4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_reint.c
34 * Lustre Metadata Target (mdt) reintegration routines
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Huang Hua <huanghua@clusterfs.com>
40 * Author: Yury Umanets <umka@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lprocfs_status.h>
46 #include "mdt_internal.h"
47 #include <lustre_lmv.h>
49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
52 ma->ma_need = MA_INODE;
57 * Get version of object by fid.
59 * Return real version or ENOENT_VERSION if object doesn't exist
61 static void mdt_obj_version_get(struct mdt_thread_info *info,
62 struct mdt_object *o, __u64 *version)
66 if (mdt_object_exists(o) && !mdt_object_remote(o) &&
67 !fid_is_obf(mdt_object_fid(o)))
68 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
70 *version = ENOENT_VERSION;
71 CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
72 PFID(mdt_object_fid(o)), *version);
76 * Check version is correct.
78 * Should be called only during replay.
80 static int mdt_version_check(struct ptlrpc_request *req,
81 __u64 version, int idx)
83 __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
86 if (!exp_connect_vbr(req->rq_export))
89 LASSERT(req_is_replay(req));
90 /** VBR: version is checked always because costs nothing */
91 LASSERT(idx < PTLRPC_NUM_VERSIONS);
92 /** Sanity check for malformed buffers */
93 if (pre_ver == NULL) {
94 CERROR("No versions in request buffer\n");
95 spin_lock(&req->rq_export->exp_lock);
96 req->rq_export->exp_vbr_failed = 1;
97 spin_unlock(&req->rq_export->exp_lock);
99 } else if (pre_ver[idx] != version) {
100 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
101 pre_ver[idx], version);
102 spin_lock(&req->rq_export->exp_lock);
103 req->rq_export->exp_vbr_failed = 1;
104 spin_unlock(&req->rq_export->exp_lock);
111 * Save pre-versions in reply.
113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
118 if (!exp_connect_vbr(req->rq_export))
121 LASSERT(!req_is_replay(req));
122 LASSERT(req->rq_repmsg != NULL);
123 reply_ver = lustre_msg_get_versions(req->rq_repmsg);
125 reply_ver[idx] = version;
129 * Save enoent version, it is needed when it is obvious that object doesn't
130 * exist, e.g. child during create.
132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
134 /* save version of file name for replay, it must be ENOENT here */
135 if (!req_is_replay(mdt_info_req(info))) {
136 info->mti_ver[idx] = ENOENT_VERSION;
137 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
142 * Get version from disk and save in reply buffer.
144 * Versions are saved in reply only during normal operations not replays.
146 void mdt_version_get_save(struct mdt_thread_info *info,
147 struct mdt_object *mto, int idx)
149 /* don't save versions during replay */
150 if (!req_is_replay(mdt_info_req(info))) {
151 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
152 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
157 * Get version from disk and check it, no save in reply.
159 int mdt_version_get_check(struct mdt_thread_info *info,
160 struct mdt_object *mto, int idx)
162 /* only check versions during replay */
163 if (!req_is_replay(mdt_info_req(info)))
166 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
167 return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
171 * Get version from disk and check if recovery or just save.
173 int mdt_version_get_check_save(struct mdt_thread_info *info,
174 struct mdt_object *mto, int idx)
178 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
179 if (req_is_replay(mdt_info_req(info)))
180 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
183 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
188 * Lookup with version checking.
190 * This checks version of 'name'. Many reint functions uses 'name' for child not
191 * FID, therefore we need to get object by name and check its version.
193 static int mdt_lookup_version_check(struct mdt_thread_info *info,
194 struct mdt_object *p,
195 const struct lu_name *lname,
196 struct lu_fid *fid, int idx)
200 rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
202 /* Check version only during replay */
203 if (!req_is_replay(mdt_info_req(info)))
206 info->mti_ver[idx] = ENOENT_VERSION;
208 struct mdt_object *child;
209 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
210 if (likely(!IS_ERR(child))) {
211 mdt_obj_version_get(info, child, &info->mti_ver[idx]);
212 mdt_object_put(info->mti_env, child);
215 vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
216 return vbrc ? vbrc : rc;
220 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
221 struct mdt_object *obj,
222 struct ldlm_enqueue_info *einfo,
225 union ldlm_policy_data *policy = &mti->mti_policy;
226 struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
227 struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
230 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
231 LASSERT(slave_locks);
233 memset(policy, 0, sizeof(*policy));
234 policy->l_inodebits.bits = einfo->ei_inodebits;
235 mdt_lock_handle_init(lh);
236 mdt_lock_reg_init(lh, einfo->ei_mode);
237 for (i = 0; i < slave_locks->ha_count; i++) {
238 if (test_bit(i, (void *)slave_locks->ha_map))
239 lh->mlh_rreg_lh = slave_locks->ha_handles[i];
241 lh->mlh_reg_lh = slave_locks->ha_handles[i];
242 mdt_object_unlock(mti, NULL, lh, decref);
243 slave_locks->ha_handles[i].cookie = 0ull;
246 return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
251 struct mdt_object *obj)
255 if (!S_ISDIR(obj->mot_header.loh_attr))
258 rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
261 return rc == -ENODATA ? 0 : rc;
267 * Lock slave stripes if necessary, the lock handles of slave stripes
268 * will be stored in einfo->ei_cbdata.
270 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
271 enum ldlm_mode mode, __u64 ibits,
272 struct ldlm_enqueue_info *einfo)
274 union ldlm_policy_data *policy = &mti->mti_policy;
276 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
278 einfo->ei_type = LDLM_IBITS;
279 einfo->ei_mode = mode;
280 einfo->ei_cb_bl = mdt_remote_blocking_ast;
281 einfo->ei_cb_local_bl = mdt_blocking_ast;
282 einfo->ei_cb_cp = ldlm_completion_ast;
283 einfo->ei_enq_slave = 1;
284 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
285 einfo->ei_inodebits = ibits;
286 memset(policy, 0, sizeof(*policy));
287 policy->l_inodebits.bits = ibits;
289 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
293 int mdt_reint_striped_lock(struct mdt_thread_info *info,
294 struct mdt_object *o,
295 struct mdt_lock_handle *lh,
297 struct ldlm_enqueue_info *einfo,
302 LASSERT(!mdt_object_remote(o));
304 memset(einfo, 0, sizeof(*einfo));
306 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
310 rc = mdt_object_striped(info, o);
313 mdt_object_unlock(info, o, lh, rc);
317 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
319 mdt_object_unlock(info, o, lh, rc);
320 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
327 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
328 struct mdt_object *o,
329 struct mdt_lock_handle *lh,
330 struct ldlm_enqueue_info *einfo, int decref)
332 if (einfo->ei_cbdata)
333 mdt_unlock_slaves(info, o, einfo, decref);
334 mdt_object_unlock(info, o, lh, decref);
338 * VBR: we save three versions in reply:
339 * 0 - parent. Check that parent version is the same during replay.
340 * 1 - name. Version of 'name' if file exists with the same name or
341 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
342 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
345 static int mdt_create(struct mdt_thread_info *info)
347 struct mdt_device *mdt = info->mti_mdt;
348 struct mdt_object *parent;
349 struct mdt_object *child;
350 struct mdt_lock_handle *lh;
351 struct mdt_body *repbody;
352 struct md_attr *ma = &info->mti_attr;
353 struct mdt_reint_record *rr = &info->mti_rr;
354 struct md_op_spec *spec = &info->mti_spec;
358 DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") "
360 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
362 if (!fid_is_md_operative(rr->rr_fid1))
365 if (S_ISDIR(ma->ma_attr.la_mode) &&
366 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
367 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
368 struct lu_ucred *uc = mdt_ucred(info);
369 struct obd_export *exp = mdt_info_req(info)->rq_export;
371 /* Only new clients can create remote dir( >= 2.4) and
372 * striped dir(>= 2.6), old client will return -ENOTSUPP */
373 if (!mdt_is_dne_client(exp))
376 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
377 if (!mdt_is_striped_client(exp))
380 if (!mdt->mdt_enable_striped_dir)
382 } else if (!mdt->mdt_enable_remote_dir) {
386 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
387 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
388 mdt->mdt_enable_remote_dir_gid != -1)
392 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
394 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
396 RETURN(PTR_ERR(parent));
398 if (!mdt_object_exists(parent))
399 GOTO(put_parent, rc = -ENOENT);
401 lh = &info->mti_lh[MDT_LH_PARENT];
402 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
403 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
405 GOTO(put_parent, rc);
407 if (!mdt_object_remote(parent)) {
408 rc = mdt_version_get_check_save(info, parent, 0);
410 GOTO(unlock_parent, rc);
414 * Check child name version during replay.
