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 ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
395 (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
399 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
400 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
401 mdt->mdt_enable_remote_dir_gid != -1)
405 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
407 parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
409 RETURN(PTR_ERR(parent));
411 if (!mdt_object_exists(parent))
412 GOTO(put_parent, rc = -ENOENT);
415 * LU-10235: check if name exists locklessly first to avoid massive
416 * lock recalls on existing directories.
418 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
419 &info->mti_tmp_fid1, 1);
421 GOTO(put_parent, rc = -EEXIST);
423 /* -ENOENT is expected here */
425 GOTO(put_parent, rc);
427 /* save version of file name for replay, it must be ENOENT here */
428 mdt_enoent_version_save(info, 1);
430 OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
432 lh = &info->mti_lh[MDT_LH_PARENT];
433 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
434 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
436 GOTO(put_parent, rc);
438 if (!mdt_object_remote(parent)) {
439 rc = mdt_version_get_check_save(info, parent, 0);
441 GOTO(unlock_parent, rc);
444 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
445 if (unlikely(IS_ERR(child)))
446 GOTO(unlock_parent, rc = PTR_ERR(child));
448 ma->ma_need = MA_INODE;
451 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
452 OBD_FAIL_MDS_REINT_CREATE_WRITE);
454 /* Version of child will be updated on disk. */
455 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
456 rc = mdt_version_get_check_save(info, child, 2);
460 /* Let lower layer know current lock mode. */
461 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
464 * Do not perform lookup sanity check. We know that name does
467 info->mti_spec.sp_cr_lookup = 0;
468 info->mti_spec.sp_feat = &dt_directory_features;
470 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
471 mdt_object_child(child), &info->mti_spec, ma);
473 rc = mdt_attr_get_complex(info, child, ma);
479 * On DNE, we need to eliminate dependey between 'mkdir a' and
480 * 'mkdir a/b' if b is a striped directory, to achieve this, two
481 * things are done below:
482 * 1. save child and slaves lock.
483 * 2. if the child is a striped directory, relock parent so to
484 * compare against with COS locks to ensure parent was
487 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
488 struct mdt_lock_handle *lhc;
489 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
492 rc = mdt_object_striped(info, child);
498 if (!mdt_object_remote(parent)) {
499 mdt_object_unlock(info, parent, lh, 1);
500 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
501 rc = mdt_reint_object_lock(info, parent, lh,
502 MDS_INODELOCK_UPDATE,
509 lhc = &info->mti_lh[MDT_LH_CHILD];
510 mdt_lock_handle_init(lhc);
511 mdt_lock_reg_init(lhc, LCK_PW);
512 rc = mdt_reint_striped_lock(info, child, lhc,
513 MDS_INODELOCK_UPDATE, einfo,
518 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
521 /* Return fid & attr to client. */
522 if (ma->ma_valid & MA_INODE)
523 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
524 mdt_object_fid(child));
526 mdt_object_put(info->mti_env, child);
528 mdt_object_unlock(info, parent, lh, rc);
530 mdt_object_put(info->mti_env, parent);
534 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
537 struct mdt_lock_handle *lh;
538 int do_vbr = ma->ma_attr.la_valid &
539 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
540 __u64 lockpart = MDS_INODELOCK_UPDATE;
541 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
546 rc = mdt_object_striped(info, mo);
552 lh = &info->mti_lh[MDT_LH_PARENT];
553 mdt_lock_reg_init(lh, LCK_PW);
555 /* Even though the new MDT will grant PERM lock to the old
556 * client, but the old client will almost ignore that during
557 * So it needs to revoke both LOOKUP and PERM lock here, so
558 * both new and old client can cancel the dcache */
559 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
560 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
562 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
567 /* all attrs are packed into mti_attr in unpack_setattr */
568 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
569 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
571 /* VBR: update version if attr changed are important for recovery */
573 /* update on-disk version of changed object */
574 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
575 rc = mdt_version_get_check_save(info, mo, 0);
577 GOTO(out_unlock, rc);
580 /* Ensure constant striping during chown(). See LU-2789. */
581 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
582 mutex_lock(&mo->mot_lov_mutex);
584 /* all attrs are packed into mti_attr in unpack_setattr */
585 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
587 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
588 mutex_unlock(&mo->mot_lov_mutex);
591 GOTO(out_unlock, rc);
592 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
595 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
600 * Check HSM flags and add HS_DIRTY flag if relevant.
