4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/mdt/mdt_reint.c
33 * Lustre Metadata Target (mdt) reintegration routines
35 * Author: Peter Braam <braam@clusterfs.com>
36 * Author: Andreas Dilger <adilger@clusterfs.com>
37 * Author: Phil Schwan <phil@clusterfs.com>
38 * Author: Huang Hua <huanghua@clusterfs.com>
39 * Author: Yury Umanets <umka@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_MDS
44 #include <lprocfs_status.h>
45 #include "mdt_internal.h"
46 #include <lustre_lmv.h>
47 #include <lustre_crypto.h>
49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
52 ma->ma_need = MA_INODE;
57 * Get version of object by fid.
59 * Return real version or ENOENT_VERSION if object doesn't exist
61 static void mdt_obj_version_get(struct mdt_thread_info *info,
62 struct mdt_object *o, __u64 *version)
66 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
67 !fid_is_obf(mdt_object_fid(o)))
68 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
70 *version = ENOENT_VERSION;
71 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
72 PFID(mdt_object_fid(o)), *version);
76 * Check version is correct.
78 * Should be called only during replay.
80 static int mdt_version_check(struct ptlrpc_request *req,
81 __u64 version, int idx)
83 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
86 if (!exp_connect_vbr(req->rq_export))
89 LASSERT(req_is_replay(req));
90 /** VBR: version is checked always because costs nothing */
91 LASSERT(idx < PTLRPC_NUM_VERSIONS);
92 /** Sanity check for malformed buffers */
93 if (pre_ver == NULL) {
94 CERROR("No versions in request buffer\n");
95 spin_lock(&req->rq_export->exp_lock);
96 req->rq_export->exp_vbr_failed = 1;
97 spin_unlock(&req->rq_export->exp_lock);
99 } else if (pre_ver[idx] != version) {
100 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
101 pre_ver[idx], version);
102 spin_lock(&req->rq_export->exp_lock);
103 req->rq_export->exp_vbr_failed = 1;
104 spin_unlock(&req->rq_export->exp_lock);
111 * Save pre-versions in reply.
113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
118 if (!exp_connect_vbr(req->rq_export))
121 LASSERT(!req_is_replay(req));
122 LASSERT(req->rq_repmsg != NULL);
123 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
125 reply_ver[idx] = version;
129 * Save enoent version, it is needed when it is obvious that object doesn't
130 * exist, e.g. child during create.
132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
134 /* save version of file name for replay, it must be ENOENT here */
135 if (!req_is_replay(mdt_info_req(info))) {
136 info->mti_ver[idx] = ENOENT_VERSION;
137 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
142 * Get version from disk and save in reply buffer.
144 * Versions are saved in reply only during normal operations not replays.
146 void mdt_version_get_save(struct mdt_thread_info *info,
147 struct mdt_object *mto, int idx)
149 /* don't save versions during replay */
150 if (!req_is_replay(mdt_info_req(info))) {
151 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
152 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
157 * Get version from disk and check it, no save in reply.
159 int mdt_version_get_check(struct mdt_thread_info *info,
160 struct mdt_object *mto, int idx)
162 /* only check versions during replay */
163 if (!req_is_replay(mdt_info_req(info)))
166 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
167 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and check if recovery or just save.
173 int mdt_version_get_check_save(struct mdt_thread_info *info,
174 struct mdt_object *mto, int idx)
178 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
179 if (req_is_replay(mdt_info_req(info)))
180 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
183 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
188 * Lookup with version checking.
190 * This checks version of 'name'. Many reint functions uses 'name' for child not
191 * FID, therefore we need to get object by name and check its version.
193 int mdt_lookup_version_check(struct mdt_thread_info *info,
194 struct mdt_object *p,
195 const struct lu_name *lname,
196 struct lu_fid *fid, int idx)
200 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
202 /* Check version only during replay */
203 if (!req_is_replay(mdt_info_req(info)))
206 info->mti_ver[idx] = ENOENT_VERSION;
208 struct mdt_object *child;
210 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
211 if (likely(!IS_ERR(child))) {
212 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
213 mdt_object_put(info->mti_env, child);
216 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
217 return vbrc ? vbrc : rc;
221 static int mdt_stripes_unlock(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 *locks = einfo->ei_cbdata;
231 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
234 memset(policy, 0, sizeof(*policy));
235 policy->l_inodebits.bits = einfo->ei_inodebits;
236 mdt_lock_reg_init(lh, einfo->ei_mode);
237 for (i = 0; i < locks->ha_count; i++) {
238 if (test_bit(i, (void *)locks->ha_map))
239 lh->mlh_rreg_lh = locks->ha_handles[i];
241 lh->mlh_reg_lh = locks->ha_handles[i];
242 mdt_object_unlock(mti, NULL, lh, decref);
243 locks->ha_handles[i].cookie = 0ull;
246 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
251 * Lock slave stripes if necessary, the lock handles of slave stripes
252 * will be stored in einfo->ei_cbdata.
254 static int mdt_stripes_lock(struct mdt_thread_info *mti, struct mdt_object *obj,
255 enum ldlm_mode mode, __u64 ibits,
256 struct ldlm_enqueue_info *einfo)
258 union ldlm_policy_data *policy = &mti->mti_policy;
260 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
261 einfo->ei_type = LDLM_IBITS;
262 einfo->ei_mode = mode;
263 einfo->ei_cb_bl = mdt_remote_blocking_ast;
264 einfo->ei_cb_local_bl = mdt_blocking_ast;
265 einfo->ei_cb_cp = ldlm_completion_ast;
266 einfo->ei_enq_slave = 1;
267 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
268 einfo->ei_inodebits = ibits;
269 einfo->ei_req_slot = 1;
270 memset(policy, 0, sizeof(*policy));
271 policy->l_inodebits.bits = ibits;
272 policy->l_inodebits.li_initiator_id = mdt_node_id(mti->mti_mdt);
274 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
278 /** lock object, and stripes if it's a striped directory
280 * object should be local, this is called in operations which modify both object
283 * \param info struct mdt_thread_info
284 * \param parent parent object, if it's NULL, find parent by mdo_lookup()
285 * \param child child object
286 * \param lh lock handle
287 * \param einfo struct ldlm_enqueue_info
288 * \param ibits MDS inode lock bits
289 * \param mode lock mode
291 * \retval 0 on success, -ev on error.
293 int mdt_object_stripes_lock(struct mdt_thread_info *info,
294 struct mdt_object *parent,
295 struct mdt_object *child,
296 struct mdt_lock_handle *lh,
297 struct ldlm_enqueue_info *einfo, __u64 ibits,
303 /* according to the protocol, child should be local, is request sent to
306 if (mdt_object_remote(child)) {
307 CERROR("%s: lock target "DFID", but it is on other MDT: rc = %d\n",
308 mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(child)),
313 memset(einfo, 0, sizeof(*einfo));
314 if (ibits & MDS_INODELOCK_LOOKUP) {
316 rc = mdt_object_check_lock(info, parent, child, lh, ibits,
319 rc = mdt_object_lock(info, child, lh, ibits, mode);
324 if (!S_ISDIR(child->mot_header.loh_attr))
327 /* lock stripes for striped directory */
328 rc = mdt_stripes_lock(info, child, lh->mlh_reg_mode, ibits, einfo);
329 if (rc == -EIO && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
332 mdt_object_unlock(info, child, lh, rc);
337 void mdt_object_stripes_unlock(struct mdt_thread_info *info,
338 struct mdt_object *obj,
339 struct mdt_lock_handle *lh,
340 struct ldlm_enqueue_info *einfo, int decref)
342 if (einfo->ei_cbdata)
343 mdt_stripes_unlock(info, obj, einfo, decref);
344 mdt_object_unlock(info, obj, lh, decref);
347 static int mdt_restripe(struct mdt_thread_info *info,
348 struct mdt_object *parent,
349 const struct lu_name *lname,
350 const struct lu_fid *tfid,
351 struct md_op_spec *spec,
354 struct mdt_device *mdt = info->mti_mdt;
355 struct lu_fid *fid = &info->mti_tmp_fid2;
356 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
357 struct lmv_user_md *lum = spec->u.sp_ea.eadata;
358 struct lu_ucred *uc = mdt_ucred(info);
359 struct lmv_mds_md_v1 *lmv;
360 struct mdt_object *child;
361 struct mdt_lock_handle *lhp;
362 struct mdt_lock_handle *lhc;
363 struct mdt_body *repbody;
368 /* we want rbac roles to have precedence over any other
369 * permission or capability checks
371 if (!mdt->mdt_enable_dir_restripe && !uc->uc_rbac_dne_ops)
375 lum->lum_hash_type |= cpu_to_le32(LMV_HASH_FLAG_FIXED);
377 rc = mdt_version_get_check_save(info, parent, 0);
381 lhp = &info->mti_lh[MDT_LH_PARENT];
382 rc = mdt_parent_lock(info, parent, lhp, lname, LCK_PW);
386 rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
388 GOTO(unlock_parent, rc);
390 if (ma->ma_valid & MA_LMV) {
391 /* don't allow restripe if parent dir layout is changing */
392 lmv = &ma->ma_lmv->lmv_md_v1;
393 if (!lmv_is_sane2(lmv))
394 GOTO(unlock_parent, rc = -EBADF);
396 if (lmv_is_layout_changing(lmv))
397 GOTO(unlock_parent, rc = -EBUSY);
401 rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
403 GOTO(unlock_parent, rc);
405 child = mdt_object_find(info->mti_env, mdt, fid);
407 GOTO(unlock_parent, rc = PTR_ERR(child));
409 if (!mdt_object_exists(child))
410 GOTO(out_child, rc = -ENOENT);
412 if (mdt_object_remote(child)) {
413 struct mdt_body *repbody;
415 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
417 GOTO(out_child, rc = -EPROTO);
419 repbody->mbo_fid1 = *fid;
420 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
421 GOTO(out_child, rc = -EREMOTE);
424 if (!S_ISDIR(lu_object_attr(&child->mot_obj)))
425 GOTO(out_child, rc = -ENOTDIR);
427 rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
431 /* race with migrate? */
432 if ((ma->ma_valid & MA_LMV) &&
433 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1))
434 GOTO(out_child, rc = -EBUSY);
437 lhc = &info->mti_lh[MDT_LH_CHILD];
438 rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
439 MDS_INODELOCK_FULL, LCK_PW);
441 GOTO(unlock_child, rc);
443 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
444 rc = mdt_version_get_check_save(info, child, 1);
446 GOTO(unlock_child, rc);
448 spin_lock(&mdt->mdt_restriper.mdr_lock);
449 if (child->mot_restriping) {
451 spin_unlock(&mdt->mdt_restriper.mdr_lock);
452 GOTO(unlock_child, rc = -EBUSY);
454 child->mot_restriping = 1;
455 spin_unlock(&mdt->mdt_restriper.mdr_lock);
458 rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
460 GOTO(restriping_clear, rc);
462 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
464 GOTO(restriping_clear, rc = -EPROTO);
466 mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
470 child->mot_restriping = 0;
472 mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
474 mdt_object_put(info->mti_env, child);
476 mdt_object_unlock(info, parent, lhp, rc);
482 * VBR: we save three versions in reply:
483 * 0 - parent. Check that parent version is the same during replay.