415 * During create replay a file may exist with same name.
417 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
418 &info->mti_tmp_fid1, 1);
420 GOTO(unlock_parent, rc = -EEXIST);
422 /* -ENOENT is expected here */
424 GOTO(unlock_parent, rc);
426 /* save version of file name for replay, it must be ENOENT here */
427 mdt_enoent_version_save(info, 1);
429 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
430 if (unlikely(IS_ERR(child)))
431 GOTO(unlock_parent, rc = PTR_ERR(child));
433 ma->ma_need = MA_INODE;
436 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
437 OBD_FAIL_MDS_REINT_CREATE_WRITE);
439 /* Version of child will be updated on disk. */
440 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
441 rc = mdt_version_get_check_save(info, child, 2);
445 /* Let lower layer know current lock mode. */
446 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
449 * Do not perform lookup sanity check. We know that name does
452 info->mti_spec.sp_cr_lookup = 0;
453 info->mti_spec.sp_feat = &dt_directory_features;
455 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
456 mdt_object_child(child), &info->mti_spec, ma);
458 rc = mdt_attr_get_complex(info, child, ma);
464 * On DNE, we need to eliminate dependey between 'mkdir a' and
465 * 'mkdir a/b' if b is a striped directory, to achieve this, two
466 * things are done below:
467 * 1. save child and slaves lock.
468 * 2. if the child is a striped directory, relock parent so to
469 * compare against with COS locks to ensure parent was
472 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
473 struct mdt_lock_handle *lhc;
474 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
477 rc = mdt_object_striped(info, child);
483 if (!mdt_object_remote(parent)) {
484 mdt_object_unlock(info, parent, lh, 1);
485 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
486 rc = mdt_reint_object_lock(info, parent, lh,
487 MDS_INODELOCK_UPDATE,
494 lhc = &info->mti_lh[MDT_LH_CHILD];
495 mdt_lock_handle_init(lhc);
496 mdt_lock_reg_init(lhc, LCK_PW);
497 rc = mdt_reint_striped_lock(info, child, lhc,
498 MDS_INODELOCK_UPDATE, einfo,
503 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
506 /* Return fid & attr to client. */
507 if (ma->ma_valid & MA_INODE)
508 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
509 mdt_object_fid(child));
511 mdt_object_put(info->mti_env, child);
513 mdt_object_unlock(info, parent, lh, rc);
515 mdt_object_put(info->mti_env, parent);
519 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
522 struct mdt_lock_handle *lh;
523 int do_vbr = ma->ma_attr.la_valid &
524 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
525 __u64 lockpart = MDS_INODELOCK_UPDATE;
526 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
531 rc = mdt_object_striped(info, mo);
537 lh = &info->mti_lh[MDT_LH_PARENT];
538 mdt_lock_reg_init(lh, LCK_PW);
540 /* Even though the new MDT will grant PERM lock to the old
541 * client, but the old client will almost ignore that during
542 * So it needs to revoke both LOOKUP and PERM lock here, so
543 * both new and old client can cancel the dcache */
544 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
545 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
547 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
552 /* all attrs are packed into mti_attr in unpack_setattr */
553 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
554 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
556 /* VBR: update version if attr changed are important for recovery */
558 /* update on-disk version of changed object */
559 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
560 rc = mdt_version_get_check_save(info, mo, 0);
562 GOTO(out_unlock, rc);
565 /* Ensure constant striping during chown(). See LU-2789. */
566 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
567 mutex_lock(&mo->mot_lov_mutex);
569 /* all attrs are packed into mti_attr in unpack_setattr */
570 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
572 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
573 mutex_unlock(&mo->mot_lov_mutex);
576 GOTO(out_unlock, rc);
577 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
580 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
585 * Check HSM flags and add HS_DIRTY flag if relevant.
587 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
588 * and is not RELEASED.
590 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
593 struct lu_ucred *uc = mdt_ucred(info);
598 /* If the file was modified, add the dirty flag */
599 ma->ma_need = MA_HSM;
600 rc = mdt_attr_get_complex(info, mo, ma);
602 CERROR("file attribute read error for "DFID": %d.\n",
603 PFID(mdt_object_fid(mo)), rc);
607 /* If an up2date copy exists in the backend, add dirty flag */
608 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
609 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
610 ma->ma_hsm.mh_flags |= HS_DIRTY;
612 /* Bump cap so that closes from non-owner writers can
613 * set the HSM state to dirty. */
614 cap_saved = uc->uc_cap;
615 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
616 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
617 uc->uc_cap = cap_saved;
619 CERROR("file attribute change error for "DFID": %d\n",
620 PFID(mdt_object_fid(mo)), rc);
626 static int mdt_reint_setattr(struct mdt_thread_info *info,
627 struct mdt_lock_handle *lhc)
629 struct mdt_device *mdt = info->mti_mdt;
630 struct md_attr *ma = &info->mti_attr;
631 struct mdt_reint_record *rr = &info->mti_rr;
632 struct ptlrpc_request *req = mdt_info_req(info);
633 struct mdt_object *mo;
634 struct mdt_body *repbody;
638 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
639 (unsigned int)ma->ma_attr.la_valid);
641 if (info->mti_dlm_req)
642 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
644 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
645 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
647 GOTO(out, rc = PTR_ERR(mo));
649 if (!mdt_object_exists(mo))
650 GOTO(out_put, rc = -ENOENT);
652 if (mdt_object_remote(mo))
653 GOTO(out_put, rc = -EREMOTE);
655 /* revoke lease lock if size is going to be changed */
656 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
657 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
658 atomic_read(&mo->mot_lease_count) > 0)) {
659 down_read(&mo->mot_open_sem);
661 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
662 lhc = &info->mti_lh[MDT_LH_LOCAL];
663 mdt_lock_reg_init(lhc, LCK_CW);
665 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
667 up_read(&mo->mot_open_sem);
671 /* revoke lease lock */
672 mdt_object_unlock(info, mo, lhc, 1);
674 up_read(&mo->mot_open_sem);
677 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
678 /* Check write access for the O_TRUNC case */
679 if (mdt_write_read(mo) < 0)
680 GOTO(out_put, rc = -ETXTBSY);
682 /* LU-10286: compatibility check for FLR.
683 * Please check the comment in mdt_finish_open() for details */
684 if (!exp_connect_flr(info->mti_exp)) {
685 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
686 if (rc < 0 && rc != -ENODATA)
689 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
690 GOTO(out_put, rc = -EOPNOTSUPP);
693 /* For truncate, the file size sent from client
694 * is believable, but the blocks are incorrect,
695 * which makes the block size in LSOM attribute
696 * inconsisent with the real block size.
698 rc = mdt_lsom_update(info, mo, true);
703 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
704 if (ma->ma_valid & MA_LOV)
705 GOTO(out_put, rc = -EPROTO);
707 /* MDT supports FMD for regular files due to Data-on-MDT */
708 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
709 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
710 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
713 rc = mdt_attr_set(info, mo, ma);
716 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
717 (ma->ma_valid & MA_INODE)) {
718 struct lu_buf *buf = &info->mti_buf;
719 struct lu_ucred *uc = mdt_ucred(info);
720 struct mdt_lock_handle *lh;
722 /* reject if either remote or striped dir is disabled */
723 if (ma->ma_valid & MA_LMV) {
724 if (!mdt->mdt_enable_remote_dir ||
725 !mdt->mdt_enable_striped_dir)
726 GOTO(out_put, rc = -EPERM);
728 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
729 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
730 mdt->mdt_enable_remote_dir_gid != -1)
731 GOTO(out_put, rc = -EPERM);
734 if (ma->ma_attr.la_valid != 0)
735 GOTO(out_put, rc = -EPROTO);
737 lh = &info->mti_lh[MDT_LH_PARENT];
738 mdt_lock_reg_init(lh, LCK_PW);
740 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
744 if (ma->ma_valid & MA_LOV) {
745 buf->lb_buf = ma->ma_lmm;
746 buf->lb_len = ma->ma_lmm_size;
748 buf->lb_buf = ma->ma_lmv;
749 buf->lb_len = ma->ma_lmv_size;
751 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
752 (ma->ma_valid & MA_LOV) ?
