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 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
656 /* revoke lease lock if size is going to be changed */
657 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
658 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
659 atomic_read(&mo->mot_lease_count) > 0)) {
660 down_read(&mo->mot_open_sem);
662 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
663 lhc = &info->mti_lh[MDT_LH_LOCAL];
664 mdt_lock_reg_init(lhc, LCK_CW);
666 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
668 up_read(&mo->mot_open_sem);
672 /* revoke lease lock */
673 mdt_object_unlock(info, mo, lhc, 1);
675 up_read(&mo->mot_open_sem);
678 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
679 /* Check write access for the O_TRUNC case */
680 if (mdt_write_read(mo) < 0)
681 GOTO(out_put, rc = -ETXTBSY);
683 /* LU-10286: compatibility check for FLR.
684 * Please check the comment in mdt_finish_open() for details */
685 if (!exp_connect_flr(info->mti_exp)) {
686 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
687 if (rc < 0 && rc != -ENODATA)
690 if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
691 GOTO(out_put, rc = -EOPNOTSUPP);
694 /* For truncate, the file size sent from client
695 * is believable, but the blocks are incorrect,
696 * which makes the block size in LSOM attribute
697 * inconsisent with the real block size.
699 rc = mdt_lsom_update(info, mo, true);
704 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
705 if (ma->ma_valid & MA_LOV)
706 GOTO(out_put, rc = -EPROTO);
708 /* MDT supports FMD for regular files due to Data-on-MDT */
709 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
710 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
711 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
714 rc = mdt_attr_set(info, mo, ma);
717 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
718 (ma->ma_valid & MA_INODE)) {
719 struct lu_buf *buf = &info->mti_buf;
720 struct lu_ucred *uc = mdt_ucred(info);
721 struct mdt_lock_handle *lh;
723 /* reject if either remote or striped dir is disabled */
724 if (ma->ma_valid & MA_LMV) {
725 if (!mdt->mdt_enable_remote_dir ||
726 !mdt->mdt_enable_striped_dir)
727 GOTO(out_put, rc = -EPERM);
729 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
730 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
731 mdt->mdt_enable_remote_dir_gid != -1)
732 GOTO(out_put, rc = -EPERM);
735 if (ma->ma_attr.la_valid != 0)
736 GOTO(out_put, rc = -EPROTO);
738 lh = &info->mti_lh[MDT_LH_PARENT];
739 mdt_lock_reg_init(lh, LCK_PW);
741 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
745 if (ma->ma_valid & MA_LOV) {
746 buf->lb_buf = ma->ma_lmm;
747 buf->lb_len = ma->ma_lmm_size;
749 buf->lb_buf = ma->ma_lmv;
750 buf->lb_len = ma->ma_lmv_size;
752 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
753 (ma->ma_valid & MA_LOV) ?
754 XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
757 mdt_object_unlock(info, mo, lh, rc);
761 GOTO(out_put, rc = -EPROTO);
764 /* If file data is modified, add the dirty flag */
765 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
766 rc = mdt_add_dirty_flag(info, mo, ma);
768 ma->ma_need = MA_INODE;
770 rc = mdt_attr_get_complex(info, mo, ma);
774 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
778 mdt_object_put(info->mti_env, mo);
781 mdt_counter_incr(req, LPROC_MDT_SETATTR);
783 mdt_client_compatibility(info);
784 rc2 = mdt_fix_reply(info);
790 static int mdt_reint_create(struct mdt_thread_info *info,
791 struct mdt_lock_handle *lhc)
793 struct ptlrpc_request *req = mdt_info_req(info);
797 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
798 RETURN(err_serious(-ESTALE));
800 if (info->mti_dlm_req)
801 ldlm_request_cancel(mdt_info_req(info),
802 info->mti_dlm_req, 0, LATF_SKIP);
804 if (!lu_name_is_valid(&info->mti_rr.rr_name))
807 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
809 mdt_counter_incr(req, LPROC_MDT_MKDIR);
817 /* Special file should stay on the same node as parent. */
818 mdt_counter_incr(req, LPROC_MDT_MKNOD);
821 CERROR("%s: Unsupported mode %o\n",
822 mdt_obd_name(info->mti_mdt),
823 info->mti_attr.ma_attr.la_mode);
824 RETURN(err_serious(-EOPNOTSUPP));
827 rc = mdt_create(info);
832 * VBR: save parent version in reply and child version getting by its name.
833 * Version of child is getting and checking during its lookup. If
835 static int mdt_reint_unlink(struct mdt_thread_info *info,
836 struct mdt_lock_handle *lhc)
838 struct mdt_reint_record *rr = &info->mti_rr;
839 struct ptlrpc_request *req = mdt_info_req(info);
840 struct md_attr *ma = &info->mti_attr;
841 struct lu_fid *child_fid = &info->mti_tmp_fid1;
842 struct mdt_object *mp;
843 struct mdt_object *mc;
844 struct mdt_lock_handle *parent_lh;
845 struct mdt_lock_handle *child_lh;
846 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
848 bool cos_incompat = false;
854 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
855 PNAME(&rr->rr_name));
857 if (info->mti_dlm_req)
858 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
860 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
861 RETURN(err_serious(-ENOENT));
863 if (!fid_is_md_operative(rr->rr_fid1))
866 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
870 if (mdt_object_remote(mp)) {
873 rc = mdt_version_get_check_save(info, mp, 0);
875 GOTO(put_parent, rc);
878 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
879 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
881 parent_lh = &info->mti_lh[MDT_LH_PARENT];
882 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
883 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
886 GOTO(put_parent, rc);
888 /* lookup child object along with version checking */
890 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
892 /* Name might not be able to find during resend of
893 * remote unlink, considering following case.
894 * dir_A is a remote directory, the name entry of
895 * dir_A is on MDT0, the directory is on MDT1,
897 * 1. client sends unlink req to MDT1.
898 * 2. MDT1 sends name delete update to MDT0.
899 * 3. name entry is being deleted in MDT0 synchronously.
900 * 4. MDT1 is restarted.
901 * 5. client resends unlink req to MDT1. So it can not
902 * find the name entry on MDT0 anymore.
