mdt_object *obj)
1531 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1534 mdt_lock_handle_init(lh);
1535 mdt_lock_reg_init(lh, LCK_EX);
1537 if (mdt_object_remote(pobj)) {
1538 /* don't bother to check if pobj and obj are on the same MDT. */
1539 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1540 &lh->mlh_rreg_lh, LCK_EX,
1541 MDS_INODELOCK_LOOKUP, false);
1542 } else if (mdt_object_remote(obj)) {
1543 struct ldlm_res_id *res = &info->mti_res_id;
1544 union ldlm_policy_data *policy = &info->mti_policy;
1545 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1546 LDLM_FL_COS_INCOMPAT;
1548 fid_build_reg_res_name(mdt_object_fid(obj), res);
1549 memset(policy, 0, sizeof(*policy));
1550 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1551 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1552 &lh->mlh_reg_lh, LCK_EX, policy, res,
1555 /* do nothing if both are local */
1563 * TODO, currently we don't save this lock because there is no place to
1564 * hold this lock handle, but to avoid race we need to save this lock.
1566 mdt_object_unlock(info, NULL, lh, 1);
1572 * operation may takes locks of linkea, or directory stripes, group them in
1575 struct mdt_sub_lock {
1576 struct mdt_object *msl_obj;
1577 struct mdt_lock_handle msl_lh;
1578 struct list_head msl_linkage;
1581 static void mdt_unlock_list(struct mdt_thread_info *info,
1582 struct list_head *list, int decref)
1584 struct mdt_sub_lock *msl;
1585 struct mdt_sub_lock *tmp;
1587 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1588 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1589 list_del(&msl->msl_linkage);
1594 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1595 struct mdt_object *obj,
1596 struct mdt_lock_handle *lh,
1597 struct ldlm_enqueue_info *einfo,
1598 struct list_head *slave_locks,
1601 if (mdt_object_remote(obj)) {
1602 mdt_unlock_list(info, slave_locks, decref);
1603 mdt_object_unlock(info, obj, lh, decref);
1605 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1610 * lock parents of links, and also check whether total locks don't exceed
1613 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1614 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1615 * \retval -ev negative errno upon error
1617 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1618 struct mdt_object *pobj,
1619 const struct md_attr *ma,
1620 struct mdt_object *obj,
1621 struct mdt_lock_handle *lhp,
1622 struct ldlm_enqueue_info *peinfo,
1623 struct list_head *parent_slave_locks,
1624 struct list_head *link_locks)
1626 struct mdt_device *mdt = info->mti_mdt;
1627 struct lu_buf *buf = &info->mti_big_buf;
1628 struct lu_name *lname = &info->mti_name;
1629 struct linkea_data ldata = { NULL };
1630 bool blocked = false;
1631 int local_lnkp_cnt = 0;
1635 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1638 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1639 if (buf->lb_buf == NULL)
1643 rc = mdt_links_read(info, obj, &ldata);
1645 if (rc == -ENOENT || rc == -ENODATA)
1650 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1651 linkea_next_entry(&ldata)) {
1652 struct mdt_object *lnkp;
1653 struct mdt_sub_lock *msl;
1657 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1660 /* check if it's also linked to parent */
1661 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1662 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1663 PFID(&fid), PNAME(lname));
1664 /* in case link is remote object, revoke LOOKUP lock */
1665 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1671 /* check if it's linked to a stripe of parent */
1672 if (ma->ma_valid & MA_LMV) {
1673 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1674 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1677 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1678 fid_le_to_cpu(stripe_fid,
1679 &lmv->lmv_stripe_fids[j]);
1680 if (lu_fid_eq(stripe_fid, &fid)) {
1681 CDEBUG(D_INFO, "skip stripe "DFID
1682 ", reovke "DNAME"\n",
1683 PFID(&fid), PNAME(lname));
1684 lnkp = mdt_object_find(info->mti_env,
1687 GOTO(out, rc = PTR_ERR(lnkp));
1693 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1695 mdt_object_put(info->mti_env, lnkp);
1700 /* Check if it's already locked */
1701 list_for_each_entry(msl, link_locks, msl_linkage) {
1702 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1704 DFID" was locked, revoke "DNAME"\n",
1705 PFID(&fid), PNAME(lname));
1706 lnkp = msl->msl_obj;
1712 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1716 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1717 PFID(&fid), PNAME(lname));
1719 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1721 CWARN("%s: cannot find obj "DFID": %ld\n",
1722 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1726 if (!mdt_object_exists(lnkp)) {
1727 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1728 PFID(&fid), PNAME(lname));
1729 mdt_object_put(info->mti_env, lnkp);
1733 if (!mdt_object_remote(lnkp))
1738 GOTO(out, rc = -ENOMEM);
1741 * we can't follow parent-child lock order like other MD
1742 * operations, use lock_try here to avoid deadlock, if the lock
1743 * cannot be taken, drop all locks taken, revoke the blocked
1744 * one, and continue processing the remaining entries, and in
1745 * the end of the loop restart from beginning.