753 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
756 mdt_object_unlock(info, mo, lh, rc);
760 GOTO(out_put, rc = -EPROTO);
763 /* If file data is modified, add the dirty flag */
764 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
765 rc = mdt_add_dirty_flag(info, mo, ma);
767 ma->ma_need = MA_INODE;
769 rc = mdt_attr_get_complex(info, mo, ma);
773 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
777 mdt_object_put(info->mti_env, mo);
780 mdt_counter_incr(req, LPROC_MDT_SETATTR);
782 mdt_client_compatibility(info);
783 rc2 = mdt_fix_reply(info);
789 static int mdt_reint_create(struct mdt_thread_info *info,
790 struct mdt_lock_handle *lhc)
792 struct ptlrpc_request *req = mdt_info_req(info);
796 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
797 RETURN(err_serious(-ESTALE));
799 if (info->mti_dlm_req)
800 ldlm_request_cancel(mdt_info_req(info),
801 info->mti_dlm_req, 0, LATF_SKIP);
803 if (!lu_name_is_valid(&info->mti_rr.rr_name))
806 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
808 mdt_counter_incr(req, LPROC_MDT_MKDIR);
816 /* Special file should stay on the same node as parent. */
817 mdt_counter_incr(req, LPROC_MDT_MKNOD);
820 CERROR("%s: Unsupported mode %o\n",
821 mdt_obd_name(info->mti_mdt),
822 info->mti_attr.ma_attr.la_mode);
823 RETURN(err_serious(-EOPNOTSUPP));
826 rc = mdt_create(info);
831 * VBR: save parent version in reply and child version getting by its name.
832 * Version of child is getting and checking during its lookup. If
834 static int mdt_reint_unlink(struct mdt_thread_info *info,
835 struct mdt_lock_handle *lhc)
837 struct mdt_reint_record *rr = &info->mti_rr;
838 struct ptlrpc_request *req = mdt_info_req(info);
839 struct md_attr *ma = &info->mti_attr;
840 struct lu_fid *child_fid = &info->mti_tmp_fid1;
841 struct mdt_object *mp;
842 struct mdt_object *mc;
843 struct mdt_lock_handle *parent_lh;
844 struct mdt_lock_handle *child_lh;
845 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
847 bool cos_incompat = false, discard = false;
853 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
854 PNAME(&rr->rr_name));
856 if (info->mti_dlm_req)
857 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
859 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
860 RETURN(err_serious(-ENOENT));
862 if (!fid_is_md_operative(rr->rr_fid1))
865 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
869 if (mdt_object_remote(mp)) {
872 rc = mdt_version_get_check_save(info, mp, 0);
874 GOTO(put_parent, rc);
878 parent_lh = &info->mti_lh[MDT_LH_PARENT];
879 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
880 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
883 GOTO(put_parent, rc);
885 /* lookup child object along with version checking */
887 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
889 /* Name might not be able to find during resend of
890 * remote unlink, considering following case.
891 * dir_A is a remote directory, the name entry of
892 * dir_A is on MDT0, the directory is on MDT1,
894 * 1. client sends unlink req to MDT1.
895 * 2. MDT1 sends name delete update to MDT0.
896 * 3. name entry is being deleted in MDT0 synchronously.
897 * 4. MDT1 is restarted.
898 * 5. client resends unlink req to MDT1. So it can not
899 * find the name entry on MDT0 anymore.
900 * In this case, MDT1 only needs to destory the local
903 if (mdt_object_remote(mp) && rc == -ENOENT &&
904 !fid_is_zero(rr->rr_fid2) &&
905 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
907 *child_fid = *rr->rr_fid2;
909 GOTO(unlock_parent, rc);
913 if (!fid_is_md_operative(child_fid))
914 GOTO(unlock_parent, rc = -EPERM);
916 /* We will lock the child regardless it is local or remote. No harm. */
917 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
919 GOTO(unlock_parent, rc = PTR_ERR(mc));
922 rc = mdt_object_striped(info, mc);
924 GOTO(unlock_parent, rc);
928 mdt_object_put(info->mti_env, mc);
929 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
934 child_lh = &info->mti_lh[MDT_LH_CHILD];
935 mdt_lock_reg_init(child_lh, LCK_EX);
936 if (info->mti_spec.sp_rm_entry) {
937 struct lu_ucred *uc = mdt_ucred(info);
939 if (!mdt_is_dne_client(req->rq_export))
940 /* Return -ENOTSUPP for old client */
941 GOTO(put_child, rc = -ENOTSUPP);
943 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
944 GOTO(put_child, rc = -EPERM);
946 ma->ma_need = MA_INODE;
948 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
949 NULL, &rr->rr_name, ma, no_name);
953 if (mdt_object_remote(mc)) {
954 struct mdt_body *repbody;
956 if (!fid_is_zero(rr->rr_fid2)) {
957 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
958 mdt_obd_name(info->mti_mdt),
959 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
960 GOTO(put_child, rc = -ENOENT);
962 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
963 mdt_obd_name(info->mti_mdt),
964 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
966 if (!mdt_is_dne_client(req->rq_export))
967 /* Return -ENOTSUPP for old client */
968 GOTO(put_child, rc = -ENOTSUPP);
970 /* Revoke the LOOKUP lock of the remote object granted by
971 * this MDT. Since the unlink will happen on another MDT,
972 * it will release the LOOKUP lock right away. Then What
973 * would happen if another client try to grab the LOOKUP
974 * lock at the same time with unlink XXX */
975 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
976 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
977 LASSERT(repbody != NULL);
978 repbody->mbo_fid1 = *mdt_object_fid(mc);
979 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
980 GOTO(unlock_child, rc = -EREMOTE);
982 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
983 * this now because a running HSM restore on the child (unlink
984 * victim) will hold the layout lock. See LU-4002. */
985 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
986 if (mdt_object_remote(mp)) {
987 /* Enqueue lookup lock from parent MDT */
988 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
989 &child_lh->mlh_rreg_lh,
990 child_lh->mlh_rreg_mode,
991 MDS_INODELOCK_LOOKUP, false);
995 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
998 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1001 GOTO(put_child, rc);
1004 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1005 * whether need MA_LOV and MA_COOKIE.
1007 ma->ma_need = MA_INODE;
1010 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1011 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1012 /* save version when object is locked */
1013 mdt_version_get_save(info, mc, 1);
1015 mutex_lock(&mc->mot_lov_mutex);
1017 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1018 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1020 mutex_unlock(&mc->mot_lov_mutex);
1022 GOTO(unlock_child, rc);
1024 if (!lu_object_is_dying(&mc->mot_header)) {
1025 rc = mdt_attr_get_complex(info, mc, ma);
1029 discard = mdt_dom_check_for_discard(info, mc);
1031 mdt_handle_last_unlink(info, mc, ma);
1034 if (ma->ma_valid & MA_INODE) {
1035 switch (ma->ma_attr.la_mode & S_IFMT) {
1037 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1045 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1048 LASSERTF(0, "bad file type %o unlinking\n",
1049 ma->ma_attr.la_mode);
1056 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1058 mdt_object_put(info->mti_env, mc);
1060 mdt_object_unlock(info, mp, parent_lh, rc);
1062 mdt_object_put(info->mti_env, mp);
1064 /* discard is just a PW DOM lock to drop the data on a client
1065 * no need to keep objects being get and locked, do that after all.
