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)
253 struct lu_device *bottom_dev;
254 struct lu_object *bottom_obj;
257 if (!S_ISDIR(obj->mot_header.loh_attr))
260 /* getxattr from bottom obj to avoid reading in shard FIDs */
261 bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
262 bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
263 mdt_object_fid(obj), NULL);
264 if (IS_ERR(bottom_obj))
265 return PTR_ERR(bottom_obj);
267 rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
269 lu_object_put(mti->mti_env, bottom_obj);
271 return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
275 * Lock slave stripes if necessary, the lock handles of slave stripes
276 * will be stored in einfo->ei_cbdata.
278 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
279 enum ldlm_mode mode, __u64 ibits,
280 struct ldlm_enqueue_info *einfo)
282 union ldlm_policy_data *policy = &mti->mti_policy;
284 LASSERT(S_ISDIR(obj->mot_header.loh_attr));
286 einfo->ei_type = LDLM_IBITS;
287 einfo->ei_mode = mode;
288 einfo->ei_cb_bl = mdt_remote_blocking_ast;
289 einfo->ei_cb_local_bl = mdt_blocking_ast;
290 einfo->ei_cb_cp = ldlm_completion_ast;
291 einfo->ei_enq_slave = 1;
292 einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
293 einfo->ei_inodebits = ibits;
294 memset(policy, 0, sizeof(*policy));
295 policy->l_inodebits.bits = ibits;
297 return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
301 int mdt_reint_striped_lock(struct mdt_thread_info *info,
302 struct mdt_object *o,
303 struct mdt_lock_handle *lh,
305 struct ldlm_enqueue_info *einfo,
310 LASSERT(!mdt_object_remote(o));
312 memset(einfo, 0, sizeof(*einfo));
314 rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
318 rc = mdt_object_striped(info, o);
321 mdt_object_unlock(info, o, lh, rc);
325 rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
327 mdt_object_unlock(info, o, lh, rc);
328 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
335 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
336 struct mdt_object *o,
337 struct mdt_lock_handle *lh,
338 struct ldlm_enqueue_info *einfo, int decref)
340 if (einfo->ei_cbdata)
341 mdt_unlock_slaves(info, o, einfo, decref);
342 mdt_object_unlock(info, o, lh, decref);
346 * VBR: we save three versions in reply:
347 * 0 - parent. Check that parent version is the same during replay.
348 * 1 - name. Version of 'name' if file exists with the same name or
349 * ENOENT_VERSION, it is needed because file may appear due to missed replays.
350 * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
353 static int mdt_create(struct mdt_thread_info *info)
355 struct mdt_device *mdt = info->mti_mdt;
356 struct mdt_object *parent;
357 struct mdt_object *child;
358 struct mdt_lock_handle *lh;
359 struct mdt_body *repbody;
360 struct md_attr *ma = &info->mti_attr;
361 struct mdt_reint_record *rr = &info->mti_rr;
362 struct md_op_spec *spec = &info->mti_spec;
366 DEBUG_REQ(D_INODE, mdt_info_req(info),
367 "Create ("DNAME"->"DFID") in "DFID,
368 PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
370 if (!fid_is_md_operative(rr->rr_fid1))
373 if (S_ISDIR(ma->ma_attr.la_mode) &&
374 spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
375 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
376 struct lu_ucred *uc = mdt_ucred(info);
377 struct obd_export *exp = mdt_info_req(info)->rq_export;
379 /* Only new clients can create remote dir( >= 2.4) and
380 * striped dir(>= 2.6), old client will return -ENOTSUPP */
381 if (!mdt_is_dne_client(exp))
384 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
385 if (!mdt_is_striped_client(exp))
388 if (!mdt->mdt_enable_striped_dir)
390 } else if (!mdt->mdt_enable_remote_dir) {
394 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
395 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
396 mdt->mdt_enable_remote_dir_gid != -1)
400 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
402 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
404 RETURN(PTR_ERR(parent));
406 if (!mdt_object_exists(parent))
407 GOTO(put_parent, rc = -ENOENT);
410 * LU-10235: check if name exists locklessly first to avoid massive
411 * lock recalls on existing directories.
413 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
414 &info->mti_tmp_fid1, 1);
416 GOTO(put_parent, rc = -EEXIST);
418 /* -ENOENT is expected here */
420 GOTO(put_parent, rc);
422 /* save version of file name for replay, it must be ENOENT here */
423 mdt_enoent_version_save(info, 1);
425 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
427 lh = &info->mti_lh[MDT_LH_PARENT];
428 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
429 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
431 GOTO(put_parent, rc);
433 if (!mdt_object_remote(parent)) {
434 rc = mdt_version_get_check_save(info, parent, 0);
436 GOTO(unlock_parent, rc);
439 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
440 if (unlikely(IS_ERR(child)))
441 GOTO(unlock_parent, rc = PTR_ERR(child));
443 ma->ma_need = MA_INODE;
446 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
447 OBD_FAIL_MDS_REINT_CREATE_WRITE);
449 /* Version of child will be updated on disk. */
450 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
451 rc = mdt_version_get_check_save(info, child, 2);
455 /* Let lower layer know current lock mode. */
456 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
459 * Do not perform lookup sanity check. We know that name does
462 info->mti_spec.sp_cr_lookup = 0;
463 info->mti_spec.sp_feat = &dt_directory_features;
465 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
466 mdt_object_child(child), &info->mti_spec, ma);
468 rc = mdt_attr_get_complex(info, child, ma);
474 * On DNE, we need to eliminate dependey between 'mkdir a' and
475 * 'mkdir a/b' if b is a striped directory, to achieve this, two
476 * things are done below:
477 * 1. save child and slaves lock.
478 * 2. if the child is a striped directory, relock parent so to
479 * compare against with COS locks to ensure parent was
482 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
483 struct mdt_lock_handle *lhc;
484 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
487 rc = mdt_object_striped(info, child);
493 if (!mdt_object_remote(parent)) {
494 mdt_object_unlock(info, parent, lh, 1);
495 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
496 rc = mdt_reint_object_lock(info, parent, lh,
497 MDS_INODELOCK_UPDATE,
504 lhc = &info->mti_lh[MDT_LH_CHILD];
505 mdt_lock_handle_init(lhc);
506 mdt_lock_reg_init(lhc, LCK_PW);
507 rc = mdt_reint_striped_lock(info, child, lhc,
508 MDS_INODELOCK_UPDATE, einfo,
513 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
516 /* Return fid & attr to client. */
517 if (ma->ma_valid & MA_INODE)
518 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
519 mdt_object_fid(child));
521 mdt_object_put(info->mti_env, child);
523 mdt_object_unlock(info, parent, lh, rc);
525 mdt_object_put(info->mti_env, parent);
529 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
532 struct mdt_lock_handle *lh;
533 int do_vbr = ma->ma_attr.la_valid &
534 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
535 __u64 lockpart = MDS_INODELOCK_UPDATE;
536 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
541 rc = mdt_object_striped(info, mo);
547 lh = &info->mti_lh[MDT_LH_PARENT];
548 mdt_lock_reg_init(lh, LCK_PW);
550 /* Even though the new MDT will grant PERM lock to the old
551 * client, but the old client will almost ignore that during
552 * So it needs to revoke both LOOKUP and PERM lock here, so
553 * both new and old client can cancel the dcache */
554 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
555 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
557 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
562 /* all attrs are packed into mti_attr in unpack_setattr */
563 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
564 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
566 /* VBR: update version if attr changed are important for recovery */
568 /* update on-disk version of changed object */
569 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
570 rc = mdt_version_get_check_save(info, mo, 0);
572 GOTO(out_unlock, rc);
575 /* Ensure constant striping during chown(). See LU-2789. */
576 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
577 mutex_lock(&mo->mot_lov_mutex);
579 /* all attrs are packed into mti_attr in unpack_setattr */
580 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
582 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
583 mutex_unlock(&mo->mot_lov_mutex);
586 GOTO(out_unlock, rc);
587 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
590 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
595 * Check HSM flags and add HS_DIRTY flag if relevant.
