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, CFS_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 */
535 if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
539 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
541 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
543 RETURN(PTR_ERR(parent));
545 if (!mdt_object_exists(parent))
546 GOTO(put_parent, rc = -ENOENT);
549 * LU-10235: check if name exists locklessly first to avoid massive
550 * lock recalls on existing directories.
552 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
553 &info->mti_tmp_fid1, 1);
556 GOTO(put_parent, rc = -EEXIST);
558 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
562 /* -ENOENT is expected here */
564 GOTO(put_parent, rc);
566 /* save version of file name for replay, it must be ENOENT here */
567 mdt_enoent_version_save(info, 1);
569 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
571 lh = &info->mti_lh[MDT_LH_PARENT];
572 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
573 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
575 GOTO(put_parent, rc);
577 if (!mdt_object_remote(parent)) {
578 rc = mdt_version_get_check_save(info, parent, 0);
580 GOTO(unlock_parent, rc);
583 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
584 if (unlikely(IS_ERR(child)))
585 GOTO(unlock_parent, rc = PTR_ERR(child));
587 ma->ma_need = MA_INODE;
590 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
591 OBD_FAIL_MDS_REINT_CREATE_WRITE);
593 /* Version of child will be updated on disk. */
594 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
595 rc = mdt_version_get_check_save(info, child, 2);
599 /* Let lower layer know current lock mode. */
600 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
603 * Do not perform lookup sanity check. We know that name does
606 info->mti_spec.sp_cr_lookup = 0;
607 info->mti_spec.sp_feat = &dt_directory_features;
609 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
610 mdt_object_child(child), &info->mti_spec, ma);
612 rc = mdt_attr_get_complex(info, child, ma);
618 * On DNE, we need to eliminate dependey between 'mkdir a' and
619 * 'mkdir a/b' if b is a striped directory, to achieve this, two
620 * things are done below:
621 * 1. save child and slaves lock.
622 * 2. if the child is a striped directory, relock parent so to
623 * compare against with COS locks to ensure parent was
626 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
627 struct mdt_lock_handle *lhc;
628 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
631 rc = mdt_object_striped(info, child);
637 if (!mdt_object_remote(parent)) {
638 mdt_object_unlock(info, parent, lh, 1);
639 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
640 rc = mdt_reint_object_lock(info, parent, lh,
641 MDS_INODELOCK_UPDATE,
648 lhc = &info->mti_lh[MDT_LH_CHILD];
649 mdt_lock_handle_init(lhc);
650 mdt_lock_reg_init(lhc, LCK_PW);
651 rc = mdt_reint_striped_lock(info, child, lhc,
652 MDS_INODELOCK_UPDATE, einfo,
657 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
660 /* Return fid & attr to client. */
661 if (ma->ma_valid & MA_INODE)
662 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
663 mdt_object_fid(child));
666 mdt_object_put(info->mti_env, child);
668 mdt_object_unlock(info, parent, lh, rc);
670 mdt_object_put(info->mti_env, parent);
674 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
677 struct mdt_lock_handle *lh;
678 int do_vbr = ma->ma_attr.la_valid &
679 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
680 __u64 lockpart = MDS_INODELOCK_UPDATE;
681 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
686 rc = mdt_object_striped(info, mo);
692 lh = &info->mti_lh[MDT_LH_PARENT];
693 mdt_lock_reg_init(lh, LCK_PW);
695 /* Even though the new MDT will grant PERM lock to the old
696 * client, but the old client will almost ignore that during
697 * So it needs to revoke both LOOKUP and PERM lock here, so
698 * both new and old client can cancel the dcache
700 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
701 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
703 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
708 /* all attrs are packed into mti_attr in unpack_setattr */
709 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
710 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
712 /* VBR: update version if attr changed are important for recovery */
714 /* update on-disk version of changed object */
715 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
716 rc = mdt_version_get_check_save(info, mo, 0);
718 GOTO(out_unlock, rc);
721 /* Ensure constant striping during chown(). See LU-2789. */
722 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
723 mutex_lock(&mo->mot_lov_mutex);
725 /* all attrs are packed into mti_attr in unpack_setattr */
726 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
728 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
729 mutex_unlock(&mo->mot_lov_mutex);
732 GOTO(out_unlock, rc);
733 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
736 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
741 * Check HSM flags and add HS_DIRTY flag if relevant.
743 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
744 * and is not RELEASED.
746 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
749 struct lu_ucred *uc = mdt_ucred(info);
754 /* If the file was modified, add the dirty flag */
755 ma->ma_need = MA_HSM;
756 rc = mdt_attr_get_complex(info, mo, ma);
758 CERROR("file attribute read error for "DFID": %d.\n",
759 PFID(mdt_object_fid(mo)), rc);
763 /* If an up2date copy exists in the backend, add dirty flag */
764 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
765 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
766 ma->ma_hsm.mh_flags |= HS_DIRTY;
768 /* Bump cap so that closes from non-owner writers can
769 * set the HSM state to dirty.
771 cap_saved = uc->uc_cap;
772 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
773 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
774 uc->uc_cap = cap_saved;
776 CERROR("file attribute change error for "DFID": %d\n",
777 PFID(mdt_object_fid(mo)), rc);
783 static int mdt_reint_setattr(struct mdt_thread_info *info,
784 struct mdt_lock_handle *lhc)
786 struct mdt_device *mdt = info->mti_mdt;
787 struct md_attr *ma = &info->mti_attr;
788 struct mdt_reint_record *rr = &info->mti_rr;
789 struct ptlrpc_request *req = mdt_info_req(info);
790 struct mdt_object *mo;
791 struct mdt_body *repbody;
792 ktime_t kstart = ktime_get();
796 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
797 (unsigned int)ma->ma_attr.la_valid);
799 if (info->mti_dlm_req)
800 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
802 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
804 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
805 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
807 GOTO(out, rc = PTR_ERR(mo));
809 if (!mdt_object_exists(mo))
810 GOTO(out_put, rc = -ENOENT);
812 if (mdt_object_remote(mo))
813 GOTO(out_put, rc = -EREMOTE);
815 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
816 /* revoke lease lock if size is going to be changed */
817 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
818 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
819 atomic_read(&mo->mot_lease_count) > 0)) {
820 down_read(&mo->mot_open_sem);
822 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
823 lhc = &info->mti_lh[MDT_LH_LOCAL];
824 mdt_lock_reg_init(lhc, LCK_CW);
826 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
828 up_read(&mo->mot_open_sem);
832 /* revoke lease lock */
833 mdt_object_unlock(info, mo, lhc, 1);
835 up_read(&mo->mot_open_sem);
838 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
839 /* Check write access for the O_TRUNC case */
840 if (mdt_write_read(mo) < 0)
841 GOTO(out_put, rc = -ETXTBSY);
843 /* LU-10286: compatibility check for FLR.
844 * Please check the comment in mdt_finish_open() for details
846 if (!exp_connect_flr(info->mti_exp) ||
847 !exp_connect_overstriping(info->mti_exp)) {
848 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
849 if (rc < 0 && rc != -ENODATA)
852 if (!exp_connect_flr(info->mti_exp)) {
854 mdt_lmm_is_flr(info->mti_big_lmm))
855 GOTO(out_put, rc = -EOPNOTSUPP);
858 if (!exp_connect_overstriping(info->mti_exp)) {
860 mdt_lmm_is_overstriping(info->mti_big_lmm))
861 GOTO(out_put, rc = -EOPNOTSUPP);
865 /* For truncate, the file size sent from client
866 * is believable, but the blocks are incorrect,
867 * which makes the block size in LSOM attribute
868 * inconsisent with the real block size.