1068 mdt_dom_discard_data(info, child_fid);
1074 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1077 static int mdt_reint_link(struct mdt_thread_info *info,
1078 struct mdt_lock_handle *lhc)
1080 struct mdt_reint_record *rr = &info->mti_rr;
1081 struct ptlrpc_request *req = mdt_info_req(info);
1082 struct md_attr *ma = &info->mti_attr;
1083 struct mdt_object *ms;
1084 struct mdt_object *mp;
1085 struct mdt_lock_handle *lhs;
1086 struct mdt_lock_handle *lhp;
1091 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1092 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1094 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1095 RETURN(err_serious(-ENOENT));
1097 if (info->mti_dlm_req)
1098 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1100 /* Invalid case so return error immediately instead of
1102 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1105 if (!fid_is_md_operative(rr->rr_fid1) ||
1106 !fid_is_md_operative(rr->rr_fid2))
1109 /* step 1: find target parent dir */
1110 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1112 RETURN(PTR_ERR(mp));
1114 rc = mdt_version_get_check_save(info, mp, 0);
1116 GOTO(put_parent, rc);
1118 /* step 2: find source */
1119 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1121 GOTO(put_parent, rc = PTR_ERR(ms));
1123 if (!mdt_object_exists(ms)) {
1124 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1125 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1126 GOTO(put_source, rc = -ENOENT);
1129 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1131 lhp = &info->mti_lh[MDT_LH_PARENT];
1132 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1133 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1136 GOTO(put_source, rc);
1138 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1140 lhs = &info->mti_lh[MDT_LH_CHILD];
1141 mdt_lock_reg_init(lhs, LCK_EX);
1142 rc = mdt_reint_object_lock(info, ms, lhs,
1143 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1146 GOTO(unlock_parent, rc);
1148 /* step 3: link it */
1149 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1150 OBD_FAIL_MDS_REINT_LINK_WRITE);
1152 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1153 rc = mdt_version_get_check_save(info, ms, 1);
1155 GOTO(unlock_source, rc);
1157 /** check target version by name during replay */
1158 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1159 &info->mti_tmp_fid1, 2);
1160 if (rc != 0 && rc != -ENOENT)
1161 GOTO(unlock_source, rc);
1162 /* save version of file name for replay, it must be ENOENT here */
1163 if (!req_is_replay(mdt_info_req(info))) {
1164 if (rc != -ENOENT) {
1165 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1166 PNAME(&rr->rr_name));
1167 GOTO(unlock_source, rc = -EEXIST);
1169 info->mti_ver[2] = ENOENT_VERSION;
1170 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1173 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1174 mdt_object_child(ms), &rr->rr_name, ma);
1177 mdt_counter_incr(req, LPROC_MDT_LINK);
1181 mdt_object_unlock(info, ms, lhs, rc);
1183 mdt_object_unlock(info, mp, lhp, rc);
1185 mdt_object_put(info->mti_env, ms);
1187 mdt_object_put(info->mti_env, mp);
1191 * lock the part of the directory according to the hash of the name
1192 * (lh->mlh_pdo_hash) in parallel directory lock.
1194 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1195 struct mdt_lock_handle *lh,
1196 struct mdt_object *obj, __u64 ibits,
1199 struct ldlm_res_id *res = &info->mti_res_id;
1200 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1201 union ldlm_policy_data *policy = &info->mti_policy;
1202 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1206 * Finish res_id initializing by name hash marking part of
1207 * directory which is taking modification.
1209 LASSERT(lh->mlh_pdo_hash != 0);
1210 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1211 memset(policy, 0, sizeof(*policy));
1212 policy->l_inodebits.bits = ibits;
1214 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1215 dlmflags |= LDLM_FL_COS_INCOMPAT;
1217 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1218 * going to be sent to client. If it is - mdt_intent_policy() path will
1219 * fix it up and turn FL_LOCAL flag off.
1221 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1222 policy, res, dlmflags,
1223 &info->mti_exp->exp_handle.h_cookie);
1228 * Get BFL lock for rename or migrate process.
1230 static int mdt_rename_lock(struct mdt_thread_info *info,
1231 struct lustre_handle *lh)
1236 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1237 struct lu_fid *fid = &info->mti_tmp_fid1;
1238 struct mdt_object *obj;
1240 /* XXX, right now, it has to use object API to
1241 * enqueue lock cross MDT, so it will enqueue
1242 * rename lock(with LUSTRE_BFL_FID) by root object */
1244 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1246 RETURN(PTR_ERR(obj));
1248 rc = mdt_remote_object_lock(info, obj,
1249 &LUSTRE_BFL_FID, lh,
1251 MDS_INODELOCK_UPDATE, false);
1252 mdt_object_put(info->mti_env, obj);
1254 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1255 union ldlm_policy_data *policy = &info->mti_policy;
1256 struct ldlm_res_id *res_id = &info->mti_res_id;
1259 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1260 memset(policy, 0, sizeof *policy);
1261 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1262 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1263 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1264 LDLM_IBITS, policy, LCK_EX, &flags,
1266 ldlm_completion_ast, NULL, NULL, 0,
1268 &info->mti_exp->exp_handle.h_cookie,
1275 static void mdt_rename_unlock(struct lustre_handle *lh)
1278 LASSERT(lustre_handle_is_used(lh));
1279 /* Cancel the single rename lock right away */
1280 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1284 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1285 const struct lu_fid *fid,
1288 struct mdt_object *dir;
1293 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1297 /* check early, the real version will be saved after locking */
1298 rc = mdt_version_get_check(info, dir, idx);
1302 if (!mdt_object_exists(dir))
1303 GOTO(out_put, rc = -ENOENT);
1305 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1306 GOTO(out_put, rc = -ENOTDIR);
1310 mdt_object_put(info->mti_env, dir);
1315 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1316 * herein we don't really check it, just do revoke.
1318 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1319 struct mdt_object *pobj,
1320 struct mdt_object *obj)
1322 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1325 mdt_lock_handle_init(lh);
1326 mdt_lock_reg_init(lh, LCK_EX);
1328 if (mdt_object_remote(pobj)) {
1329 /* don't bother to check if pobj and obj are on the same MDT. */
1330 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1331 &lh->mlh_rreg_lh, LCK_EX,
1332 MDS_INODELOCK_LOOKUP, false);
1333 } else if (mdt_object_remote(obj)) {
1334 struct ldlm_res_id *res = &info->mti_res_id;
1335 union ldlm_policy_data *policy = &info->mti_policy;
1336 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1337 LDLM_FL_COS_INCOMPAT;
1339 fid_build_reg_res_name(mdt_object_fid(obj), res);
1340 memset(policy, 0, sizeof(*policy));
1341 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1342 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1343 &lh->mlh_reg_lh, LCK_EX, policy, res,
1346 /* do nothing if both are local */
1354 * TODO, currently we don't save this lock because there is no place to
1355 * hold this lock handle, but to avoid race we need to save this lock.