1747 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1749 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1750 MDS_INODELOCK_UPDATE, true);
1751 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1753 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1754 PFID(&fid), PNAME(lname));
1756 mdt_unlock_list(info, link_locks, 1);
1757 /* also unlock parent locks to avoid deadlock */
1759 mdt_migrate_object_unlock(info, pobj, lhp,
1766 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1767 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1768 MDS_INODELOCK_UPDATE);
1770 mdt_object_put(info->mti_env, lnkp);
1775 if (mdt_object_remote(lnkp)) {
1776 struct ldlm_lock *lock;
1779 * for remote object, set lock cb_atomic,
1780 * so lock can be released in blocking_ast()
1781 * immediately, then the next lock_try will
1782 * have better chance of success.
1784 lock = ldlm_handle2lock(
1785 &msl->msl_lh.mlh_rreg_lh);
1786 LASSERT(lock != NULL);
1787 lock_res_and_lock(lock);
1788 ldlm_set_atomic_cb(lock);
1789 unlock_res_and_lock(lock);
1790 LDLM_LOCK_PUT(lock);
1793 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1798 INIT_LIST_HEAD(&msl->msl_linkage);
1799 msl->msl_obj = lnkp;
1800 list_add_tail(&msl->msl_linkage, link_locks);
1802 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1806 GOTO(out, rc = -EBUSY);
1811 mdt_unlock_list(info, link_locks, rc);
1812 } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1813 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1816 * parent may have 3 local objects: master object and 2 stripes
1817 * (if it's being migrated too); source may have 1 local objects
1818 * as regular file; target has 1 local object.
1819 * Note, source may have 2 local locks if it is directory but it
1820 * can't have hardlinks, so it is not considered here.
1827 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1828 struct mdt_object *obj,
1829 const struct md_attr *ma,
1830 struct list_head *slave_locks)
1832 struct mdt_device *mdt = info->mti_mdt;
1833 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1834 struct lu_fid *fid = &info->mti_tmp_fid1;
1835 struct mdt_object *slave;
1836 struct mdt_sub_lock *msl;
1841 LASSERT(mdt_object_remote(obj));
1842 LASSERT(ma->ma_valid & MA_LMV);
1845 if (!lmv_is_sane(lmv))
1848 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1849 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1851 if (!fid_is_sane(fid))
1854 slave = mdt_object_find(info->mti_env, mdt, fid);
1856 GOTO(out, rc = PTR_ERR(slave));
1860 mdt_object_put(info->mti_env, slave);
1861 GOTO(out, rc = -ENOMEM);
1864 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1865 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1866 MDS_INODELOCK_UPDATE, true);
1869 mdt_object_put(info->mti_env, slave);
1873 INIT_LIST_HEAD(&msl->msl_linkage);
1874 msl->msl_obj = slave;
1875 list_add_tail(&msl->msl_linkage, slave_locks);
1881 mdt_unlock_list(info, slave_locks, rc);
1885 /* lock parent and its stripes */
1886 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1887 struct mdt_object *obj,
1888 const struct md_attr *ma,
1889 struct mdt_lock_handle *lh,
1890 struct ldlm_enqueue_info *einfo,
1891 struct list_head *slave_locks)
1895 if (mdt_object_remote(obj)) {
1896 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1897 &lh->mlh_rreg_lh, LCK_PW,
1898 MDS_INODELOCK_UPDATE, false);
1903 * if obj is remote and striped, lock its stripes explicitly
1904 * because it's not striped in LOD layer on this MDT.
1906 if (ma->ma_valid & MA_LMV) {
1907 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1909 mdt_object_unlock(info, obj, lh, rc);
1912 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1920 * in migration, object may be remote, and we need take full lock of it and its
1921 * stripes if it's directory, besides, object may be a remote object on its
1922 * parent, revoke its LOOKUP lock on where its parent is located.
1924 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1925 struct mdt_object *pobj,
1926 struct mdt_object *obj,
1927 struct mdt_lock_handle *lh,
1928 struct ldlm_enqueue_info *einfo,
1929 struct list_head *slave_locks)
1933 if (mdt_object_remote(obj)) {
1934 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1938 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1939 &lh->mlh_rreg_lh, LCK_EX,
1940 MDS_INODELOCK_FULL, false);
1945 * if obj is remote and striped, lock its stripes explicitly
1946 * because it's not striped in LOD layer on this MDT.