484 * 1 - name. Version of 'name' if file exists with the same name or
485 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
486 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
489 static int mdt_create(struct mdt_thread_info *info)
491 struct mdt_device *mdt = info->mti_mdt;
492 struct mdt_object *parent;
493 struct mdt_object *child;
494 struct mdt_lock_handle *lh;
495 struct mdt_body *repbody;
496 struct md_attr *ma = &info->mti_attr;
497 struct mdt_reint_record *rr = &info->mti_rr;
498 struct md_op_spec *spec = &info->mti_spec;
499 struct lu_ucred *uc = mdt_ucred(info);
500 bool restripe = false;
504 DEBUG_REQ(D_INODE, mdt_info_req(info),
505 "Create ("DNAME"->"DFID") in "DFID,
506 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
508 if (!fid_is_md_operative(rr->rr_fid1))
511 /* MDS_OPEN_DEFAULT_LMV means eadata is parent default LMV, which is set
512 * if client maintains inherited default LMV
514 if (S_ISDIR(ma->ma_attr.la_mode) &&
515 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0 &&
516 !(spec->sp_cr_flags & MDS_OPEN_DEFAULT_LMV)) {
517 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
518 struct obd_export *exp = mdt_info_req(info)->rq_export;
520 /* Only new clients can create remote dir( >= 2.4) and
521 * striped dir(>= 2.6), old client will return -ENOTSUPP
523 if (!mdt_is_dne_client(exp))
526 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
527 if (!mdt_is_striped_client(exp))
530 if (!mdt->mdt_enable_striped_dir)
532 } else if (!mdt->mdt_enable_remote_dir) {
536 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
537 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) >=
541 /* we want rbac roles to have precedence over any other
542 * permission or capability checks
544 if (!uc->uc_rbac_dne_ops ||
545 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
546 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
547 mdt->mdt_enable_remote_dir_gid != -1))
550 /* restripe if later found dir exists, MDS_OPEN_CREAT means
551 * this is create only, don't try restripe.
553 if (mdt->mdt_enable_dir_restripe &&
554 le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
555 !(spec->sp_cr_flags & MDS_OPEN_CREAT))
559 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
561 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
563 RETURN(PTR_ERR(parent));
565 if (!mdt_object_exists(parent))
566 GOTO(put_parent, rc = -ENOENT);
568 rc = mdt_check_enc(info, parent);
570 GOTO(put_parent, rc);
572 if (!uc->uc_rbac_fscrypt_admin &&
573 parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
574 GOTO(put_parent, rc = -EPERM);
577 * LU-10235: check if name exists locklessly first to avoid massive
578 * lock recalls on existing directories.
580 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
581 &info->mti_tmp_fid1, 1);
584 GOTO(put_parent, rc = -EEXIST);
586 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
590 /* -ENOENT is expected here */
592 GOTO(put_parent, rc);
594 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_CREATE_AFTER_LOOKUP, cfs_fail_val);
596 /* save version of file name for replay, it must be ENOENT here */
597 mdt_enoent_version_save(info, 1);
599 CFS_RACE(OBD_FAIL_MDS_CREATE_RACE);
601 lh = &info->mti_lh[MDT_LH_PARENT];
602 rc = mdt_parent_lock(info, parent, lh, &rr->rr_name, LCK_PW);
604 GOTO(put_parent, rc);
606 if (!mdt_object_remote(parent)) {
607 rc = mdt_version_get_check_save(info, parent, 0);
609 GOTO(unlock_parent, rc);
613 * now repeat the lookup having a LDLM lock on the parent dir,
614 * as another thread could create the same name. notice this
615 * lookup is supposed to hit cache in OSD and be cheap if the
616 * directory is not being modified concurrently.
618 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), &rr->rr_name,
619 &info->mti_tmp_fid1, &info->mti_spec);
620 if (unlikely(rc == 0))
621 GOTO(unlock_parent, rc = -EEXIST);
623 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
624 if (unlikely(IS_ERR(child)))
625 GOTO(unlock_parent, rc = PTR_ERR(child));
627 ma->ma_need = MA_INODE;
630 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
631 OBD_FAIL_MDS_REINT_CREATE_WRITE);
633 /* Version of child will be updated on disk. */
634 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
635 rc = mdt_version_get_check_save(info, child, 2);
639 if (parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD ||
640 (rr->rr_name.ln_namelen == strlen(dot_fscrypt_name) &&
641 strncmp(rr->rr_name.ln_name, dot_fscrypt_name,
642 rr->rr_name.ln_namelen) == 0))
643 child->mot_obj.lo_header->loh_attr |= LOHA_FSCRYPT_MD;
646 * Do not perform lookup sanity check. We know that name does
649 info->mti_spec.sp_cr_lookup = 0;
650 if (mdt_object_remote(parent))
651 info->mti_spec.sp_cr_lookup = 1;
652 info->mti_spec.sp_feat = &dt_directory_features;
654 /* set jobid xattr name from sysfs parameter */
655 strncpy(info->mti_spec.sp_cr_job_xattr, mdt->mdt_job_xattr,
658 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
659 mdt_object_child(child), &info->mti_spec, ma);
661 rc = mdt_attr_get_complex(info, child, ma);
666 /* save child locks to eliminate dependey between 'mkdir a' and
667 * 'mkdir a/b' if b is a remote directory
669 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
670 struct mdt_lock_handle *lhc;
671 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
673 lhc = &info->mti_lh[MDT_LH_CHILD];
674 rc = mdt_object_stripes_lock(info, parent, child, lhc, einfo,
675 MDS_INODELOCK_UPDATE, LCK_PW);
679 mdt_object_stripes_unlock(info, child, lhc, einfo, rc);
682 /* Return fid & attr to client. */
683 if (ma->ma_valid & MA_INODE)
684 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
685 mdt_object_fid(child));
688 mdt_object_put(info->mti_env, child);
690 mdt_object_unlock(info, parent, lh, rc);
692 mdt_object_put(info->mti_env, parent);
696 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
699 struct mdt_lock_handle *lh;
700 int do_vbr = ma->ma_attr.la_valid &
701 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
702 __u64 lockpart = MDS_INODELOCK_UPDATE;
703 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
707 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
708 lockpart |= MDS_INODELOCK_PERM;
709 /* Clear xattr cache on clients, so the virtual project ID xattr
710 * can get the new project ID
712 if (ma->ma_attr.la_valid & LA_PROJID)
713 lockpart |= MDS_INODELOCK_XATTR;
715 lh = &info->mti_lh[MDT_LH_PARENT];
716 rc = mdt_object_stripes_lock(info, NULL, mo, lh, einfo, lockpart,
721 /* all attrs are packed into mti_attr in unpack_setattr */
722 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
723 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
725 /* VBR: update version if attr changed are important for recovery */
727 /* update on-disk version of changed object */
728 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
729 rc = mdt_version_get_check_save(info, mo, 0);
731 GOTO(out_unlock, rc);
734 /* Ensure constant striping during chown(). See LU-2789. */
735 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
736 mutex_lock(&mo->mot_lov_mutex);
738 /* all attrs are packed into mti_attr in unpack_setattr */
739 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
741 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
742 mutex_unlock(&mo->mot_lov_mutex);
745 GOTO(out_unlock, rc);
746 mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
749 mdt_object_stripes_unlock(info, mo, lh, einfo, rc);
754 * Check HSM flags and add HS_DIRTY flag if relevant.
756 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
757 * and is not RELEASED.
759 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
762 struct lu_ucred *uc = mdt_ucred(info);
763 kernel_cap_t cap_saved;
767 /* If the file was modified, add the dirty flag */
768 ma->ma_need = MA_HSM;
769 rc = mdt_attr_get_complex(info, mo, ma);
771 CERROR("file attribute read error for "DFID": %d.\n",
772 PFID(mdt_object_fid(mo)), rc);
776 /* If an up2date copy exists in the backend, add dirty flag */
777 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
778 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
779 ma->ma_hsm.mh_flags |= HS_DIRTY;
781 /* Bump cap so that closes from non-owner writers can
782 * set the HSM state to dirty.
784 cap_saved = uc->uc_cap;
785 cap_raise(uc->uc_cap, CAP_FOWNER);
786 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
787 uc->uc_cap = cap_saved;
789 CERROR("file attribute change error for "DFID": %d\n",
790 PFID(mdt_object_fid(mo)), rc);
796 static int mdt_reint_setattr(struct mdt_thread_info *info,
797 struct mdt_lock_handle *lhc)
799 struct mdt_device *mdt = info->mti_mdt;
800 struct md_attr *ma = &info->mti_attr;
801 struct mdt_reint_record *rr = &info->mti_rr;
802 struct ptlrpc_request *req = mdt_info_req(info);
803 struct mdt_object *mo;
804 struct mdt_body *repbody;
805 ktime_t kstart = ktime_get();
809 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
810 (unsigned int)ma->ma_attr.la_valid);
812 if (info->mti_dlm_req)
813 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
815 CFS_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
817 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
818 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
820 GOTO(out, rc = PTR_ERR(mo));
822 if (!mdt_object_exists(mo))
823 GOTO(out_put, rc = -ENOENT);
825 if (mdt_object_remote(mo))
826 GOTO(out_put, rc = -EREMOTE);
828 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
829 /* revoke lease lock if size is going to be changed */
830 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
831 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
832 atomic_read(&mo->mot_lease_count) > 0)) {
833 down_read(&mo->mot_open_sem);
835 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
836 lhc = &info->mti_lh[MDT_LH_LOCAL];
837 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN,
840 up_read(&mo->mot_open_sem);
844 /* revoke lease lock */
845 mdt_object_unlock(info, mo, lhc, 1);
847 up_read(&mo->mot_open_sem);
850 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
851 /* Check write access for the O_TRUNC case */
852 if (mdt_write_read(mo) < 0)
853 GOTO(out_put, rc = -ETXTBSY);
855 /* LU-10286: compatibility check for FLR.