1357 mdt_object_unlock(info, NULL, lh, 1);
1363 * operation may takes locks of linkea, or directory stripes, group them in
1366 struct mdt_sub_lock {
1367 struct mdt_object *msl_obj;
1368 struct mdt_lock_handle msl_lh;
1369 struct list_head msl_linkage;
1372 static void mdt_unlock_list(struct mdt_thread_info *info,
1373 struct list_head *list, int decref)
1375 struct mdt_sub_lock *msl;
1376 struct mdt_sub_lock *tmp;
1378 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1379 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1380 list_del(&msl->msl_linkage);
1385 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1386 struct mdt_object *obj,
1387 struct mdt_lock_handle *lh,
1388 struct ldlm_enqueue_info *einfo,
1389 struct list_head *slave_locks,
1392 if (mdt_object_remote(obj)) {
1393 mdt_unlock_list(info, slave_locks, decref);
1394 mdt_object_unlock(info, obj, lh, decref);
1396 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1401 * lock parents of links, and also check whether total locks don't exceed
1404 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1405 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1406 * \retval -ev negative errno upon error
1408 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1409 struct mdt_object *pobj,
1410 const struct md_attr *ma,
1411 struct mdt_object *obj,
1412 struct mdt_lock_handle *lhp,
1413 struct ldlm_enqueue_info *peinfo,
1414 struct list_head *parent_slave_locks,
1415 struct list_head *link_locks)
1417 struct mdt_device *mdt = info->mti_mdt;
1418 struct lu_buf *buf = &info->mti_big_buf;
1419 struct lu_name *lname = &info->mti_name;
1420 struct linkea_data ldata = { NULL };
1421 bool blocked = false;
1422 int local_lnkp_cnt = 0;
1427 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1430 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1431 if (buf->lb_buf == NULL)
1435 rc = mdt_links_read(info, obj, &ldata);
1437 if (rc == -ENOENT || rc == -ENODATA)
1442 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1443 linkea_next_entry(&ldata)) {
1444 struct mdt_object *lnkp;
1445 struct mdt_sub_lock *msl;
1449 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1452 /* check if it's also linked to parent */
1453 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1454 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1455 PFID(&fid), PNAME(lname));
1456 /* in case link is remote object, revoke LOOKUP lock */
1457 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1463 /* check if it's linked to a stripe of parent */
1464 if (ma->ma_valid & MA_LMV) {
1465 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1466 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1469 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1470 fid_le_to_cpu(stripe_fid,
1471 &lmv->lmv_stripe_fids[j]);
1472 if (lu_fid_eq(stripe_fid, &fid)) {
1473 CDEBUG(D_INFO, "skip stripe "DFID
1474 ", reovke "DNAME"\n",
1475 PFID(&fid), PNAME(lname));
1476 lnkp = mdt_object_find(info->mti_env,
1479 GOTO(out, rc = PTR_ERR(lnkp));
1485 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1487 mdt_object_put(info->mti_env, lnkp);
1492 /* Check if it's already locked */
1493 list_for_each_entry(msl, link_locks, msl_linkage) {
1494 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1496 DFID" was locked, revoke "DNAME"\n",
1497 PFID(&fid), PNAME(lname));
1498 lnkp = msl->msl_obj;
1504 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1508 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1509 PFID(&fid), PNAME(lname));
1511 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1513 CWARN("%s: cannot find obj "DFID": %ld\n",
1514 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1518 if (!mdt_object_exists(lnkp)) {
1519 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1520 PFID(&fid), PNAME(lname));
1521 mdt_object_put(info->mti_env, lnkp);
1525 if (!mdt_object_remote(lnkp))
1530 GOTO(out, rc = -ENOMEM);
1533 * we can't follow parent-child lock order like other MD
1534 * operations, use lock_try here to avoid deadlock, if the lock
1535 * cannot be taken, drop all locks taken, revoke the blocked
1536 * one, and continue processing the remaining entries, and in
1537 * the end of the loop restart from beginning.
1539 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1541 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1542 MDS_INODELOCK_UPDATE, true);
1543 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1545 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1546 PFID(&fid), PNAME(lname));
1548 mdt_unlock_list(info, link_locks, 1);
1549 /* also unlock parent locks to avoid deadlock */
1551 mdt_migrate_object_unlock(info, pobj, lhp,
1558 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1559 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1560 MDS_INODELOCK_UPDATE);
1562 mdt_object_put(info->mti_env, lnkp);
1567 if (mdt_object_remote(lnkp)) {
1568 struct ldlm_lock *lock;
1571 * for remote object, set lock cb_atomic,
1572 * so lock can be released in blocking_ast()
1573 * immediately, then the next lock_try will
1574 * have better chance of success.
1576 lock = ldlm_handle2lock(
1577 &msl->msl_lh.mlh_rreg_lh);
1578 LASSERT(lock != NULL);
1579 lock_res_and_lock(lock);
1580 ldlm_set_atomic_cb(lock);
1581 unlock_res_and_lock(lock);
1582 LDLM_LOCK_PUT(lock);
1585 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1590 INIT_LIST_HEAD(&msl->msl_linkage);
1591 msl->msl_obj = lnkp;
1592 list_add_tail(&msl->msl_linkage, link_locks);
1594 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1598 GOTO(out, rc = -EBUSY);
1603 mdt_unlock_list(info, link_locks, rc);
1604 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1606 * parent may have 3 local objects: master object and 2 stripes
1607 * (if it's being migrated too); source may have 2 local
1608 * objects: master and 1 stripe; target has 1 local object.
1614 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1615 struct mdt_object *obj,
1616 const struct md_attr *ma,
1617 struct list_head *slave_locks)
1619 struct mdt_device *mdt = info->mti_mdt;
1620 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1621 struct lu_fid *fid = &info->mti_tmp_fid1;
1622 struct mdt_object *slave;
1623 struct mdt_sub_lock *msl;
1629 LASSERT(mdt_object_remote(obj));
1630 LASSERT(ma->ma_valid & MA_LMV);
1633 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1636 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1639 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1640 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1642 slave = mdt_object_find(info->mti_env, mdt, fid);
1644 GOTO(out, rc = PTR_ERR(slave));
1648 mdt_object_put(info->mti_env, slave);
1649 GOTO(out, rc = -ENOMEM);
1652 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1653 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1654 MDS_INODELOCK_UPDATE, true);
1657 mdt_object_put(info->mti_env, slave);
1661 INIT_LIST_HEAD(&msl->msl_linkage);
1662 msl->msl_obj = slave;
1663 list_add_tail(&msl->msl_linkage, slave_locks);
1669 mdt_unlock_list(info, slave_locks, rc);
1673 /* lock parent and its stripes */
1674 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1675 struct mdt_object *obj,
1676 const struct md_attr *ma,
1677 struct mdt_lock_handle *lh,
1678 struct ldlm_enqueue_info *einfo,
1679 struct list_head *slave_locks)
1683 if (mdt_object_remote(obj)) {
1684 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1685 &lh->mlh_rreg_lh, LCK_PW,
1686 MDS_INODELOCK_UPDATE, false);
1691 * if obj is remote and striped, lock its stripes explicitly
1692 * because it's not striped in LOD layer on this MDT.
1694 if (ma->ma_valid & MA_LMV) {
1695 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1697 mdt_object_unlock(info, obj, lh, rc);
1700 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1708 * in migration, object may be remote, and we need take full lock of it and its
1709 * stripes if it's directory, besides, object may be a remote object on its
1710 * parent, revoke its LOOKUP lock on where its parent is located.
1712 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1713 struct mdt_object *pobj,
1714 struct mdt_object *obj,
1715 struct mdt_lock_handle *lh,
1716 struct ldlm_enqueue_info *einfo,
1717 struct list_head *slave_locks)
1721 if (mdt_object_remote(obj)) {
1722 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1726 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1727 &lh->mlh_rreg_lh, LCK_EX,
1728 MDS_INODELOCK_FULL, false);
1733 * if obj is remote and striped, lock its stripes explicitly
1734 * because it's not striped in LOD layer on this MDT.
1736 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1737 struct md_attr *ma = &info->mti_attr;
1739 ma->ma_lmv = info->mti_big_lmm;
1740 ma->ma_lmv_size = info->mti_big_lmmsize;
1742 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1744 mdt_object_unlock(info, obj, lh, rc);
1748 if (ma->ma_valid & MA_LMV) {
1749 rc = mdt_lock_remote_slaves(info, obj, ma,
1752 mdt_object_unlock(info, obj, lh, rc);
1756 if (mdt_object_remote(pobj)) {
1757 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1762 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1770 * lookup source by name, if parent is striped directory, we need to find the
1771 * corresponding stripe where source is located, and then lookup there.
1773 * besides, if parent is migrating too, and file is already in target stripe,
1774 * this should be a redo of 'lfs migrate' on client side.
1776 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1777 struct mdt_object *pobj,
1778 const struct md_attr *ma,
1779 const struct lu_name *lname,
1780 struct mdt_object **spobj,
1781 struct mdt_object **sobj)
1783 const struct lu_env *env = info->mti_env;
1784 struct lu_fid *fid = &info->mti_tmp_fid1;
1785 struct mdt_object *stripe;
1788 if (ma->ma_valid & MA_LMV) {
1789 /* if parent is striped, lookup on corresponding stripe */
1790 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1791 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1792 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1793 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1794 LMV_HASH_FLAG_MIGRATION;
1797 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1798 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1801 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1807 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1808 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1810 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1812 stripe = mdt_object_find(env, info->mti_mdt, fid);
1814 return PTR_ERR(stripe);
1817 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1819 if (rc == -ENOENT && is_migrating) {
1821 * if parent is migrating, and lookup child failed on
1822 * source stripe, lookup again on target stripe, if it
1823 * exists, it means previous migration was interrupted,
1824 * and current file was migrated already.