870 rc = mdt_lsom_update(info, mo, true);
875 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
876 if (ma->ma_valid & MA_LOV)
877 GOTO(out_put, rc = -EPROTO);
879 /* MDT supports FMD for regular files due to Data-on-MDT */
880 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
881 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
882 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
885 if (ma->ma_attr.la_valid & LA_MTIME) {
886 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
888 ma->ma_valid |= MA_PFID;
892 rc = mdt_attr_set(info, mo, ma);
895 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
896 (ma->ma_valid & MA_INODE)) {
897 struct lu_buf *buf = &info->mti_buf;
898 struct lu_ucred *uc = mdt_ucred(info);
899 struct mdt_lock_handle *lh;
901 __u64 lockpart = MDS_INODELOCK_XATTR;
903 /* reject if either remote or striped dir is disabled */
904 if (ma->ma_valid & MA_LMV) {
905 if (!mdt->mdt_enable_remote_dir ||
906 !mdt->mdt_enable_striped_dir)
907 GOTO(out_put, rc = -EPERM);
909 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
910 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
911 mdt->mdt_enable_remote_dir_gid != -1)
912 GOTO(out_put, rc = -EPERM);
915 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
916 GOTO(out_put, rc = -ENOTDIR);
918 if (ma->ma_attr.la_valid != 0)
919 GOTO(out_put, rc = -EPROTO);
921 if (ma->ma_valid & MA_LOV) {
922 buf->lb_buf = ma->ma_lmm;
923 buf->lb_len = ma->ma_lmm_size;
924 name = XATTR_NAME_LOV;
926 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
929 buf->lb_len = ma->ma_lmv_size;
930 name = XATTR_NAME_DEFAULT_LMV;
931 /* force client to update dir default layout */
932 lockpart |= MDS_INODELOCK_LOOKUP;
935 lh = &info->mti_lh[MDT_LH_PARENT];
936 mdt_lock_reg_init(lh, LCK_PW);
938 rc = mdt_object_lock(info, mo, lh, lockpart);
942 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
945 mdt_object_unlock(info, mo, lh, rc);
949 GOTO(out_put, rc = -EPROTO);
952 /* If file data is modified, add the dirty flag */
953 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
954 rc = mdt_add_dirty_flag(info, mo, ma);
956 ma->ma_need = MA_INODE;
958 rc = mdt_attr_get_complex(info, mo, ma);
962 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
966 mdt_object_put(info->mti_env, mo);
969 mdt_counter_incr(req, LPROC_MDT_SETATTR,
970 ktime_us_delta(ktime_get(), kstart));
972 mdt_client_compatibility(info);
973 rc2 = mdt_fix_reply(info);
979 static int mdt_reint_create(struct mdt_thread_info *info,
980 struct mdt_lock_handle *lhc)
982 struct ptlrpc_request *req = mdt_info_req(info);
983 ktime_t kstart = ktime_get();
987 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
988 RETURN(err_serious(-ESTALE));
990 if (info->mti_dlm_req)
991 ldlm_request_cancel(mdt_info_req(info),
992 info->mti_dlm_req, 0, LATF_SKIP);
994 if (!lu_name_is_valid(&info->mti_rr.rr_name))
997 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1007 CERROR("%s: Unsupported mode %o\n",
1008 mdt_obd_name(info->mti_mdt),
1009 info->mti_attr.ma_attr.la_mode);
1010 RETURN(err_serious(-EOPNOTSUPP));
1013 rc = mdt_create(info);
1015 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1016 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1017 ktime_us_delta(ktime_get(), kstart));
1019 /* Special file should stay on the same node as parent*/
1020 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1021 ktime_us_delta(ktime_get(), kstart));
1028 * VBR: save parent version in reply and child version getting by its name.
1029 * Version of child is getting and checking during its lookup. If
1031 static int mdt_reint_unlink(struct mdt_thread_info *info,
1032 struct mdt_lock_handle *lhc)
1034 struct mdt_reint_record *rr = &info->mti_rr;
1035 struct ptlrpc_request *req = mdt_info_req(info);
1036 struct md_attr *ma = &info->mti_attr;
1037 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1038 struct mdt_object *mp;
1039 struct mdt_object *mc;
1040 struct mdt_lock_handle *parent_lh;
1041 struct mdt_lock_handle *child_lh;
1042 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1044 bool cos_incompat = false;
1046 ktime_t kstart = ktime_get();
1050 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1051 PNAME(&rr->rr_name));
1053 if (info->mti_dlm_req)
1054 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1056 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1057 RETURN(err_serious(-ENOENT));
1059 if (!fid_is_md_operative(rr->rr_fid1))
1062 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1064 RETURN(PTR_ERR(mp));
1066 if (mdt_object_remote(mp)) {
1067 cos_incompat = true;
1069 rc = mdt_version_get_check_save(info, mp, 0);
1071 GOTO(put_parent, rc);
1075 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1076 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1077 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1080 GOTO(put_parent, rc);
1082 /* lookup child object along with version checking */
1083 fid_zero(child_fid);
1084 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1086 /* Name might not be able to find during resend of
1087 * remote unlink, considering following case.
1088 * dir_A is a remote directory, the name entry of
1089 * dir_A is on MDT0, the directory is on MDT1,
1091 * 1. client sends unlink req to MDT1.
1092 * 2. MDT1 sends name delete update to MDT0.
1093 * 3. name entry is being deleted in MDT0 synchronously.
1094 * 4. MDT1 is restarted.
1095 * 5. client resends unlink req to MDT1. So it can not
1096 * find the name entry on MDT0 anymore.
1097 * In this case, MDT1 only needs to destory the local
1100 if (mdt_object_remote(mp) && rc == -ENOENT &&
1101 !fid_is_zero(rr->rr_fid2) &&
1102 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1104 *child_fid = *rr->rr_fid2;
1106 GOTO(unlock_parent, rc);
1110 if (!fid_is_md_operative(child_fid))
1111 GOTO(unlock_parent, rc = -EPERM);
1113 /* We will lock the child regardless it is local or remote. No harm. */
1114 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1116 GOTO(unlock_parent, rc = PTR_ERR(mc));
1118 if (!cos_incompat) {
1119 rc = mdt_object_striped(info, mc);
1121 GOTO(put_child, rc);
1125 mdt_object_put(info->mti_env, mc);
1126 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1131 child_lh = &info->mti_lh[MDT_LH_CHILD];
1132 mdt_lock_reg_init(child_lh, LCK_EX);
1133 if (info->mti_spec.sp_rm_entry) {
1134 struct lu_ucred *uc = mdt_ucred(info);
1136 if (!mdt_is_dne_client(req->rq_export))
1137 /* Return -ENOTSUPP for old client */
1138 GOTO(put_child, rc = -ENOTSUPP);
1140 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
1141 GOTO(put_child, rc = -EPERM);
1143 ma->ma_need = MA_INODE;
1145 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1146 NULL, &rr->rr_name, ma, no_name);
1147 GOTO(put_child, rc);
1150 if (mdt_object_remote(mc)) {
1151 struct mdt_body *repbody;
1153 if (!fid_is_zero(rr->rr_fid2)) {
1154 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1155 mdt_obd_name(info->mti_mdt),
1156 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1157 GOTO(put_child, rc = -ENOENT);
1159 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1160 mdt_obd_name(info->mti_mdt),
1161 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1163 if (!mdt_is_dne_client(req->rq_export))
1164 /* Return -ENOTSUPP for old client */
1165 GOTO(put_child, rc = -ENOTSUPP);
1167 /* Revoke the LOOKUP lock of the remote object granted by
1168 * this MDT. Since the unlink will happen on another MDT,
1169 * it will release the LOOKUP lock right away. Then What
1170 * would happen if another client try to grab the LOOKUP
1171 * lock at the same time with unlink XXX
1173 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1174 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1175 LASSERT(repbody != NULL);
1176 repbody->mbo_fid1 = *mdt_object_fid(mc);
1177 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1178 GOTO(unlock_child, rc = -EREMOTE);
1180 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1181 * this now because a running HSM restore on the child (unlink
1182 * victim) will hold the layout lock. See LU-4002.