856 * Please check the comment in mdt_finish_open() for details
858 if (!exp_connect_flr(info->mti_exp) ||
859 !exp_connect_overstriping(info->mti_exp)) {
860 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
861 if (rc < 0 && rc != -ENODATA)
864 if (!exp_connect_flr(info->mti_exp)) {
866 mdt_lmm_is_flr(info->mti_big_lmm))
867 GOTO(out_put, rc = -EOPNOTSUPP);
870 if (!exp_connect_overstriping(info->mti_exp)) {
872 mdt_lmm_is_overstriping(info->mti_big_lmm))
873 GOTO(out_put, rc = -EOPNOTSUPP);
877 /* For truncate, the file size sent from client
878 * is believable, but the blocks are incorrect,
879 * which makes the block size in LSOM attribute
880 * inconsisent with the real block size.
882 rc = mdt_lsom_update(info, mo, true);
887 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
888 if (ma->ma_valid & MA_LOV)
889 GOTO(out_put, rc = -EPROTO);
891 /* MDT supports FMD for regular files due to Data-on-MDT */
892 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
893 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
894 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
897 if (ma->ma_attr.la_valid & LA_MTIME) {
898 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
900 ma->ma_valid |= MA_PFID;
904 rc = mdt_attr_set(info, mo, ma);
907 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
908 (ma->ma_valid & MA_INODE)) {
909 struct lu_buf *buf = &info->mti_buf;
910 struct lu_ucred *uc = mdt_ucred(info);
911 struct mdt_lock_handle *lh;
914 /* reject if either remote or striped dir is disabled */
915 if (ma->ma_valid & MA_LMV) {
916 if (!mdt->mdt_enable_remote_dir ||
917 !mdt->mdt_enable_striped_dir)
918 GOTO(out_put, rc = -EPERM);
920 /* we want rbac roles to have precedence over any other
921 * permission or capability checks
923 if (!uc->uc_rbac_dne_ops ||
924 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
925 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
926 mdt->mdt_enable_remote_dir_gid != -1))
927 GOTO(out_put, rc = -EPERM);
930 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
931 GOTO(out_put, rc = -ENOTDIR);
933 if (ma->ma_attr.la_valid != 0)
934 GOTO(out_put, rc = -EPROTO);
936 lh = &info->mti_lh[MDT_LH_PARENT];
937 if (ma->ma_valid & MA_LOV) {
938 buf->lb_buf = ma->ma_lmm;
939 buf->lb_len = ma->ma_lmm_size;
940 name = XATTR_NAME_LOV;
941 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
944 buf->lb_buf = &ma->ma_lmv->lmv_user_md;
945 buf->lb_len = ma->ma_lmv_size;
946 name = XATTR_NAME_DEFAULT_LMV;
948 if (unlikely(fid_is_root(mdt_object_fid(mo)))) {
949 rc = mdt_object_lock(info, mo, lh,
950 MDS_INODELOCK_XATTR |
951 MDS_INODELOCK_LOOKUP,
954 struct lu_fid *pfid = &info->mti_tmp_fid1;
955 struct lu_name *pname = &info->mti_name;
956 const char dotdot[] = "..";
957 struct mdt_object *pobj;
960 pname->ln_name = dotdot;
961 pname->ln_namelen = sizeof(dotdot);
962 rc = mdo_lookup(info->mti_env,
963 mdt_object_child(mo), pname,
968 pobj = mdt_object_find(info->mti_env,
969 info->mti_mdt, pfid);
971 GOTO(out_put, rc = PTR_ERR(pobj));
973 rc = mdt_object_check_lock(info, pobj, mo, lh,
974 MDS_INODELOCK_XATTR |
975 MDS_INODELOCK_LOOKUP,
977 mdt_object_put(info->mti_env, pobj);
984 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
987 mdt_object_unlock(info, mo, lh, rc);
991 GOTO(out_put, rc = -EPROTO);
994 /* If file data is modified, add the dirty flag */
995 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
996 rc = mdt_add_dirty_flag(info, mo, ma);
998 ma->ma_need = MA_INODE;
1000 rc = mdt_attr_get_complex(info, mo, ma);
1004 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1008 mdt_object_put(info->mti_env, mo);
1011 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1012 ktime_us_delta(ktime_get(), kstart));
1014 mdt_client_compatibility(info);
1018 static int mdt_reint_create(struct mdt_thread_info *info,
1019 struct mdt_lock_handle *lhc)
1021 struct ptlrpc_request *req = mdt_info_req(info);
1022 ktime_t kstart = ktime_get();
1026 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1027 RETURN(err_serious(-ESTALE));
1029 if (info->mti_dlm_req)
1030 ldlm_request_cancel(mdt_info_req(info),
1031 info->mti_dlm_req, 0, LATF_SKIP);
1033 if (!lu_name_is_valid(&info->mti_rr.rr_name))
1036 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1046 CERROR("%s: Unsupported mode %o\n",
1047 mdt_obd_name(info->mti_mdt),
1048 info->mti_attr.ma_attr.la_mode);
1049 RETURN(err_serious(-EOPNOTSUPP));
1052 rc = mdt_create(info);
1054 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1055 mdt_counter_incr(req, LPROC_MDT_MKDIR,
1056 ktime_us_delta(ktime_get(), kstart));
1058 /* Special file should stay on the same node as parent*/
1059 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1060 ktime_us_delta(ktime_get(), kstart));
1067 * VBR: save parent version in reply and child version getting by its name.
1068 * Version of child is getting and checking during its lookup. If
1070 static int mdt_reint_unlink(struct mdt_thread_info *info,
1071 struct mdt_lock_handle *lhc)
1073 struct mdt_reint_record *rr = &info->mti_rr;
1074 struct ptlrpc_request *req = mdt_info_req(info);
1075 struct md_attr *ma = &info->mti_attr;
1076 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1077 struct mdt_object *mp;
1078 struct mdt_object *mc;
1079 struct mdt_lock_handle *parent_lh;
1080 struct mdt_lock_handle *child_lh;
1081 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
1082 struct lu_ucred *uc = mdt_ucred(info);
1084 ktime_t kstart = ktime_get();
1088 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1089 PNAME(&rr->rr_name));
1091 if (info->mti_dlm_req)
1092 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1094 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1095 RETURN(err_serious(-ENOENT));
1097 if (!fid_is_md_operative(rr->rr_fid1))
1100 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1102 RETURN(PTR_ERR(mp));
1104 if (!mdt_object_remote(mp)) {
1105 rc = mdt_version_get_check_save(info, mp, 0);
1107 GOTO(put_parent, rc);
1110 if (!uc->uc_rbac_fscrypt_admin &&
1111 mp->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
1112 GOTO(put_parent, rc = -EPERM);
1114 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
1115 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1116 parent_lh = &info->mti_lh[MDT_LH_PARENT];
1117 rc = mdt_parent_lock(info, mp, parent_lh, &rr->rr_name, LCK_PW);
1119 GOTO(put_parent, rc);
1121 if (info->mti_spec.sp_rm_entry) {
1122 if (!mdt_is_dne_client(req->rq_export))
1123 /* Return -ENOTSUPP for old client */
1124 GOTO(unlock_parent, rc = -ENOTSUPP);
1126 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1127 GOTO(unlock_parent, rc = -EPERM);
1129 ma->ma_need = MA_INODE;
1131 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1132 NULL, &rr->rr_name, ma, no_name);
1133 GOTO(unlock_parent, rc);
1136 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1137 *child_fid = *rr->rr_fid2;
1139 /* lookup child object along with version checking */
1140 fid_zero(child_fid);
1141 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1144 /* Name might not be able to find during resend of
1145 * remote unlink, considering following case.
1146 * dir_A is a remote directory, the name entry of
1147 * dir_A is on MDT0, the directory is on MDT1,
1149 * 1. client sends unlink req to MDT1.
1150 * 2. MDT1 sends name delete update to MDT0.
1151 * 3. name entry is being deleted in MDT0 synchronously.
1152 * 4. MDT1 is restarted.
1153 * 5. client resends unlink req to MDT1. So it can not
1154 * find the name entry on MDT0 anymore.
1155 * In this case, MDT1 only needs to destory the local
1158 if (mdt_object_remote(mp) && rc == -ENOENT &&
1159 !fid_is_zero(rr->rr_fid2) &&
1160 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1162 *child_fid = *rr->rr_fid2;
1164 GOTO(unlock_parent, rc);
1169 if (!fid_is_md_operative(child_fid))
1170 GOTO(unlock_parent, rc = -EPERM);
1172 /* We will lock the child regardless it is local or remote. No harm. */
1173 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1175 GOTO(unlock_parent, rc = PTR_ERR(mc));
1177 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1178 /* In this case, child fid is embedded in the request, and we do
1179 * not have a proper name as rr_name contains an encoded
1180 * hash. So find name that matches provided hash.
1182 if (!find_name_matching_hash(info, &rr->rr_name,
1184 GOTO(put_child, rc = -ENOENT);
1187 child_lh = &info->mti_lh[MDT_LH_CHILD];
1188 if (mdt_object_remote(mc)) {
1189 struct mdt_body *repbody;
1191 if (!fid_is_zero(rr->rr_fid2)) {
1192 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1193 mdt_obd_name(info->mti_mdt),
1194 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1195 GOTO(put_child, rc = -ENOENT);
1197 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1198 mdt_obd_name(info->mti_mdt),
1199 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1201 if (!mdt_is_dne_client(req->rq_export))
1202 /* Return -ENOTSUPP for old client */
1203 GOTO(put_child, rc = -ENOTSUPP);
1205 /* Revoke the LOOKUP lock of the remote object granted by
1206 * this MDT. Since the unlink will happen on another MDT,
1207 * it will release the LOOKUP lock right away. Then What
1208 * would happen if another client try to grab the LOOKUP
1209 * lock at the same time with unlink XXX
1211 rc = mdt_object_lookup_lock(info, NULL, mc, child_lh, LCK_EX);
1213 GOTO(put_child, rc);
1215 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1216 LASSERT(repbody != NULL);
1217 repbody->mbo_fid1 = *mdt_object_fid(mc);
1218 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1219 GOTO(unlock_child, rc = -EREMOTE);
1221 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1222 * this now because a running HSM restore on the child (unlink
1223 * victim) will hold the layout lock. See LU-4002.