597 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
598 * and is not RELEASED.
600 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
603 struct lu_ucred *uc = mdt_ucred(info);
608 /* If the file was modified, add the dirty flag */
609 ma->ma_need = MA_HSM;
610 rc = mdt_attr_get_complex(info, mo, ma);
612 CERROR("file attribute read error for "DFID": %d.\n",
613 PFID(mdt_object_fid(mo)), rc);
617 /* If an up2date copy exists in the backend, add dirty flag */
618 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
619 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
620 ma->ma_hsm.mh_flags |= HS_DIRTY;
622 /* Bump cap so that closes from non-owner writers can
623 * set the HSM state to dirty. */
624 cap_saved = uc->uc_cap;
625 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
626 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
627 uc->uc_cap = cap_saved;
629 CERROR("file attribute change error for "DFID": %d\n",
630 PFID(mdt_object_fid(mo)), rc);
636 static int mdt_reint_setattr(struct mdt_thread_info *info,
637 struct mdt_lock_handle *lhc)
639 struct mdt_device *mdt = info->mti_mdt;
640 struct md_attr *ma = &info->mti_attr;
641 struct mdt_reint_record *rr = &info->mti_rr;
642 struct ptlrpc_request *req = mdt_info_req(info);
643 struct mdt_object *mo;
644 struct mdt_body *repbody;
648 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
649 (unsigned int)ma->ma_attr.la_valid);
651 if (info->mti_dlm_req)
652 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
654 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
656 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
657 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
659 GOTO(out, rc = PTR_ERR(mo));
661 if (!mdt_object_exists(mo))
662 GOTO(out_put, rc = -ENOENT);
664 if (mdt_object_remote(mo))
665 GOTO(out_put, rc = -EREMOTE);
667 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
668 /* revoke lease lock if size is going to be changed */
669 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
670 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
671 atomic_read(&mo->mot_lease_count) > 0)) {
672 down_read(&mo->mot_open_sem);
674 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
675 lhc = &info->mti_lh[MDT_LH_LOCAL];
676 mdt_lock_reg_init(lhc, LCK_CW);
678 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
680 up_read(&mo->mot_open_sem);
684 /* revoke lease lock */
685 mdt_object_unlock(info, mo, lhc, 1);
687 up_read(&mo->mot_open_sem);
690 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
691 /* Check write access for the O_TRUNC case */
692 if (mdt_write_read(mo) < 0)
693 GOTO(out_put, rc = -ETXTBSY);
695 /* LU-10286: compatibility check for FLR.
696 * Please check the comment in mdt_finish_open() for details */
697 if (!exp_connect_flr(info->mti_exp) ||
698 !exp_connect_overstriping(info->mti_exp)) {
699 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
700 if (rc < 0 && rc != -ENODATA)
703 if (!exp_connect_flr(info->mti_exp)) {
705 mdt_lmm_is_flr(info->mti_big_lmm))
706 GOTO(out_put, rc = -EOPNOTSUPP);
709 if (!exp_connect_overstriping(info->mti_exp)) {
711 mdt_lmm_is_overstriping(info->mti_big_lmm))
712 GOTO(out_put, rc = -EOPNOTSUPP);
716 /* For truncate, the file size sent from client
717 * is believable, but the blocks are incorrect,
718 * which makes the block size in LSOM attribute
719 * inconsisent with the real block size.
721 rc = mdt_lsom_update(info, mo, true);
726 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
727 if (ma->ma_valid & MA_LOV)
728 GOTO(out_put, rc = -EPROTO);
730 /* MDT supports FMD for regular files due to Data-on-MDT */
731 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
732 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
733 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
736 rc = mdt_attr_set(info, mo, ma);
739 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
740 (ma->ma_valid & MA_INODE)) {
741 struct lu_buf *buf = &info->mti_buf;
742 struct lu_ucred *uc = mdt_ucred(info);
743 struct mdt_lock_handle *lh;
745 __u64 lockpart = MDS_INODELOCK_XATTR;
747 /* reject if either remote or striped dir is disabled */
748 if (ma->ma_valid & MA_LMV) {
749 if (!mdt->mdt_enable_remote_dir ||
750 !mdt->mdt_enable_striped_dir)
751 GOTO(out_put, rc = -EPERM);
753 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
754 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
755 mdt->mdt_enable_remote_dir_gid != -1)
756 GOTO(out_put, rc = -EPERM);
759 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
760 GOTO(out_put, rc = -ENOTDIR);
762 if (ma->ma_attr.la_valid != 0)
763 GOTO(out_put, rc = -EPROTO);
765 if (ma->ma_valid & MA_LOV) {
766 buf->lb_buf = ma->ma_lmm;
767 buf->lb_len = ma->ma_lmm_size;
768 name = XATTR_NAME_LOV;
770 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
773 buf->lb_len = ma->ma_lmv_size;
774 name = XATTR_NAME_DEFAULT_LMV;
775 /* force client to update dir default layout */
776 lockpart |= MDS_INODELOCK_LOOKUP;
779 lh = &info->mti_lh[MDT_LH_PARENT];
780 mdt_lock_reg_init(lh, LCK_PW);
782 rc = mdt_object_lock(info, mo, lh, lockpart);
786 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
789 mdt_object_unlock(info, mo, lh, rc);
793 GOTO(out_put, rc = -EPROTO);
796 /* If file data is modified, add the dirty flag */
797 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
798 rc = mdt_add_dirty_flag(info, mo, ma);
800 ma->ma_need = MA_INODE;
802 rc = mdt_attr_get_complex(info, mo, ma);
806 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
810 mdt_object_put(info->mti_env, mo);
813 mdt_counter_incr(req, LPROC_MDT_SETATTR);
815 mdt_client_compatibility(info);
816 rc2 = mdt_fix_reply(info);
822 static int mdt_reint_create(struct mdt_thread_info *info,
823 struct mdt_lock_handle *lhc)
825 struct ptlrpc_request *req = mdt_info_req(info);
829 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
830 RETURN(err_serious(-ESTALE));
832 if (info->mti_dlm_req)
833 ldlm_request_cancel(mdt_info_req(info),
834 info->mti_dlm_req, 0, LATF_SKIP);
836 if (!lu_name_is_valid(&info->mti_rr.rr_name))
839 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
841 mdt_counter_incr(req, LPROC_MDT_MKDIR);
849 /* Special file should stay on the same node as parent. */
850 mdt_counter_incr(req, LPROC_MDT_MKNOD);
853 CERROR("%s: Unsupported mode %o\n",
854 mdt_obd_name(info->mti_mdt),
855 info->mti_attr.ma_attr.la_mode);
856 RETURN(err_serious(-EOPNOTSUPP));
859 rc = mdt_create(info);
864 * VBR: save parent version in reply and child version getting by its name.