1184 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1185 if (mdt_object_remote(mp)) {
1186 /* Enqueue lookup lock from parent MDT */
1187 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1188 &child_lh->mlh_rreg_lh,
1189 child_lh->mlh_rreg_mode,
1190 MDS_INODELOCK_LOOKUP, false);
1192 GOTO(put_child, rc);
1194 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1197 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1200 GOTO(put_child, rc);
1203 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1204 * whether need MA_LOV and MA_COOKIE.
1206 ma->ma_need = MA_INODE;
1209 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1210 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1211 /* save version when object is locked */
1212 mdt_version_get_save(info, mc, 1);
1214 mutex_lock(&mc->mot_lov_mutex);
1216 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1217 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1219 mutex_unlock(&mc->mot_lov_mutex);
1221 GOTO(unlock_child, rc);
1223 if (!lu_object_is_dying(&mc->mot_header)) {
1224 rc = mdt_attr_get_complex(info, mc, ma);
1227 } else if (mdt_dom_check_for_discard(info, mc)) {
1228 mdt_dom_discard_data(info, mc);
1230 mdt_handle_last_unlink(info, mc, ma);
1233 if (ma->ma_valid & MA_INODE) {
1234 switch (ma->ma_attr.la_mode & S_IFMT) {
1236 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1237 ktime_us_delta(ktime_get(), kstart));
1245 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1246 ktime_us_delta(ktime_get(), kstart));
1249 LASSERTF(0, "bad file type %o unlinking\n",
1250 ma->ma_attr.la_mode);
1257 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1259 mdt_object_put(info->mti_env, mc);
1261 mdt_object_unlock(info, mp, parent_lh, rc);
1263 mdt_object_put(info->mti_env, mp);
1264 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1269 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1272 static int mdt_reint_link(struct mdt_thread_info *info,
1273 struct mdt_lock_handle *lhc)
1275 struct mdt_reint_record *rr = &info->mti_rr;
1276 struct ptlrpc_request *req = mdt_info_req(info);
1277 struct md_attr *ma = &info->mti_attr;
1278 struct mdt_object *ms;
1279 struct mdt_object *mp;
1280 struct mdt_lock_handle *lhs;
1281 struct mdt_lock_handle *lhp;
1282 ktime_t kstart = ktime_get();
1287 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1288 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1290 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1291 RETURN(err_serious(-ENOENT));
1293 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1294 req->rq_no_reply = 1;
1295 RETURN(err_serious(-ENOENT));
1298 if (info->mti_dlm_req)
1299 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1301 /* Invalid case so return error immediately instead of
1304 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1307 if (!fid_is_md_operative(rr->rr_fid1) ||
1308 !fid_is_md_operative(rr->rr_fid2))
1311 /* step 1: find target parent dir */
1312 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1314 RETURN(PTR_ERR(mp));
1316 rc = mdt_version_get_check_save(info, mp, 0);
1318 GOTO(put_parent, rc);
1320 /* step 2: find source */
1321 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1323 GOTO(put_parent, rc = PTR_ERR(ms));
1325 if (!mdt_object_exists(ms)) {
1326 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1327 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1328 GOTO(put_source, rc = -ENOENT);
1331 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1333 lhp = &info->mti_lh[MDT_LH_PARENT];
1334 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1335 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1338 GOTO(put_source, rc);
1340 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1342 lhs = &info->mti_lh[MDT_LH_CHILD];
1343 mdt_lock_reg_init(lhs, LCK_EX);
1344 rc = mdt_reint_object_lock(info, ms, lhs,
1345 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1348 GOTO(unlock_parent, rc);
1350 /* step 3: link it */
1351 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1352 OBD_FAIL_MDS_REINT_LINK_WRITE);
1354 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1355 rc = mdt_version_get_check_save(info, ms, 1);
1357 GOTO(unlock_source, rc);
1359 /** check target version by name during replay */
1360 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1361 &info->mti_tmp_fid1, 2);
1362 if (rc != 0 && rc != -ENOENT)
1363 GOTO(unlock_source, rc);
1364 /* save version of file name for replay, it must be ENOENT here */
1365 if (!req_is_replay(mdt_info_req(info))) {
1366 if (rc != -ENOENT) {
1367 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1368 PNAME(&rr->rr_name));
1369 GOTO(unlock_source, rc = -EEXIST);
1371 info->mti_ver[2] = ENOENT_VERSION;
1372 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1375 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1376 mdt_object_child(ms), &rr->rr_name, ma);
1379 mdt_counter_incr(req, LPROC_MDT_LINK,
1380 ktime_us_delta(ktime_get(), kstart));
1384 mdt_object_unlock(info, ms, lhs, rc);
1386 mdt_object_unlock(info, mp, lhp, rc);
1388 mdt_object_put(info->mti_env, ms);
1390 mdt_object_put(info->mti_env, mp);
1394 * lock the part of the directory according to the hash of the name
1395 * (lh->mlh_pdo_hash) in parallel directory lock.
1397 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1398 struct mdt_lock_handle *lh,
1399 struct mdt_object *obj, __u64 ibits,
1402 struct ldlm_res_id *res = &info->mti_res_id;
1403 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1404 union ldlm_policy_data *policy = &info->mti_policy;
1405 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1409 * Finish res_id initializing by name hash marking part of
1410 * directory which is taking modification.
1412 LASSERT(lh->mlh_pdo_hash != 0);
1413 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1414 memset(policy, 0, sizeof(*policy));
1415 policy->l_inodebits.bits = ibits;
1417 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1418 dlmflags |= LDLM_FL_COS_INCOMPAT;
1420 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1421 * going to be sent to client. If it is - mdt_intent_policy() path will
1422 * fix it up and turn FL_LOCAL flag off.
1424 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1425 policy, res, dlmflags,
1426 &info->mti_exp->exp_handle.h_cookie);
1431 * Get BFL lock for rename or migrate process.