1225 rc = mdt_object_stripes_lock(info, mp, mc, child_lh, einfo,
1226 MDS_INODELOCK_LOOKUP |
1227 MDS_INODELOCK_UPDATE, LCK_EX);
1229 GOTO(put_child, rc);
1232 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1233 * whether need MA_LOV and MA_COOKIE.
1235 ma->ma_need = MA_INODE;
1238 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1239 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1240 /* save version when object is locked */
1241 mdt_version_get_save(info, mc, 1);
1243 mutex_lock(&mc->mot_lov_mutex);
1245 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1246 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1248 mutex_unlock(&mc->mot_lov_mutex);
1250 GOTO(unlock_child, rc);
1252 if (!lu_object_is_dying(&mc->mot_header)) {
1253 rc = mdt_attr_get_complex(info, mc, ma);
1256 } else if (mdt_dom_check_for_discard(info, mc)) {
1257 mdt_dom_discard_data(info, mc);
1259 mdt_handle_last_unlink(info, mc, ma);
1262 if (ma->ma_valid & MA_INODE) {
1263 switch (ma->ma_attr.la_mode & S_IFMT) {
1265 mdt_counter_incr(req, LPROC_MDT_RMDIR,
1266 ktime_us_delta(ktime_get(), kstart));
1274 mdt_counter_incr(req, LPROC_MDT_UNLINK,
1275 ktime_us_delta(ktime_get(), kstart));
1278 LASSERTF(0, "bad file type %o unlinking\n",
1279 ma->ma_attr.la_mode);
1286 /* after unlink the object is gone, no need to keep lock */
1287 mdt_object_stripes_unlock(info, mc, child_lh, einfo, 1);
1289 if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1290 info->mti_big_buf.lb_buf)
1291 lu_buf_free(&info->mti_big_buf);
1292 mdt_object_put(info->mti_env, mc);
1294 mdt_object_unlock(info, mp, parent_lh, rc);
1296 mdt_object_put(info->mti_env, mp);
1297 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1302 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1305 static int mdt_reint_link(struct mdt_thread_info *info,
1306 struct mdt_lock_handle *lhc)
1308 struct mdt_reint_record *rr = &info->mti_rr;
1309 struct ptlrpc_request *req = mdt_info_req(info);
1310 struct md_attr *ma = &info->mti_attr;
1311 struct mdt_object *ms;
1312 struct mdt_object *mp;
1313 struct mdt_lock_handle *lhs;
1314 struct mdt_lock_handle *lhp;
1315 ktime_t kstart = ktime_get();
1319 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1320 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1322 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1323 RETURN(err_serious(-ENOENT));
1325 if (CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1326 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1327 req->rq_no_reply = 1;
1328 RETURN(err_serious(-ENOENT));
1331 if (info->mti_dlm_req)
1332 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1334 /* Invalid case so return error immediately instead of
1337 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1340 if (!fid_is_md_operative(rr->rr_fid1) ||
1341 !fid_is_md_operative(rr->rr_fid2))
1344 /* step 1: find target parent dir */
1345 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1347 RETURN(PTR_ERR(mp));
1349 rc = mdt_version_get_check_save(info, mp, 0);
1351 GOTO(put_parent, rc);
1353 rc = mdt_check_enc(info, mp);
1355 GOTO(put_parent, rc);
1357 /* step 2: find source */
1358 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1360 GOTO(put_parent, rc = PTR_ERR(ms));
1362 if (!mdt_object_exists(ms)) {
1363 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1364 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1365 GOTO(put_source, rc = -ENOENT);
1368 CFS_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1370 lhp = &info->mti_lh[MDT_LH_PARENT];
1371 rc = mdt_parent_lock(info, mp, lhp, &rr->rr_name, LCK_PW);
1373 GOTO(put_source, rc);
1375 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1377 lhs = &info->mti_lh[MDT_LH_CHILD];
1378 rc = mdt_object_lock(info, ms, lhs,
1379 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1382 GOTO(unlock_parent, rc);
1384 /* step 3: link it */
1385 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1386 OBD_FAIL_MDS_REINT_LINK_WRITE);
1388 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1389 rc = mdt_version_get_check_save(info, ms, 1);
1391 GOTO(unlock_source, rc);
1393 /** check target version by name during replay */
1394 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1395 &info->mti_tmp_fid1, 2);
1396 if (rc != 0 && rc != -ENOENT)
1397 GOTO(unlock_source, rc);
1398 /* save version of file name for replay, it must be ENOENT here */
1399 if (!req_is_replay(mdt_info_req(info))) {
1400 if (rc != -ENOENT) {
1401 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1402 PNAME(&rr->rr_name));
1403 GOTO(unlock_source, rc = -EEXIST);
1405 info->mti_ver[2] = ENOENT_VERSION;
1406 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1409 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1410 mdt_object_child(ms), &rr->rr_name, ma);
1413 mdt_counter_incr(req, LPROC_MDT_LINK,
1414 ktime_us_delta(ktime_get(), kstart));
1418 mdt_object_unlock(info, ms, lhs, rc);
1420 mdt_object_unlock(info, mp, lhp, rc);
1422 mdt_object_put(info->mti_env, ms);
1424 mdt_object_put(info->mti_env, mp);
1429 * Get BFL lock for rename or migrate process.
1431 static int mdt_rename_lock(struct mdt_thread_info *info,
1432 struct mdt_lock_handle *lh)
1434 struct lu_fid *fid = &info->mti_tmp_fid1;
1435 struct mdt_object *obj;
1436 __u64 ibits = MDS_INODELOCK_UPDATE;
1441 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1443 RETURN(PTR_ERR(obj));
1445 mdt_lock_reg_init(lh, LCK_EX);
1446 rc = mdt_object_lock_internal(info, obj, &LUSTRE_BFL_FID, lh,
1448 mdt_object_put(info->mti_env, obj);
1452 static void mdt_rename_unlock(struct mdt_thread_info *info,
1453 struct mdt_lock_handle *lh)
1456 /* Cancel the single rename lock right away */
1457 mdt_object_unlock(info, NULL, lh, 1);
1461 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1462 const struct lu_fid *fid,
1465 struct mdt_object *dir;
1469 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1473 /* check early, the real version will be saved after locking */
1474 rc = mdt_version_get_check(info, dir, idx);
1478 if (!mdt_object_exists(dir))
1479 GOTO(out_put, rc = -ENOENT);
1481 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1482 GOTO(out_put, rc = -ENOTDIR);
1486 mdt_object_put(info->mti_env, dir);
1491 * lock rename source object.
1493 * Both source and its parent object may be located on remote MDTs, and even on
1494 * different MDTs, which means source object is a remote object on parent.
1496 * \retval 0 on success
1497 * \retval -ev negative errno upon error
1499 static int mdt_rename_source_lock(struct mdt_thread_info *info,
1500 struct mdt_object *parent,
1501 struct mdt_object *child,
1502 struct mdt_lock_handle *lh,
1503 struct mdt_lock_handle *lh_lookup,
1508 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
1509 /* if @obj is remote object, LOOKUP lock needs to be taken from
1512 rc = mdt_is_remote_object(info, parent, child);
1517 rc = mdt_object_lookup_lock(info, parent, child, lh_lookup,
1522 ibits &= ~MDS_INODELOCK_LOOKUP;
1525 rc = mdt_object_lock(info, child, lh, ibits, LCK_EX);
1526 if (unlikely(rc && !(ibits & MDS_INODELOCK_LOOKUP)))
1527 mdt_object_unlock(info, NULL, lh_lookup, rc);
1532 static void mdt_rename_source_unlock(struct mdt_thread_info *info,
1533 struct mdt_object *obj,
1534 struct mdt_lock_handle *lh,
1535 struct mdt_lock_handle *lh_lookup,
1538 mdt_object_unlock(info, obj, lh, decref);
1539 mdt_object_unlock(info, NULL, lh_lookup, decref);
1542 /* migration takes UPDATE lock of link parent, and LOOKUP lock of link */
1543 struct mdt_link_lock {
1544 struct mdt_object *mll_obj;
1545 struct mdt_lock_handle mll_lh;
1546 struct list_head mll_linkage;
1549 static inline int mdt_migrate_link_lock_add(struct mdt_thread_info *info,
1550 struct mdt_object *o,
1551 struct mdt_lock_handle *lh,
1552 struct list_head *list)
1554 struct mdt_link_lock *mll;
1560 INIT_LIST_HEAD(&mll->mll_linkage);
1561 mdt_object_get(info->mti_env, o);
1564 memset(lh, 0, sizeof(*lh));
1565 list_add_tail(&mll->mll_linkage, list);
1570 static inline void mdt_migrate_link_lock_del(struct mdt_thread_info *info,
1571 struct mdt_link_lock *mll,
1574 mdt_object_unlock(info, mll->mll_obj, &mll->mll_lh, decref);
1575 mdt_object_put(info->mti_env, mll->mll_obj);
1576 list_del(&mll->mll_linkage);
1580 static void mdt_migrate_links_unlock(struct mdt_thread_info *info,
1581 struct list_head *list, int decref)
1583 struct mdt_link_lock *mll;
1584 struct mdt_link_lock *tmp;
1586 list_for_each_entry_safe(mll, tmp, list, mll_linkage)
1587 mdt_migrate_link_lock_del(info, mll, decref);
1590 /* take link parent UPDATE lock.
1591 * \retval 0 \a lnkp is already locked, no lock taken.
1593 * -ev negative errno.