865 * Version of child is getting and checking during its lookup. If
867 static int mdt_reint_unlink(struct mdt_thread_info *info,
868 struct mdt_lock_handle *lhc)
870 struct mdt_reint_record *rr = &info->mti_rr;
871 struct ptlrpc_request *req = mdt_info_req(info);
872 struct md_attr *ma = &info->mti_attr;
873 struct lu_fid *child_fid = &info->mti_tmp_fid1;
874 struct mdt_object *mp;
875 struct mdt_object *mc;
876 struct mdt_lock_handle *parent_lh;
877 struct mdt_lock_handle *child_lh;
878 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
880 bool cos_incompat = false;
886 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
887 PNAME(&rr->rr_name));
889 if (info->mti_dlm_req)
890 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
892 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
893 RETURN(err_serious(-ENOENT));
895 if (!fid_is_md_operative(rr->rr_fid1))
898 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
902 if (mdt_object_remote(mp)) {
905 rc = mdt_version_get_check_save(info, mp, 0);
907 GOTO(put_parent, rc);
911 parent_lh = &info->mti_lh[MDT_LH_PARENT];
912 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
913 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
916 GOTO(put_parent, rc);
918 /* lookup child object along with version checking */
920 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
922 /* Name might not be able to find during resend of
923 * remote unlink, considering following case.
924 * dir_A is a remote directory, the name entry of
925 * dir_A is on MDT0, the directory is on MDT1,
927 * 1. client sends unlink req to MDT1.
928 * 2. MDT1 sends name delete update to MDT0.
929 * 3. name entry is being deleted in MDT0 synchronously.
930 * 4. MDT1 is restarted.
931 * 5. client resends unlink req to MDT1. So it can not
932 * find the name entry on MDT0 anymore.
933 * In this case, MDT1 only needs to destory the local
936 if (mdt_object_remote(mp) && rc == -ENOENT &&
937 !fid_is_zero(rr->rr_fid2) &&
938 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
940 *child_fid = *rr->rr_fid2;
942 GOTO(unlock_parent, rc);
946 if (!fid_is_md_operative(child_fid))
947 GOTO(unlock_parent, rc = -EPERM);
949 /* We will lock the child regardless it is local or remote. No harm. */
950 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
952 GOTO(unlock_parent, rc = PTR_ERR(mc));
955 rc = mdt_object_striped(info, mc);
961 mdt_object_put(info->mti_env, mc);
962 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
967 child_lh = &info->mti_lh[MDT_LH_CHILD];
968 mdt_lock_reg_init(child_lh, LCK_EX);
969 if (info->mti_spec.sp_rm_entry) {
970 struct lu_ucred *uc = mdt_ucred(info);
972 if (!mdt_is_dne_client(req->rq_export))
973 /* Return -ENOTSUPP for old client */
974 GOTO(put_child, rc = -ENOTSUPP);
976 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
977 GOTO(put_child, rc = -EPERM);
979 ma->ma_need = MA_INODE;
981 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
982 NULL, &rr->rr_name, ma, no_name);
986 if (mdt_object_remote(mc)) {
987 struct mdt_body *repbody;
989 if (!fid_is_zero(rr->rr_fid2)) {
990 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
991 mdt_obd_name(info->mti_mdt),
992 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
993 GOTO(put_child, rc = -ENOENT);
995 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
996 mdt_obd_name(info->mti_mdt),
997 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
999 if (!mdt_is_dne_client(req->rq_export))
1000 /* Return -ENOTSUPP for old client */
1001 GOTO(put_child, rc = -ENOTSUPP);
1003 /* Revoke the LOOKUP lock of the remote object granted by
1004 * this MDT. Since the unlink will happen on another MDT,
1005 * it will release the LOOKUP lock right away. Then What
1006 * would happen if another client try to grab the LOOKUP
1007 * lock at the same time with unlink XXX */
1008 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1009 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1010 LASSERT(repbody != NULL);
1011 repbody->mbo_fid1 = *mdt_object_fid(mc);
1012 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1013 GOTO(unlock_child, rc = -EREMOTE);
1015 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1016 * this now because a running HSM restore on the child (unlink
1017 * victim) will hold the layout lock. See LU-4002. */
1018 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1019 if (mdt_object_remote(mp)) {
1020 /* Enqueue lookup lock from parent MDT */
1021 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1022 &child_lh->mlh_rreg_lh,
1023 child_lh->mlh_rreg_mode,
1024 MDS_INODELOCK_LOOKUP, false);
1026 GOTO(put_child, rc);
1028 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1031 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1034 GOTO(put_child, rc);
1037 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1038 * whether need MA_LOV and MA_COOKIE.
1040 ma->ma_need = MA_INODE;
1043 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1044 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1045 /* save version when object is locked */
1046 mdt_version_get_save(info, mc, 1);
1048 mutex_lock(&mc->mot_lov_mutex);
1050 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1051 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1053 mutex_unlock(&mc->mot_lov_mutex);
1055 GOTO(unlock_child, rc);
1057 if (!lu_object_is_dying(&mc->mot_header)) {
1058 rc = mdt_attr_get_complex(info, mc, ma);
1061 } else if (mdt_dom_check_for_discard(info, mc)) {
1062 mdt_dom_discard_data(info, mc);
1064 mdt_handle_last_unlink(info, mc, ma);
1067 if (ma->ma_valid & MA_INODE) {
1068 switch (ma->ma_attr.la_mode & S_IFMT) {
1070 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1078 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1081 LASSERTF(0, "bad file type %o unlinking\n",
1082 ma->ma_attr.la_mode);
1089 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1091 mdt_object_put(info->mti_env, mc);
1093 mdt_object_unlock(info, mp, parent_lh, rc);
1095 mdt_object_put(info->mti_env, mp);
1096 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1101 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1104 static int mdt_reint_link(struct mdt_thread_info *info,
1105 struct mdt_lock_handle *lhc)
1107 struct mdt_reint_record *rr = &info->mti_rr;
1108 struct ptlrpc_request *req = mdt_info_req(info);
1109 struct md_attr *ma = &info->mti_attr;
1110 struct mdt_object *ms;
1111 struct mdt_object *mp;
1112 struct mdt_lock_handle *lhs;
1113 struct mdt_lock_handle *lhp;
1118 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1119 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1121 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1122 RETURN(err_serious(-ENOENT));
1124 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1125 req->rq_no_reply = 1;
1126 RETURN(err_serious(-ENOENT));
1129 if (info->mti_dlm_req)
1130 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1132 /* Invalid case so return error immediately instead of
1134 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1137 if (!fid_is_md_operative(rr->rr_fid1) ||
1138 !fid_is_md_operative(rr->rr_fid2))
1141 /* step 1: find target parent dir */
1142 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1144 RETURN(PTR_ERR(mp));
1146 rc = mdt_version_get_check_save(info, mp, 0);
1148 GOTO(put_parent, rc);
1150 /* step 2: find source */
1151 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1153 GOTO(put_parent, rc = PTR_ERR(ms));
1155 if (!mdt_object_exists(ms)) {
1156 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1157 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1158 GOTO(put_source, rc = -ENOENT);
1161 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1163 lhp = &info->mti_lh[MDT_LH_PARENT];
1164 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1165 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1168 GOTO(put_source, rc);
1170 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1172 lhs = &info->mti_lh[MDT_LH_CHILD];
1173 mdt_lock_reg_init(lhs, LCK_EX);
1174 rc = mdt_reint_object_lock(info, ms, lhs,
1175 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1178 GOTO(unlock_parent, rc);
1180 /* step 3: link it */
1181 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1182 OBD_FAIL_MDS_REINT_LINK_WRITE);
1184 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1185 rc = mdt_version_get_check_save(info, ms, 1);
1187 GOTO(unlock_source, rc);
1189 /** check target version by name during replay */
1190 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1191 &info->mti_tmp_fid1, 2);
1192 if (rc != 0 && rc != -ENOENT)
1193 GOTO(unlock_source, rc);
1194 /* save version of file name for replay, it must be ENOENT here */
1195 if (!req_is_replay(mdt_info_req(info))) {
1196 if (rc != -ENOENT) {
1197 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1198 PNAME(&rr->rr_name));
1199 GOTO(unlock_source, rc = -EEXIST);
1201 info->mti_ver[2] = ENOENT_VERSION;
1202 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1205 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1206 mdt_object_child(ms), &rr->rr_name, ma);
1209 mdt_counter_incr(req, LPROC_MDT_LINK);
1213 mdt_object_unlock(info, ms, lhs, rc);
1215 mdt_object_unlock(info, mp, lhp, rc);
1217 mdt_object_put(info->mti_env, ms);
1219 mdt_object_put(info->mti_env, mp);
1223 * lock the part of the directory according to the hash of the name
1224 * (lh->mlh_pdo_hash) in parallel directory lock.