903 * In this case, MDT1 only needs to destory the local
906 if (mdt_object_remote(mp) && rc == -ENOENT &&
907 !fid_is_zero(rr->rr_fid2) &&
908 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
910 *child_fid = *rr->rr_fid2;
912 GOTO(unlock_parent, rc);
916 if (!fid_is_md_operative(child_fid))
917 GOTO(unlock_parent, rc = -EPERM);
919 /* We will lock the child regardless it is local or remote. No harm. */
920 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
922 GOTO(unlock_parent, rc = PTR_ERR(mc));
925 rc = mdt_object_striped(info, mc);
931 mdt_object_put(info->mti_env, mc);
932 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
937 child_lh = &info->mti_lh[MDT_LH_CHILD];
938 mdt_lock_reg_init(child_lh, LCK_EX);
939 if (info->mti_spec.sp_rm_entry) {
940 struct lu_ucred *uc = mdt_ucred(info);
942 if (!mdt_is_dne_client(req->rq_export))
943 /* Return -ENOTSUPP for old client */
944 GOTO(put_child, rc = -ENOTSUPP);
946 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
947 GOTO(put_child, rc = -EPERM);
949 ma->ma_need = MA_INODE;
951 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
952 NULL, &rr->rr_name, ma, no_name);
956 if (mdt_object_remote(mc)) {
957 struct mdt_body *repbody;
959 if (!fid_is_zero(rr->rr_fid2)) {
960 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
961 mdt_obd_name(info->mti_mdt),
962 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
963 GOTO(put_child, rc = -ENOENT);
965 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
966 mdt_obd_name(info->mti_mdt),
967 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
969 if (!mdt_is_dne_client(req->rq_export))
970 /* Return -ENOTSUPP for old client */
971 GOTO(put_child, rc = -ENOTSUPP);
973 /* Revoke the LOOKUP lock of the remote object granted by
974 * this MDT. Since the unlink will happen on another MDT,
975 * it will release the LOOKUP lock right away. Then What
976 * would happen if another client try to grab the LOOKUP
977 * lock at the same time with unlink XXX */
978 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
979 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
980 LASSERT(repbody != NULL);
981 repbody->mbo_fid1 = *mdt_object_fid(mc);
982 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
983 GOTO(unlock_child, rc = -EREMOTE);
985 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
986 * this now because a running HSM restore on the child (unlink
987 * victim) will hold the layout lock. See LU-4002. */
988 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
989 if (mdt_object_remote(mp)) {
990 /* Enqueue lookup lock from parent MDT */
991 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
992 &child_lh->mlh_rreg_lh,
993 child_lh->mlh_rreg_mode,
994 MDS_INODELOCK_LOOKUP, false);
998 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1001 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1004 GOTO(put_child, rc);
1007 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1008 * whether need MA_LOV and MA_COOKIE.
1010 ma->ma_need = MA_INODE;
1013 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1014 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1015 /* save version when object is locked */
1016 mdt_version_get_save(info, mc, 1);
1018 mutex_lock(&mc->mot_lov_mutex);
1020 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1021 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1023 mutex_unlock(&mc->mot_lov_mutex);
1025 GOTO(unlock_child, rc);
1027 if (!lu_object_is_dying(&mc->mot_header)) {
1028 rc = mdt_attr_get_complex(info, mc, ma);
1031 } else if (mdt_dom_check_for_discard(info, mc)) {
1032 mdt_dom_discard_data(info, mc);
1034 mdt_handle_last_unlink(info, mc, ma);
1037 if (ma->ma_valid & MA_INODE) {
1038 switch (ma->ma_attr.la_mode & S_IFMT) {
1040 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1048 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1051 LASSERTF(0, "bad file type %o unlinking\n",
1052 ma->ma_attr.la_mode);
1059 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1061 mdt_object_put(info->mti_env, mc);
1063 mdt_object_unlock(info, mp, parent_lh, rc);
1065 mdt_object_put(info->mti_env, mp);
1066 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1071 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1074 static int mdt_reint_link(struct mdt_thread_info *info,
1075 struct mdt_lock_handle *lhc)
1077 struct mdt_reint_record *rr = &info->mti_rr;
1078 struct ptlrpc_request *req = mdt_info_req(info);
1079 struct md_attr *ma = &info->mti_attr;
1080 struct mdt_object *ms;
1081 struct mdt_object *mp;
1082 struct mdt_lock_handle *lhs;
1083 struct mdt_lock_handle *lhp;
1088 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1089 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1091 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1092 RETURN(err_serious(-ENOENT));
1094 if (info->mti_dlm_req)
1095 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1097 /* Invalid case so return error immediately instead of
1099 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1102 if (!fid_is_md_operative(rr->rr_fid1) ||
1103 !fid_is_md_operative(rr->rr_fid2))
1106 /* step 1: find target parent dir */
1107 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1109 RETURN(PTR_ERR(mp));
1111 rc = mdt_version_get_check_save(info, mp, 0);
1113 GOTO(put_parent, rc);
1115 /* step 2: find source */
1116 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1118 GOTO(put_parent, rc = PTR_ERR(ms));
1120 if (!mdt_object_exists(ms)) {
1121 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1122 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1123 GOTO(put_source, rc = -ENOENT);
1126 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1128 lhp = &info->mti_lh[MDT_LH_PARENT];
1129 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1130 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1133 GOTO(put_source, rc);
1135 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1137 lhs = &info->mti_lh[MDT_LH_CHILD];
1138 mdt_lock_reg_init(lhs, LCK_EX);
1139 rc = mdt_reint_object_lock(info, ms, lhs,
1140 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1143 GOTO(unlock_parent, rc);
1145 /* step 3: link it */
1146 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1147 OBD_FAIL_MDS_REINT_LINK_WRITE);
1149 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1150 rc = mdt_version_get_check_save(info, ms, 1);
1152 GOTO(unlock_source, rc);
1154 /** check target version by name during replay */
1155 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1156 &info->mti_tmp_fid1, 2);
1157 if (rc != 0 && rc != -ENOENT)
1158 GOTO(unlock_source, rc);
1159 /* save version of file name for replay, it must be ENOENT here */
1160 if (!req_is_replay(mdt_info_req(info))) {
1161 if (rc != -ENOENT) {
1162 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1163 PNAME(&rr->rr_name));
1164 GOTO(unlock_source, rc = -EEXIST);
1166 info->mti_ver[2] = ENOENT_VERSION;
1167 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1170 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1171 mdt_object_child(ms), &rr->rr_name, ma);
1174 mdt_counter_incr(req, LPROC_MDT_LINK);
1178 mdt_object_unlock(info, ms, lhs, rc);
1180 mdt_object_unlock(info, mp, lhp, rc);
1182 mdt_object_put(info->mti_env, ms);
1184 mdt_object_put(info->mti_env, mp);
1188 * lock the part of the directory according to the hash of the name
1189 * (lh->mlh_pdo_hash) in parallel directory lock.