1595 static int mdt_migrate_link_parent_lock(struct mdt_thread_info *info,
1596 struct mdt_object *lnkp,
1597 struct list_head *update_locks,
1600 const struct lu_fid *fid = mdt_object_fid(lnkp);
1601 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1602 struct mdt_link_lock *entry;
1608 /* check if it's already locked */
1609 list_for_each_entry(entry, update_locks, mll_linkage) {
1610 if (lu_fid_eq(mdt_object_fid(entry->mll_obj), fid)) {
1611 CDEBUG(D_INFO, "skip "DFID" lock\n", PFID(fid));
1616 /* link parent UPDATE lock */
1617 CDEBUG(D_INFO, "lock "DFID"\n", PFID(fid));
1620 /* revoke lock instead of take in *blocked* mode */
1621 rc = mdt_object_lock(info, lnkp, lhl, MDS_INODELOCK_UPDATE,
1626 if (mdt_object_remote(lnkp)) {
1627 struct ldlm_lock *lock;
1630 * for remote object, set lock cb_atomic, so lock can be
1631 * released in blocking_ast() immediately, then the next
1632 * lock_try will have better chance of success.
1634 lock = ldlm_handle2lock(&lhl->mlh_rreg_lh);
1635 LASSERT(lock != NULL);
1636 lock_res_and_lock(lock);
1637 ldlm_set_atomic_cb(lock);
1638 unlock_res_and_lock(lock);
1639 LDLM_LOCK_PUT(lock);
1642 mdt_object_unlock(info, lnkp, lhl, 1);
1647 * we can't follow parent-child lock order like other MD
1648 * operations, use lock_try here to avoid deadlock, if the lock
1649 * cannot be taken, drop all locks taken, revoke the blocked
1650 * one, and continue processing the remaining entries, and in
1651 * the end of the loop restart from beginning.
1653 * don't lock with PDO mode in case two links are under the same
1654 * parent and their hash values are different.
1656 rc = mdt_object_lock_try(info, lnkp, lhl, &ibits, MDS_INODELOCK_UPDATE,
1661 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1662 CDEBUG(D_INFO, "busy lock on "DFID"\n", PFID(fid));
1667 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, update_locks);
1669 mdt_object_unlock(info, lnkp, lhl, 1);
1676 /* take link LOOKUP lock.
1677 * \retval 0 \a lnkp is already locked, no lock taken.
1679 * -ev negative errno.
1681 static int mdt_migrate_link_lock(struct mdt_thread_info *info,
1682 struct mdt_object *lnkp,
1683 struct mdt_object *spobj,
1684 struct mdt_object *obj,
1685 struct list_head *lookup_locks)
1687 const struct lu_fid *fid = mdt_object_fid(lnkp);
1688 struct mdt_lock_handle *lhl = &info->mti_lh[MDT_LH_LOCAL];
1689 struct mdt_link_lock *entry;
1694 /* check if it's already locked by source */
1695 rc = mdt_fids_different_target(info, fid, mdt_object_fid(spobj));
1697 CDEBUG(D_INFO, "skip lookup lock on source parent "DFID"\n",
1702 /* check if it's already locked by other links */
1703 list_for_each_entry(entry, lookup_locks, mll_linkage) {
1704 rc = mdt_fids_different_target(info, fid,
1705 mdt_object_fid(entry->mll_obj));
1707 CDEBUG(D_INFO, "skip lookup lock on parent "DFID"\n",
1713 rc = mdt_object_lookup_lock(info, lnkp, obj, lhl, LCK_EX);
1717 /* don't take local LOOKUP lock, because later we will lock other ibits
1718 * of sobj (which is on local MDT), and lock the same object twice may
1719 * deadlock, just revoke this lock.
1721 if (!mdt_object_remote(lnkp))
1722 GOTO(unlock, rc = 0);
1724 rc = mdt_migrate_link_lock_add(info, lnkp, lhl, lookup_locks);
1730 mdt_object_unlock(info, lnkp, lhl, 1);
1735 * take UPDATE lock of link parents and LOOKUP lock of links, also check whether
1736 * total local lock count exceeds RS_MAX_LOCKS.
1738 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1739 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1740 * \retval -ev negative errno upon error
1742 static int mdt_migrate_links_lock(struct mdt_thread_info *info,
1743 struct mdt_object *spobj,
1744 struct mdt_object *tpobj,
1745 struct mdt_object *obj,
1746 struct mdt_lock_handle *lhsp,
1747 struct mdt_lock_handle *lhtp,
1748 struct list_head *link_locks)
1750 struct mdt_device *mdt = info->mti_mdt;
1751 struct lu_buf *buf = &info->mti_big_buf;
1752 struct lu_name *lname = &info->mti_name;
1753 struct linkea_data ldata = { NULL };
1754 int local_lock_cnt = 0;
1755 bool blocked = false;
1757 struct mdt_object *lnkp;
1759 LIST_HEAD(update_locks);
1760 LIST_HEAD(lookup_locks);
1764 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1767 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1768 if (buf->lb_buf == NULL)
1772 rc = mdt_links_read(info, obj, &ldata);
1774 if (rc == -ENOENT || rc == -ENODATA)
1779 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1780 linkea_next_entry(&ldata)) {
1781 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1784 /* check if link parent is source parent too */
1785 if (lu_fid_eq(mdt_object_fid(spobj), &fid)) {
1787 "skip lock on source parent "DFID"/"DNAME"\n",
1788 PFID(&fid), PNAME(lname));
1792 /* check if link parent is target parent too */
1793 if (tpobj != spobj && lu_fid_eq(mdt_object_fid(tpobj), &fid)) {
1795 "skip lock on target parent "DFID"/"DNAME"\n",
1796 PFID(&fid), PNAME(lname));
1800 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1802 CWARN("%s: cannot find obj "DFID": %ld\n",
1803 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1807 if (!mdt_object_exists(lnkp)) {
1808 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1809 PFID(&fid), PNAME(lname));
1810 mdt_object_put(info->mti_env, lnkp);
1815 rc = mdt_migrate_link_parent_lock(info, lnkp, &update_locks,
1817 if (!saved && blocked) {
1818 /* unlock all locks taken to avoid deadlock */
1819 mdt_migrate_links_unlock(info, &update_locks, 1);
1820 mdt_object_unlock(info, spobj, lhsp, 1);
1822 mdt_object_unlock(info, tpobj, lhtp, 1);
1826 mdt_object_put(info->mti_env, lnkp);
1830 if (rc == 1 && !mdt_object_remote(lnkp))
1833 rc = mdt_migrate_link_lock(info, lnkp, spobj, obj,
1836 mdt_object_put(info->mti_env, lnkp);
1839 if (rc == 1 && !mdt_object_remote(lnkp))
1841 mdt_object_put(info->mti_env, lnkp);
1845 GOTO(out, rc = -EBUSY);
1849 list_splice(&update_locks, link_locks);
1850 list_splice(&lookup_locks, link_locks);
1852 mdt_migrate_links_unlock(info, link_locks, rc);
1853 } else if (local_lock_cnt > RS_MAX_LOCKS - 5) {
1855 * parent may have 3 local objects: master object and 2 stripes
1856 * (if it's being migrated too); source may have 1 local objects
1857 * as regular file; target has 1 local object.
1858 * Note, source may have 2 local locks if it is directory but it
1859 * can't have hardlinks, so it is not considered here.
1861 CDEBUG(D_INFO, "Too many local locks (%d), migrate in sync mode\n",
1869 * lookup source by name, if parent is striped directory, we need to find the
1870 * corresponding stripe where source is located, and then lookup there.
1872 * besides, if parent is migrating too, and file is already in target stripe,
1873 * this should be a redo of 'lfs migrate' on client side.