1433 static int mdt_rename_lock(struct mdt_thread_info *info,
1434 struct lustre_handle *lh)
1439 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1440 struct lu_fid *fid = &info->mti_tmp_fid1;
1441 struct mdt_object *obj;
1443 /* XXX, right now, it has to use object API to
1444 * enqueue lock cross MDT, so it will enqueue
1445 * rename lock(with LUSTRE_BFL_FID) by root object
1448 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1450 RETURN(PTR_ERR(obj));
1452 rc = mdt_remote_object_lock(info, obj,
1453 &LUSTRE_BFL_FID, lh,
1455 MDS_INODELOCK_UPDATE, false);
1456 mdt_object_put(info->mti_env, obj);
1458 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1459 union ldlm_policy_data *policy = &info->mti_policy;
1460 struct ldlm_res_id *res_id = &info->mti_res_id;
1463 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1464 memset(policy, 0, sizeof(*policy));
1465 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1466 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1467 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1468 LDLM_IBITS, policy, LCK_EX, &flags,
1470 ldlm_completion_ast, NULL, NULL, 0,
1472 &info->mti_exp->exp_handle.h_cookie,
1479 static void mdt_rename_unlock(struct lustre_handle *lh)
1482 LASSERT(lustre_handle_is_used(lh));
1483 /* Cancel the single rename lock right away */
1484 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1488 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1489 const struct lu_fid *fid,
1492 struct mdt_object *dir;
1496 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1500 /* check early, the real version will be saved after locking */
1501 rc = mdt_version_get_check(info, dir, idx);
1505 if (!mdt_object_exists(dir))
1506 GOTO(out_put, rc = -ENOENT);
1508 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1509 GOTO(out_put, rc = -ENOTDIR);
1513 mdt_object_put(info->mti_env, dir);
1518 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1519 * herein we don't really check it, just do revoke.
1521 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1522 struct mdt_object *pobj,
1523 struct mdt_object *obj)
1525 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1528 mdt_lock_handle_init(lh);
1529 mdt_lock_reg_init(lh, LCK_EX);
1531 if (mdt_object_remote(pobj)) {
1532 /* don't bother to check if pobj and obj are on the same MDT. */
1533 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1534 &lh->mlh_rreg_lh, LCK_EX,
1535 MDS_INODELOCK_LOOKUP, false);
1536 } else if (mdt_object_remote(obj)) {
1537 struct ldlm_res_id *res = &info->mti_res_id;
1538 union ldlm_policy_data *policy = &info->mti_policy;
1539 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1540 LDLM_FL_COS_INCOMPAT;
1542 fid_build_reg_res_name(mdt_object_fid(obj), res);
1543 memset(policy, 0, sizeof(*policy));
1544 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1545 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1546 &lh->mlh_reg_lh, LCK_EX, policy, res,
1549 /* do nothing if both are local */
1557 * TODO, currently we don't save this lock because there is no place to
1558 * hold this lock handle, but to avoid race we need to save this lock.
1560 mdt_object_unlock(info, NULL, lh, 1);
1566 * operation may takes locks of linkea, or directory stripes, group them in
1569 struct mdt_sub_lock {
1570 struct mdt_object *msl_obj;
1571 struct mdt_lock_handle msl_lh;
1572 struct list_head msl_linkage;
1575 static void mdt_unlock_list(struct mdt_thread_info *info,
1576 struct list_head *list, int decref)
1578 struct mdt_sub_lock *msl;
1579 struct mdt_sub_lock *tmp;
1581 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1582 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1583 list_del(&msl->msl_linkage);
1588 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1589 struct mdt_object *obj,
1590 struct mdt_lock_handle *lh,
1591 struct ldlm_enqueue_info *einfo,
1592 struct list_head *slave_locks,
1595 if (mdt_object_remote(obj)) {
1596 mdt_unlock_list(info, slave_locks, decref);
1597 mdt_object_unlock(info, obj, lh, decref);
1599 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1604 * lock parents of links, and also check whether total locks don't exceed
1607 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1608 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1609 * \retval -ev negative errno upon error
1611 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1612 struct mdt_object *pobj,
1613 const struct md_attr *ma,
1614 struct mdt_object *obj,
1615 struct mdt_lock_handle *lhp,
1616 struct ldlm_enqueue_info *peinfo,
1617 struct list_head *parent_slave_locks,
1618 struct list_head *link_locks)
1620 struct mdt_device *mdt = info->mti_mdt;
1621 struct lu_buf *buf = &info->mti_big_buf;
1622 struct lu_name *lname = &info->mti_name;
1623 struct linkea_data ldata = { NULL };
1624 bool blocked = false;
1625 int local_lnkp_cnt = 0;
1629 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1632 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1633 if (buf->lb_buf == NULL)
1637 rc = mdt_links_read(info, obj, &ldata);
1639 if (rc == -ENOENT || rc == -ENODATA)
1644 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1645 linkea_next_entry(&ldata)) {
1646 struct mdt_object *lnkp;
1647 struct mdt_sub_lock *msl;
1651 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1654 /* check if it's also linked to parent */
1655 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1656 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1657 PFID(&fid), PNAME(lname));
1658 /* in case link is remote object, revoke LOOKUP lock */
1659 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1665 /* check if it's linked to a stripe of parent */
1666 if (ma->ma_valid & MA_LMV) {
1667 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1668 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1671 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1672 fid_le_to_cpu(stripe_fid,
1673 &lmv->lmv_stripe_fids[j]);
1674 if (lu_fid_eq(stripe_fid, &fid)) {
1675 CDEBUG(D_INFO, "skip stripe "DFID
1676 ", reovke "DNAME"\n",
1677 PFID(&fid), PNAME(lname));
1678 lnkp = mdt_object_find(info->mti_env,
1681 GOTO(out, rc = PTR_ERR(lnkp));
1687 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1689 mdt_object_put(info->mti_env, lnkp);
1694 /* Check if it's already locked */
1695 list_for_each_entry(msl, link_locks, msl_linkage) {
1696 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1698 DFID" was locked, revoke "DNAME"\n",
1699 PFID(&fid), PNAME(lname));
1700 lnkp = msl->msl_obj;
1706 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1710 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1711 PFID(&fid), PNAME(lname));
1713 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1715 CWARN("%s: cannot find obj "DFID": %ld\n",
1716 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1720 if (!mdt_object_exists(lnkp)) {
1721 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1722 PFID(&fid), PNAME(lname));
1723 mdt_object_put(info->mti_env, lnkp);
1727 if (!mdt_object_remote(lnkp))
1732 GOTO(out, rc = -ENOMEM);
1735 * we can't follow parent-child lock order like other MD
1736 * operations, use lock_try here to avoid deadlock, if the lock
1737 * cannot be taken, drop all locks taken, revoke the blocked
1738 * one, and continue processing the remaining entries, and in
1739 * the end of the loop restart from beginning.
1741 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1743 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1744 MDS_INODELOCK_UPDATE, true);
1745 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1747 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1748 PFID(&fid), PNAME(lname));
1750 mdt_unlock_list(info, link_locks, 1);
1751 /* also unlock parent locks to avoid deadlock */
1753 mdt_migrate_object_unlock(info, pobj, lhp,
1760 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1761 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1762 MDS_INODELOCK_UPDATE);
1764 mdt_object_put(info->mti_env, lnkp);
1769 if (mdt_object_remote(lnkp)) {
1770 struct ldlm_lock *lock;
1773 * for remote object, set lock cb_atomic,
1774 * so lock can be released in blocking_ast()
1775 * immediately, then the next lock_try will
1776 * have better chance of success.
1778 lock = ldlm_handle2lock(
1779 &msl->msl_lh.mlh_rreg_lh);
1780 LASSERT(lock != NULL);
1781 lock_res_and_lock(lock);
1782 ldlm_set_atomic_cb(lock);
1783 unlock_res_and_lock(lock);
1784 LDLM_LOCK_PUT(lock);
1787 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1792 INIT_LIST_HEAD(&msl->msl_linkage);
1793 msl->msl_obj = lnkp;
1794 list_add_tail(&msl->msl_linkage, link_locks);
1796 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1800 GOTO(out, rc = -EBUSY);
1805 mdt_unlock_list(info, link_locks, rc);
1806 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1808 * parent may have 3 local objects: master object and 2 stripes
1809 * (if it's being migrated too); source may have 2 local
1810 * objects: master and 1 stripe; target has 1 local object.