1948 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1949 struct md_attr *ma = &info->mti_attr;
1951 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1953 mdt_object_unlock(info, obj, lh, rc);
1957 if (ma->ma_valid & MA_LMV) {
1958 rc = mdt_lock_remote_slaves(info, obj, ma,
1961 mdt_object_unlock(info, obj, lh, rc);
1965 if (mdt_object_remote(pobj)) {
1966 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1971 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1979 * lookup source by name, if parent is striped directory, we need to find the
1980 * corresponding stripe where source is located, and then lookup there.
1982 * besides, if parent is migrating too, and file is already in target stripe,
1983 * this should be a redo of 'lfs migrate' on client side.
1985 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1986 struct mdt_object *pobj,
1987 const struct md_attr *ma,
1988 const struct lu_name *lname,
1989 struct mdt_object **spobj,
1990 struct mdt_object **sobj)
1992 const struct lu_env *env = info->mti_env;
1993 struct lu_fid *fid = &info->mti_tmp_fid1;
1994 struct mdt_object *stripe;
1997 if (ma->ma_valid & MA_LMV) {
1998 /* if parent is striped, lookup on corresponding stripe */
1999 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2001 if (!lmv_is_sane(lmv))
2004 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2009 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2011 stripe = mdt_object_find(env, info->mti_mdt, fid);
2013 return PTR_ERR(stripe);
2016 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2018 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2020 * if parent layout is changeing, and lookup child
2021 * failed on source stripe, lookup again on target
2022 * stripe, if it exists, it means previous migration
2023 * was interrupted, and current file was migrated
2026 mdt_object_put(env, stripe);
2028 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2033 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2035 stripe = mdt_object_find(env, info->mti_mdt, fid);
2037 return PTR_ERR(stripe);
2040 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2041 fid, &info->mti_spec);
2042 mdt_object_put(env, stripe);
2043 return rc ?: -EALREADY;
2045 mdt_object_put(env, stripe);
2050 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2056 mdt_object_get(env, stripe);
2061 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2062 if (IS_ERR(*sobj)) {
2063 mdt_object_put(env, stripe);
2064 rc = PTR_ERR(*sobj);
2072 /* end lease and close file for regular file */
2073 static int mdd_migrate_close(struct mdt_thread_info *info,
2074 struct mdt_object *obj)
2076 struct close_data *data;
2077 struct mdt_body *repbody;
2078 struct ldlm_lock *lease;
2083 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2085 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2089 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2094 lease = ldlm_handle2lock(&data->cd_handle);
2098 /* check if the lease was already canceled */
2099 lock_res_and_lock(lease);
2100 rc = ldlm_is_cancel(lease);
2101 unlock_res_and_lock(lease);
2105 LDLM_DEBUG(lease, DFID" lease broken",
2106 PFID(mdt_object_fid(obj)));
2110 * cancel server side lease, client side counterpart should have been
2111 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2113 ldlm_lock_cancel(lease);
2114 ldlm_reprocess_all(lease->l_resource, lease);
2115 LDLM_LOCK_PUT(lease);
2118 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2119 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2120 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2126 * migrate file in below steps:
2127 * 1. lock parent and its stripes
2128 * 2. lookup source by name
2129 * 3. lock parents of source links if source is not directory
2130 * 4. reject if source is in HSM
2131 * 5. take source open_sem and close file if source is regular file
2132 * 6. lock source and its stripes if it's directory
2133 * 7. lock target so subsequent change to it can trigger COS
2135 * 9. unlock above locks
2136 * 10. sync device if source has links
2138 int mdt_reint_migrate(struct mdt_thread_info *info,
2139 struct mdt_lock_handle *unused)
2141 const struct lu_env *env = info->mti_env;
2142 struct mdt_device *mdt = info->mti_mdt;
2143 struct ptlrpc_request *req = mdt_info_req(info);
2144 struct mdt_reint_record *rr = &info->mti_rr;
2145 struct lu_ucred *uc = mdt_ucred(info);
2146 struct md_attr *ma = &info->mti_attr;
2147 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2148 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2149 struct mdt_object *pobj;
2150 struct mdt_object *spobj = NULL;
2151 struct mdt_object *sobj = NULL;
2152 struct mdt_object *tobj;
2153 struct lustre_handle rename_lh = { 0 };
2154 struct mdt_lock_handle *lhp;
2155 struct mdt_lock_handle *lhs;
2156 struct mdt_lock_handle *lht;
2157 LIST_HEAD(parent_slave_locks);
2158 LIST_HEAD(child_slave_locks);
2159 LIST_HEAD(link_locks);
2160 int lock_retries = 5;
2161 bool open_sem_locked = false;
2162 bool do_sync = false;
2166 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2167 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2169 if (info->mti_dlm_req)
2170 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2172 if (!fid_is_md_operative(rr->rr_fid1) ||
2173 !fid_is_md_operative(rr->rr_fid2))
2176 /* don't allow migrate . or .. */
2177 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2180 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2183 if (uc && !md_capable(uc, CAP_SYS_ADMIN) &&
2184 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2185 mdt->mdt_enable_remote_dir_gid != -1)
2189 * Note: do not enqueue rename lock for replay request, because
2190 * if other MDT holds rename lock, but being blocked to wait for
2191 * this MDT to finish its recovery, and the failover MDT can not
2192 * get rename lock, which will cause deadlock.