602 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
603 * and is not RELEASED.
605 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
608 struct lu_ucred *uc = mdt_ucred(info);
613 /* If the file was modified, add the dirty flag */
614 ma->ma_need = MA_HSM;
615 rc = mdt_attr_get_complex(info, mo, ma);
617 CERROR("file attribute read error for "DFID": %d.\n",
618 PFID(mdt_object_fid(mo)), rc);
622 /* If an up2date copy exists in the backend, add dirty flag */
623 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
624 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
625 ma->ma_hsm.mh_flags |= HS_DIRTY;
627 /* Bump cap so that closes from non-owner writers can
628 * set the HSM state to dirty. */
629 cap_saved = uc->uc_cap;
630 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
631 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
632 uc->uc_cap = cap_saved;
634 CERROR("file attribute change error for "DFID": %d\n",
635 PFID(mdt_object_fid(mo)), rc);
641 static int mdt_reint_setattr(struct mdt_thread_info *info,
642 struct mdt_lock_handle *lhc)
644 struct mdt_device *mdt = info->mti_mdt;
645 struct md_attr *ma = &info->mti_attr;
646 struct mdt_reint_record *rr = &info->mti_rr;
647 struct ptlrpc_request *req = mdt_info_req(info);
648 struct mdt_object *mo;
649 struct mdt_body *repbody;
653 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
654 (unsigned int)ma->ma_attr.la_valid);
656 if (info->mti_dlm_req)
657 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
659 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
661 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
662 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
664 GOTO(out, rc = PTR_ERR(mo));
666 if (!mdt_object_exists(mo))
667 GOTO(out_put, rc = -ENOENT);
669 if (mdt_object_remote(mo))
670 GOTO(out_put, rc = -EREMOTE);
672 ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
673 /* revoke lease lock if size is going to be changed */
674 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
675 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
676 atomic_read(&mo->mot_lease_count) > 0)) {
677 down_read(&mo->mot_open_sem);
679 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
680 lhc = &info->mti_lh[MDT_LH_LOCAL];
681 mdt_lock_reg_init(lhc, LCK_CW);
683 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
685 up_read(&mo->mot_open_sem);
689 /* revoke lease lock */
690 mdt_object_unlock(info, mo, lhc, 1);
692 up_read(&mo->mot_open_sem);
695 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
696 /* Check write access for the O_TRUNC case */
697 if (mdt_write_read(mo) < 0)
698 GOTO(out_put, rc = -ETXTBSY);
700 /* LU-10286: compatibility check for FLR.
701 * Please check the comment in mdt_finish_open() for details */
702 if (!exp_connect_flr(info->mti_exp) ||
703 !exp_connect_overstriping(info->mti_exp)) {
704 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
705 if (rc < 0 && rc != -ENODATA)
708 if (!exp_connect_flr(info->mti_exp)) {
710 mdt_lmm_is_flr(info->mti_big_lmm))
711 GOTO(out_put, rc = -EOPNOTSUPP);
714 if (!exp_connect_overstriping(info->mti_exp)) {
716 mdt_lmm_is_overstriping(info->mti_big_lmm))
717 GOTO(out_put, rc = -EOPNOTSUPP);
721 /* For truncate, the file size sent from client
722 * is believable, but the blocks are incorrect,
723 * which makes the block size in LSOM attribute
724 * inconsisent with the real block size.