1816 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1817 struct mdt_object *obj,
1818 const struct md_attr *ma,
1819 struct list_head *slave_locks)
1821 struct mdt_device *mdt = info->mti_mdt;
1822 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1823 struct lu_fid *fid = &info->mti_tmp_fid1;
1824 struct mdt_object *slave;
1825 struct mdt_sub_lock *msl;
1830 LASSERT(mdt_object_remote(obj));
1831 LASSERT(ma->ma_valid & MA_LMV);
1834 if (!lmv_is_sane(lmv))
1837 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1838 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1840 if (!fid_is_sane(fid))
1843 slave = mdt_object_find(info->mti_env, mdt, fid);
1845 GOTO(out, rc = PTR_ERR(slave));
1849 mdt_object_put(info->mti_env, slave);
1850 GOTO(out, rc = -ENOMEM);
1853 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1854 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1855 MDS_INODELOCK_UPDATE, true);
1858 mdt_object_put(info->mti_env, slave);
1862 INIT_LIST_HEAD(&msl->msl_linkage);
1863 msl->msl_obj = slave;
1864 list_add_tail(&msl->msl_linkage, slave_locks);
1870 mdt_unlock_list(info, slave_locks, rc);
1874 /* lock parent and its stripes */
1875 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1876 struct mdt_object *obj,
1877 const struct md_attr *ma,
1878 struct mdt_lock_handle *lh,
1879 struct ldlm_enqueue_info *einfo,
1880 struct list_head *slave_locks)
1884 if (mdt_object_remote(obj)) {
1885 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1886 &lh->mlh_rreg_lh, LCK_PW,
1887 MDS_INODELOCK_UPDATE, false);
1892 * if obj is remote and striped, lock its stripes explicitly
1893 * because it's not striped in LOD layer on this MDT.
1895 if (ma->ma_valid & MA_LMV) {
1896 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1898 mdt_object_unlock(info, obj, lh, rc);
1901 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1909 * in migration, object may be remote, and we need take full lock of it and its
1910 * stripes if it's directory, besides, object may be a remote object on its
1911 * parent, revoke its LOOKUP lock on where its parent is located.
1913 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1914 struct mdt_object *pobj,
1915 struct mdt_object *obj,
1916 struct mdt_lock_handle *lh,
1917 struct ldlm_enqueue_info *einfo,
1918 struct list_head *slave_locks)
1922 if (mdt_object_remote(obj)) {
1923 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1927 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1928 &lh->mlh_rreg_lh, LCK_EX,
1929 MDS_INODELOCK_FULL, false);
1934 * if obj is remote and striped, lock its stripes explicitly
1935 * because it's not striped in LOD layer on this MDT.
1937 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1938 struct md_attr *ma = &info->mti_attr;
1940 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1942 mdt_object_unlock(info, obj, lh, rc);
1946 if (ma->ma_valid & MA_LMV) {
1947 rc = mdt_lock_remote_slaves(info, obj, ma,
1950 mdt_object_unlock(info, obj, lh, rc);
1954 if (mdt_object_remote(pobj)) {
1955 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1960 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1968 * lookup source by name, if parent is striped directory, we need to find the
1969 * corresponding stripe where source is located, and then lookup there.
1971 * besides, if parent is migrating too, and file is already in target stripe,
1972 * this should be a redo of 'lfs migrate' on client side.
1974 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1975 struct mdt_object *pobj,
1976 const struct md_attr *ma,
1977 const struct lu_name *lname,
1978 struct mdt_object **spobj,
1979 struct mdt_object **sobj)
1981 const struct lu_env *env = info->mti_env;
1982 struct lu_fid *fid = &info->mti_tmp_fid1;
1983 struct mdt_object *stripe;
1986 if (ma->ma_valid & MA_LMV) {
1987 /* if parent is striped, lookup on corresponding stripe */
1988 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1990 if (!lmv_is_sane(lmv))
1993 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1998 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2000 stripe = mdt_object_find(env, info->mti_mdt, fid);
2002 return PTR_ERR(stripe);
2005 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2007 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2009 * if parent layout is changeing, and lookup child
2010 * failed on source stripe, lookup again on target
2011 * stripe, if it exists, it means previous migration
2012 * was interrupted, and current file was migrated
2015 mdt_object_put(env, stripe);
2017 rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2022 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2024 stripe = mdt_object_find(env, info->mti_mdt, fid);
2026 return PTR_ERR(stripe);
2029 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2030 fid, &info->mti_spec);
2031 mdt_object_put(env, stripe);
2032 return rc ?: -EALREADY;
2034 mdt_object_put(env, stripe);
2039 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2045 mdt_object_get(env, stripe);
2050 *sobj = mdt_object_find(env, info->mti_mdt, fid);
2051 if (IS_ERR(*sobj)) {
2052 mdt_object_put(env, stripe);
2053 rc = PTR_ERR(*sobj);
2061 /* end lease and close file for regular file */
2062 static int mdd_migrate_close(struct mdt_thread_info *info,
2063 struct mdt_object *obj)
2065 struct close_data *data;
2066 struct mdt_body *repbody;
2067 struct ldlm_lock *lease;
2072 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2074 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2078 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2083 lease = ldlm_handle2lock(&data->cd_handle);
2087 /* check if the lease was already canceled */
2088 lock_res_and_lock(lease);
2089 rc = ldlm_is_cancel(lease);
2090 unlock_res_and_lock(lease);
2094 LDLM_DEBUG(lease, DFID" lease broken",
2095 PFID(mdt_object_fid(obj)));
2099 * cancel server side lease, client side counterpart should have been
2100 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2102 ldlm_lock_cancel(lease);
2103 ldlm_reprocess_all(lease->l_resource, lease);
2104 LDLM_LOCK_PUT(lease);
2107 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2108 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2109 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2115 * migrate file in below steps:
2116 * 1. lock parent and its stripes
2117 * 2. lookup source by name
2118 * 3. lock parents of source links if source is not directory
2119 * 4. reject if source is in HSM
2120 * 5. take source open_sem and close file if source is regular file
2121 * 6. lock source and its stripes if it's directory
2122 * 7. lock target so subsequent change to it can trigger COS
2124 * 9. unlock above locks
2125 * 10. sync device if source has links
2127 int mdt_reint_migrate(struct mdt_thread_info *info,
2128 struct mdt_lock_handle *unused)
2130 const struct lu_env *env = info->mti_env;
2131 struct mdt_device *mdt = info->mti_mdt;
2132 struct ptlrpc_request *req = mdt_info_req(info);
2133 struct mdt_reint_record *rr = &info->mti_rr;
2134 struct lu_ucred *uc = mdt_ucred(info);
2135 struct md_attr *ma = &info->mti_attr;
2136 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2137 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2138 struct mdt_object *pobj;
2139 struct mdt_object *spobj = NULL;
2140 struct mdt_object *sobj = NULL;
2141 struct mdt_object *tobj;
2142 struct lustre_handle rename_lh = { 0 };
2143 struct mdt_lock_handle *lhp;
2144 struct mdt_lock_handle *lhs;
2145 struct mdt_lock_handle *lht;
2146 LIST_HEAD(parent_slave_locks);
2147 LIST_HEAD(child_slave_locks);
2148 LIST_HEAD(link_locks);
2149 int lock_retries = 5;
2150 bool open_sem_locked = false;
2151 bool do_sync = false;
2155 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2156 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2158 if (info->mti_dlm_req)
2159 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2161 if (!fid_is_md_operative(rr->rr_fid1) ||
2162 !fid_is_md_operative(rr->rr_fid2))
2165 /* don't allow migrate . or .. */
2166 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2169 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2172 if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2173 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2174 mdt->mdt_enable_remote_dir_gid != -1)
2178 * Note: do not enqueue rename lock for replay request, because
2179 * if other MDT holds rename lock, but being blocked to wait for
2180 * this MDT to finish its recovery, and the failover MDT can not
2181 * get rename lock, which will cause deadlock.