1226 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1227 struct mdt_lock_handle *lh,
1228 struct mdt_object *obj, __u64 ibits,
1231 struct ldlm_res_id *res = &info->mti_res_id;
1232 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1233 union ldlm_policy_data *policy = &info->mti_policy;
1234 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1238 * Finish res_id initializing by name hash marking part of
1239 * directory which is taking modification.
1241 LASSERT(lh->mlh_pdo_hash != 0);
1242 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1243 memset(policy, 0, sizeof(*policy));
1244 policy->l_inodebits.bits = ibits;
1246 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1247 dlmflags |= LDLM_FL_COS_INCOMPAT;
1249 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1250 * going to be sent to client. If it is - mdt_intent_policy() path will
1251 * fix it up and turn FL_LOCAL flag off.
1253 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1254 policy, res, dlmflags,
1255 &info->mti_exp->exp_handle.h_cookie);
1260 * Get BFL lock for rename or migrate process.
1262 static int mdt_rename_lock(struct mdt_thread_info *info,
1263 struct lustre_handle *lh)
1268 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1269 struct lu_fid *fid = &info->mti_tmp_fid1;
1270 struct mdt_object *obj;
1272 /* XXX, right now, it has to use object API to
1273 * enqueue lock cross MDT, so it will enqueue
1274 * rename lock(with LUSTRE_BFL_FID) by root object */
1276 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1278 RETURN(PTR_ERR(obj));
1280 rc = mdt_remote_object_lock(info, obj,
1281 &LUSTRE_BFL_FID, lh,
1283 MDS_INODELOCK_UPDATE, false);
1284 mdt_object_put(info->mti_env, obj);
1286 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1287 union ldlm_policy_data *policy = &info->mti_policy;
1288 struct ldlm_res_id *res_id = &info->mti_res_id;
1291 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1292 memset(policy, 0, sizeof *policy);
1293 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1294 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1295 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1296 LDLM_IBITS, policy, LCK_EX, &flags,
1298 ldlm_completion_ast, NULL, NULL, 0,
1300 &info->mti_exp->exp_handle.h_cookie,
1307 static void mdt_rename_unlock(struct lustre_handle *lh)
1310 LASSERT(lustre_handle_is_used(lh));
1311 /* Cancel the single rename lock right away */
1312 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1316 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1317 const struct lu_fid *fid,
1320 struct mdt_object *dir;
1325 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1329 /* check early, the real version will be saved after locking */
1330 rc = mdt_version_get_check(info, dir, idx);
1334 if (!mdt_object_exists(dir))
1335 GOTO(out_put, rc = -ENOENT);
1337 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1338 GOTO(out_put, rc = -ENOTDIR);
1342 mdt_object_put(info->mti_env, dir);
1347 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1348 * herein we don't really check it, just do revoke.
1350 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1351 struct mdt_object *pobj,
1352 struct mdt_object *obj)
1354 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1357 mdt_lock_handle_init(lh);
1358 mdt_lock_reg_init(lh, LCK_EX);
1360 if (mdt_object_remote(pobj)) {
1361 /* don't bother to check if pobj and obj are on the same MDT. */
1362 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1363 &lh->mlh_rreg_lh, LCK_EX,
1364 MDS_INODELOCK_LOOKUP, false);
1365 } else if (mdt_object_remote(obj)) {
1366 struct ldlm_res_id *res = &info->mti_res_id;
1367 union ldlm_policy_data *policy = &info->mti_policy;
1368 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1369 LDLM_FL_COS_INCOMPAT;
1371 fid_build_reg_res_name(mdt_object_fid(obj), res);
1372 memset(policy, 0, sizeof(*policy));
1373 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1374 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1375 &lh->mlh_reg_lh, LCK_EX, policy, res,
1378 /* do nothing if both are local */
1386 * TODO, currently we don't save this lock because there is no place to
1387 * hold this lock handle, but to avoid race we need to save this lock.
1389 mdt_object_unlock(info, NULL, lh, 1);
1395 * operation may takes locks of linkea, or directory stripes, group them in
1398 struct mdt_sub_lock {
1399 struct mdt_object *msl_obj;
1400 struct mdt_lock_handle msl_lh;
1401 struct list_head msl_linkage;
1404 static void mdt_unlock_list(struct mdt_thread_info *info,
1405 struct list_head *list, int decref)
1407 struct mdt_sub_lock *msl;
1408 struct mdt_sub_lock *tmp;
1410 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1411 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1412 list_del(&msl->msl_linkage);
1417 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1418 struct mdt_object *obj,
1419 struct mdt_lock_handle *lh,
1420 struct ldlm_enqueue_info *einfo,
1421 struct list_head *slave_locks,
1424 if (mdt_object_remote(obj)) {
1425 mdt_unlock_list(info, slave_locks, decref);
1426 mdt_object_unlock(info, obj, lh, decref);
1428 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1433 * lock parents of links, and also check whether total locks don't exceed
1436 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1437 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1438 * \retval -ev negative errno upon error
1440 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1441 struct mdt_object *pobj,
1442 const struct md_attr *ma,
1443 struct mdt_object *obj,
1444 struct mdt_lock_handle *lhp,
1445 struct ldlm_enqueue_info *peinfo,
1446 struct list_head *parent_slave_locks,
1447 struct list_head *link_locks)
1449 struct mdt_device *mdt = info->mti_mdt;
1450 struct lu_buf *buf = &info->mti_big_buf;
1451 struct lu_name *lname = &info->mti_name;
1452 struct linkea_data ldata = { NULL };
1453 bool blocked = false;
1454 int local_lnkp_cnt = 0;
1459 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1462 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1463 if (buf->lb_buf == NULL)
1467 rc = mdt_links_read(info, obj, &ldata);
1469 if (rc == -ENOENT || rc == -ENODATA)
1474 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1475 linkea_next_entry(&ldata)) {
1476 struct mdt_object *lnkp;
1477 struct mdt_sub_lock *msl;
1481 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1484 /* check if it's also linked to parent */
1485 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1486 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1487 PFID(&fid), PNAME(lname));
1488 /* in case link is remote object, revoke LOOKUP lock */
1489 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1495 /* check if it's linked to a stripe of parent */
1496 if (ma->ma_valid & MA_LMV) {
1497 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1498 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1501 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1502 fid_le_to_cpu(stripe_fid,
1503 &lmv->lmv_stripe_fids[j]);
1504 if (lu_fid_eq(stripe_fid, &fid)) {
1505 CDEBUG(D_INFO, "skip stripe "DFID
1506 ", reovke "DNAME"\n",
1507 PFID(&fid), PNAME(lname));
1508 lnkp = mdt_object_find(info->mti_env,
1511 GOTO(out, rc = PTR_ERR(lnkp));
1517 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1519 mdt_object_put(info->mti_env, lnkp);
1524 /* Check if it's already locked */
1525 list_for_each_entry(msl, link_locks, msl_linkage) {
1526 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1528 DFID" was locked, revoke "DNAME"\n",
1529 PFID(&fid), PNAME(lname));
1530 lnkp = msl->msl_obj;
1536 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1540 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1541 PFID(&fid), PNAME(lname));
1543 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1545 CWARN("%s: cannot find obj "DFID": %ld\n",
1546 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1550 if (!mdt_object_exists(lnkp)) {
1551 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1552 PFID(&fid), PNAME(lname));
1553 mdt_object_put(info->mti_env, lnkp);
1557 if (!mdt_object_remote(lnkp))
1562 GOTO(out, rc = -ENOMEM);
1565 * we can't follow parent-child lock order like other MD
1566 * operations, use lock_try here to avoid deadlock, if the lock
1567 * cannot be taken, drop all locks taken, revoke the blocked
1568 * one, and continue processing the remaining entries, and in
1569 * the end of the loop restart from beginning.