726 rc = mdt_lsom_update(info, mo, true);
731 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
732 if (ma->ma_valid & MA_LOV)
733 GOTO(out_put, rc = -EPROTO);
735 /* MDT supports FMD for regular files due to Data-on-MDT */
736 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
737 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
738 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
741 rc = mdt_attr_set(info, mo, ma);
744 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
745 (ma->ma_valid & MA_INODE)) {
746 struct lu_buf *buf = &info->mti_buf;
747 struct lu_ucred *uc = mdt_ucred(info);
748 struct mdt_lock_handle *lh;
750 __u64 lockpart = MDS_INODELOCK_XATTR;
752 /* reject if either remote or striped dir is disabled */
753 if (ma->ma_valid & MA_LMV) {
754 if (!mdt->mdt_enable_remote_dir ||
755 !mdt->mdt_enable_striped_dir)
756 GOTO(out_put, rc = -EPERM);
758 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
759 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
760 mdt->mdt_enable_remote_dir_gid != -1)
761 GOTO(out_put, rc = -EPERM);
764 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
765 GOTO(out_put, rc = -ENOTDIR);
767 if (ma->ma_attr.la_valid != 0)
768 GOTO(out_put, rc = -EPROTO);
770 if (ma->ma_valid & MA_LOV) {
771 buf->lb_buf = ma->ma_lmm;
772 buf->lb_len = ma->ma_lmm_size;
773 name = XATTR_NAME_LOV;
775 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
778 buf->lb_len = ma->ma_lmv_size;
779 name = XATTR_NAME_DEFAULT_LMV;
780 /* force client to update dir default layout */
781 lockpart |= MDS_INODELOCK_LOOKUP;
784 lh = &info->mti_lh[MDT_LH_PARENT];
785 mdt_lock_reg_init(lh, LCK_PW);
787 rc = mdt_object_lock(info, mo, lh, lockpart);
791 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
794 mdt_object_unlock(info, mo, lh, rc);
798 GOTO(out_put, rc = -EPROTO);
801 /* If file data is modified, add the dirty flag */
802 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
803 rc = mdt_add_dirty_flag(info, mo, ma);
805 ma->ma_need = MA_INODE;
807 rc = mdt_attr_get_complex(info, mo, ma);
811 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
815 mdt_object_put(info->mti_env, mo);
818 mdt_counter_incr(req, LPROC_MDT_SETATTR);
820 mdt_client_compatibility(info);
821 rc2 = mdt_fix_reply(info);
827 static int mdt_reint_create(struct mdt_thread_info *info,
828 struct mdt_lock_handle *lhc)
830 struct ptlrpc_request *req = mdt_info_req(info);
834 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
835 RETURN(err_serious(-ESTALE));
837 if (info->mti_dlm_req)
838 ldlm_request_cancel(mdt_info_req(info),
839 info->mti_dlm_req, 0, LATF_SKIP);
841 if (!lu_name_is_valid(&info->mti_rr.rr_name))
844 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
846 mdt_counter_incr(req, LPROC_MDT_MKDIR);
854 /* Special file should stay on the same node as parent. */
855 mdt_counter_incr(req, LPROC_MDT_MKNOD);
858 CERROR("%s: Unsupported mode %o\n",
859 mdt_obd_name(info->mti_mdt),
860 info->mti_attr.ma_attr.la_mode);
861 RETURN(err_serious(-EOPNOTSUPP));
864 rc = mdt_create(info);
869 * VBR: save parent version in reply and child version getting by its name.