2183 * req is NULL if this is called by directory auto-split.
2185 if (req && !req_is_replay(req)) {
2186 rc = mdt_rename_lock(info, &rename_lh);
2188 CERROR("%s: can't lock FS for rename: rc = %d\n",
2189 mdt_obd_name(info->mti_mdt), rc);
2194 /* pobj is master object of parent */
2195 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2197 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2200 rc = mdt_version_get_check(info, pobj, 0);
2202 GOTO(put_parent, rc);
2205 if (!mdt_object_exists(pobj))
2206 GOTO(put_parent, rc = -ENOENT);
2208 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2209 GOTO(put_parent, rc = -ENOTDIR);
2211 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2213 GOTO(put_parent, rc);
2216 /* lock parent object */
2217 lhp = &info->mti_lh[MDT_LH_PARENT];
2218 mdt_lock_reg_init(lhp, LCK_PW);
2219 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2220 &parent_slave_locks);
2222 GOTO(put_parent, rc);
2225 * spobj is the corresponding stripe against name if pobj is striped
2226 * directory, which is the real parent, and no need to lock, because
2227 * we've taken full lock of pobj.
2229 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2231 GOTO(unlock_parent, rc);
2233 /* lock parents of source links, and revoke LOOKUP lock of links */
2234 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2235 &parent_slave_locks, &link_locks);
2236 if (rc == -EBUSY && lock_retries-- > 0) {
2237 mdt_object_put(env, sobj);
2238 mdt_object_put(env, spobj);
2243 GOTO(put_source, rc);
2246 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2247 * with one request, if total lock count exceeds this limit, we will
2248 * drop all locks after migration, and synchronous device in the end.
2252 /* TODO: DoM migration is not supported, migrate dirent only */
2253 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2254 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2256 GOTO(unlock_links, rc);
2258 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2259 info->mti_spec.sp_migrate_nsonly = 1;
2262 /* if migration HSM is allowed */
2263 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2264 ma->ma_need = MA_HSM;
2266 rc = mdt_attr_get_complex(info, sobj, ma);
2268 GOTO(unlock_links, rc);
2270 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2271 GOTO(unlock_links, rc = -EOPNOTSUPP);
2274 /* end lease and close file for regular file */
2275 if (info->mti_spec.sp_migrate_close) {
2276 /* try to hold open_sem so that nobody else can open the file */
2277 if (!down_write_trylock(&sobj->mot_open_sem)) {
2279 mdd_migrate_close(info, sobj);
2280 GOTO(unlock_links, rc = -EBUSY);
2282 open_sem_locked = true;
2283 rc = mdd_migrate_close(info, sobj);
2285 GOTO(unlock_open_sem, rc);
2290 lhs = &info->mti_lh[MDT_LH_OLD];
2291 mdt_lock_reg_init(lhs, LCK_EX);
2292 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2293 &child_slave_locks);
2295 GOTO(unlock_open_sem, rc);
2298 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2300 GOTO(unlock_source, rc = PTR_ERR(tobj));
2302 lht = &info->mti_lh[MDT_LH_NEW];
2303 mdt_lock_reg_init(lht, LCK_EX);
2304 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2306 GOTO(put_target, rc);
2308 /* Don't do lookup sanity check. We know name doesn't exist. */
2309 info->mti_spec.sp_cr_lookup = 0;
2310 info->mti_spec.sp_feat = &dt_directory_features;
2312 rc = mdo_migrate(env, mdt_object_child(pobj),
2313 mdt_object_child(sobj), &rr->rr_name,
2314 mdt_object_child(tobj),
2315 &info->mti_spec, ma);
2317 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2318 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2321 mdt_object_unlock(info, tobj, lht, rc);
2323 mdt_object_put(env, tobj);
2325 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2326 &child_slave_locks, rc);
2328 if (open_sem_locked)
2329 up_write(&sobj->mot_open_sem);
2331 /* if we've got too many locks to save into RPC,
2332 * then just commit before the locks are released
2335 mdt_device_sync(env, mdt);
2336 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2338 mdt_object_put(env, sobj);
2339 mdt_object_put(env, spobj);
2341 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2342 &parent_slave_locks, rc);
2344 mdt_object_put(env, pobj);
2346 if (lustre_handle_is_used(&rename_lh))
2347 mdt_rename_unlock(&rename_lh);
2352 static int mdt_object_lock_save(struct mdt_thread_info *info,
2353 struct mdt_object *dir,
2354 struct mdt_lock_handle *lh,
2355 int idx, bool cos_incompat)
2359 /* we lock the target dir if it is local */
2360 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2365 /* get and save correct version after locking */
2366 mdt_version_get_save(info, dir, idx);
2371 * determine lock order of sobj and tobj
2373 * there are two situations we need to lock tobj before sobj:
2374 * 1. sobj is child of tobj
2375 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2376 * larger than that of tobj
2378 * \retval 1 lock tobj before sobj
2379 * \retval 0 lock sobj before tobj
2380 * \retval -ev negative errno upon error
2382 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2383 struct mdt_object *sobj,
2384 struct mdt_object *tobj)
2386 struct md_attr *ma = &info->mti_attr;
2387 struct lu_fid *spfid = &info->mti_tmp_fid1;
2388 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2389 struct lmv_mds_md_v1 *lmv;
2394 /* sobj and tobj are the same */
2398 if (fid_is_root(mdt_object_fid(sobj)))
2401 if (fid_is_root(mdt_object_fid(tobj)))
2404 /* check whether sobj is child of tobj */
2405 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2406 mdt_object_fid(tobj));
2413 /* check whether sobj and tobj are children of the same parent */
2414 rc = mdt_attr_get_pfid(info, sobj, spfid);
2418 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2422 if (!lu_fid_eq(spfid, tpfid))
2425 /* check whether sobj and tobj are sibling stripes */
2426 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2430 if (!(ma->ma_valid & MA_LMV))
2433 lmv = &ma->ma_lmv->lmv_md_v1;
2434 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2436 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2439 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2443 if (!(ma->ma_valid & MA_LMV))
2446 lmv = &ma->ma_lmv->lmv_md_v1;
2447 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2449 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2451 /* check stripe index of sobj and tobj */
2452 if (sindex == tindex)
2455 return sindex < tindex ? 0 : 1;
2459 * lock rename source object.
2461 * Both source and source parent may be remote, and source may be a remote
2462 * object on source parent, to avoid overriding lock handle, store remote
2463 * LOOKUP lock separately in @lhr.