1875 * \retval 1 tpobj stripe index is less than spobj stripe index
1876 * \retval 0 tpobj stripe index is larger than or equal to spobj stripe index
1877 * \retval -ev negative errno upon error
1879 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1880 struct mdt_object *pobj,
1881 const struct md_attr *ma,
1882 const struct lu_name *lname,
1883 struct mdt_object **spobj,
1884 struct mdt_object **tpobj,
1885 struct mdt_object **sobj)
1887 const struct lu_env *env = info->mti_env;
1888 struct lu_fid *fid = &info->mti_tmp_fid1;
1893 if (ma->ma_valid & MA_LMV) {
1894 /* if parent is striped, lookup on corresponding stripe */
1895 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1896 struct lu_fid *fid2 = &info->mti_tmp_fid2;
1898 if (!lmv_is_sane(lmv))
1901 spindex = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1906 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[spindex]);
1908 *spobj = mdt_object_find(env, info->mti_mdt, fid2);
1909 if (IS_ERR(*spobj)) {
1910 rc = PTR_ERR(*spobj);
1915 if (!mdt_object_exists(*spobj))
1916 GOTO(spobj_put, rc = -ENOENT);
1919 rc = mdo_lookup(env, mdt_object_child(*spobj), lname, fid,
1921 if ((rc == -ENOENT || rc == 0) && lmv_is_layout_changing(lmv)) {
1922 /* fail check here to let top dir migration succeed. */
1923 if (CFS_FAIL_CHECK_RESET(OBD_FAIL_MIGRATE_ENTRIES, 0))
1924 GOTO(spobj_put, rc = -EIO);
1927 * if parent layout is changeing, and lookup child
1928 * failed on source stripe, lookup again on target
1929 * stripe, if it exists, it means previous migration
1930 * was interrupted, and current file was migrated
1933 tpindex = lmv_name_to_stripe_index(lmv, lname->ln_name,
1936 GOTO(spobj_put, rc = tpindex);
1938 fid_le_to_cpu(fid2, &lmv->lmv_stripe_fids[tpindex]);
1940 *tpobj = mdt_object_find(env, info->mti_mdt, fid2);
1941 if (IS_ERR(*tpobj)) {
1942 rc = PTR_ERR(*tpobj);
1944 GOTO(spobj_put, rc);
1947 if (!mdt_object_exists(*tpobj))
1948 GOTO(tpobj_put, rc = -ENOENT);
1950 if (rc == -ENOENT) {
1952 rc = mdo_lookup(env, mdt_object_child(*tpobj),
1953 lname, fid, &info->mti_spec);
1954 GOTO(tpobj_put, rc = rc ?: -EALREADY);
1957 GOTO(spobj_put, rc);
1961 mdt_object_get(env, *tpobj);
1965 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1972 mdt_object_get(env, pobj);
1973 mdt_object_get(env, pobj);
1976 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1977 if (IS_ERR(*sobj)) {
1978 rc = PTR_ERR(*sobj);
1980 GOTO(tpobj_put, rc);
1983 if (!mdt_object_exists(*sobj))
1984 GOTO(sobj_put, rc = -ENOENT);
1986 return (tpindex < spindex);
1989 mdt_object_put(env, *sobj);
1992 mdt_object_put(env, *tpobj);
1995 mdt_object_put(env, *spobj);
2001 /* end lease and close file for regular file */
2002 static int mdd_migrate_close(struct mdt_thread_info *info,
2003 struct mdt_object *obj)
2005 struct close_data *data;
2006 struct mdt_body *repbody;
2007 struct ldlm_lock *lease;
2012 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2014 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2018 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2023 lease = ldlm_handle2lock(&data->cd_handle);
2027 /* check if the lease was already canceled */
2028 lock_res_and_lock(lease);
2029 rc = ldlm_is_cancel(lease);
2030 unlock_res_and_lock(lease);
2034 LDLM_DEBUG(lease, DFID" lease broken",
2035 PFID(mdt_object_fid(obj)));
2039 * cancel server side lease, client side counterpart should have been
2040 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2042 ldlm_lock_cancel(lease);
2043 ldlm_reprocess_all(lease->l_resource,
2044 lease->l_policy_data.l_inodebits.bits);
2045 LDLM_LOCK_PUT(lease);
2048 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2049 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2050 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2055 /* LFSCK used to clear hash type and MIGRATION flag upon migration failure */
2056 static inline bool lmv_is_failed_migration(const struct lmv_mds_md_v1 *lmv)
2058 return le32_to_cpu(lmv->lmv_hash_type) ==
2059 (LMV_HASH_TYPE_UNKNOWN | LMV_HASH_FLAG_BAD_TYPE) &&
2060 lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_migrate_hash)) &&
2061 le32_to_cpu(lmv->lmv_migrate_offset) > 0 &&
2062 le32_to_cpu(lmv->lmv_migrate_offset) <
2063 le32_to_cpu(lmv->lmv_stripe_count);
2067 * migrate file in below steps:
2068 * 1. lock source and target stripes
2069 * 2. lookup source by name
2070 * 3. lock parents of source links if source is not directory
2071 * 4. reject if source is in HSM
2072 * 5. take source open_sem and close file if source is regular file
2073 * 6. lock source, and its stripes if it's directory
2075 * 8. lock target so subsequent change to it can trigger COS
2076 * 9. unlock above locks
2077 * 10. sync device if source has too many links
2079 int mdt_reint_migrate(struct mdt_thread_info *info,
2080 struct mdt_lock_handle *unused)
2082 const struct lu_env *env = info->mti_env;
2083 struct mdt_device *mdt = info->mti_mdt;
2084 struct ptlrpc_request *req = mdt_info_req(info);
2085 struct mdt_reint_record *rr = &info->mti_rr;
2086 struct lu_ucred *uc = mdt_ucred(info);
2087 struct md_attr *ma = &info->mti_attr;
2088 struct mdt_object *pobj;
2089 struct mdt_object *spobj;
2090 struct mdt_object *tpobj;
2091 struct mdt_object *sobj;
2092 struct mdt_object *tobj;
2093 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2094 struct mdt_lock_handle *lhsp;
2095 struct mdt_lock_handle *lhtp;
2096 struct mdt_lock_handle *lhs;
2097 struct mdt_lock_handle *lhl;
2098 LIST_HEAD(link_locks);
2099 int lock_retries = 5;
2100 bool reverse = false;
2101 bool open_sem_locked = false;
2102 bool do_sync = false;
2103 bool is_plain_dir = false;
2107 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2108 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2110 if (info->mti_dlm_req)
2111 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2113 if (!fid_is_md_operative(rr->rr_fid1) ||
2114 !fid_is_md_operative(rr->rr_fid2))
2117 /* don't allow migrate . or .. */
2118 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2121 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2124 /* we want rbac roles to have precedence over any other
2125 * permission or capability checks
2127 if (uc && (!uc->uc_rbac_dne_ops ||
2128 (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2129 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2130 mdt->mdt_enable_remote_dir_gid != -1)))
2134 * Note: do not enqueue rename lock for replay request, because
2135 * if other MDT holds rename lock, but being blocked to wait for
2136 * this MDT to finish its recovery, and the failover MDT can not
2137 * get rename lock, which will cause deadlock.
2139 * req is NULL if this is called by directory auto-split.
2141 if (req && !req_is_replay(req)) {
2142 rc = mdt_rename_lock(info, rename_lh);
2144 CERROR("%s: can't lock FS for rename: rc = %d\n",
2145 mdt_obd_name(info->mti_mdt), rc);
2150 /* pobj is master object of parent */
2151 pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2153 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2156 rc = mdt_version_get_check(info, pobj, 0);
2158 GOTO(put_parent, rc);
2161 if (!mdt_object_exists(pobj))
2162 GOTO(put_parent, rc = -ENOENT);
2164 if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2165 GOTO(put_parent, rc = -ENOTDIR);
2167 rc = mdt_check_enc(info, pobj);
2169 GOTO(put_parent, rc);
2171 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2173 GOTO(put_parent, rc);
2175 if (CFS_FAIL_CHECK(OBD_FAIL_MIGRATE_BAD_HASH) &&
2176 (ma->ma_valid & MA_LMV) &&
2177 lmv_is_migrating(&ma->ma_lmv->lmv_md_v1)) {
2178 struct lu_buf *buf = &info->mti_buf;
2179 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2180 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2182 lmv->lmv_hash_type = cpu_to_le32(LMV_HASH_TYPE_UNKNOWN |
2183 LMV_HASH_FLAG_BAD_TYPE);
2184 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2186 buf->lb_len = sizeof(*lmv);
2187 rc = mo_xattr_set(env, mdt_object_child(pobj), buf,
2188 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2189 mo_invalidate(env, mdt_object_child(pobj));
2190 GOTO(put_parent, rc);
2193 /* @spobj is the parent stripe of @sobj if @pobj is striped directory,
2194 * if @pobj is migrating too, tpobj is the target parent stripe.
2196 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &tpobj,
2199 GOTO(put_parent, rc);
2202 /* parent unchanged, this happens in dir restripe */
2203 if (info->mti_spec.sp_migrate_nsonly && spobj == tpobj)
2204 GOTO(put_source, rc = -EALREADY);
2209 lhsp = &info->mti_lh[MDT_LH_PARENT];
2210 lhtp = &info->mti_lh[MDT_LH_CHILD];
2211 /* lock spobj and tpobj in stripe index order */
2213 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name, LCK_PW);
2215 GOTO(put_source, rc);
2217 LASSERT(spobj != tpobj);
2218 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW);
2220 GOTO(unlock_parent, rc);
2222 rc = mdt_parent_lock(info, spobj, lhsp, &rr->rr_name, LCK_PW);
2224 GOTO(put_source, rc);
2226 if (tpobj != spobj) {
2227 rc = mdt_parent_lock(info, tpobj, lhtp, &rr->rr_name,
2230 GOTO(unlock_parent, rc);
2234 /* if inode is not migrated, or is dir, no need to lock links */
2235 if (!info->mti_spec.sp_migrate_nsonly &&
2236 !S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2237 /* lock link parents, and take LOOKUP lock of links */
2238 rc = mdt_migrate_links_lock(info, spobj, tpobj, sobj, lhsp,
2240 if (rc == -EBUSY && lock_retries-- > 0) {
2241 LASSERT(list_empty(&link_locks));
2246 GOTO(put_source, rc);
2249 * RS_MAX_LOCKS is the limit of number of locks that can be
2250 * saved along with one request, if total lock count exceeds
2251 * this limit, we will drop all locks after migration, and
2252 * trigger commit in the end.
2258 lhs = &info->mti_lh[MDT_LH_OLD];
2259 lhl = &info->mti_lh[MDT_LH_LOOKUP];
2260 rc = mdt_rename_source_lock(info, spobj, sobj, lhs, lhl,
2261 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR |
2262 MDS_INODELOCK_OPEN);
2264 GOTO(unlock_links, rc);
2266 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2267 /* TODO: DoM migration is not supported, migrate dirent only */
2268 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2270 GOTO(unlock_source, rc);
2272 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2273 info->mti_spec.sp_migrate_nsonly = 1;
2274 } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2275 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2277 GOTO(unlock_source, rc);
2279 if (!(ma->ma_valid & MA_LMV))
2280 is_plain_dir = true;
2281 else if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2282 /* race with restripe/auto-split */
2283 GOTO(unlock_source, rc = -EBUSY);
2284 else if (lmv_is_failed_migration(&ma->ma_lmv->lmv_md_v1)) {
2285 struct lu_buf *buf = &info->mti_buf;
2286 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2287 __u32 version = le32_to_cpu(lmv->lmv_layout_version);
2289 /* migration failed before, and LFSCK cleared hash type
2290 * and flags, fake it to resume migration.