1191 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1192 struct mdt_lock_handle *lh,
1193 struct mdt_object *obj, __u64 ibits,
1196 struct ldlm_res_id *res = &info->mti_res_id;
1197 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1198 union ldlm_policy_data *policy = &info->mti_policy;
1199 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1203 * Finish res_id initializing by name hash marking part of
1204 * directory which is taking modification.
1206 LASSERT(lh->mlh_pdo_hash != 0);
1207 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1208 memset(policy, 0, sizeof(*policy));
1209 policy->l_inodebits.bits = ibits;
1211 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1212 dlmflags |= LDLM_FL_COS_INCOMPAT;
1214 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1215 * going to be sent to client. If it is - mdt_intent_policy() path will
1216 * fix it up and turn FL_LOCAL flag off.
1218 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1219 policy, res, dlmflags,
1220 &info->mti_exp->exp_handle.h_cookie);
1225 * Get BFL lock for rename or migrate process.
1227 static int mdt_rename_lock(struct mdt_thread_info *info,
1228 struct lustre_handle *lh)
1233 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1234 struct lu_fid *fid = &info->mti_tmp_fid1;
1235 struct mdt_object *obj;
1237 /* XXX, right now, it has to use object API to
1238 * enqueue lock cross MDT, so it will enqueue
1239 * rename lock(with LUSTRE_BFL_FID) by root object */
1241 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1243 RETURN(PTR_ERR(obj));
1245 rc = mdt_remote_object_lock(info, obj,
1246 &LUSTRE_BFL_FID, lh,
1248 MDS_INODELOCK_UPDATE, false);
1249 mdt_object_put(info->mti_env, obj);
1251 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1252 union ldlm_policy_data *policy = &info->mti_policy;
1253 struct ldlm_res_id *res_id = &info->mti_res_id;
1256 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1257 memset(policy, 0, sizeof *policy);
1258 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1259 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1260 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1261 LDLM_IBITS, policy, LCK_EX, &flags,
1263 ldlm_completion_ast, NULL, NULL, 0,
1265 &info->mti_exp->exp_handle.h_cookie,
1272 static void mdt_rename_unlock(struct lustre_handle *lh)
1275 LASSERT(lustre_handle_is_used(lh));
1276 /* Cancel the single rename lock right away */
1277 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1281 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1282 const struct lu_fid *fid,
1285 struct mdt_object *dir;
1290 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1294 /* check early, the real version will be saved after locking */
1295 rc = mdt_version_get_check(info, dir, idx);
1299 if (!mdt_object_exists(dir))
1300 GOTO(out_put, rc = -ENOENT);
1302 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1303 GOTO(out_put, rc = -ENOTDIR);
1307 mdt_object_put(info->mti_env, dir);
1312 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1313 * herein we don't really check it, just do revoke.
1315 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1316 struct mdt_object *pobj,
1317 struct mdt_object *obj)
1319 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1322 mdt_lock_handle_init(lh);
1323 mdt_lock_reg_init(lh, LCK_EX);
1325 if (mdt_object_remote(pobj)) {
1326 /* don't bother to check if pobj and obj are on the same MDT. */
1327 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1328 &lh->mlh_rreg_lh, LCK_EX,
1329 MDS_INODELOCK_LOOKUP, false);
1330 } else if (mdt_object_remote(obj)) {
1331 struct ldlm_res_id *res = &info->mti_res_id;
1332 union ldlm_policy_data *policy = &info->mti_policy;
1333 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1334 LDLM_FL_COS_INCOMPAT;
1336 fid_build_reg_res_name(mdt_object_fid(obj), res);
1337 memset(policy, 0, sizeof(*policy));
1338 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1339 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1340 &lh->mlh_reg_lh, LCK_EX, policy, res,
1343 /* do nothing if both are local */
1351 * TODO, currently we don't save this lock because there is no place to
1352 * hold this lock handle, but to avoid race we need to save this lock.
1354 mdt_object_unlock(info, NULL, lh, 1);
1360 * operation may takes locks of linkea, or directory stripes, group them in
1363 struct mdt_sub_lock {
1364 struct mdt_object *msl_obj;
1365 struct mdt_lock_handle msl_lh;
1366 struct list_head msl_linkage;
1369 static void mdt_unlock_list(struct mdt_thread_info *info,
1370 struct list_head *list, int decref)
1372 struct mdt_sub_lock *msl;
1373 struct mdt_sub_lock *tmp;
1375 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1376 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1377 list_del(&msl->msl_linkage);
1383 * lock parents of links, and also check whether total locks don't exceed
1386 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1387 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1388 * \retval -ev negative errno upon error
1390 static int mdt_lock_links(struct mdt_thread_info *info,
1391 struct mdt_object *pobj,
1392 const struct md_attr *ma,
1393 struct mdt_object *obj,
1394 struct list_head *link_locks)
1396 struct mdt_device *mdt = info->mti_mdt;
1397 struct lu_buf *buf = &info->mti_big_buf;
1398 struct lu_name *lname = &info->mti_name;
1399 struct linkea_data ldata = { NULL };
1400 bool blocked = false;
1402 int local_lnkp_cnt = 0;
1407 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1410 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1411 if (buf->lb_buf == NULL)
1415 rc = mdt_links_read(info, obj, &ldata);
1417 if (rc == -ENOENT || rc == -ENODATA)
1423 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1424 linkea_next_entry(&ldata)) {
1425 struct mdt_object *lnkp;
1426 struct mdt_sub_lock *msl;
1430 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1433 /* check if it's also linked to parent */
1434 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1435 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1436 PFID(&fid), PNAME(lname));
1437 /* in case link is remote object, revoke LOOKUP lock */
1438 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1444 /* check if it's linked to a stripe of parent */
1445 if (ma->ma_valid & MA_LMV) {
1446 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1447 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1450 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1451 fid_le_to_cpu(stripe_fid,
1452 &lmv->lmv_stripe_fids[j]);
1453 if (lu_fid_eq(stripe_fid, &fid)) {
1454 CDEBUG(D_INFO, "skip stripe "DFID
1455 ", reovke "DNAME"\n",
1456 PFID(&fid), PNAME(lname));
1457 lnkp = mdt_object_find(info->mti_env,
1460 GOTO(out, rc = PTR_ERR(lnkp));
1466 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1468 mdt_object_put(info->mti_env, lnkp);
1473 /* Check if it's already locked */
1474 list_for_each_entry(msl, link_locks, msl_linkage) {
1475 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1477 DFID" was locked, revoke "DNAME"\n",
1478 PFID(&fid), PNAME(lname));
1479 lnkp = msl->msl_obj;
1485 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1489 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1490 PFID(&fid), PNAME(lname));
1492 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1494 CWARN("%s: cannot find obj "DFID": %ld\n",
1495 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1499 if (!mdt_object_exists(lnkp)) {
1500 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1501 PFID(&fid), PNAME(lname));
1502 mdt_object_put(info->mti_env, lnkp);
1506 if (!mdt_object_remote(lnkp))
1511 GOTO(out, rc = -ENOMEM);
1514 * we can't follow parent-child lock order like other MD
1515 * operations, use lock_try here to avoid deadlock, if the lock
1516 * cannot be taken, drop all locks taken, revoke the blocked
1517 * one, and continue processing the remaining entries, and in
1518 * the end of the loop restart from beginning.