2194 * req is NULL if this is called by directory auto-split.
2196 if (req && !req_is_replay(req)) {
2197 rc = mdt_rename_lock(info, &rename_lh);
2199 CERROR("%s: can't lock FS for rename: rc = %d\n",
2200 mdt_obd_name(info->mti_mdt), rc);
2205 /* pobj is master object of parent */
2206 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2208 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2211 rc = mdt_version_get_check(info, pobj, 0);
2213 GOTO(put_parent, rc);
2216 if (!mdt_object_exists(pobj))
2217 GOTO(put_parent, rc = -ENOENT);
2219 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2220 GOTO(put_parent, rc = -ENOTDIR);
2222 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2224 GOTO(put_parent, rc);
2227 /* lock parent object */
2228 lhp = &info->mti_lh[MDT_LH_PARENT];
2229 mdt_lock_reg_init(lhp, LCK_PW);
2230 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2231 &parent_slave_locks);
2233 GOTO(put_parent, rc);
2236 * spobj is the corresponding stripe against name if pobj is striped
2237 * directory, which is the real parent, and no need to lock, because
2238 * we've taken full lock of pobj.
2240 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2242 GOTO(unlock_parent, rc);
2244 /* lock parents of source links, and revoke LOOKUP lock of links */
2245 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2246 &parent_slave_locks, &link_locks);
2247 if (rc == -EBUSY && lock_retries-- > 0) {
2248 mdt_object_put(env, sobj);
2249 mdt_object_put(env, spobj);
2254 GOTO(put_source, rc);
2257 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2258 * with one request, if total lock count exceeds this limit, we will
2259 * drop all locks after migration, and synchronous device in the end.
2263 /* TODO: DoM migration is not supported, migrate dirent only */
2264 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2265 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2267 GOTO(unlock_links, rc);
2269 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2270 info->mti_spec.sp_migrate_nsonly = 1;
2273 /* if migration HSM is allowed */
2274 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2275 ma->ma_need = MA_HSM;
2277 rc = mdt_attr_get_complex(info, sobj, ma);
2279 GOTO(unlock_links, rc);
2281 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2282 GOTO(unlock_links, rc = -EOPNOTSUPP);
2285 /* end lease and close file for regular file */
2286 if (info->mti_spec.sp_migrate_close) {
2287 /* try to hold open_sem so that nobody else can open the file */
2288 if (!down_write_trylock(&sobj->mot_open_sem)) {
2290 mdd_migrate_close(info, sobj);
2291 GOTO(unlock_links, rc = -EBUSY);
2293 open_sem_locked = true;
2294 rc = mdd_migrate_close(info, sobj);
2296 GOTO(unlock_open_sem, rc);
2301 lhs = &info->mti_lh[MDT_LH_OLD];
2302 mdt_lock_reg_init(lhs, LCK_EX);
2303 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2304 &child_slave_locks);
2306 GOTO(unlock_open_sem, rc);
2309 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2311 GOTO(unlock_source, rc = PTR_ERR(tobj));
2313 lht = &info->mti_lh[MDT_LH_NEW];
2314 mdt_lock_reg_init(lht, LCK_EX);
2315 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2317 GOTO(put_target, rc);
2319 /* Don't do lookup sanity check. We know name doesn't exist. */
2320 info->mti_spec.sp_cr_lookup = 0;
2321 info->mti_spec.sp_feat = &dt_directory_features;
2323 rc = mdo_migrate(env, mdt_object_child(pobj),
2324 mdt_object_child(sobj), &rr->rr_name,
2325 mdt_object_child(tobj),
2326 &info->mti_spec, ma);
2328 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2329 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2332 mdt_object_unlock(info, tobj, lht, rc);
2334 mdt_object_put(env, tobj);
2336 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2337 &child_slave_locks, rc);
2339 if (open_sem_locked)
2340 up_write(&sobj->mot_open_sem);