870 * Version of child is getting and checking during its lookup. If
872 static int mdt_reint_unlink(struct mdt_thread_info *info,
873 struct mdt_lock_handle *lhc)
875 struct mdt_reint_record *rr = &info->mti_rr;
876 struct ptlrpc_request *req = mdt_info_req(info);
877 struct md_attr *ma = &info->mti_attr;
878 struct lu_fid *child_fid = &info->mti_tmp_fid1;
879 struct mdt_object *mp;
880 struct mdt_object *mc;
881 struct mdt_lock_handle *parent_lh;
882 struct mdt_lock_handle *child_lh;
883 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
885 bool cos_incompat = false;
891 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
892 PNAME(&rr->rr_name));
894 if (info->mti_dlm_req)
895 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
897 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
898 RETURN(err_serious(-ENOENT));
900 if (!fid_is_md_operative(rr->rr_fid1))
903 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
907 if (mdt_object_remote(mp)) {
910 rc = mdt_version_get_check_save(info, mp, 0);
912 GOTO(put_parent, rc);
916 parent_lh = &info->mti_lh[MDT_LH_PARENT];
917 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
918 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
921 GOTO(put_parent, rc);
923 /* lookup child object along with version checking */
925 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
927 /* Name might not be able to find during resend of
928 * remote unlink, considering following case.
929 * dir_A is a remote directory, the name entry of
930 * dir_A is on MDT0, the directory is on MDT1,
932 * 1. client sends unlink req to MDT1.
933 * 2. MDT1 sends name delete update to MDT0.
934 * 3. name entry is being deleted in MDT0 synchronously.
935 * 4. MDT1 is restarted.
936 * 5. client resends unlink req to MDT1. So it can not
937 * find the name entry on MDT0 anymore.
938 * In this case, MDT1 only needs to destory the local
941 if (mdt_object_remote(mp) && rc == -ENOENT &&
942 !fid_is_zero(rr->rr_fid2) &&
943 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
945 *child_fid = *rr->rr_fid2;
947 GOTO(unlock_parent, rc);
951 if (!fid_is_md_operative(child_fid))
952 GOTO(unlock_parent, rc = -EPERM);
954 /* We will lock the child regardless it is local or remote. No harm. */
955 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
957 GOTO(unlock_parent, rc = PTR_ERR(mc));
960 rc = mdt_object_striped(info, mc);
966 mdt_object_put(info->mti_env, mc);
967 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
972 child_lh = &info->mti_lh[MDT_LH_CHILD];
973 mdt_lock_reg_init(child_lh, LCK_EX);
974 if (info->mti_spec.sp_rm_entry) {
975 struct lu_ucred *uc = mdt_ucred(info);
977 if (!mdt_is_dne_client(req->rq_export))
978 /* Return -ENOTSUPP for old client */
979 GOTO(put_child, rc = -ENOTSUPP);
981 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
982 GOTO(put_child, rc = -EPERM);
984 ma->ma_need = MA_INODE;
986 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
987 NULL, &rr->rr_name, ma, no_name);
991 if (mdt_object_remote(mc)) {
992 struct mdt_body *repbody;
994 if (!fid_is_zero(rr->rr_fid2)) {
995 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
996 mdt_obd_name(info->mti_mdt),
997 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
998 GOTO(put_child, rc = -ENOENT);
1000 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1001 mdt_obd_name(info->mti_mdt),
1002 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1004 if (!mdt_is_dne_client(req->rq_export))
1005 /* Return -ENOTSUPP for old client */
1006 GOTO(put_child, rc = -ENOTSUPP);
1008 /* Revoke the LOOKUP lock of the remote object granted by
1009 * this MDT. Since the unlink will happen on another MDT,
1010 * it will release the LOOKUP lock right away. Then What
1011 * would happen if another client try to grab the LOOKUP
1012 * lock at the same time with unlink XXX */
1013 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1014 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1015 LASSERT(repbody != NULL);
1016 repbody->mbo_fid1 = *mdt_object_fid(mc);
1017 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1018 GOTO(unlock_child, rc = -EREMOTE);
1020 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1021 * this now because a running HSM restore on the child (unlink
1022 * victim) will hold the layout lock. See LU-4002. */
1023 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1024 if (mdt_object_remote(mp)) {
1025 /* Enqueue lookup lock from parent MDT */
1026 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1027 &child_lh->mlh_rreg_lh,
1028 child_lh->mlh_rreg_mode,
1029 MDS_INODELOCK_LOOKUP, false);
1031 GOTO(put_child, rc);
1033 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1036 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1039 GOTO(put_child, rc);
1042 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1043 * whether need MA_LOV and MA_COOKIE.