1520 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1522 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1523 MDS_INODELOCK_UPDATE, true);
1524 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1527 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1528 PFID(&fid), PNAME(lname), retries);
1530 mdt_unlock_list(info, link_locks, 1);
1532 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1533 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1534 MDS_INODELOCK_UPDATE);
1536 mdt_object_put(info->mti_env, lnkp);
1541 if (mdt_object_remote(lnkp)) {
1542 struct ldlm_lock *lock;
1545 * for remote object, set lock cb_atomic,
1546 * so lock can be released in blocking_ast()
1547 * immediately, then the next lock_try will
1548 * have better chance of success.
1550 lock = ldlm_handle2lock(
1551 &msl->msl_lh.mlh_rreg_lh);
1552 LASSERT(lock != NULL);
1553 lock_res_and_lock(lock);
1554 ldlm_set_atomic_cb(lock);
1555 unlock_res_and_lock(lock);
1556 LDLM_LOCK_PUT(lock);
1559 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1564 INIT_LIST_HEAD(&msl->msl_linkage);
1565 msl->msl_obj = lnkp;
1566 list_add_tail(&msl->msl_linkage, link_locks);
1568 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1573 if (--retries > 0) {
1574 mdt_unlock_list(info, link_locks, rc);
1584 mdt_unlock_list(info, link_locks, rc);
1585 } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1586 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1589 * parent may have 3 local objects: master object and 2 stripes
1590 * (if it's being migrated too); source may have 1 local object
1591 * if regular file: master and 1 stripe; target has 1 local
1593 * Note, if source is directory it may also have 2 local objects
1594 * but can't has hardlinks, so consider only regular files here.
1601 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1602 struct mdt_object *obj,
1603 const struct md_attr *ma,
1604 struct list_head *slave_locks)
1606 struct mdt_device *mdt = info->mti_mdt;
1607 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1608 struct lu_fid *fid = &info->mti_tmp_fid1;
1609 struct mdt_object *slave;
1610 struct mdt_sub_lock *msl;
1616 LASSERT(mdt_object_remote(obj));
1617 LASSERT(ma->ma_valid & MA_LMV);
1620 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1623 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1626 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1627 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1629 if (!fid_is_sane(fid))
1632 slave = mdt_object_find(info->mti_env, mdt, fid);
1634 GOTO(out, rc = PTR_ERR(slave));
1638 mdt_object_put(info->mti_env, slave);
1639 GOTO(out, rc = -ENOMEM);
1642 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1643 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1644 MDS_INODELOCK_UPDATE, true);
1647 mdt_object_put(info->mti_env, slave);
1651 INIT_LIST_HEAD(&msl->msl_linkage);
1652 msl->msl_obj = slave;
1653 list_add_tail(&msl->msl_linkage, slave_locks);
1659 mdt_unlock_list(info, slave_locks, rc);
1663 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1664 struct mdt_object *obj,
1665 struct mdt_lock_handle *lh,
1666 struct ldlm_enqueue_info *einfo,
1667 struct list_head *slave_locks,
1670 if (mdt_object_remote(obj)) {
1671 mdt_unlock_list(info, slave_locks, decref);
1672 mdt_object_unlock(info, obj, lh, decref);
1674 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1678 /* lock parent and its stripes */
1679 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1680 struct mdt_object *obj,
1681 const struct md_attr *ma,
1682 struct mdt_lock_handle *lh,
1683 struct ldlm_enqueue_info *einfo,
1684 struct list_head *slave_locks)
1688 if (mdt_object_remote(obj)) {
1689 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1690 &lh->mlh_rreg_lh, LCK_PW,
1691 MDS_INODELOCK_UPDATE, false);
1696 * if obj is remote and striped, lock its stripes explicitly
1697 * because it's not striped in LOD layer on this MDT.
1699 if (ma->ma_valid & MA_LMV) {
1700 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1702 mdt_object_unlock(info, obj, lh, rc);
1705 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1713 * in migration, object may be remote, and we need take full lock of it and its
1714 * stripes if it's directory, besides, object may be a remote object on its
1715 * parent, revoke its LOOKUP lock on where its parent is located.
1717 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1718 struct mdt_object *pobj,
1719 struct mdt_object *obj,
1720 struct mdt_lock_handle *lh,
1721 struct ldlm_enqueue_info *einfo,
1722 struct list_head *slave_locks)
1726 if (mdt_object_remote(obj)) {
1727 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1731 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1732 &lh->mlh_rreg_lh, LCK_EX,
1733 MDS_INODELOCK_FULL, false);
1738 * if obj is remote and striped, lock its stripes explicitly
1739 * because it's not striped in LOD layer on this MDT.
1741 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1742 struct md_attr *ma = &info->mti_attr;
1744 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1746 mdt_object_unlock(info, obj, lh, rc);
1750 if (ma->ma_valid & MA_LMV) {
1751 rc = mdt_lock_remote_slaves(info, obj, ma,
1754 mdt_object_unlock(info, obj, lh, rc);
1758 if (mdt_object_remote(pobj)) {
1759 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1764 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1772 * lookup source by name, if parent is striped directory, we need to find the
1773 * corresponding stripe where source is located, and then lookup there.
1775 * besides, if parent is migrating too, and file is already in target stripe,
1776 * this should be a redo of 'lfs migrate' on client side.