1826 mdt_object_put(env, stripe);
1828 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1829 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1831 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1837 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1839 stripe = mdt_object_find(env, info->mti_mdt, fid);
1841 return PTR_ERR(stripe);
1844 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1845 fid, &info->mti_spec);
1846 mdt_object_put(env, stripe);
1847 return rc ?: -EALREADY;
1849 mdt_object_put(env, stripe);
1854 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1860 mdt_object_get(env, stripe);
1865 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1866 if (IS_ERR(*sobj)) {
1867 mdt_object_put(env, stripe);
1868 rc = PTR_ERR(*sobj);
1876 /* end lease and close file for regular file */
1877 static int mdd_migrate_close(struct mdt_thread_info *info,
1878 struct mdt_object *obj)
1880 struct close_data *data;
1881 struct mdt_body *repbody;
1882 struct ldlm_lock *lease;
1887 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1889 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1893 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1898 lease = ldlm_handle2lock(&data->cd_handle);
1902 /* check if the lease was already canceled */
1903 lock_res_and_lock(lease);
1904 rc = ldlm_is_cancel(lease);
1905 unlock_res_and_lock(lease);
1909 LDLM_DEBUG(lease, DFID" lease broken",
1910 PFID(mdt_object_fid(obj)));
1914 * cancel server side lease, client side counterpart should have been
1915 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1917 ldlm_lock_cancel(lease);
1918 ldlm_reprocess_all(lease->l_resource);
1919 LDLM_LOCK_PUT(lease);
1922 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1923 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1924 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1930 * migrate file in below steps:
1931 * 1. lock parent and its stripes
1932 * 2. lookup source by name
1933 * 3. lock parents of source links if source is not directory
1934 * 4. reject if source is in HSM
1935 * 5. take source open_sem and close file if source is regular file
1936 * 6. lock source and its stripes if it's directory
1937 * 7. lock target so subsequent change to it can trigger COS
1939 * 9. unlock above locks
1940 * 10. sync device if source has links
1942 static int mdt_reint_migrate(struct mdt_thread_info *info,
1943 struct mdt_lock_handle *unused)
1945 const struct lu_env *env = info->mti_env;
1946 struct mdt_device *mdt = info->mti_mdt;
1947 struct ptlrpc_request *req = mdt_info_req(info);
1948 struct mdt_reint_record *rr = &info->mti_rr;
1949 struct lu_ucred *uc = mdt_ucred(info);
1950 struct md_attr *ma = &info->mti_attr;
1951 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1952 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1953 struct mdt_object *pobj;
1954 struct mdt_object *spobj = NULL;
1955 struct mdt_object *sobj = NULL;
1956 struct mdt_object *tobj;
1957 struct lustre_handle rename_lh = { 0 };
1958 struct mdt_lock_handle *lhp;
1959 struct mdt_lock_handle *lhs;
1960 struct mdt_lock_handle *lht;
1961 LIST_HEAD(parent_slave_locks);
1962 LIST_HEAD(child_slave_locks);
1963 LIST_HEAD(link_locks);
1964 int lock_retries = 5;
1965 bool open_sem_locked = false;
1966 bool do_sync = false;
1970 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1971 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1973 if (info->mti_dlm_req)
1974 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1976 if (!fid_is_md_operative(rr->rr_fid1) ||
1977 !fid_is_md_operative(rr->rr_fid2))
1980 /* don't allow migrate . or .. */
1981 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1984 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1987 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1988 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1989 mdt->mdt_enable_remote_dir_gid != -1)
1993 * Note: do not enqueue rename lock for replay request, because
1994 * if other MDT holds rename lock, but being blocked to wait for
1995 * this MDT to finish its recovery, and the failover MDT can not
1996 * get rename lock, which will cause deadlock.
1998 if (!req_is_replay(req)) {
1999 rc = mdt_rename_lock(info, &rename_lh);
2001 CERROR("%s: can't lock FS for rename: rc = %d\n",
2002 mdt_obd_name(info->mti_mdt), rc);
2007 /* pobj is master object of parent */
2008 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2010 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2012 if (unlikely(!info->mti_big_lmm)) {
2013 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2014 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2015 if (!info->mti_big_lmm)
2016 GOTO(put_parent, rc = -ENOMEM);
2019 ma->ma_lmv = info->mti_big_lmm;
2020 ma->ma_lmv_size = info->mti_big_lmmsize;
2022 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2024 GOTO(put_parent, rc);
2027 /* lock parent object */
2028 lhp = &info->mti_lh[MDT_LH_PARENT];
2029 mdt_lock_reg_init(lhp, LCK_PW);
2030 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2031 &parent_slave_locks);
2033 GOTO(put_parent, rc);
2036 * spobj is the corresponding stripe against name if pobj is striped
2037 * directory, which is the real parent, and no need to lock, because
2038 * we've taken full lock of pobj.
2040 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2042 GOTO(unlock_parent, rc);
2044 /* lock parents of source links, and revoke LOOKUP lock of links */
2045 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2046 &parent_slave_locks, &link_locks);
2047 if (rc == -EBUSY && lock_retries-- > 0) {
2048 mdt_object_put(env, sobj);
2049 mdt_object_put(env, spobj);
2054 GOTO(put_source, rc);
2057 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2058 * with one request, if total lock count exceeds this limit, we will
2059 * drop all locks after migration, and synchronous device in the end.
2063 /* TODO: DoM migration is not supported yet */
2064 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2065 ma->ma_lmm = info->mti_big_lmm;
2066 ma->ma_lmm_size = info->mti_big_lmmsize;
2068 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2070 GOTO(put_source, rc);
2072 if (ma->ma_valid & MA_LOV &&
2073 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2074 GOTO(put_source, rc = -EOPNOTSUPP);
2077 /* if migration HSM is allowed */
2078 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2079 ma->ma_need = MA_HSM;
2081 rc = mdt_attr_get_complex(info, sobj, ma);
2083 GOTO(unlock_links, rc);
2085 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2086 GOTO(unlock_links, rc = -EOPNOTSUPP);
2089 /* end lease and close file for regular file */
2090 if (info->mti_spec.sp_migrate_close) {
2091 /* try to hold open_sem so that nobody else can open the file */
2092 if (!down_write_trylock(&sobj->mot_open_sem)) {
2094 mdd_migrate_close(info, sobj);
2095 GOTO(unlock_links, rc = -EBUSY);
2097 open_sem_locked = true;
2098 rc = mdd_migrate_close(info, sobj);
2100 GOTO(unlock_open_sem, rc);
2105 lhs = &info->mti_lh[MDT_LH_OLD];
2106 mdt_lock_reg_init(lhs, LCK_EX);
2107 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2108 &child_slave_locks);
2110 GOTO(unlock_open_sem, rc);
2113 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2115 GOTO(unlock_source, rc = PTR_ERR(tobj));
2117 lht = &info->mti_lh[MDT_LH_NEW];
2118 mdt_lock_reg_init(lht, LCK_EX);
2119 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2121 GOTO(put_target, rc);
2123 /* Don't do lookup sanity check. We know name doesn't exist. */
2124 info->mti_spec.sp_cr_lookup = 0;
2125 info->mti_spec.sp_feat = &dt_directory_features;
2127 rc = mdo_migrate(env, mdt_object_child(pobj),
2128 mdt_object_child(sobj), &rr->rr_name,
2129 mdt_object_child(tobj), &info->mti_spec, ma);
2132 mdt_object_unlock(info, tobj, lht, rc);
2134 mdt_object_put(env, tobj);
2136 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2137 &child_slave_locks, rc);
2139 if (open_sem_locked)
2140 up_write(&sobj->mot_open_sem);
2142 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2144 mdt_object_put(env, sobj);
2145 mdt_object_put(env, spobj);
2147 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2148 &parent_slave_locks, rc);
2150 mdt_object_put(env, pobj);
2152 if (lustre_handle_is_used(&rename_lh))
2153 mdt_rename_unlock(&rename_lh);
2156 mdt_device_sync(env, mdt);
2161 static int mdt_object_lock_save(struct mdt_thread_info *info,
2162 struct mdt_object *dir,
2163 struct mdt_lock_handle *lh,
2164 int idx, bool cos_incompat)
2168 /* we lock the target dir if it is local */
2169 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2174 /* get and save correct version after locking */
2175 mdt_version_get_save(info, dir, idx);
2180 * determine lock order of sobj and tobj
2182 * there are two situations we need to lock tobj before sobj:
2183 * 1. sobj is child of tobj
2184 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2185 * larger than that of tobj
2187 * \retval 1 lock tobj before sobj
2188 * \retval 0 lock sobj before tobj
2189 * \retval -ev negative errno upon error
2191 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2192 struct mdt_object *sobj,
2193 struct mdt_object *tobj)
2195 struct md_attr *ma = &info->mti_attr;
2196 struct lu_fid *spfid = &info->mti_tmp_fid1;
2197 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2198 struct lmv_mds_md_v1 *lmv;
2203 /* sobj and tobj are the same */
2207 if (fid_is_root(mdt_object_fid(sobj)))
2210 if (fid_is_root(mdt_object_fid(tobj)))
2213 /* check whether sobj is child of tobj */
2214 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2215 mdt_object_fid(tobj));
2222 /* check whether sobj and tobj are children of the same parent */
2223 rc = mdt_attr_get_pfid(info, sobj, spfid);
2227 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2231 if (!lu_fid_eq(spfid, tpfid))
2234 /* check whether sobj and tobj are sibling stripes */
2235 ma->ma_need = MA_LMV;
2237 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2238 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2239 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2243 if (!(ma->ma_valid & MA_LMV))
2246 lmv = &ma->ma_lmv->lmv_md_v1;
2247 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2249 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2252 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2256 if (!(ma->ma_valid & MA_LMV))
2259 lmv = &ma->ma_lmv->lmv_md_v1;
2260 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2262 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2264 /* check stripe index of sobj and tobj */
2265 if (sindex == tindex)
2268 return sindex < tindex ? 0 : 1;
2272 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2273 * 2 - srcdir child; 3 - tgtdir child.