1571 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1573 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1574 MDS_INODELOCK_UPDATE, true);
1575 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1577 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1578 PFID(&fid), PNAME(lname));
1580 mdt_unlock_list(info, link_locks, 1);
1581 /* also unlock parent locks to avoid deadlock */
1583 mdt_migrate_object_unlock(info, pobj, lhp,
1590 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1591 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1592 MDS_INODELOCK_UPDATE);
1594 mdt_object_put(info->mti_env, lnkp);
1599 if (mdt_object_remote(lnkp)) {
1600 struct ldlm_lock *lock;
1603 * for remote object, set lock cb_atomic,
1604 * so lock can be released in blocking_ast()
1605 * immediately, then the next lock_try will
1606 * have better chance of success.
1608 lock = ldlm_handle2lock(
1609 &msl->msl_lh.mlh_rreg_lh);
1610 LASSERT(lock != NULL);
1611 lock_res_and_lock(lock);
1612 ldlm_set_atomic_cb(lock);
1613 unlock_res_and_lock(lock);
1614 LDLM_LOCK_PUT(lock);
1617 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1622 INIT_LIST_HEAD(&msl->msl_linkage);
1623 msl->msl_obj = lnkp;
1624 list_add_tail(&msl->msl_linkage, link_locks);
1626 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1630 GOTO(out, rc = -EBUSY);
1635 mdt_unlock_list(info, link_locks, rc);
1636 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1638 * parent may have 3 local objects: master object and 2 stripes
1639 * (if it's being migrated too); source may have 2 local
1640 * objects: master and 1 stripe; target has 1 local object.
1646 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1647 struct mdt_object *obj,
1648 const struct md_attr *ma,
1649 struct list_head *slave_locks)
1651 struct mdt_device *mdt = info->mti_mdt;
1652 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1653 struct lu_fid *fid = &info->mti_tmp_fid1;
1654 struct mdt_object *slave;
1655 struct mdt_sub_lock *msl;
1661 LASSERT(mdt_object_remote(obj));
1662 LASSERT(ma->ma_valid & MA_LMV);
1665 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1668 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1671 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1672 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1674 if (!fid_is_sane(fid))
1677 slave = mdt_object_find(info->mti_env, mdt, fid);
1679 GOTO(out, rc = PTR_ERR(slave));
1683 mdt_object_put(info->mti_env, slave);
1684 GOTO(out, rc = -ENOMEM);
1687 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1688 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1689 MDS_INODELOCK_UPDATE, true);
1692 mdt_object_put(info->mti_env, slave);
1696 INIT_LIST_HEAD(&msl->msl_linkage);
1697 msl->msl_obj = slave;
1698 list_add_tail(&msl->msl_linkage, slave_locks);
1704 mdt_unlock_list(info, slave_locks, rc);
1708 /* lock parent and its stripes */
1709 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1710 struct mdt_object *obj,
1711 const struct md_attr *ma,
1712 struct mdt_lock_handle *lh,
1713 struct ldlm_enqueue_info *einfo,
1714 struct list_head *slave_locks)
1718 if (mdt_object_remote(obj)) {
1719 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1720 &lh->mlh_rreg_lh, LCK_PW,
1721 MDS_INODELOCK_UPDATE, false);
1726 * if obj is remote and striped, lock its stripes explicitly
1727 * because it's not striped in LOD layer on this MDT.
1729 if (ma->ma_valid & MA_LMV) {
1730 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1732 mdt_object_unlock(info, obj, lh, rc);
1735 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1743 * in migration, object may be remote, and we need take full lock of it and its
1744 * stripes if it's directory, besides, object may be a remote object on its
1745 * parent, revoke its LOOKUP lock on where its parent is located.
1747 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1748 struct mdt_object *pobj,
1749 struct mdt_object *obj,
1750 struct mdt_lock_handle *lh,
1751 struct ldlm_enqueue_info *einfo,
1752 struct list_head *slave_locks)
1756 if (mdt_object_remote(obj)) {
1757 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1761 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1762 &lh->mlh_rreg_lh, LCK_EX,
1763 MDS_INODELOCK_FULL, false);
1768 * if obj is remote and striped, lock its stripes explicitly
1769 * because it's not striped in LOD layer on this MDT.
1771 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1772 struct md_attr *ma = &info->mti_attr;
1774 ma->ma_lmv = info->mti_big_lmm;
1775 ma->ma_lmv_size = info->mti_big_lmmsize;
1777 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1779 mdt_object_unlock(info, obj, lh, rc);
1783 if (ma->ma_valid & MA_LMV) {
1784 rc = mdt_lock_remote_slaves(info, obj, ma,
1787 mdt_object_unlock(info, obj, lh, rc);
1791 if (mdt_object_remote(pobj)) {
1792 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1797 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1805 * lookup source by name, if parent is striped directory, we need to find the
1806 * corresponding stripe where source is located, and then lookup there.
1808 * besides, if parent is migrating too, and file is already in target stripe,
1809 * this should be a redo of 'lfs migrate' on client side.
1811 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1812 struct mdt_object *pobj,
1813 const struct md_attr *ma,
1814 const struct lu_name *lname,
1815 struct mdt_object **spobj,
1816 struct mdt_object **sobj)
1818 const struct lu_env *env = info->mti_env;
1819 struct lu_fid *fid = &info->mti_tmp_fid1;
1820 struct mdt_object *stripe;
1823 if (ma->ma_valid & MA_LMV) {
1824 /* if parent is striped, lookup on corresponding stripe */
1825 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1826 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1827 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1828 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1829 LMV_HASH_FLAG_MIGRATION;
1832 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1833 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1836 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1842 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1843 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1845 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1847 stripe = mdt_object_find(env, info->mti_mdt, fid);
1849 return PTR_ERR(stripe);
1852 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1854 if (rc == -ENOENT && is_migrating) {
1856 * if parent is migrating, and lookup child failed on
1857 * source stripe, lookup again on target stripe, if it
1858 * exists, it means previous migration was interrupted,
1859 * and current file was migrated already.