2292 lmv->lmv_hash_type =
2293 cpu_to_le32(LMV_HASH_TYPE_FNV_1A_64 |
2294 LMV_HASH_FLAG_MIGRATION |
2295 LMV_HASH_FLAG_BAD_TYPE |
2296 LMV_HASH_FLAG_FIXED);
2297 lmv->lmv_layout_version = cpu_to_le32(version + 1);
2299 buf->lb_len = sizeof(*lmv);
2300 rc = mo_xattr_set(env, mdt_object_child(sobj), buf,
2301 XATTR_NAME_LMV, LU_XATTR_REPLACE);
2302 mo_invalidate(env, mdt_object_child(sobj));
2303 GOTO(unlock_source, rc = -EALREADY);
2307 /* if migration HSM is allowed */
2308 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2309 ma->ma_need = MA_HSM;
2311 rc = mdt_attr_get_complex(info, sobj, ma);
2313 GOTO(unlock_source, rc);
2315 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2316 GOTO(unlock_source, rc = -EOPNOTSUPP);
2319 /* end lease and close file for regular file */
2320 if (info->mti_spec.sp_migrate_close) {
2321 /* try to hold open_sem so that nobody else can open the file */
2322 if (!down_write_trylock(&sobj->mot_open_sem)) {
2324 mdd_migrate_close(info, sobj);
2325 GOTO(unlock_source, rc = -EBUSY);
2327 open_sem_locked = true;
2328 rc = mdd_migrate_close(info, sobj);
2329 if (rc && rc != -ESTALE)
2330 GOTO(unlock_open_sem, rc);
2334 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2336 GOTO(unlock_open_sem, rc = PTR_ERR(tobj));
2338 /* Don't do lookup sanity check. We know name doesn't exist. */
2339 info->mti_spec.sp_cr_lookup = 0;
2340 info->mti_spec.sp_feat = &dt_directory_features;
2342 rc = mdo_migrate(env, mdt_object_child(spobj),
2343 mdt_object_child(tpobj), mdt_object_child(sobj),
2344 mdt_object_child(tobj), &rr->rr_name,
2345 &info->mti_spec, ma);
2347 GOTO(put_target, rc);
2349 /* save target locks for directory */
2350 if (S_ISDIR(lu_object_attr(&sobj->mot_obj)) &&
2351 !info->mti_spec.sp_migrate_nsonly) {
2352 struct mdt_lock_handle *lht = &info->mti_lh[MDT_LH_NEW];
2353 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
2355 /* in case sobj becomes a stripe of tobj, unlock sobj here,
2356 * otherwise stripes lock may deadlock.
2359 mdt_rename_source_unlock(info, sobj, lhs, lhl, 1);
2361 rc = mdt_object_stripes_lock(info, tpobj, tobj, lht, einfo,
2362 MDS_INODELOCK_UPDATE, LCK_PW);
2364 GOTO(put_target, rc);
2366 mdt_object_stripes_unlock(info, tobj, lht, einfo, 0);
2369 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2370 LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2374 mdt_object_put(env, tobj);
2376 if (open_sem_locked)
2377 up_write(&sobj->mot_open_sem);
2379 mdt_rename_source_unlock(info, sobj, lhs, lhl, rc);
2381 /* if we've got too many locks to save into RPC,
2382 * then just commit before the locks are released
2385 mdt_device_sync(env, mdt);
2386 mdt_migrate_links_unlock(info, &link_locks, do_sync ? 1 : rc);
2388 mdt_object_unlock(info, spobj, lhsp, rc);
2389 mdt_object_unlock(info, tpobj, lhtp, rc);
2391 mdt_object_put(env, sobj);
2392 mdt_object_put(env, spobj);
2393 mdt_object_put(env, tpobj);
2395 mo_invalidate(env, mdt_object_child(pobj));
2396 mdt_object_put(env, pobj);
2398 mdt_rename_unlock(info, rename_lh);
2401 CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
2402 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
2403 PNAME(&rr->rr_name), rc);
2409 * determine lock order of sobj and tobj
2411 * there are two situations we need to lock tobj before sobj:
2412 * 1. sobj is child of tobj
2413 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2414 * larger than that of tobj
2416 * \retval 1 lock tobj before sobj
2417 * \retval 0 lock sobj before tobj
2418 * \retval -ev negative errno upon error
2420 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2421 struct mdt_object *sobj,
2422 struct mdt_object *tobj)
2424 struct md_attr *ma = &info->mti_attr;
2425 struct lu_fid *spfid = &info->mti_tmp_fid1;
2426 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2427 struct lmv_mds_md_v1 *lmv;
2432 /* sobj and tobj are the same */
2436 if (fid_is_root(mdt_object_fid(sobj)))
2439 if (fid_is_root(mdt_object_fid(tobj)))
2442 /* check whether sobj is child of tobj */
2443 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2444 mdt_object_fid(tobj));
2451 /* check whether sobj and tobj are children of the same parent */
2452 rc = mdt_attr_get_pfid(info, sobj, spfid);
2456 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2460 if (!lu_fid_eq(spfid, tpfid))
2463 /* check whether sobj and tobj are sibling stripes */
2464 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2468 if (!(ma->ma_valid & MA_LMV))
2471 lmv = &ma->ma_lmv->lmv_md_v1;
2472 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2474 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2477 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2481 if (!(ma->ma_valid & MA_LMV))
2484 lmv = &ma->ma_lmv->lmv_md_v1;
2485 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2487 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2489 /* check stripe index of sobj and tobj */
2490 if (sindex == tindex)
2493 return sindex < tindex ? 0 : 1;
2496 /* Helper function for mdt_reint_rename so we don't need to opencode
2497 * two different order lockings
2499 static int mdt_lock_two_dirs(struct mdt_thread_info *info,
2500 struct mdt_object *mfirstdir,
2501 struct mdt_lock_handle *lh_firstdirp,
2502 const struct lu_name *firstname,
2503 struct mdt_object *mseconddir,
2504 struct mdt_lock_handle *lh_seconddirp,
2505 const struct lu_name *secondname)
2509 rc = mdt_parent_lock(info, mfirstdir, lh_firstdirp, firstname, LCK_PW);
2513 mdt_version_get_save(info, mfirstdir, 0);
2514 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2516 if (mfirstdir != mseconddir) {
2517 rc = mdt_parent_lock(info, mseconddir, lh_seconddirp,
2518 secondname, LCK_PW);
2519 } else if (!mdt_object_remote(mseconddir)) {
2520 if (lh_firstdirp->mlh_pdo_hash !=
2521 lh_seconddirp->mlh_pdo_hash) {
2522 rc = mdt_object_pdo_lock(info, mseconddir,
2523 lh_seconddirp, secondname,
2525 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2528 mdt_version_get_save(info, mseconddir, 1);
2531 mdt_object_unlock(info, mfirstdir, lh_firstdirp, rc);
2537 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2538 * 2 - srcdir child; 3 - tgtdir child.
2539 * Update on disk version of srcdir child.
2541 static int mdt_reint_rename(struct mdt_thread_info *info,
2542 struct mdt_lock_handle *unused)
2544 struct mdt_device *mdt = info->mti_mdt;
2545 struct mdt_reint_record *rr = &info->mti_rr;
2546 struct md_attr *ma = &info->mti_attr;
2547 struct ptlrpc_request *req = mdt_info_req(info);
2548 struct mdt_object *msrcdir = NULL;
2549 struct mdt_object *mtgtdir = NULL;
2550 struct mdt_object *mold;
2551 struct mdt_object *mnew = NULL;
2552 struct mdt_lock_handle *rename_lh = &info->mti_lh[MDT_LH_RMT];
2553 struct mdt_lock_handle *lh_srcdirp;
2554 struct mdt_lock_handle *lh_tgtdirp;
2555 struct mdt_lock_handle *lh_oldp = NULL;
2556 struct mdt_lock_handle *lh_lookup = NULL;
2557 struct mdt_lock_handle *lh_newp = NULL;
2558 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2559 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2560 struct lu_ucred *uc = mdt_ucred(info);
2561 bool reverse = false, discard = false;
2562 ktime_t kstart = ktime_get();
2563 enum mdt_stat_idx msi = 0;
2567 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2568 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2569 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2571 if (info->mti_dlm_req)
2572 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2574 if (!fid_is_md_operative(rr->rr_fid1) ||
2575 !fid_is_md_operative(rr->rr_fid2))
2578 /* find both parents. */
2579 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2580 if (IS_ERR(msrcdir))
2581 RETURN(PTR_ERR(msrcdir));
2583 rc = mdt_check_enc(info, msrcdir);
2585 GOTO(out_put_srcdir, rc);
2587 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2589 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2591 mdt_object_get(info->mti_env, mtgtdir);
2593 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2594 if (IS_ERR(mtgtdir))
2595 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2598 rc = mdt_check_enc(info, mtgtdir);
2600 GOTO(out_put_tgtdir, rc);
2602 if (!uc->uc_rbac_fscrypt_admin &&
2603 mtgtdir->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD)
2604 GOTO(out_put_tgtdir, rc = -EPERM);
2607 * Note: do not enqueue rename lock for replay request, because
2608 * if other MDT holds rename lock, but being blocked to wait for
2609 * this MDT to finish its recovery, and the failover MDT can not
2610 * get rename lock, which will cause deadlock.
2612 if (!req_is_replay(req)) {
2613 bool remote = mdt_object_remote(msrcdir);
2616 * Normally rename RPC is handled on the MDT with the target
2617 * directory (if target exists, it's on the MDT with the
2618 * target), if the source directory is remote, it's a hint that
2619 * source is remote too (this may not be true, but it won't
2620 * cause any issue), return -EXDEV early to avoid taking
2623 if (!mdt->mdt_enable_remote_rename && remote)
2624 GOTO(out_put_tgtdir, rc = -EXDEV);
2626 /* This might be further relaxed in the future for regular file
2627 * renames in different source and target parents. Start with
2628 * only same-directory renames for simplicity and because this
2629 * is by far the most the common use case.
2631 * Striped directories should be considered "remote".
2633 if (msrcdir != mtgtdir || remote ||
2634 (S_ISDIR(ma->ma_attr.la_mode) &&
2635 !mdt->mdt_enable_parallel_rename_dir) ||
2636 (!S_ISDIR(ma->ma_attr.la_mode) &&
2637 !mdt->mdt_enable_parallel_rename_file)) {
2638 rc = mdt_rename_lock(info, rename_lh);
2640 CERROR("%s: cannot lock for rename: rc = %d\n",
2641 mdt_obd_name(mdt), rc);
2642 GOTO(out_put_tgtdir, rc);
2645 if (S_ISDIR(ma->ma_attr.la_mode))
2646 msi = LPROC_MDT_RENAME_PAR_DIR;
2648 msi = LPROC_MDT_RENAME_PAR_FILE;
2651 "%s: samedir parallel rename "DFID"/"DNAME"\n",
2652 mdt_obd_name(mdt), PFID(rr->rr_fid1),
2653 PNAME(&rr->rr_name));
2657 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2659 GOTO(out_unlock_rename, rc);
2662 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2663 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
2664 CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2666 /* lock parents in the proper order. */
2667 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2668 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2669 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2670 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2672 /* In case of same dir local rename we must sort by the hash,
2673 * otherwise a lock deadlock is possible when renaming
2674 * a to b and b to a at the same time LU-15285
2676 if (!mdt_object_remote(mtgtdir) && mtgtdir == msrcdir)
2677 reverse = lh_srcdirp->mlh_pdo_hash > lh_tgtdirp->mlh_pdo_hash;
2678 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)))
2682 rc = mdt_lock_two_dirs(info, mtgtdir, lh_tgtdirp,
2683 &rr->rr_tgt_name, msrcdir, lh_srcdirp,
2686 rc = mdt_lock_two_dirs(info, msrcdir, lh_srcdirp, &rr->rr_name,
2687 mtgtdir, lh_tgtdirp, &rr->rr_tgt_name);
2690 GOTO(out_unlock_rename, rc);
2692 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2693 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2695 /* find mold object. */
2697 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2699 GOTO(out_unlock_parents, rc);
2701 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2702 GOTO(out_unlock_parents, rc = -EINVAL);
2704 if (!fid_is_md_operative(old_fid))
2705 GOTO(out_unlock_parents, rc = -EPERM);
2707 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2709 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2711 if (!mdt_object_exists(mold)) {
2712 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2714 "object does not exist");
2715 GOTO(out_put_old, rc = -ENOENT);
2718 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2719 GOTO(out_put_old, rc = -EXDEV);
2721 /* Check if @mtgtdir is subdir of @mold, before locking child
2722 * to avoid reverse locking.