2465 * \retval 0 on success
2466 * \retval -ev negative errno upon error
2468 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2469 struct mdt_object *parent,
2470 struct mdt_object *child,
2471 struct mdt_lock_handle *lhc,
2472 struct mdt_lock_handle *lhr,
2478 rc = mdt_is_remote_object(info, parent, child);
2483 /* enqueue remote LOOKUP lock from the parent MDT */
2484 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2486 if (mdt_object_remote(parent)) {
2487 rc = mdt_remote_object_lock(info, parent,
2488 mdt_object_fid(child),
2495 LASSERT(mdt_object_remote(child));
2496 rc = mdt_object_local_lock(info, child, lhr,
2497 &rmt_ibits, 0, true);
2502 ibits &= ~MDS_INODELOCK_LOOKUP;
2505 if (mdt_object_remote(child)) {
2506 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2513 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2518 mdt_object_unlock(info, child, lhr, rc);
2524 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2525 * 2 - srcdir child; 3 - tgtdir child.
2526 * Update on disk version of srcdir child.
2528 static int mdt_reint_rename(struct mdt_thread_info *info,
2529 struct mdt_lock_handle *unused)
2531 struct mdt_device *mdt = info->mti_mdt;
2532 struct mdt_reint_record *rr = &info->mti_rr;
2533 struct md_attr *ma = &info->mti_attr;
2534 struct ptlrpc_request *req = mdt_info_req(info);
2535 struct mdt_object *msrcdir = NULL;
2536 struct mdt_object *mtgtdir = NULL;
2537 struct mdt_object *mold;
2538 struct mdt_object *mnew = NULL;
2539 struct lustre_handle rename_lh = { 0 };
2540 struct mdt_lock_handle *lh_srcdirp;
2541 struct mdt_lock_handle *lh_tgtdirp;
2542 struct mdt_lock_handle *lh_oldp = NULL;
2543 struct mdt_lock_handle *lh_rmt = NULL;
2544 struct mdt_lock_handle *lh_newp = NULL;
2545 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2546 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2548 bool reverse = false, discard = false;
2550 ktime_t kstart = ktime_get();
2554 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2555 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2556 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2558 if (info->mti_dlm_req)
2559 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2561 if (!fid_is_md_operative(rr->rr_fid1) ||
2562 !fid_is_md_operative(rr->rr_fid2))
2565 /* find both parents. */
2566 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2567 if (IS_ERR(msrcdir))
2568 RETURN(PTR_ERR(msrcdir));
2570 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2572 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2574 mdt_object_get(info->mti_env, mtgtdir);
2576 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2577 if (IS_ERR(mtgtdir))
2578 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2582 * Note: do not enqueue rename lock for replay request, because
2583 * if other MDT holds rename lock, but being blocked to wait for
2584 * this MDT to finish its recovery, and the failover MDT can not
2585 * get rename lock, which will cause deadlock.
2587 if (!req_is_replay(req)) {
2589 * Normally rename RPC is handled on the MDT with the target
2590 * directory (if target exists, it's on the MDT with the
2591 * target), if the source directory is remote, it's a hint that
2592 * source is remote too (this may not be true, but it won't
2593 * cause any issue), return -EXDEV early to avoid taking
2596 if (!mdt->mdt_enable_remote_rename &&
2597 mdt_object_remote(msrcdir))
2598 GOTO(out_put_tgtdir, rc = -EXDEV);
2600 rc = mdt_rename_lock(info, &rename_lh);
2602 CERROR("%s: can't lock FS for rename: rc = %d\n",
2603 mdt_obd_name(mdt), rc);
2604 GOTO(out_put_tgtdir, rc);
2608 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2610 GOTO(out_unlock_rename, rc);
2614 /* source needs to be looked up after locking source parent, otherwise
2615 * this rename may race with unlink source, and cause rename hang, see
2616 * sanityn.sh 55b, so check parents first, if later we found source is
2617 * remote, relock parents.
2619 cos_incompat = (mdt_object_remote(msrcdir) ||
2620 mdt_object_remote(mtgtdir));
2622 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2624 /* lock parents in the proper order. */
2625 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2626 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2629 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2630 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2633 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2636 GOTO(out_unlock_rename, rc);
2638 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2640 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2643 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2644 GOTO(out_unlock_rename, rc);
2647 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2650 GOTO(out_unlock_rename, rc);
2652 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2654 if (mtgtdir != msrcdir) {
2655 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2657 } else if (!mdt_object_remote(mtgtdir) &&
2658 lh_srcdirp->mlh_pdo_hash !=
2659 lh_tgtdirp->mlh_pdo_hash) {
2660 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2661 MDS_INODELOCK_UPDATE,
2663 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2666 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2667 GOTO(out_unlock_rename, rc);
2671 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2672 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2674 /* find mold object. */
2676 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2678 GOTO(out_unlock_parents, rc);
2680 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2681 GOTO(out_unlock_parents, rc = -EINVAL);
2683 if (!fid_is_md_operative(old_fid))
2684 GOTO(out_unlock_parents, rc = -EPERM);
2686 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2688 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2690 if (!mdt_object_exists(mold)) {
2691 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2693 "object does not exist");
2694 GOTO(out_put_old, rc = -ENOENT);
2697 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2698 GOTO(out_put_old, rc = -EXDEV);
2700 /* Check if @mtgtdir is subdir of @mold, before locking child
2701 * to avoid reverse locking.
2703 if (mtgtdir != msrcdir) {
2704 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2709 GOTO(out_put_old, rc);
2713 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2714 /* save version after locking */
2715 mdt_version_get_save(info, mold, 2);
2717 if (!cos_incompat && mdt_object_remote(mold)) {
2718 cos_incompat = true;
2719 mdt_object_put(info->mti_env, mold);
2720 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2721 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2725 /* find mnew object:
2726 * mnew target object may not exist now
2727 * lookup with version checking
2730 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2733 /* the new_fid should have been filled at this moment */
2734 if (lu_fid_eq(old_fid, new_fid))
2735 GOTO(out_put_old, rc);
2737 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2738 lu_fid_eq(new_fid, rr->rr_fid2))
2739 GOTO(out_put_old, rc = -EINVAL);
2741 if (!fid_is_md_operative(new_fid))
2742 GOTO(out_put_old, rc = -EPERM);
2744 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2746 GOTO(out_put_old, rc = PTR_ERR(mnew));
2748 if (!mdt_object_exists(mnew)) {
2749 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2751 "object does not exist");
2752 GOTO(out_put_new, rc = -ENOENT);
2755 if (mdt_object_remote(mnew)) {
2756 struct mdt_body *repbody;
2758 /* Always send rename req to the target child MDT */
2759 repbody = req_capsule_server_get(info->mti_pill,
2761 LASSERT(repbody != NULL);
2762 repbody->mbo_fid1 = *new_fid;
2763 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2764 GOTO(out_put_new, rc = -EXDEV);
2766 /* Before locking the target dir, check we do not replace
2767 * a dir with a non-dir, otherwise it may deadlock with
2768 * link op which tries to create a link in this dir
2769 * back to this non-dir.
2771 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2772 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2773 GOTO(out_put_new, rc = -EISDIR);
2775 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2776 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2777 mdt_lock_reg_init(lh_oldp, LCK_EX);
2778 mdt_lock_reg_init(lh_rmt, LCK_EX);
2779 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2780 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2781 lh_rmt, lock_ibits, cos_incompat);
2783 GOTO(out_put_new, rc);
2785 /* Check if @msrcdir is subdir of @mnew, before locking child
2786 * to avoid reverse locking.