2274 * Update on disk version of srcdir child.
2276 static int mdt_reint_rename(struct mdt_thread_info *info,
2277 struct mdt_lock_handle *unused)
2279 struct mdt_device *mdt = info->mti_mdt;
2280 struct mdt_reint_record *rr = &info->mti_rr;
2281 struct md_attr *ma = &info->mti_attr;
2282 struct ptlrpc_request *req = mdt_info_req(info);
2283 struct mdt_object *msrcdir = NULL;
2284 struct mdt_object *mtgtdir = NULL;
2285 struct mdt_object *mold;
2286 struct mdt_object *mnew = NULL;
2287 struct lustre_handle rename_lh = { 0 };
2288 struct mdt_lock_handle *lh_srcdirp;
2289 struct mdt_lock_handle *lh_tgtdirp;
2290 struct mdt_lock_handle *lh_oldp = NULL;
2291 struct mdt_lock_handle *lh_newp = NULL;
2292 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2293 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2295 bool reverse = false;
2296 bool cos_incompat, discard = false;
2300 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2301 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2302 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2304 if (info->mti_dlm_req)
2305 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2307 if (!fid_is_md_operative(rr->rr_fid1) ||
2308 !fid_is_md_operative(rr->rr_fid2))
2311 /* find both parents. */
2312 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2313 if (IS_ERR(msrcdir))
2314 RETURN(PTR_ERR(msrcdir));
2316 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2318 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2320 mdt_object_get(info->mti_env, mtgtdir);
2322 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2323 if (IS_ERR(mtgtdir))
2324 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2328 * Note: do not enqueue rename lock for replay request, because
2329 * if other MDT holds rename lock, but being blocked to wait for
2330 * this MDT to finish its recovery, and the failover MDT can not
2331 * get rename lock, which will cause deadlock.
2333 if (!req_is_replay(req)) {
2335 * Normally rename RPC is handled on the MDT with the target
2336 * directory (if target exists, it's on the MDT with the
2337 * target), if the source directory is remote, it's a hint that
2338 * source is remote too (this may not be true, but it won't
2339 * cause any issue), return -EXDEV early to avoid taking
2342 if (!mdt->mdt_enable_remote_rename &&
2343 mdt_object_remote(msrcdir))
2344 GOTO(out_put_tgtdir, rc = -EXDEV);
2346 rc = mdt_rename_lock(info, &rename_lh);
2348 CERROR("%s: can't lock FS for rename: rc = %d\n",
2349 mdt_obd_name(mdt), rc);
2350 GOTO(out_put_tgtdir, rc);
2354 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2356 GOTO(out_unlock_rename, rc);
2360 /* source needs to be looked up after locking source parent, otherwise
2361 * this rename may race with unlink source, and cause rename hang, see
2362 * sanityn.sh 55b, so check parents first, if later we found source is
2363 * remote, relock parents. */
2364 cos_incompat = (mdt_object_remote(msrcdir) ||
2365 mdt_object_remote(mtgtdir));
2367 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2369 /* lock parents in the proper order. */
2370 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2371 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2374 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2375 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2378 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2381 GOTO(out_unlock_rename, rc);
2383 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2385 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2388 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2389 GOTO(out_unlock_rename, rc);
2392 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2395 GOTO(out_unlock_rename, rc);
2397 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2399 if (mtgtdir != msrcdir) {
2400 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2402 } else if (!mdt_object_remote(mtgtdir) &&
2403 lh_srcdirp->mlh_pdo_hash !=
2404 lh_tgtdirp->mlh_pdo_hash) {
2405 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2406 MDS_INODELOCK_UPDATE,
2408 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2411 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2412 GOTO(out_unlock_rename, rc);
2416 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2417 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2419 /* find mold object. */
2421 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2423 GOTO(out_unlock_parents, rc);
2425 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2426 GOTO(out_unlock_parents, rc = -EINVAL);
2428 if (!fid_is_md_operative(old_fid))
2429 GOTO(out_unlock_parents, rc = -EPERM);
2431 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2433 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2435 if (!mdt_object_exists(mold)) {
2436 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2438 "object does not exist");
2439 GOTO(out_put_old, rc = -ENOENT);
2442 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2443 GOTO(out_put_old, rc = -EXDEV);
2445 /* Check if @mtgtdir is subdir of @mold, before locking child
2446 * to avoid reverse locking. */
2447 if (mtgtdir != msrcdir) {
2448 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2453 GOTO(out_put_old, rc);
2457 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2458 /* save version after locking */
2459 mdt_version_get_save(info, mold, 2);
2461 if (!cos_incompat && mdt_object_remote(mold)) {
2462 cos_incompat = true;
2463 mdt_object_put(info->mti_env, mold);
2464 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2465 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2469 /* find mnew object:
2470 * mnew target object may not exist now
2471 * lookup with version checking */
2473 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2476 /* the new_fid should have been filled at this moment */
2477 if (lu_fid_eq(old_fid, new_fid))
2478 GOTO(out_put_old, rc);
2480 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2481 lu_fid_eq(new_fid, rr->rr_fid2))
2482 GOTO(out_put_old, rc = -EINVAL);
2484 if (!fid_is_md_operative(new_fid))
2485 GOTO(out_put_old, rc = -EPERM);
2487 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2489 GOTO(out_put_old, rc = PTR_ERR(mnew));
2491 if (!mdt_object_exists(mnew)) {
2492 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2494 "object does not exist");
2495 GOTO(out_put_new, rc = -ENOENT);
2498 if (mdt_object_remote(mnew)) {
2499 struct mdt_body *repbody;
2501 /* Always send rename req to the target child MDT */
2502 repbody = req_capsule_server_get(info->mti_pill,
2504 LASSERT(repbody != NULL);
2505 repbody->mbo_fid1 = *new_fid;
2506 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2507 GOTO(out_put_new, rc = -EXDEV);
2509 /* Before locking the target dir, check we do not replace
2510 * a dir with a non-dir, otherwise it may deadlock with
2511 * link op which tries to create a link in this dir
2512 * back to this non-dir. */