1778 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1779 struct mdt_object *pobj,
1780 const struct md_attr *ma,
1781 const struct lu_name *lname,
1782 struct mdt_object **spobj,
1783 struct mdt_object **sobj)
1785 const struct lu_env *env = info->mti_env;
1786 struct lu_fid *fid = &info->mti_tmp_fid1;
1787 struct mdt_object *stripe;
1790 if (ma->ma_valid & MA_LMV) {
1791 /* if parent is striped, lookup on corresponding stripe */
1792 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1793 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1794 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1795 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1796 LMV_HASH_FLAG_MIGRATION;
1799 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1800 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1803 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1809 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1810 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1812 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1814 stripe = mdt_object_find(env, info->mti_mdt, fid);
1816 return PTR_ERR(stripe);
1819 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1821 if (rc == -ENOENT && is_migrating) {
1823 * if parent is migrating, and lookup child failed on
1824 * source stripe, lookup again on target stripe, if it
1825 * exists, it means previous migration was interrupted,
1826 * and current file was migrated already.
1828 mdt_object_put(env, stripe);
1830 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1831 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1833 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1839 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1841 stripe = mdt_object_find(env, info->mti_mdt, fid);
1843 return PTR_ERR(stripe);
1846 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1847 fid, &info->mti_spec);
1848 mdt_object_put(env, stripe);
1849 return rc ?: -EALREADY;
1851 mdt_object_put(env, stripe);
1856 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1862 mdt_object_get(env, stripe);
1867 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1868 if (IS_ERR(*sobj)) {
1869 mdt_object_put(env, stripe);
1870 rc = PTR_ERR(*sobj);
1878 /* end lease and close file for regular file */
1879 static int mdd_migrate_close(struct mdt_thread_info *info,
1880 struct mdt_object *obj)
1882 struct close_data *data;
1883 struct mdt_body *repbody;
1884 struct ldlm_lock *lease;
1889 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1891 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1895 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1900 lease = ldlm_handle2lock(&data->cd_handle);
1904 /* check if the lease was already canceled */
1905 lock_res_and_lock(lease);
1906 rc = ldlm_is_cancel(lease);
1907 unlock_res_and_lock(lease);
1911 LDLM_DEBUG(lease, DFID" lease broken",
1912 PFID(mdt_object_fid(obj)));
1916 * cancel server side lease, client side counterpart should have been
1917 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1919 ldlm_lock_cancel(lease);
1920 ldlm_reprocess_all(lease->l_resource, lease);
1921 LDLM_LOCK_PUT(lease);
1924 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1925 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1926 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1932 * migrate file in below steps:
1933 * 1. lock parent and its stripes
1934 * 2. lookup source by name
1935 * 3. lock parents of source links if source is not directory
1936 * 4. reject if source is in HSM
1937 * 5. take source open_sem and close file if source is regular file
1938 * 6. lock source and its stripes if it's directory
1939 * 7. lock target so subsequent change to it can trigger COS
1941 * 9. unlock above locks
1942 * 10. sync device if source has links
1944 static int mdt_reint_migrate(struct mdt_thread_info *info,
1945 struct mdt_lock_handle *unused)
1947 const struct lu_env *env = info->mti_env;
1948 struct mdt_device *mdt = info->mti_mdt;
1949 struct ptlrpc_request *req = mdt_info_req(info);
1950 struct mdt_reint_record *rr = &info->mti_rr;
1951 struct lu_ucred *uc = mdt_ucred(info);
1952 struct md_attr *ma = &info->mti_attr;
1953 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1954 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1955 struct mdt_object *pobj;
1956 struct mdt_object *spobj = NULL;
1957 struct mdt_object *sobj = NULL;
1958 struct mdt_object *tobj;
1959 struct lustre_handle rename_lh = { 0 };
1960 struct mdt_lock_handle *lhp;
1961 struct mdt_lock_handle *lhs;
1962 struct mdt_lock_handle *lht;
1963 LIST_HEAD(parent_slave_locks);
1964 LIST_HEAD(child_slave_locks);
1965 LIST_HEAD(link_locks);
1966 bool open_sem_locked = false;
1967 bool do_sync = false;
1971 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1972 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1974 if (info->mti_dlm_req)
1975 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1977 if (!fid_is_md_operative(rr->rr_fid1) ||
1978 !fid_is_md_operative(rr->rr_fid2))
1981 /* don't allow migrate . or .. */
1982 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1985 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1988 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1989 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1990 mdt->mdt_enable_remote_dir_gid != -1)
1994 * Note: do not enqueue rename lock for replay request, because
1995 * if other MDT holds rename lock, but being blocked to wait for
1996 * this MDT to finish its recovery, and the failover MDT can not
1997 * get rename lock, which will cause deadlock.
1999 if (!req_is_replay(req)) {
2000 rc = mdt_rename_lock(info, &rename_lh);
2002 CERROR("%s: can't lock FS for rename: rc = %d\n",
2003 mdt_obd_name(info->mti_mdt), rc);
2008 /* pobj is master object of parent */
2009 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2011 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2013 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2015 GOTO(put_parent, rc);
2017 /* lock parent object */
2018 lhp = &info->mti_lh[MDT_LH_PARENT];
2019 mdt_lock_reg_init(lhp, LCK_PW);
2020 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2021 &parent_slave_locks);
2023 GOTO(put_parent, rc);
2026 * spobj is the corresponding stripe against name if pobj is striped
2027 * directory, which is the real parent, and no need to lock, because
2028 * we've taken full lock of pobj.
2030 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2032 GOTO(unlock_parent, rc);
2034 /* lock parents of source links, and revoke LOOKUP lock of links */
2035 rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2037 GOTO(put_source, rc);
2040 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2041 * with one request, if total lock count exceeds this limit, we will
2042 * drop all locks after migration, and synchronous device in the end.