1861 mdt_object_put(env, stripe);
1863 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1864 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1866 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1872 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1874 stripe = mdt_object_find(env, info->mti_mdt, fid);
1876 return PTR_ERR(stripe);
1879 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1880 fid, &info->mti_spec);
1881 mdt_object_put(env, stripe);
1882 return rc ?: -EALREADY;
1884 mdt_object_put(env, stripe);
1889 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1895 mdt_object_get(env, stripe);
1900 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1901 if (IS_ERR(*sobj)) {
1902 mdt_object_put(env, stripe);
1903 rc = PTR_ERR(*sobj);
1911 /* end lease and close file for regular file */
1912 static int mdd_migrate_close(struct mdt_thread_info *info,
1913 struct mdt_object *obj)
1915 struct close_data *data;
1916 struct mdt_body *repbody;
1917 struct ldlm_lock *lease;
1922 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1924 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1928 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1933 lease = ldlm_handle2lock(&data->cd_handle);
1937 /* check if the lease was already canceled */
1938 lock_res_and_lock(lease);
1939 rc = ldlm_is_cancel(lease);
1940 unlock_res_and_lock(lease);
1944 LDLM_DEBUG(lease, DFID" lease broken",
1945 PFID(mdt_object_fid(obj)));
1949 * cancel server side lease, client side counterpart should have been
1950 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1952 ldlm_lock_cancel(lease);
1953 ldlm_reprocess_all(lease->l_resource, lease);
1954 LDLM_LOCK_PUT(lease);
1957 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1958 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1959 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1965 * migrate file in below steps:
1966 * 1. lock parent and its stripes
1967 * 2. lookup source by name
1968 * 3. lock parents of source links if source is not directory
1969 * 4. reject if source is in HSM
1970 * 5. take source open_sem and close file if source is regular file
1971 * 6. lock source and its stripes if it's directory
1972 * 7. lock target so subsequent change to it can trigger COS
1974 * 9. unlock above locks
1975 * 10. sync device if source has links
1977 static int mdt_reint_migrate(struct mdt_thread_info *info,
1978 struct mdt_lock_handle *unused)
1980 const struct lu_env *env = info->mti_env;
1981 struct mdt_device *mdt = info->mti_mdt;
1982 struct ptlrpc_request *req = mdt_info_req(info);
1983 struct mdt_reint_record *rr = &info->mti_rr;
1984 struct lu_ucred *uc = mdt_ucred(info);
1985 struct md_attr *ma = &info->mti_attr;
1986 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1987 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1988 struct mdt_object *pobj;
1989 struct mdt_object *spobj = NULL;
1990 struct mdt_object *sobj = NULL;
1991 struct mdt_object *tobj;
1992 struct lustre_handle rename_lh = { 0 };
1993 struct mdt_lock_handle *lhp;
1994 struct mdt_lock_handle *lhs;
1995 struct mdt_lock_handle *lht;
1996 LIST_HEAD(parent_slave_locks);
1997 LIST_HEAD(child_slave_locks);
1998 LIST_HEAD(link_locks);
1999 int lock_retries = 5;
2000 bool open_sem_locked = false;
2001 bool do_sync = false;
2005 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2006 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2008 if (info->mti_dlm_req)
2009 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2011 if (!fid_is_md_operative(rr->rr_fid1) ||
2012 !fid_is_md_operative(rr->rr_fid2))
2015 /* don't allow migrate . or .. */
2016 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2019 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2022 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2023 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2024 mdt->mdt_enable_remote_dir_gid != -1)
2028 * Note: do not enqueue rename lock for replay request, because
2029 * if other MDT holds rename lock, but being blocked to wait for
2030 * this MDT to finish its recovery, and the failover MDT can not
2031 * get rename lock, which will cause deadlock.
2033 if (!req_is_replay(req)) {
2034 rc = mdt_rename_lock(info, &rename_lh);
2036 CERROR("%s: can't lock FS for rename: rc = %d\n",
2037 mdt_obd_name(info->mti_mdt), rc);
2042 /* pobj is master object of parent */
2043 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2045 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2047 if (unlikely(!info->mti_big_lmm)) {
2048 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2049 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2050 if (!info->mti_big_lmm)
2051 GOTO(put_parent, rc = -ENOMEM);
2054 ma->ma_lmv = info->mti_big_lmm;
2055 ma->ma_lmv_size = info->mti_big_lmmsize;
2057 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2059 GOTO(put_parent, rc);
2062 /* lock parent object */
2063 lhp = &info->mti_lh[MDT_LH_PARENT];
2064 mdt_lock_reg_init(lhp, LCK_PW);
2065 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2066 &parent_slave_locks);
2068 GOTO(put_parent, rc);
2071 * spobj is the corresponding stripe against name if pobj is striped
2072 * directory, which is the real parent, and no need to lock, because
2073 * we've taken full lock of pobj.
2075 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2077 GOTO(unlock_parent, rc);
2079 /* lock parents of source links, and revoke LOOKUP lock of links */
2080 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2081 &parent_slave_locks, &link_locks);
2082 if (rc == -EBUSY && lock_retries-- > 0) {
2083 mdt_object_put(env, sobj);
2084 mdt_object_put(env, spobj);
2089 GOTO(put_source, rc);
2092 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2093 * with one request, if total lock count exceeds this limit, we will
2094 * drop all locks after migration, and synchronous device in the end.
2098 /* TODO: DoM migration is not supported yet */
2099 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2100 ma->ma_lmm = info->mti_big_lmm;
2101 ma->ma_lmm_size = info->mti_big_lmmsize;
2103 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2105 GOTO(put_source, rc);
2107 if (ma->ma_valid & MA_LOV &&
2108 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2109 GOTO(put_source, rc = -EOPNOTSUPP);
2112 /* if migration HSM is allowed */
2113 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2114 ma->ma_need = MA_HSM;
2116 rc = mdt_attr_get_complex(info, sobj, ma);
2118 GOTO(unlock_links, rc);
2120 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2121 GOTO(unlock_links, rc = -EOPNOTSUPP);
2124 /* end lease and close file for regular file */
2125 if (info->mti_spec.sp_migrate_close) {
2126 /* try to hold open_sem so that nobody else can open the file */
2127 if (!down_write_trylock(&sobj->mot_open_sem)) {
2129 mdd_migrate_close(info, sobj);
2130 GOTO(unlock_links, rc = -EBUSY);
2132 open_sem_locked = true;
2133 rc = mdd_migrate_close(info, sobj);
2135 GOTO(unlock_open_sem, rc);
2140 lhs = &info->mti_lh[MDT_LH_OLD];
2141 mdt_lock_reg_init(lhs, LCK_EX);
2142 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2143 &child_slave_locks);
2145 GOTO(unlock_open_sem, rc);
2148 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2150 GOTO(unlock_source, rc = PTR_ERR(tobj));
2152 lht = &info->mti_lh[MDT_LH_NEW];
2153 mdt_lock_reg_init(lht, LCK_EX);
2154 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2156 GOTO(put_target, rc);
2158 /* Don't do lookup sanity check. We know name doesn't exist. */
2159 info->mti_spec.sp_cr_lookup = 0;
2160 info->mti_spec.sp_feat = &dt_directory_features;
2162 rc = mdo_migrate(env, mdt_object_child(pobj),
2163 mdt_object_child(sobj), &rr->rr_name,
2164 mdt_object_child(tobj), &info->mti_spec, ma);
2167 mdt_object_unlock(info, tobj, lht, rc);
2169 mdt_object_put(env, tobj);
2171 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2172 &child_slave_locks, rc);
2174 if (open_sem_locked)
2175 up_write(&sobj->mot_open_sem);
2177 /* if we've got too many locks to save into RPC,
2178 * then just commit before the locks are released */
2180 mdt_device_sync(env, mdt);
2181 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2183 mdt_object_put(env, sobj);
2184 mdt_object_put(env, spobj);
2186 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2187 &parent_slave_locks, rc);
2189 mdt_object_put(env, pobj);
2191 if (lustre_handle_is_used(&rename_lh))
2192 mdt_rename_unlock(&rename_lh);
2197 static int mdt_object_lock_save(struct mdt_thread_info *info,
2198 struct mdt_object *dir,
2199 struct mdt_lock_handle *lh,
2200 int idx, bool cos_incompat)
2204 /* we lock the target dir if it is local */
2205 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2210 /* get and save correct version after locking */
2211 mdt_version_get_save(info, dir, idx);
2216 * determine lock order of sobj and tobj
2218 * there are two situations we need to lock tobj before sobj:
2219 * 1. sobj is child of tobj
2220 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2221 * larger than that of tobj
2223 * \retval 1 lock tobj before sobj
2224 * \retval 0 lock sobj before tobj
2225 * \retval -ev negative errno upon error
2227 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2228 struct mdt_object *sobj,
2229 struct mdt_object *tobj)
2231 struct md_attr *ma = &info->mti_attr;
2232 struct lu_fid *spfid = &info->mti_tmp_fid1;
2233 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2234 struct lmv_mds_md_v1 *lmv;
2239 /* sobj and tobj are the same */
2243 if (fid_is_root(mdt_object_fid(sobj)))
2246 if (fid_is_root(mdt_object_fid(tobj)))
2249 /* check whether sobj is child of tobj */
2250 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2251 mdt_object_fid(tobj));
2258 /* check whether sobj and tobj are children of the same parent */
2259 rc = mdt_attr_get_pfid(info, sobj, spfid);
2263 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2267 if (!lu_fid_eq(spfid, tpfid))
2270 /* check whether sobj and tobj are sibling stripes */
2271 ma->ma_need = MA_LMV;
2273 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2274 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2275 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2279 if (!(ma->ma_valid & MA_LMV))
2282 lmv = &ma->ma_lmv->lmv_md_v1;
2283 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2285 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2288 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2292 if (!(ma->ma_valid & MA_LMV))
2295 lmv = &ma->ma_lmv->lmv_md_v1;
2296 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2298 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2300 /* check stripe index of sobj and tobj */
2301 if (sindex == tindex)
2304 return sindex < tindex ? 0 : 1;
2308 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2309 * 2 - srcdir child; 3 - tgtdir child.