2513 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2514 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2515 GOTO(out_put_new, rc = -EISDIR);
2517 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2518 mdt_lock_reg_init(lh_oldp, LCK_EX);
2519 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2520 if (mdt_object_remote(msrcdir)) {
2521 /* Enqueue lookup lock from the parent MDT */
2522 rc = mdt_remote_object_lock(info, msrcdir,
2523 mdt_object_fid(mold),
2524 &lh_oldp->mlh_rreg_lh,
2525 lh_oldp->mlh_rreg_mode,
2526 MDS_INODELOCK_LOOKUP,
2529 GOTO(out_put_new, rc);
2531 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2534 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2537 GOTO(out_unlock_old, rc);
2539 /* Check if @msrcdir is subdir of @mnew, before locking child
2540 * to avoid reverse locking. */
2541 if (mtgtdir != msrcdir) {
2542 rc = mdo_is_subdir(info->mti_env,
2543 mdt_object_child(msrcdir), new_fid);
2547 GOTO(out_unlock_old, rc);
2551 /* We used to acquire MDS_INODELOCK_FULL here but we
2552 * can't do this now because a running HSM restore on
2553 * the rename onto victim will hold the layout
2554 * lock. See LU-4002. */
2556 lh_newp = &info->mti_lh[MDT_LH_NEW];
2557 mdt_lock_reg_init(lh_newp, LCK_EX);
2558 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2559 MDS_INODELOCK_LOOKUP |
2560 MDS_INODELOCK_UPDATE,
2563 GOTO(out_unlock_old, rc);
2565 /* get and save version after locking */
2566 mdt_version_get_save(info, mnew, 3);
2567 } else if (rc != -EREMOTE && rc != -ENOENT) {
2568 GOTO(out_put_old, rc);
2570 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2571 mdt_lock_reg_init(lh_oldp, LCK_EX);
2572 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2573 if (mdt_object_remote(msrcdir)) {
2574 /* Enqueue lookup lock from the parent MDT */
2575 rc = mdt_remote_object_lock(info, msrcdir,
2576 mdt_object_fid(mold),
2577 &lh_oldp->mlh_rreg_lh,
2578 lh_oldp->mlh_rreg_mode,
2579 MDS_INODELOCK_LOOKUP,
2582 GOTO(out_put_old, rc);
2584 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2587 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2590 GOTO(out_unlock_old, rc);
2592 mdt_enoent_version_save(info, 3);
2595 /* step 5: rename it */
2596 mdt_reint_init_ma(info, ma);
2598 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2599 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2602 mutex_lock(&mnew->mot_lov_mutex);
2604 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2605 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2606 mnew != NULL ? mdt_object_child(mnew) : NULL,
2607 &rr->rr_tgt_name, ma);
2610 mutex_unlock(&mnew->mot_lov_mutex);
2612 /* handle last link of tgt object */
2614 mdt_counter_incr(req, LPROC_MDT_RENAME);
2616 mdt_handle_last_unlink(info, mnew, ma);
2617 discard = mdt_dom_check_for_discard(info, mnew);
2620 mdt_rename_counter_tally(info, info->mti_mdt, req,
2626 mdt_object_unlock(info, mnew, lh_newp, rc);
2628 mdt_object_unlock(info, mold, lh_oldp, rc);
2631 mdt_object_put(info->mti_env, mnew);
2633 mdt_object_put(info->mti_env, mold);
2635 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2636 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2638 if (lustre_handle_is_used(&rename_lh))
2639 mdt_rename_unlock(&rename_lh);
2641 mdt_object_put(info->mti_env, mtgtdir);
2643 mdt_object_put(info->mti_env, msrcdir);
2645 /* If 'discard' is set then new_fid must exits.
2646 * DOM data discard need neither object nor lock,
2647 * so do this at the end.
2650 mdt_dom_discard_data(info, new_fid);
2655 static int mdt_reint_resync(struct mdt_thread_info *info,
2656 struct mdt_lock_handle *lhc)
2658 struct mdt_reint_record *rr = &info->mti_rr;
2659 struct ptlrpc_request *req = mdt_info_req(info);
2660 struct md_attr *ma = &info->mti_attr;
2661 struct mdt_object *mo;
2662 struct ldlm_lock *lease;
2663 struct mdt_body *repbody;
2664 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2669 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2671 if (info->mti_dlm_req)
2672 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2674 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2676 GOTO(out, rc = PTR_ERR(mo));
2678 if (!mdt_object_exists(mo))
2679 GOTO(out_obj, rc = -ENOENT);
2681 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2682 GOTO(out_obj, rc = -EINVAL);
2684 if (mdt_object_remote(mo))
2685 GOTO(out_obj, rc = -EREMOTE);
2687 lease = ldlm_handle2lock(rr->rr_lease_handle);
2689 GOTO(out_obj, rc = -ESTALE);
2691 /* It's really necessary to grab open_sem and check if the lease lock
2692 * has been lost. There would exist a concurrent writer coming in and
2693 * generating some dirty data in memory cache, the writeback would fail
2694 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2695 if (!down_write_trylock(&mo->mot_open_sem))
2696 GOTO(out_put_lease, rc = -EBUSY);
2698 lock_res_and_lock(lease);
2699 lease_broken = ldlm_is_cancel(lease);
2700 unlock_res_and_lock(lease);
2702 GOTO(out_unlock, rc = -EBUSY);
2704 /* the file has yet opened by anyone else after we took the lease. */
2705 layout.mlc_opc = MD_LAYOUT_RESYNC;
2706 rc = mdt_layout_change(info, mo, &layout);
2708 GOTO(out_unlock, rc);
2710 ma->ma_need = MA_INODE;
2712 rc = mdt_attr_get_complex(info, mo, ma);
2714 GOTO(out_unlock, rc);
2716 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2717 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2721 up_write(&mo->mot_open_sem);
2723 LDLM_LOCK_PUT(lease);
2725 mdt_object_put(info->mti_env, mo);
2727 mdt_client_compatibility(info);
2728 rc2 = mdt_fix_reply(info);
2734 struct mdt_reinter {
2735 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2736 enum lprocfs_extra_opc mr_extra_opc;
2739 static const struct mdt_reinter mdt_reinters[] = {
2741 .mr_handler = &mdt_reint_setattr,
2742 .mr_extra_opc = MDS_REINT_SETATTR,
2745 .mr_handler = &mdt_reint_create,
2746 .mr_extra_opc = MDS_REINT_CREATE,
2749 .mr_handler = &mdt_reint_link,
2750 .mr_extra_opc = MDS_REINT_LINK,
2753 .mr_handler = &mdt_reint_unlink,
2754 .mr_extra_opc = MDS_REINT_UNLINK,
2757 .mr_handler = &mdt_reint_rename,
2758 .mr_extra_opc = MDS_REINT_RENAME,
2761 .mr_handler = &mdt_reint_open,
2762 .mr_extra_opc = MDS_REINT_OPEN,
2764 [REINT_SETXATTR] = {
2765 .mr_handler = &mdt_reint_setxattr,
2766 .mr_extra_opc = MDS_REINT_SETXATTR,
2769 .mr_handler = &mdt_reint_unlink,
2770 .mr_extra_opc = MDS_REINT_UNLINK,
2773 .mr_handler = &mdt_reint_migrate,
2774 .mr_extra_opc = MDS_REINT_RENAME,
2777 .mr_handler = &mdt_reint_resync,
2778 .mr_extra_opc = MDS_REINT_RESYNC,
2782 int mdt_reint_rec(struct mdt_thread_info *info,
2783 struct mdt_lock_handle *lhc)
2785 const struct mdt_reinter *mr;
2789 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2792 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2793 if (mr->mr_handler == NULL)
2796 rc = (*mr->mr_handler)(info, lhc);
2798 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2799 PTLRPC_LAST_CNTR + mr->mr_extra_opc);