2046 /* TODO: DoM migration is not supported yet */
2047 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2048 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2050 GOTO(put_source, rc);
2052 if (ma->ma_valid & MA_LOV &&
2053 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2054 GOTO(put_source, rc = -EOPNOTSUPP);
2057 /* if migration HSM is allowed */
2058 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2059 ma->ma_need = MA_HSM;
2061 rc = mdt_attr_get_complex(info, sobj, ma);
2063 GOTO(unlock_links, rc);
2065 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2066 GOTO(unlock_links, rc = -EOPNOTSUPP);
2069 /* end lease and close file for regular file */
2070 if (info->mti_spec.sp_migrate_close) {
2071 /* try to hold open_sem so that nobody else can open the file */
2072 if (!down_write_trylock(&sobj->mot_open_sem)) {
2074 mdd_migrate_close(info, sobj);
2075 GOTO(unlock_links, rc = -EBUSY);
2077 open_sem_locked = true;
2078 rc = mdd_migrate_close(info, sobj);
2080 GOTO(unlock_open_sem, rc);
2085 lhs = &info->mti_lh[MDT_LH_OLD];
2086 mdt_lock_reg_init(lhs, LCK_EX);
2087 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2088 &child_slave_locks);
2090 GOTO(unlock_open_sem, rc);
2093 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2095 GOTO(unlock_source, rc = PTR_ERR(tobj));
2097 lht = &info->mti_lh[MDT_LH_NEW];
2098 mdt_lock_reg_init(lht, LCK_EX);
2099 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2101 GOTO(put_target, rc);
2103 /* Don't do lookup sanity check. We know name doesn't exist. */
2104 info->mti_spec.sp_cr_lookup = 0;
2105 info->mti_spec.sp_feat = &dt_directory_features;
2107 rc = mdo_migrate(env, mdt_object_child(pobj),
2108 mdt_object_child(sobj), &rr->rr_name,
2109 mdt_object_child(tobj), &info->mti_spec, ma);
2112 mdt_object_unlock(info, tobj, lht, rc || do_sync);
2114 mdt_object_put(env, tobj);
2116 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2117 &child_slave_locks, rc || do_sync);
2119 if (open_sem_locked)
2120 up_write(&sobj->mot_open_sem);
2122 mdt_unlock_list(info, &link_locks, rc || do_sync);
2124 mdt_object_put(env, sobj);
2125 mdt_object_put(env, spobj);
2127 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2128 &parent_slave_locks, rc || do_sync);
2130 mdt_object_put(env, pobj);
2132 if (lustre_handle_is_used(&rename_lh))
2133 mdt_rename_unlock(&rename_lh);
2136 mdt_device_sync(env, mdt);
2141 static int mdt_object_lock_save(struct mdt_thread_info *info,
2142 struct mdt_object *dir,
2143 struct mdt_lock_handle *lh,
2144 int idx, bool cos_incompat)
2148 /* we lock the target dir if it is local */
2149 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2154 /* get and save correct version after locking */
2155 mdt_version_get_save(info, dir, idx);
2160 * determine lock order of sobj and tobj
2162 * there are two situations we need to lock tobj before sobj:
2163 * 1. sobj is child of tobj
2164 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2165 * larger than that of tobj
2167 * \retval 1 lock tobj before sobj
2168 * \retval 0 lock sobj before tobj
2169 * \retval -ev negative errno upon error
2171 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2172 struct mdt_object *sobj,
2173 struct mdt_object *tobj)
2175 struct md_attr *ma = &info->mti_attr;
2176 struct lu_fid *spfid = &info->mti_tmp_fid1;
2177 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2178 struct lmv_mds_md_v1 *lmv;
2183 /* sobj and tobj are the same */
2187 if (fid_is_root(mdt_object_fid(sobj)))
2190 if (fid_is_root(mdt_object_fid(tobj)))
2193 /* check whether sobj is child of tobj */
2194 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2195 mdt_object_fid(tobj));
2202 /* check whether sobj and tobj are children of the same parent */
2203 rc = mdt_attr_get_pfid(info, sobj, spfid);
2207 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2211 if (!lu_fid_eq(spfid, tpfid))
2214 /* check whether sobj and tobj are sibling stripes */
2215 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2219 if (!(ma->ma_valid & MA_LMV))
2222 lmv = &ma->ma_lmv->lmv_md_v1;
2223 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2225 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2228 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2232 if (!(ma->ma_valid & MA_LMV))
2235 lmv = &ma->ma_lmv->lmv_md_v1;
2236 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2238 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2240 /* check stripe index of sobj and tobj */
2241 if (sindex == tindex)
2244 return sindex < tindex ? 0 : 1;
2248 * lock rename source object.
2250 * Both source and source parent may be remote, and source may be a remote
2251 * object on source parent, to avoid overriding lock handle, store remote
2252 * LOOKUP lock separately in @lhr.
2254 * \retval 0 on success
2255 * \retval -ev negative errno upon error
2257 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2258 struct mdt_object *parent,
2259 struct mdt_object *child,
2260 struct mdt_lock_handle *lhc,
2261 struct mdt_lock_handle *lhr,
2267 rc = mdt_is_remote_object(info, parent, child);
2272 /* enqueue remote LOOKUP lock from the parent MDT */
2273 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2275 if (mdt_object_remote(parent)) {
2276 rc = mdt_remote_object_lock(info, parent,
2277 mdt_object_fid(child),
2284 LASSERT(mdt_object_remote(child));
2285 rc = mdt_object_local_lock(info, child, lhr,
2286 &rmt_ibits, 0, true);
2291 ibits &= ~MDS_INODELOCK_LOOKUP;
2294 if (mdt_object_remote(child)) {
2295 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2302 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2307 mdt_object_unlock(info, child, lhr, rc);
2313 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2314 * 2 - srcdir child; 3 - tgtdir child.
2315 * Update on disk version of srcdir child.
2317 static int mdt_reint_rename(struct mdt_thread_info *info,
2318 struct mdt_lock_handle *unused)
2320 struct mdt_device *mdt = info->mti_mdt;
2321 struct mdt_reint_record *rr = &info->mti_rr;
2322 struct md_attr *ma = &info->mti_attr;
2323 struct ptlrpc_request *req = mdt_info_req(info);
2324 struct mdt_object *msrcdir = NULL;
2325 struct mdt_object *mtgtdir = NULL;
2326 struct mdt_object *mold;
2327 struct mdt_object *mnew = NULL;
2328 struct lustre_handle rename_lh = { 0 };
2329 struct mdt_lock_handle *lh_srcdirp;
2330 struct mdt_lock_handle *lh_tgtdirp;
2331 struct mdt_lock_handle *lh_oldp = NULL;
2332 struct mdt_lock_handle *lh_rmt = NULL;
2333 struct mdt_lock_handle *lh_newp = NULL;
2334 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2335 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2337 bool reverse = false, discard = false;
2342 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2343 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2344 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2346 if (info->mti_dlm_req)
2347 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2349 if (!fid_is_md_operative(rr->rr_fid1) ||
2350 !fid_is_md_operative(rr->rr_fid2))
2353 /* find both parents. */
2354 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2355 if (IS_ERR(msrcdir))
2356 RETURN(PTR_ERR(msrcdir));
2358 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2360 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2362 mdt_object_get(info->mti_env, mtgtdir);
2364 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2365 if (IS_ERR(mtgtdir))
2366 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2370 * Note: do not enqueue rename lock for replay request, because
2371 * if other MDT holds rename lock, but being blocked to wait for
2372 * this MDT to finish its recovery, and the failover MDT can not
2373 * get rename lock, which will cause deadlock.