2342 /* if we've got too many locks to save into RPC,
2343 * then just commit before the locks are released
2346 mdt_device_sync(env, mdt);
2347 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2349 mdt_object_put(env, sobj);
2350 mdt_object_put(env, spobj);
2352 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2353 &parent_slave_locks, rc);
2355 mdt_object_put(env, pobj);
2357 if (lustre_handle_is_used(&rename_lh))
2358 mdt_rename_unlock(&rename_lh);
2363 static int mdt_object_lock_save(struct mdt_thread_info *info,
2364 struct mdt_object *dir,
2365 struct mdt_lock_handle *lh,
2366 int idx, bool cos_incompat)
2370 /* we lock the target dir if it is local */
2371 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2376 /* get and save correct version after locking */
2377 mdt_version_get_save(info, dir, idx);
2382 * determine lock order of sobj and tobj
2384 * there are two situations we need to lock tobj before sobj:
2385 * 1. sobj is child of tobj
2386 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2387 * larger than that of tobj
2389 * \retval 1 lock tobj before sobj
2390 * \retval 0 lock sobj before tobj
2391 * \retval -ev negative errno upon error
2393 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2394 struct mdt_object *sobj,
2395 struct mdt_object *tobj)
2397 struct md_attr *ma = &info->mti_attr;
2398 struct lu_fid *spfid = &info->mti_tmp_fid1;
2399 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2400 struct lmv_mds_md_v1 *lmv;
2405 /* sobj and tobj are the same */
2409 if (fid_is_root(mdt_object_fid(sobj)))
2412 if (fid_is_root(mdt_object_fid(tobj)))
2415 /* check whether sobj is child of tobj */
2416 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2417 mdt_object_fid(tobj));
2424 /* check whether sobj and tobj are children of the same parent */
2425 rc = mdt_attr_get_pfid(info, sobj, spfid);
2429 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2433 if (!lu_fid_eq(spfid, tpfid))
2436 /* check whether sobj and tobj are sibling stripes */
2437 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2441 if (!(ma->ma_valid & MA_LMV))
2444 lmv = &ma->ma_lmv->lmv_md_v1;
2445 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2447 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2450 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2454 if (!(ma->ma_valid & MA_LMV))
2457 lmv = &ma->ma_lmv->lmv_md_v1;
2458 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2460 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2462 /* check stripe index of sobj and tobj */
2463 if (sindex == tindex)
2466 return sindex < tindex ? 0 : 1;
2470 * lock rename source object.
2472 * Both source and source parent may be remote, and source may be a remote
2473 * object on source parent, to avoid overriding lock handle, store remote
2474 * LOOKUP lock separately in @lhr.
2476 * \retval 0 on success
2477 * \retval -ev negative errno upon error
2479 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2480 struct mdt_object *parent,
2481 struct mdt_object *child,
2482 struct mdt_lock_handle *lhc,
2483 struct mdt_lock_handle *lhr,
2489 rc = mdt_is_remote_object(info, parent, child);
2494 /* enqueue remote LOOKUP lock from the parent MDT */
2495 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2497 if (mdt_object_remote(parent)) {
2498 rc = mdt_remote_object_lock(info, parent,
2499 mdt_object_fid(child),
2506 LASSERT(mdt_object_remote(child));
2507 rc = mdt_object_local_lock(info, child, lhr,
2508 &rmt_ibits, 0, true);
2513 ibits &= ~MDS_INODELOCK_LOOKUP;
2516 if (mdt_object_remote(child)) {
2517 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2524 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2529 mdt_object_unlock(info, child, lhr, rc);
2535 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2536 * 2 - srcdir child; 3 - tgtdir child.
2537 * Update on disk version of srcdir child.