1045 ma->ma_need = MA_INODE;
1048 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1049 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1050 /* save version when object is locked */
1051 mdt_version_get_save(info, mc, 1);
1053 mutex_lock(&mc->mot_lov_mutex);
1055 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1056 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1058 mutex_unlock(&mc->mot_lov_mutex);
1060 GOTO(unlock_child, rc);
1062 if (!lu_object_is_dying(&mc->mot_header)) {
1063 rc = mdt_attr_get_complex(info, mc, ma);
1066 } else if (mdt_dom_check_for_discard(info, mc)) {
1067 mdt_dom_discard_data(info, mc);
1069 mdt_handle_last_unlink(info, mc, ma);
1072 if (ma->ma_valid & MA_INODE) {
1073 switch (ma->ma_attr.la_mode & S_IFMT) {
1075 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1083 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1086 LASSERTF(0, "bad file type %o unlinking\n",
1087 ma->ma_attr.la_mode);
1094 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1096 mdt_object_put(info->mti_env, mc);
1098 mdt_object_unlock(info, mp, parent_lh, rc);
1100 mdt_object_put(info->mti_env, mp);
1101 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1106 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1109 static int mdt_reint_link(struct mdt_thread_info *info,
1110 struct mdt_lock_handle *lhc)
1112 struct mdt_reint_record *rr = &info->mti_rr;
1113 struct ptlrpc_request *req = mdt_info_req(info);
1114 struct md_attr *ma = &info->mti_attr;
1115 struct mdt_object *ms;
1116 struct mdt_object *mp;
1117 struct mdt_lock_handle *lhs;
1118 struct mdt_lock_handle *lhp;
1123 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1124 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1126 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1127 RETURN(err_serious(-ENOENT));
1129 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1130 req->rq_no_reply = 1;
1131 RETURN(err_serious(-ENOENT));
1134 if (info->mti_dlm_req)
1135 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1137 /* Invalid case so return error immediately instead of
1139 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1142 if (!fid_is_md_operative(rr->rr_fid1) ||
1143 !fid_is_md_operative(rr->rr_fid2))
1146 /* step 1: find target parent dir */
1147 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1149 RETURN(PTR_ERR(mp));
1151 rc = mdt_version_get_check_save(info, mp, 0);
1153 GOTO(put_parent, rc);
1155 /* step 2: find source */
1156 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1158 GOTO(put_parent, rc = PTR_ERR(ms));
1160 if (!mdt_object_exists(ms)) {
1161 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1162 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1163 GOTO(put_source, rc = -ENOENT);
1166 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1168 lhp = &info->mti_lh[MDT_LH_PARENT];
1169 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1170 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1173 GOTO(put_source, rc);
1175 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1177 lhs = &info->mti_lh[MDT_LH_CHILD];
1178 mdt_lock_reg_init(lhs, LCK_EX);
1179 rc = mdt_reint_object_lock(info, ms, lhs,
1180 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1183 GOTO(unlock_parent, rc);
1185 /* step 3: link it */
1186 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1187 OBD_FAIL_MDS_REINT_LINK_WRITE);
1189 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1190 rc = mdt_version_get_check_save(info, ms, 1);
1192 GOTO(unlock_source, rc);
1194 /** check target version by name during replay */
1195 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1196 &info->mti_tmp_fid1, 2);
1197 if (rc != 0 && rc != -ENOENT)
1198 GOTO(unlock_source, rc);
1199 /* save version of file name for replay, it must be ENOENT here */
1200 if (!req_is_replay(mdt_info_req(info))) {
1201 if (rc != -ENOENT) {
1202 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1203 PNAME(&rr->rr_name));
1204 GOTO(unlock_source, rc = -EEXIST);
1206 info->mti_ver[2] = ENOENT_VERSION;
1207 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1210 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1211 mdt_object_child(ms), &rr->rr_name, ma);
1214 mdt_counter_incr(req, LPROC_MDT_LINK);
1218 mdt_object_unlock(info, ms, lhs, rc);
1220 mdt_object_unlock(info, mp, lhp, rc);
1222 mdt_object_put(info->mti_env, ms);
1224 mdt_object_put(info->mti_env, mp);
1228 * lock the part of the directory according to the hash of the name
1229 * (lh->mlh_pdo_hash) in parallel directory lock.