2724 if (mtgtdir != msrcdir) {
2725 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2730 GOTO(out_put_old, rc);
2734 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2735 /* save version after locking */
2736 mdt_version_get_save(info, mold, 2);
2738 /* find mnew object:
2739 * mnew target object may not exist now
2740 * lookup with version checking
2743 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2746 /* the new_fid should have been filled at this moment */
2747 if (lu_fid_eq(old_fid, new_fid))
2748 GOTO(out_put_old, rc);
2750 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2751 lu_fid_eq(new_fid, rr->rr_fid2))
2752 GOTO(out_put_old, rc = -EINVAL);
2754 if (!fid_is_md_operative(new_fid))
2755 GOTO(out_put_old, rc = -EPERM);
2757 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2759 GOTO(out_put_old, rc = PTR_ERR(mnew));
2761 if (!mdt_object_exists(mnew)) {
2762 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2764 "object does not exist");
2765 GOTO(out_put_new, rc = -ENOENT);
2768 if (mdt_object_remote(mnew)) {
2769 struct mdt_body *repbody;
2771 /* Always send rename req to the target child MDT */
2772 repbody = req_capsule_server_get(info->mti_pill,
2774 LASSERT(repbody != NULL);
2775 repbody->mbo_fid1 = *new_fid;
2776 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2777 GOTO(out_put_new, rc = -EXDEV);
2779 /* Before locking the target dir, check we do not replace
2780 * a dir with a non-dir, otherwise it may deadlock with
2781 * link op which tries to create a link in this dir
2782 * back to this non-dir.
2784 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2785 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2786 GOTO(out_put_new, rc = -EISDIR);
2788 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2789 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2790 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2792 MDS_INODELOCK_LOOKUP |
2793 MDS_INODELOCK_XATTR);
2795 GOTO(out_put_new, rc);
2797 /* Check if @msrcdir is subdir of @mnew, before locking child
2798 * to avoid reverse locking.
2800 if (mtgtdir != msrcdir) {
2801 rc = mdo_is_subdir(info->mti_env,
2802 mdt_object_child(msrcdir), new_fid);
2806 GOTO(out_unlock_old, rc);
2810 /* We used to acquire MDS_INODELOCK_FULL here but we
2811 * can't do this now because a running HSM restore on
2812 * the rename onto victim will hold the layout
2813 * lock. See LU-4002.
2816 lh_newp = &info->mti_lh[MDT_LH_NEW];
2817 rc = mdt_object_check_lock(info, mtgtdir, mnew, lh_newp,
2818 MDS_INODELOCK_LOOKUP |
2819 MDS_INODELOCK_UPDATE, LCK_EX);
2821 GOTO(out_unlock_new, rc);
2823 /* get and save version after locking */
2824 mdt_version_get_save(info, mnew, 3);
2825 } else if (rc != -ENOENT) {
2826 GOTO(out_put_old, rc);
2828 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2829 lh_lookup = &info->mti_lh[MDT_LH_LOOKUP];
2830 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2832 MDS_INODELOCK_LOOKUP |
2833 MDS_INODELOCK_XATTR);
2835 GOTO(out_put_old, rc);
2837 mdt_enoent_version_save(info, 3);
2840 /* step 5: rename it */
2841 mdt_reint_init_ma(info, ma);
2843 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2844 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2847 mutex_lock(&mnew->mot_lov_mutex);
2849 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2850 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2851 mnew != NULL ? mdt_object_child(mnew) : NULL,
2852 &rr->rr_tgt_name, ma);
2855 mutex_unlock(&mnew->mot_lov_mutex);
2857 /* handle last link of tgt object */
2860 mdt_handle_last_unlink(info, mnew, ma);
2861 discard = mdt_dom_check_for_discard(info, mnew);
2863 mdt_rename_counter_tally(info, info->mti_mdt, req,
2864 msrcdir, mtgtdir, msi,
2865 ktime_us_delta(ktime_get(), kstart));
2871 /* mnew is gone, no need to keep lock */
2872 mdt_object_unlock(info, mnew, lh_newp, 1);
2874 mdt_object_unlock(info, NULL, lh_lookup, rc);
2875 mdt_object_unlock(info, mold, lh_oldp, rc);
2877 if (mnew && !discard)
2878 mdt_object_put(info->mti_env, mnew);
2880 mdt_object_put(info->mti_env, mold);
2882 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2883 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2885 mdt_rename_unlock(info, rename_lh);
2887 mdt_object_put(info->mti_env, mtgtdir);
2889 mdt_object_put(info->mti_env, msrcdir);
2891 /* The DoM discard can be done right in the place above where it is
2892 * assigned, meanwhile it is done here after rename unlock due to
2893 * compatibility with old clients, for them the discard blocks
2894 * the main thread until completion. Check LU-11359 for details.
2897 mdt_dom_discard_data(info, mnew);
2898 mdt_object_put(info->mti_env, mnew);
2900 CFS_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
2904 static int mdt_reint_resync(struct mdt_thread_info *info,
2905 struct mdt_lock_handle *lhc)
2907 struct mdt_reint_record *rr = &info->mti_rr;
2908 struct ptlrpc_request *req = mdt_info_req(info);
2909 struct md_attr *ma = &info->mti_attr;
2910 struct mdt_object *mo;
2911 struct ldlm_lock *lease;
2912 struct mdt_body *repbody;
2913 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2918 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2920 if (info->mti_dlm_req)
2921 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2923 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2925 GOTO(out, rc = PTR_ERR(mo));
2927 if (!mdt_object_exists(mo))
2928 GOTO(out_obj, rc = -ENOENT);
2930 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2931 GOTO(out_obj, rc = -EINVAL);
2933 if (mdt_object_remote(mo))
2934 GOTO(out_obj, rc = -EREMOTE);
2936 lease = ldlm_handle2lock(rr->rr_lease_handle);
2938 GOTO(out_obj, rc = -ESTALE);
2940 /* It's really necessary to grab open_sem and check if the lease lock
2941 * has been lost. There would exist a concurrent writer coming in and
2942 * generating some dirty data in memory cache, the writeback would fail
2943 * after the layout version is increased by MDS_REINT_RESYNC RPC.
2945 if (!down_write_trylock(&mo->mot_open_sem))
2946 GOTO(out_put_lease, rc = -EBUSY);
2948 lock_res_and_lock(lease);
2949 lease_broken = ldlm_is_cancel(lease);
2950 unlock_res_and_lock(lease);
2952 GOTO(out_unlock, rc = -EBUSY);
2954 /* the file has yet opened by anyone else after we took the lease. */
2955 layout.mlc_opc = MD_LAYOUT_RESYNC;
2956 lhc = &info->mti_lh[MDT_LH_LOCAL];
2957 rc = mdt_layout_change(info, mo, lhc, &layout);
2959 GOTO(out_unlock, rc);
2961 mdt_object_unlock(info, mo, lhc, 0);
2963 ma->ma_need = MA_INODE;
2965 rc = mdt_attr_get_complex(info, mo, ma);
2967 GOTO(out_unlock, rc);
2969 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2970 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2974 up_write(&mo->mot_open_sem);
2976 LDLM_LOCK_PUT(lease);
2978 mdt_object_put(info->mti_env, mo);
2980 mdt_client_compatibility(info);
2984 struct mdt_reinter {
2985 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2986 enum lprocfs_extra_opc mr_extra_opc;
2989 static const struct mdt_reinter mdt_reinters[] = {
2991 .mr_handler = &mdt_reint_setattr,
2992 .mr_extra_opc = MDS_REINT_SETATTR,
2995 .mr_handler = &mdt_reint_create,
2996 .mr_extra_opc = MDS_REINT_CREATE,
2999 .mr_handler = &mdt_reint_link,
3000 .mr_extra_opc = MDS_REINT_LINK,
3003 .mr_handler = &mdt_reint_unlink,
3004 .mr_extra_opc = MDS_REINT_UNLINK,
3007 .mr_handler = &mdt_reint_rename,
3008 .mr_extra_opc = MDS_REINT_RENAME,
3011 .mr_handler = &mdt_reint_open,
3012 .mr_extra_opc = MDS_REINT_OPEN,
3014 [REINT_SETXATTR] = {
3015 .mr_handler = &mdt_reint_setxattr,
3016 .mr_extra_opc = MDS_REINT_SETXATTR,
3019 .mr_handler = &mdt_reint_unlink,
3020 .mr_extra_opc = MDS_REINT_UNLINK,
3023 .mr_handler = &mdt_reint_migrate,
3024 .mr_extra_opc = MDS_REINT_RENAME,
3027 .mr_handler = &mdt_reint_resync,
3028 .mr_extra_opc = MDS_REINT_RESYNC,
3032 int mdt_reint_rec(struct mdt_thread_info *info,
3033 struct mdt_lock_handle *lhc)
3035 const struct mdt_reinter *mr;
3039 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3042 mr = &mdt_reinters[info->mti_rr.rr_opcode];
3043 if (mr->mr_handler == NULL)
3046 rc = (*mr->mr_handler)(info, lhc);
3048 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3049 PTLRPC_LAST_CNTR + mr->mr_extra_opc);