2375 if (!req_is_replay(req)) {
2377 * Normally rename RPC is handled on the MDT with the target
2378 * directory (if target exists, it's on the MDT with the
2379 * target), if the source directory is remote, it's a hint that
2380 * source is remote too (this may not be true, but it won't
2381 * cause any issue), return -EXDEV early to avoid taking
2384 if (!mdt->mdt_enable_remote_rename &&
2385 mdt_object_remote(msrcdir))
2386 GOTO(out_put_tgtdir, rc = -EXDEV);
2388 rc = mdt_rename_lock(info, &rename_lh);
2390 CERROR("%s: can't lock FS for rename: rc = %d\n",
2391 mdt_obd_name(mdt), rc);
2392 GOTO(out_put_tgtdir, rc);
2396 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2398 GOTO(out_unlock_rename, rc);
2402 /* source needs to be looked up after locking source parent, otherwise
2403 * this rename may race with unlink source, and cause rename hang, see
2404 * sanityn.sh 55b, so check parents first, if later we found source is
2405 * remote, relock parents. */
2406 cos_incompat = (mdt_object_remote(msrcdir) ||
2407 mdt_object_remote(mtgtdir));
2409 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2411 /* lock parents in the proper order. */
2412 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2413 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2415 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2416 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2418 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2419 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2422 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2425 GOTO(out_unlock_rename, rc);
2427 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2429 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2432 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2433 GOTO(out_unlock_rename, rc);
2436 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2439 GOTO(out_unlock_rename, rc);
2441 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2443 if (mtgtdir != msrcdir) {
2444 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2446 } else if (!mdt_object_remote(mtgtdir) &&
2447 lh_srcdirp->mlh_pdo_hash !=
2448 lh_tgtdirp->mlh_pdo_hash) {
2449 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2450 MDS_INODELOCK_UPDATE,
2452 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2455 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2456 GOTO(out_unlock_rename, rc);
2460 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2461 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2463 /* find mold object. */
2465 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2467 GOTO(out_unlock_parents, rc);
2469 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2470 GOTO(out_unlock_parents, rc = -EINVAL);
2472 if (!fid_is_md_operative(old_fid))
2473 GOTO(out_unlock_parents, rc = -EPERM);
2475 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2477 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2479 if (!mdt_object_exists(mold)) {
2480 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2482 "object does not exist");
2483 GOTO(out_put_old, rc = -ENOENT);
2486 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2487 GOTO(out_put_old, rc = -EXDEV);
2489 /* Check if @mtgtdir is subdir of @mold, before locking child
2490 * to avoid reverse locking. */
2491 if (mtgtdir != msrcdir) {
2492 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2497 GOTO(out_put_old, rc);
2501 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2502 /* save version after locking */
2503 mdt_version_get_save(info, mold, 2);
2505 if (!cos_incompat && mdt_object_remote(mold)) {
2506 cos_incompat = true;
2507 mdt_object_put(info->mti_env, mold);
2508 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2509 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2513 /* find mnew object:
2514 * mnew target object may not exist now
2515 * lookup with version checking */
2517 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2520 /* the new_fid should have been filled at this moment */
2521 if (lu_fid_eq(old_fid, new_fid))
2522 GOTO(out_put_old, rc);
2524 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2525 lu_fid_eq(new_fid, rr->rr_fid2))
2526 GOTO(out_put_old, rc = -EINVAL);
2528 if (!fid_is_md_operative(new_fid))
2529 GOTO(out_put_old, rc = -EPERM);
2531 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2533 GOTO(out_put_old, rc = PTR_ERR(mnew));
2535 if (!mdt_object_exists(mnew)) {
2536 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2538 "object does not exist");
2539 GOTO(out_put_new, rc = -ENOENT);
2542 if (mdt_object_remote(mnew)) {
2543 struct mdt_body *repbody;
2545 /* Always send rename req to the target child MDT */
2546 repbody = req_capsule_server_get(info->mti_pill,
2548 LASSERT(repbody != NULL);
2549 repbody->mbo_fid1 = *new_fid;
2550 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2551 GOTO(out_put_new, rc = -EXDEV);
2553 /* Before locking the target dir, check we do not replace
2554 * a dir with a non-dir, otherwise it may deadlock with
2555 * link op which tries to create a link in this dir
2556 * back to this non-dir. */
2557 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2558 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2559 GOTO(out_put_new, rc = -EISDIR);
2561 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2562 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2563 mdt_lock_reg_init(lh_oldp, LCK_EX);
2564 mdt_lock_reg_init(lh_rmt, LCK_EX);
2565 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2566 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2567 lh_rmt, lock_ibits, cos_incompat);
2569 GOTO(out_put_new, rc);
2571 /* Check if @msrcdir is subdir of @mnew, before locking child
2572 * to avoid reverse locking. */
2573 if (mtgtdir != msrcdir) {
2574 rc = mdo_is_subdir(info->mti_env,
2575 mdt_object_child(msrcdir), new_fid);
2579 GOTO(out_unlock_old, rc);
2583 /* We used to acquire MDS_INODELOCK_FULL here but we
2584 * can't do this now because a running HSM restore on
2585 * the rename onto victim will hold the layout
2586 * lock. See LU-4002. */
2588 lh_newp = &info->mti_lh[MDT_LH_NEW];