2539 static int mdt_reint_rename(struct mdt_thread_info *info,
2540 struct mdt_lock_handle *unused)
2542 struct mdt_device *mdt = info->mti_mdt;
2543 struct mdt_reint_record *rr = &info->mti_rr;
2544 struct md_attr *ma = &info->mti_attr;
2545 struct ptlrpc_request *req = mdt_info_req(info);
2546 struct mdt_object *msrcdir = NULL;
2547 struct mdt_object *mtgtdir = NULL;
2548 struct mdt_object *mold;
2549 struct mdt_object *mnew = NULL;
2550 struct lustre_handle rename_lh = { 0 };
2551 struct mdt_lock_handle *lh_srcdirp;
2552 struct mdt_lock_handle *lh_tgtdirp;
2553 struct mdt_lock_handle *lh_oldp = NULL;
2554 struct mdt_lock_handle *lh_rmt = NULL;
2555 struct mdt_lock_handle *lh_newp = NULL;
2556 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2557 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2559 bool reverse = false, discard = false;
2561 ktime_t kstart = ktime_get();
2565 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2566 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2567 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2569 if (info->mti_dlm_req)
2570 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2572 if (!fid_is_md_operative(rr->rr_fid1) ||
2573 !fid_is_md_operative(rr->rr_fid2))
2576 /* find both parents. */
2577 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2578 if (IS_ERR(msrcdir))
2579 RETURN(PTR_ERR(msrcdir));
2581 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2583 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2585 mdt_object_get(info->mti_env, mtgtdir);
2587 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2588 if (IS_ERR(mtgtdir))
2589 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2593 * Note: do not enqueue rename lock for replay request, because
2594 * if other MDT holds rename lock, but being blocked to wait for
2595 * this MDT to finish its recovery, and the failover MDT can not
2596 * get rename lock, which will cause deadlock.
2598 if (!req_is_replay(req)) {
2600 * Normally rename RPC is handled on the MDT with the target
2601 * directory (if target exists, it's on the MDT with the
2602 * target), if the source directory is remote, it's a hint that
2603 * source is remote too (this may not be true, but it won't
2604 * cause any issue), return -EXDEV early to avoid taking
2607 if (!mdt->mdt_enable_remote_rename &&
2608 mdt_object_remote(msrcdir))
2609 GOTO(out_put_tgtdir, rc = -EXDEV);
2611 /* This may be further relaxed in the future for different
2612 * source and target parents. Start with only same-directory
2613 * renames of non-directory files for simplicity and because
2614 * this is by far the most common use case.
2616 if (msrcdir != mtgtdir || S_ISDIR(ma->ma_attr.la_mode)) {
2617 rc = mdt_rename_lock(info, &rename_lh);
2619 CERROR("%s: cannot lock for rename: rc = %d\n",
2620 mdt_obd_name(mdt), rc);
2621 GOTO(out_put_tgtdir, rc);
2624 CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2625 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2626 PNAME(&rr->rr_name));
2630 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2632 GOTO(out_unlock_rename, rc);
2636 /* source needs to be looked up after locking source parent, otherwise
2637 * this rename may race with unlink source, and cause rename hang, see
2638 * sanityn.sh 55b, so check parents first, if later we found source is
2639 * remote, relock parents.
2641 cos_incompat = (mdt_object_remote(msrcdir) ||
2642 mdt_object_remote(mtgtdir));
2644 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2646 /* lock parents in the proper order. */
2647 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2648 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2650 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2651 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2653 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2654 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2657 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2660 GOTO(out_unlock_rename, rc);
2662 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2664 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2667 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2668 GOTO(out_unlock_rename, rc);
2671 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2674 GOTO(out_unlock_rename, rc);
2676 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2678 if (mtgtdir != msrcdir) {
2679 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2681 } else if (!mdt_object_remote(mtgtdir) &&
2682 lh_srcdirp->mlh_pdo_hash !=
2683 lh_tgtdirp->mlh_pdo_hash) {
2684 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2685 MDS_INODELOCK_UPDATE,
2687 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2690 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2691 GOTO(out_unlock_rename, rc);
2695 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2696 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2698 /* find mold object. */
2700 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2702 GOTO(out_unlock_parents, rc);
2704 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2705 GOTO(out_unlock_parents, rc = -EINVAL);
2707 if (!fid_is_md_operative(old_fid))
2708 GOTO(out_unlock_parents, rc = -EPERM);
2710 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2712 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2714 if (!mdt_object_exists(mold)) {
2715 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2717 "object does not exist");
2718 GOTO(out_put_old, rc = -ENOENT);
2721 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2722 GOTO(out_put_old, rc = -EXDEV);
2724 /* Check if @mtgtdir is subdir of @mold, before locking child
2725 * to avoid reverse locking.