2310 * Update on disk version of srcdir child.
2312 static int mdt_reint_rename(struct mdt_thread_info *info,
2313 struct mdt_lock_handle *unused)
2315 struct mdt_device *mdt = info->mti_mdt;
2316 struct mdt_reint_record *rr = &info->mti_rr;
2317 struct md_attr *ma = &info->mti_attr;
2318 struct ptlrpc_request *req = mdt_info_req(info);
2319 struct mdt_object *msrcdir = NULL;
2320 struct mdt_object *mtgtdir = NULL;
2321 struct mdt_object *mold;
2322 struct mdt_object *mnew = NULL;
2323 struct lustre_handle rename_lh = { 0 };
2324 struct mdt_lock_handle *lh_srcdirp;
2325 struct mdt_lock_handle *lh_tgtdirp;
2326 struct mdt_lock_handle *lh_oldp = NULL;
2327 struct mdt_lock_handle *lh_newp = NULL;
2328 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2329 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2331 bool reverse = false, discard = false;
2336 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2337 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2338 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2340 if (info->mti_dlm_req)
2341 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2343 if (!fid_is_md_operative(rr->rr_fid1) ||
2344 !fid_is_md_operative(rr->rr_fid2))
2347 /* find both parents. */
2348 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2349 if (IS_ERR(msrcdir))
2350 RETURN(PTR_ERR(msrcdir));
2352 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2354 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2356 mdt_object_get(info->mti_env, mtgtdir);
2358 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2359 if (IS_ERR(mtgtdir))
2360 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2364 * Note: do not enqueue rename lock for replay request, because
2365 * if other MDT holds rename lock, but being blocked to wait for
2366 * this MDT to finish its recovery, and the failover MDT can not
2367 * get rename lock, which will cause deadlock.
2369 if (!req_is_replay(req)) {
2371 * Normally rename RPC is handled on the MDT with the target
2372 * directory (if target exists, it's on the MDT with the
2373 * target), if the source directory is remote, it's a hint that
2374 * source is remote too (this may not be true, but it won't
2375 * cause any issue), return -EXDEV early to avoid taking
2378 if (!mdt->mdt_enable_remote_rename &&
2379 mdt_object_remote(msrcdir))
2380 GOTO(out_put_tgtdir, rc = -EXDEV);
2382 rc = mdt_rename_lock(info, &rename_lh);
2384 CERROR("%s: can't lock FS for rename: rc = %d\n",
2385 mdt_obd_name(mdt), rc);
2386 GOTO(out_put_tgtdir, rc);
2390 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2392 GOTO(out_unlock_rename, rc);
2396 /* source needs to be looked up after locking source parent, otherwise
2397 * this rename may race with unlink source, and cause rename hang, see
2398 * sanityn.sh 55b, so check parents first, if later we found source is
2399 * remote, relock parents. */
2400 cos_incompat = (mdt_object_remote(msrcdir) ||
2401 mdt_object_remote(mtgtdir));
2403 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2405 /* lock parents in the proper order. */
2406 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2407 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2410 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2411 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2414 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2417 GOTO(out_unlock_rename, rc);
2419 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2421 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2424 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2425 GOTO(out_unlock_rename, rc);
2428 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2431 GOTO(out_unlock_rename, rc);
2433 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2435 if (mtgtdir != msrcdir) {
2436 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2438 } else if (!mdt_object_remote(mtgtdir) &&
2439 lh_srcdirp->mlh_pdo_hash !=
2440 lh_tgtdirp->mlh_pdo_hash) {
2441 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2442 MDS_INODELOCK_UPDATE,
2444 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2447 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2448 GOTO(out_unlock_rename, rc);
2452 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2453 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2455 /* find mold object. */
2457 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2459 GOTO(out_unlock_parents, rc);
2461 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2462 GOTO(out_unlock_parents, rc = -EINVAL);
2464 if (!fid_is_md_operative(old_fid))
2465 GOTO(out_unlock_parents, rc = -EPERM);
2467 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2469 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2471 if (!mdt_object_exists(mold)) {
2472 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2474 "object does not exist");
2475 GOTO(out_put_old, rc = -ENOENT);
2478 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2479 GOTO(out_put_old, rc = -EXDEV);
2481 /* Check if @mtgtdir is subdir of @mold, before locking child
2482 * to avoid reverse locking. */
2483 if (mtgtdir != msrcdir) {
2484 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2489 GOTO(out_put_old, rc);
2493 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2494 /* save version after locking */
2495 mdt_version_get_save(info, mold, 2);
2497 if (!cos_incompat && mdt_object_remote(mold)) {
2498 cos_incompat = true;
2499 mdt_object_put(info->mti_env, mold);
2500 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2501 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2505 /* find mnew object:
2506 * mnew target object may not exist now
2507 * lookup with version checking */
2509 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2512 /* the new_fid should have been filled at this moment */
2513 if (lu_fid_eq(old_fid, new_fid))
2514 GOTO(out_put_old, rc);
2516 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2517 lu_fid_eq(new_fid, rr->rr_fid2))
2518 GOTO(out_put_old, rc = -EINVAL);
2520 if (!fid_is_md_operative(new_fid))
2521 GOTO(out_put_old, rc = -EPERM);
2523 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2525 GOTO(out_put_old, rc = PTR_ERR(mnew));
2527 if (!mdt_object_exists(mnew)) {
2528 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2530 "object does not exist");
2531 GOTO(out_put_new, rc = -ENOENT);
2534 if (mdt_object_remote(mnew)) {
2535 struct mdt_body *repbody;
2537 /* Always send rename req to the target child MDT */
2538 repbody = req_capsule_server_get(info->mti_pill,
2540 LASSERT(repbody != NULL);
2541 repbody->mbo_fid1 = *new_fid;
2542 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2543 GOTO(out_put_new, rc = -EXDEV);
2545 /* Before locking the target dir, check we do not replace
2546 * a dir with a non-dir, otherwise it may deadlock with
2547 * link op which tries to create a link in this dir
2548 * back to this non-dir. */
2549 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2550 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2551 GOTO(out_put_new, rc = -EISDIR);