2589 mdt_lock_reg_init(lh_newp, LCK_EX);
2590 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2591 if (mdt_object_remote(mtgtdir)) {
2592 rc = mdt_remote_object_lock(info, mtgtdir,
2593 mdt_object_fid(mnew),
2594 &lh_newp->mlh_rreg_lh,
2595 lh_newp->mlh_rreg_mode,
2596 MDS_INODELOCK_LOOKUP,
2599 GOTO(out_unlock_old, rc);
2601 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2603 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2606 GOTO(out_unlock_new, rc);
2608 /* get and save version after locking */
2609 mdt_version_get_save(info, mnew, 3);
2610 } else if (rc != -ENOENT) {
2611 GOTO(out_put_old, rc);
2613 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2614 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2615 mdt_lock_reg_init(lh_oldp, LCK_EX);
2616 mdt_lock_reg_init(lh_rmt, LCK_EX);
2617 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2618 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2619 lh_rmt, lock_ibits, cos_incompat);
2621 GOTO(out_put_old, rc);
2623 mdt_enoent_version_save(info, 3);
2626 /* step 5: rename it */
2627 mdt_reint_init_ma(info, ma);
2629 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2630 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2633 mutex_lock(&mnew->mot_lov_mutex);
2635 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2636 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2637 mnew != NULL ? mdt_object_child(mnew) : NULL,
2638 &rr->rr_tgt_name, ma);
2641 mutex_unlock(&mnew->mot_lov_mutex);
2643 /* handle last link of tgt object */
2645 mdt_counter_incr(req, LPROC_MDT_RENAME);
2647 mdt_handle_last_unlink(info, mnew, ma);
2648 discard = mdt_dom_check_for_discard(info, mnew);
2650 mdt_rename_counter_tally(info, info->mti_mdt, req,
2657 mdt_object_unlock(info, mnew, lh_newp, rc);
2659 mdt_object_unlock(info, NULL, lh_rmt, rc);
2660 mdt_object_unlock(info, mold, lh_oldp, rc);
2662 if (mnew && !discard)
2663 mdt_object_put(info->mti_env, mnew);
2665 mdt_object_put(info->mti_env, mold);
2667 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2668 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2670 if (lustre_handle_is_used(&rename_lh))
2671 mdt_rename_unlock(&rename_lh);
2673 mdt_object_put(info->mti_env, mtgtdir);
2675 mdt_object_put(info->mti_env, msrcdir);
2677 /* The DoM discard can be done right in the place above where it is
2678 * assigned, meanwhile it is done here after rename unlock due to
2679 * compatibility with old clients, for them the discard blocks
2680 * the main thread until completion. Check LU-11359 for details.
2683 mdt_dom_discard_data(info, mnew);
2684 mdt_object_put(info->mti_env, mnew);
2689 static int mdt_reint_resync(struct mdt_thread_info *info,
2690 struct mdt_lock_handle *lhc)
2692 struct mdt_reint_record *rr = &info->mti_rr;
2693 struct ptlrpc_request *req = mdt_info_req(info);
2694 struct md_attr *ma = &info->mti_attr;
2695 struct mdt_object *mo;
2696 struct ldlm_lock *lease;
2697 struct mdt_body *repbody;
2698 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2704 DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2706 if (info->mti_dlm_req)
2707 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2709 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2711 GOTO(out, rc = PTR_ERR(mo));
2713 if (!mdt_object_exists(mo))
2714 GOTO(out_obj, rc = -ENOENT);
2716 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2717 GOTO(out_obj, rc = -EINVAL);
2719 if (mdt_object_remote(mo))
2720 GOTO(out_obj, rc = -EREMOTE);
2722 lease = ldlm_handle2lock(rr->rr_lease_handle);
2724 GOTO(out_obj, rc = -ESTALE);
2726 /* It's really necessary to grab open_sem and check if the lease lock
2727 * has been lost. There would exist a concurrent writer coming in and
2728 * generating some dirty data in memory cache, the writeback would fail
2729 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2730 if (!down_write_trylock(&mo->mot_open_sem))
2731 GOTO(out_put_lease, rc = -EBUSY);
2733 lock_res_and_lock(lease);
2734 lease_broken = ldlm_is_cancel(lease);
2735 unlock_res_and_lock(lease);
2737 GOTO(out_unlock, rc = -EBUSY);
2739 /* the file has yet opened by anyone else after we took the lease. */
2740 layout.mlc_opc = MD_LAYOUT_RESYNC;
2741 lhc = &info->mti_lh[MDT_LH_LOCAL];
2742 rc = mdt_layout_change(info, mo, lhc, &layout);
2744 GOTO(out_unlock, rc);
2746 mdt_object_unlock(info, mo, lhc, 0);
2748 ma->ma_need = MA_INODE;
2750 rc = mdt_attr_get_complex(info, mo, ma);
2752 GOTO(out_unlock, rc);
2754 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2755 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2759 up_write(&mo->mot_open_sem);
2761 LDLM_LOCK_PUT(lease);
2763 mdt_object_put(info->mti_env, mo);
2765 mdt_client_compatibility(info);
2766 rc2 = mdt_fix_reply(info);
2772 struct mdt_reinter {
2773 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2774 enum lprocfs_extra_opc mr_extra_opc;
2777 static const struct mdt_reinter mdt_reinters[] = {
2779 .mr_handler = &mdt_reint_setattr,
2780 .mr_extra_opc = MDS_REINT_SETATTR,
2783 .mr_handler = &mdt_reint_create,
2784 .mr_extra_opc = MDS_REINT_CREATE,
2787 .mr_handler = &mdt_reint_link,
2788 .mr_extra_opc = MDS_REINT_LINK,
2791 .mr_handler = &mdt_reint_unlink,
2792 .mr_extra_opc = MDS_REINT_UNLINK,
2795 .mr_handler = &mdt_reint_rename,
2796 .mr_extra_opc = MDS_REINT_RENAME,
2799 .mr_handler = &mdt_reint_open,
2800 .mr_extra_opc = MDS_REINT_OPEN,
2802 [REINT_SETXATTR] = {
2803 .mr_handler = &mdt_reint_setxattr,
2804 .mr_extra_opc = MDS_REINT_SETXATTR,
2807 .mr_handler = &mdt_reint_unlink,
2808 .mr_extra_opc = MDS_REINT_UNLINK,
2811 .mr_handler = &mdt_reint_migrate,
2812 .mr_extra_opc = MDS_REINT_RENAME,
2815 .mr_handler = &mdt_reint_resync,
2816 .mr_extra_opc = MDS_REINT_RESYNC,
2820 int mdt_reint_rec(struct mdt_thread_info *info,
2821 struct mdt_lock_handle *lhc)
2823 const struct mdt_reinter *mr;
2827 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2830 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2831 if (mr->mr_handler == NULL)
2834 rc = (*mr->mr_handler)(info, lhc);
2836 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2837 PTLRPC_LAST_CNTR + mr->mr_extra_opc);