2727 if (mtgtdir != msrcdir) {
2728 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2733 GOTO(out_put_old, rc);
2737 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2738 /* save version after locking */
2739 mdt_version_get_save(info, mold, 2);
2741 if (!cos_incompat && mdt_object_remote(mold)) {
2742 cos_incompat = true;
2743 mdt_object_put(info->mti_env, mold);
2744 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2745 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2749 /* find mnew object:
2750 * mnew target object may not exist now
2751 * lookup with version checking
2754 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2757 /* the new_fid should have been filled at this moment */
2758 if (lu_fid_eq(old_fid, new_fid))
2759 GOTO(out_put_old, rc);
2761 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2762 lu_fid_eq(new_fid, rr->rr_fid2))
2763 GOTO(out_put_old, rc = -EINVAL);
2765 if (!fid_is_md_operative(new_fid))
2766 GOTO(out_put_old, rc = -EPERM);
2768 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2770 GOTO(out_put_old, rc = PTR_ERR(mnew));
2772 if (!mdt_object_exists(mnew)) {
2773 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2775 "object does not exist");
2776 GOTO(out_put_new, rc = -ENOENT);
2779 if (mdt_object_remote(mnew)) {
2780 struct mdt_body *repbody;
2782 /* Always send rename req to the target child MDT */
2783 repbody = req_capsule_server_get(info->mti_pill,
2785 LASSERT(repbody != NULL);
2786 repbody->mbo_fid1 = *new_fid;
2787 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2788 GOTO(out_put_new, rc = -EXDEV);
2790 /* Before locking the target dir, check we do not replace
2791 * a dir with a non-dir, otherwise it may deadlock with
2792 * link op which tries to create a link in this dir
2793 * back to this non-dir.
2795 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2796 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2797 GOTO(out_put_new, rc = -EISDIR);
2799 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2800 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2801 mdt_lock_reg_init(lh_oldp, LCK_EX);
2802 mdt_lock_reg_init(lh_rmt, LCK_EX);
2803 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2804 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2805 lh_rmt, lock_ibits, cos_incompat);
2807 GOTO(out_put_new, rc);
2809 /* Check if @msrcdir is subdir of @mnew, before locking child
2810 * to avoid reverse locking.
2812 if (mtgtdir != msrcdir) {
2813 rc = mdo_is_subdir(info->mti_env,
2814 mdt_object_child(msrcdir), new_fid);
2818 GOTO(out_unlock_old, rc);
2822 /* We used to acquire MDS_INODELOCK_FULL here but we
2823 * can't do this now because a running HSM restore on
2824 * the rename onto victim will hold the layout
2825 * lock. See LU-4002.
2828 lh_newp = &info->mti_lh[MDT_LH_NEW];
2829 mdt_lock_reg_init(lh_newp, LCK_EX);
2830 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2831 if (mdt_object_remote(mtgtdir)) {
2832 rc = mdt_remote_object_lock(info, mtgtdir,
2833 mdt_object_fid(mnew),
2834 &lh_newp->mlh_rreg_lh,
2835 lh_newp->mlh_rreg_mode,
2836 MDS_INODELOCK_LOOKUP,
2839 GOTO(out_unlock_old, rc);
2841 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2843 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2846 GOTO(out_unlock_new, rc);
2848 /* get and save version after locking */
2849 mdt_version_get_save(info, mnew, 3);
2850 } else if (rc != -ENOENT) {
2851 GOTO(out_put_old, rc);
2853 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2854 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2855 mdt_lock_reg_init(lh_oldp, LCK_EX);
2856 mdt_lock_reg_init(lh_rmt, LCK_EX);
2857 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2858 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2859 lh_rmt, lock_ibits, cos_incompat);
2861 GOTO(out_put_old, rc);
2863 mdt_enoent_version_save(info, 3);
2866 /* step 5: rename it */
2867 mdt_reint_init_ma(info, ma);
2869 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2870 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2873 mutex_lock(&mnew->mot_lov_mutex);
2875 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2876 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2877 mnew != NULL ? mdt_object_child(mnew) : NULL,
2878 &rr->rr_tgt_name, ma);
2881 mutex_unlock(&mnew->mot_lov_mutex);
2883 /* handle last link of tgt object */
2885 mdt_counter_incr(req, LPROC_MDT_RENAME,
2886 ktime_us_delta(ktime_get(), kstart));
2888 mdt_handle_last_unlink(info, mnew, ma);
2889 discard = mdt_dom_check_for_discard(info, mnew);
2891 mdt_rename_counter_tally(info, info->mti_mdt, req,
2893 ktime_us_delta(ktime_get(), kstart));
2899 mdt_object_unlock(info, mnew, lh_newp, rc);
2901 mdt_object_unlock(info, NULL, lh_rmt, rc);
2902 mdt_object_unlock(info, mold, lh_oldp, rc);
2904 if (mnew && !discard)
2905 mdt_object_put(info->mti_env, mnew);
2907 mdt_object_put(info->mti_env, mold);
2909 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2910 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2912 if (lustre_handle_is_used(&rename_lh))
2913 mdt_rename_unlock(&rename_lh);
2915 mdt_object_put(info->mti_env, mtgtdir);
2917 mdt_object_put(info->mti_env, msrcdir);
2919 /* The DoM discard can be done right in the place above where it is
2920 * assigned, meanwhile it is done here after rename unlock due to
2921 * compatibility with old clients, for them the discard blocks
2922 * the main thread until completion. Check LU-11359 for details.