2553 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2554 mdt_lock_reg_init(lh_oldp, LCK_EX);
2555 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2556 if (mdt_object_remote(msrcdir)) {
2557 /* Enqueue lookup lock from the parent MDT */
2558 rc = mdt_remote_object_lock(info, msrcdir,
2559 mdt_object_fid(mold),
2560 &lh_oldp->mlh_rreg_lh,
2561 lh_oldp->mlh_rreg_mode,
2562 MDS_INODELOCK_LOOKUP,
2565 GOTO(out_put_new, rc);
2567 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2570 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2573 GOTO(out_unlock_old, rc);
2575 /* Check if @msrcdir is subdir of @mnew, before locking child
2576 * to avoid reverse locking. */
2577 if (mtgtdir != msrcdir) {
2578 rc = mdo_is_subdir(info->mti_env,
2579 mdt_object_child(msrcdir), new_fid);
2583 GOTO(out_unlock_old, rc);
2587 /* We used to acquire MDS_INODELOCK_FULL here but we
2588 * can't do this now because a running HSM restore on
2589 * the rename onto victim will hold the layout
2590 * lock. See LU-4002. */
2592 lh_newp = &info->mti_lh[MDT_LH_NEW];
2593 mdt_lock_reg_init(lh_newp, LCK_EX);
2594 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2595 MDS_INODELOCK_LOOKUP |
2596 MDS_INODELOCK_UPDATE,
2599 GOTO(out_unlock_old, rc);
2601 /* get and save version after locking */
2602 mdt_version_get_save(info, mnew, 3);
2603 } else if (rc != -EREMOTE && rc != -ENOENT) {
2604 GOTO(out_put_old, rc);
2606 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2607 mdt_lock_reg_init(lh_oldp, LCK_EX);
2608 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2609 if (mdt_object_remote(msrcdir)) {
2610 /* Enqueue lookup lock from the parent MDT */
2611 rc = mdt_remote_object_lock(info, msrcdir,
2612 mdt_object_fid(mold),
2613 &lh_oldp->mlh_rreg_lh,
2614 lh_oldp->mlh_rreg_mode,
2615 MDS_INODELOCK_LOOKUP,
2618 GOTO(out_put_old, rc);
2620 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2623 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2626 GOTO(out_unlock_old, rc);
2628 mdt_enoent_version_save(info, 3);
2631 /* step 5: rename it */
2632 mdt_reint_init_ma(info, ma);
2634 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2635 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2638 mutex_lock(&mnew->mot_lov_mutex);
2640 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2641 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2642 mnew != NULL ? mdt_object_child(mnew) : NULL,
2643 &rr->rr_tgt_name, ma);
2646 mutex_unlock(&mnew->mot_lov_mutex);
2648 /* handle last link of tgt object */
2650 mdt_counter_incr(req, LPROC_MDT_RENAME);
2652 mdt_handle_last_unlink(info, mnew, ma);
2653 discard = mdt_dom_check_for_discard(info, mnew);
2655 mdt_rename_counter_tally(info, info->mti_mdt, req,
2661 mdt_object_unlock(info, mnew, lh_newp, rc);
2663 mdt_object_unlock(info, mold, lh_oldp, rc);
2665 if (mnew && !discard)
2666 mdt_object_put(info->mti_env, mnew);
2668 mdt_object_put(info->mti_env, mold);
2670 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2671 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2673 if (lustre_handle_is_used(&rename_lh))
2674 mdt_rename_unlock(&rename_lh);
2676 mdt_object_put(info->mti_env, mtgtdir);
2678 mdt_object_put(info->mti_env, msrcdir);
2680 /* The DoM discard can be done right in the place above where it is
2681 * assigned, meanwhile it is done here after rename unlock due to
2682 * compatibility with old clients, for them the discard blocks
2683 * the main thread until completion. Check LU-11359 for details.
2686 mdt_dom_discard_data(info, mnew);
2687 mdt_object_put(info->mti_env, mnew);
2692 static int mdt_reint_resync(struct mdt_thread_info *info,
2693 struct mdt_lock_handle *lhc)
2695 struct mdt_reint_record *rr = &info->mti_rr;
2696 struct ptlrpc_request *req = mdt_info_req(info);
2697 struct md_attr *ma = &info->mti_attr;
2698 struct mdt_object *mo;
2699 struct ldlm_lock *lease;
2700 struct mdt_body *repbody;
2701 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2707 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2709 if (info->mti_dlm_req)
2710 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2712 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2714 GOTO(out, rc = PTR_ERR(mo));
2716 if (!mdt_object_exists(mo))
2717 GOTO(out_obj, rc = -ENOENT);
2719 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2720 GOTO(out_obj, rc = -EINVAL);
2722 if (mdt_object_remote(mo))
2723 GOTO(out_obj, rc = -EREMOTE);
2725 lease = ldlm_handle2lock(rr->rr_lease_handle);
2727 GOTO(out_obj, rc = -ESTALE);
2729 /* It's really necessary to grab open_sem and check if the lease lock
2730 * has been lost. There would exist a concurrent writer coming in and
2731 * generating some dirty data in memory cache, the writeback would fail
2732 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2733 if (!down_write_trylock(&mo->mot_open_sem))
2734 GOTO(out_put_lease, rc = -EBUSY);
2736 lock_res_and_lock(lease);
2737 lease_broken = ldlm_is_cancel(lease);
2738 unlock_res_and_lock(lease);
2740 GOTO(out_unlock, rc = -EBUSY);
2742 /* the file has yet opened by anyone else after we took the lease. */
2743 layout.mlc_opc = MD_LAYOUT_RESYNC;
2744 lhc = &info->mti_lh[MDT_LH_LOCAL];
2745 rc = mdt_layout_change(info, mo, lhc, &layout);
2747 GOTO(out_unlock, rc);
2749 mdt_object_unlock(info, mo, lhc, 0);
2751 ma->ma_need = MA_INODE;
2753 rc = mdt_attr_get_complex(info, mo, ma);
2755 GOTO(out_unlock, rc);
2757 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2758 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2762 up_write(&mo->mot_open_sem);
2764 LDLM_LOCK_PUT(lease);
2766 mdt_object_put(info->mti_env, mo);
2768 mdt_client_compatibility(info);
2769 rc2 = mdt_fix_reply(info);
2775 struct mdt_reinter {
2776 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2777 enum lprocfs_extra_opc mr_extra_opc;
2780 static const struct mdt_reinter mdt_reinters[] = {
2782 .mr_handler = &mdt_reint_setattr,
2783 .mr_extra_opc = MDS_REINT_SETATTR,
2786 .mr_handler = &mdt_reint_create,
2787 .mr_extra_opc = MDS_REINT_CREATE,
2790 .mr_handler = &mdt_reint_link,
2791 .mr_extra_opc = MDS_REINT_LINK,
2794 .mr_handler = &mdt_reint_unlink,
2795 .mr_extra_opc = MDS_REINT_UNLINK,
2798 .mr_handler = &mdt_reint_rename,
2799 .mr_extra_opc = MDS_REINT_RENAME,
2802 .mr_handler = &mdt_reint_open,
2803 .mr_extra_opc = MDS_REINT_OPEN,
2805 [REINT_SETXATTR] = {
2806 .mr_handler = &mdt_reint_setxattr,
2807 .mr_extra_opc = MDS_REINT_SETXATTR,
2810 .mr_handler = &mdt_reint_unlink,
2811 .mr_extra_opc = MDS_REINT_UNLINK,
2814 .mr_handler = &mdt_reint_migrate,
2815 .mr_extra_opc = MDS_REINT_RENAME,
2818 .mr_handler = &mdt_reint_resync,
2819 .mr_extra_opc = MDS_REINT_RESYNC,
2823 int mdt_reint_rec(struct mdt_thread_info *info,
2824 struct mdt_lock_handle *lhc)
2826 const struct mdt_reinter *mr;
2830 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2833 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2834 if (mr->mr_handler == NULL)
2837 rc = (*mr->mr_handler)(info, lhc);
2839 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2840 PTLRPC_LAST_CNTR + mr->mr_extra_opc);