2788 if (mtgtdir != msrcdir) {
2789 rc = mdo_is_subdir(info->mti_env,
2790 mdt_object_child(msrcdir), new_fid);
2794 GOTO(out_unlock_old, rc);
2798 /* We used to acquire MDS_INODELOCK_FULL here but we
2799 * can't do this now because a running HSM restore on
2800 * the rename onto victim will hold the layout
2801 * lock. See LU-4002.
2804 lh_newp = &info->mti_lh[MDT_LH_NEW];
2805 mdt_lock_reg_init(lh_newp, LCK_EX);
2806 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2807 if (mdt_object_remote(mtgtdir)) {
2808 rc = mdt_remote_object_lock(info, mtgtdir,
2809 mdt_object_fid(mnew),
2810 &lh_newp->mlh_rreg_lh,
2811 lh_newp->mlh_rreg_mode,
2812 MDS_INODELOCK_LOOKUP,
2815 GOTO(out_unlock_old, rc);
2817 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2819 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2822 GOTO(out_unlock_new, rc);
2824 /* get and save version after locking */
2825 mdt_version_get_save(info, mnew, 3);
2826 } else if (rc != -ENOENT) {
2827 GOTO(out_put_old, rc);
2829 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2830 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2831 mdt_lock_reg_init(lh_oldp, LCK_EX);
2832 mdt_lock_reg_init(lh_rmt, LCK_EX);
2833 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2834 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2835 lh_rmt, lock_ibits, cos_incompat);
2837 GOTO(out_put_old, rc);
2839 mdt_enoent_version_save(info, 3);
2842 /* step 5: rename it */
2843 mdt_reint_init_ma(info, ma);
2845 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2846 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2849 mutex_lock(&mnew->mot_lov_mutex);
2851 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2852 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2853 mnew != NULL ? mdt_object_child(mnew) : NULL,
2854 &rr->rr_tgt_name, ma);
2857 mutex_unlock(&mnew->mot_lov_mutex);
2859 /* handle last link of tgt object */
2861 mdt_counter_incr(req, LPROC_MDT_RENAME,
2862 ktime_us_delta(ktime_get(), kstart));
2864 mdt_handle_last_unlink(info, mnew, ma);
2865 discard = mdt_dom_check_for_discard(info, mnew);
2867 mdt_rename_counter_tally(info, info->mti_mdt, req,
2869 ktime_us_delta(ktime_get(), kstart));
2875 mdt_object_unlock(info, mnew, lh_newp, rc);
2877 mdt_object_unlock(info, NULL, lh_rmt, rc);
2878 mdt_object_unlock(info, mold, lh_oldp, rc);
2880 if (mnew && !discard)
2881 mdt_object_put(info->mti_env, mnew);
2883 mdt_object_put(info->mti_env, mold);
2885 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2886 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2888 if (lustre_handle_is_used(&rename_lh))
2889 mdt_rename_unlock(&rename_lh);
2891 mdt_object_put(info->mti_env, mtgtdir);
2893 mdt_object_put(info->mti_env, msrcdir);
2895 /* The DoM discard can be done right in the place above where it is
2896 * assigned, meanwhile it is done here after rename unlock due to
2897 * compatibility with old clients, for them the discard blocks
2898 * the main thread until completion. Check LU-11359 for details.
2901 mdt_dom_discard_data(info, mnew);
2902 mdt_object_put(info->mti_env, mnew);
2907 static int mdt_reint_resync(struct mdt_thread_info *info,
2908 struct mdt_lock_handle *lhc)
2910 struct mdt_reint_record *rr = &info->mti_rr;
2911 struct ptlrpc_request *req = mdt_info_req(info);
2912 struct md_attr *ma = &info->mti_attr;
2913 struct mdt_object *mo;
2914 struct ldlm_lock *lease;
2915 struct mdt_body *repbody;
2916 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2921 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2923 if (info->mti_dlm_req)
2924 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2926 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2928 GOTO(out, rc = PTR_ERR(mo));
2930 if (!mdt_object_exists(mo))
2931 GOTO(out_obj, rc = -ENOENT);
2933 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2934 GOTO(out_obj, rc = -EINVAL);
2936 if (mdt_object_remote(mo))
2937 GOTO(out_obj, rc = -EREMOTE);
2939 lease = ldlm_handle2lock(rr->rr_lease_handle);
2941 GOTO(out_obj, rc = -ESTALE);
2943 /* It's really necessary to grab open_sem and check if the lease lock
2944 * has been lost. There would exist a concurrent writer coming in and
2945 * generating some dirty data in memory cache, the writeback would fail
2946 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2948 if (!down_write_trylock(&mo->mot_open_sem))
2949 GOTO(out_put_lease, rc = -EBUSY);
2951 lock_res_and_lock(lease);
2952 lease_broken = ldlm_is_cancel(lease);
2953 unlock_res_and_lock(lease);
2955 GOTO(out_unlock, rc = -EBUSY);
2957 /* the file has yet opened by anyone else after we took the lease. */
2958 layout.mlc_opc = MD_LAYOUT_RESYNC;
2959 lhc = &info->mti_lh[MDT_LH_LOCAL];
2960 rc = mdt_layout_change(info, mo, lhc, &layout);
2962 GOTO(out_unlock, rc);
2964 mdt_object_unlock(info, mo, lhc, 0);
2966 ma->ma_need = MA_INODE;
2968 rc = mdt_attr_get_complex(info, mo, ma);
2970 GOTO(out_unlock, rc);
2972 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2973 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2977 up_write(&mo->mot_open_sem);
2979 LDLM_LOCK_PUT(lease);
2981 mdt_object_put(info->mti_env, mo);
2983 mdt_client_compatibility(info);
2984 rc2 = mdt_fix_reply(info);
2990 struct mdt_reinter {
2991 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2992 enum lprocfs_extra_opc mr_extra_opc;
2995 static const struct mdt_reinter mdt_reinters[] = {
2997 .mr_handler = &mdt_reint_setattr,
2998 .mr_extra_opc = MDS_REINT_SETATTR,
3001 .mr_handler = &mdt_reint_create,
3002 .mr_extra_opc = MDS_REINT_CREATE,
3005 .mr_handler = &mdt_reint_link,
3006 .mr_extra_opc = MDS_REINT_LINK,
3009 .mr_handler = &mdt_reint_unlink,
3010 .mr_extra_opc = MDS_REINT_UNLINK,
3013 .mr_handler = &mdt_reint_rename,
3014 .mr_extra_opc = MDS_REINT_RENAME,
3017 .mr_handler = &mdt_reint_open,
3018 .mr_extra_opc = MDS_REINT_OPEN,
3020 [REINT_SETXATTR] = {
3021 .mr_handler = &mdt_reint_setxattr,
3022 .mr_extra_opc = MDS_REINT_SETXATTR,
3025 .mr_handler = &mdt_reint_unlink,
3026 .mr_extra_opc = MDS_REINT_UNLINK,
3029 .mr_handler = &mdt_reint_migrate,
3030 .mr_extra_opc = MDS_REINT_RENAME,
3033 .mr_handler = &mdt_reint_resync,
3034 .mr_extra_opc = MDS_REINT_RESYNC,
3038 int mdt_reint_rec(struct mdt_thread_info *info,
3039 struct mdt_lock_handle *lhc)
3041 const struct mdt_reinter *mr;
3045 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3048 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3049 if (mr->mr_handler == NULL)
3052 rc = (*mr->mr_handler)(info, lhc);
3054 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3055 PTLRPC_LAST_CNTR + mr->mr_extra_opc);