2925 mdt_dom_discard_data(info, mnew);
2926 mdt_object_put(info->mti_env, mnew);
2931 static int mdt_reint_resync(struct mdt_thread_info *info,
2932 struct mdt_lock_handle *lhc)
2934 struct mdt_reint_record *rr = &info->mti_rr;
2935 struct ptlrpc_request *req = mdt_info_req(info);
2936 struct md_attr *ma = &info->mti_attr;
2937 struct mdt_object *mo;
2938 struct ldlm_lock *lease;
2939 struct mdt_body *repbody;
2940 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2945 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2947 if (info->mti_dlm_req)
2948 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2950 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2952 GOTO(out, rc = PTR_ERR(mo));
2954 if (!mdt_object_exists(mo))
2955 GOTO(out_obj, rc = -ENOENT);
2957 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2958 GOTO(out_obj, rc = -EINVAL);
2960 if (mdt_object_remote(mo))
2961 GOTO(out_obj, rc = -EREMOTE);
2963 lease = ldlm_handle2lock(rr->rr_lease_handle);
2965 GOTO(out_obj, rc = -ESTALE);
2967 /* It's really necessary to grab open_sem and check if the lease lock
2968 * has been lost. There would exist a concurrent writer coming in and
2969 * generating some dirty data in memory cache, the writeback would fail
2970 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2972 if (!down_write_trylock(&mo->mot_open_sem))
2973 GOTO(out_put_lease, rc = -EBUSY);
2975 lock_res_and_lock(lease);
2976 lease_broken = ldlm_is_cancel(lease);
2977 unlock_res_and_lock(lease);
2979 GOTO(out_unlock, rc = -EBUSY);
2981 /* the file has yet opened by anyone else after we took the lease. */
2982 layout.mlc_opc = MD_LAYOUT_RESYNC;
2983 lhc = &info->mti_lh[MDT_LH_LOCAL];
2984 rc = mdt_layout_change(info, mo, lhc, &layout);
2986 GOTO(out_unlock, rc);
2988 mdt_object_unlock(info, mo, lhc, 0);
2990 ma->ma_need = MA_INODE;
2992 rc = mdt_attr_get_complex(info, mo, ma);
2994 GOTO(out_unlock, rc);
2996 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2997 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3001 up_write(&mo->mot_open_sem);
3003 LDLM_LOCK_PUT(lease);
3005 mdt_object_put(info->mti_env, mo);
3007 mdt_client_compatibility(info);
3008 rc2 = mdt_fix_reply(info);
3014 struct mdt_reinter {
3015 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3016 enum lprocfs_extra_opc mr_extra_opc;
3019 static const struct mdt_reinter mdt_reinters[] = {
3021 .mr_handler = &mdt_reint_setattr,
3022 .mr_extra_opc = MDS_REINT_SETATTR,
3025 .mr_handler = &mdt_reint_create,
3026 .mr_extra_opc = MDS_REINT_CREATE,
3029 .mr_handler = &mdt_reint_link,
3030 .mr_extra_opc = MDS_REINT_LINK,
3033 .mr_handler = &mdt_reint_unlink,
3034 .mr_extra_opc = MDS_REINT_UNLINK,
3037 .mr_handler = &mdt_reint_rename,
3038 .mr_extra_opc = MDS_REINT_RENAME,
3041 .mr_handler = &mdt_reint_open,
3042 .mr_extra_opc = MDS_REINT_OPEN,
3044 [REINT_SETXATTR] = {
3045 .mr_handler = &mdt_reint_setxattr,
3046 .mr_extra_opc = MDS_REINT_SETXATTR,
3049 .mr_handler = &mdt_reint_unlink,
3050 .mr_extra_opc = MDS_REINT_UNLINK,
3053 .mr_handler = &mdt_reint_migrate,
3054 .mr_extra_opc = MDS_REINT_RENAME,
3057 .mr_handler = &mdt_reint_resync,
3058 .mr_extra_opc = MDS_REINT_RESYNC,
3062 int mdt_reint_rec(struct mdt_thread_info *info,
3063 struct mdt_lock_handle *lhc)
3065 const struct mdt_reinter *mr;
3069 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3072 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3073 if (mr->mr_handler == NULL)
3076 rc = (*mr->mr_handler)(info, lhc);
3078 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3079 PTLRPC_LAST_CNTR + mr->mr_extra_opc);