1231 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1232 struct mdt_lock_handle *lh,
1233 struct mdt_object *obj, __u64 ibits,
1236 struct ldlm_res_id *res = &info->mti_res_id;
1237 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1238 union ldlm_policy_data *policy = &info->mti_policy;
1239 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1243 * Finish res_id initializing by name hash marking part of
1244 * directory which is taking modification.
1246 LASSERT(lh->mlh_pdo_hash != 0);
1247 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1248 memset(policy, 0, sizeof(*policy));
1249 policy->l_inodebits.bits = ibits;
1251 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1252 dlmflags |= LDLM_FL_COS_INCOMPAT;
1254 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1255 * going to be sent to client. If it is - mdt_intent_policy() path will
1256 * fix it up and turn FL_LOCAL flag off.
1258 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1259 policy, res, dlmflags,
1260 &info->mti_exp->exp_handle.h_cookie);
1265 * Get BFL lock for rename or migrate process.
1267 static int mdt_rename_lock(struct mdt_thread_info *info,
1268 struct lustre_handle *lh)
1273 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1274 struct lu_fid *fid = &info->mti_tmp_fid1;
1275 struct mdt_object *obj;
1277 /* XXX, right now, it has to use object API to
1278 * enqueue lock cross MDT, so it will enqueue
1279 * rename lock(with LUSTRE_BFL_FID) by root object */
1281 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1283 RETURN(PTR_ERR(obj));
1285 rc = mdt_remote_object_lock(info, obj,
1286 &LUSTRE_BFL_FID, lh,
1288 MDS_INODELOCK_UPDATE, false);
1289 mdt_object_put(info->mti_env, obj);
1291 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1292 union ldlm_policy_data *policy = &info->mti_policy;
1293 struct ldlm_res_id *res_id = &info->mti_res_id;
1296 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1297 memset(policy, 0, sizeof *policy);
1298 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1299 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1300 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1301 LDLM_IBITS, policy, LCK_EX, &flags,
1303 ldlm_completion_ast, NULL, NULL, 0,
1305 &info->mti_exp->exp_handle.h_cookie,
1312 static void mdt_rename_unlock(struct lustre_handle *lh)
1315 LASSERT(lustre_handle_is_used(lh));
1316 /* Cancel the single rename lock right away */
1317 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1321 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1322 const struct lu_fid *fid,
1325 struct mdt_object *dir;
1330 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1334 /* check early, the real version will be saved after locking */
1335 rc = mdt_version_get_check(info, dir, idx);
1339 if (!mdt_object_exists(dir))
1340 GOTO(out_put, rc = -ENOENT);
1342 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1343 GOTO(out_put, rc = -ENOTDIR);
1347 mdt_object_put(info->mti_env, dir);
1352 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1353 * herein we don't really check it, just do revoke.
1355 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1356 struct mdt_object *pobj,
1357 struct mdt_object *obj)
1359 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1362 mdt_lock_handle_init(lh);
1363 mdt_lock_reg_init(lh, LCK_EX);
1365 if (mdt_object_remote(pobj)) {
1366 /* don't bother to check if pobj and obj are on the same MDT. */
1367 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1368 &lh->mlh_rreg_lh, LCK_EX,
1369 MDS_INODELOCK_LOOKUP, false);
1370 } else if (mdt_object_remote(obj)) {
1371 struct ldlm_res_id *res = &info->mti_res_id;
1372 union ldlm_policy_data *policy = &info->mti_policy;
1373 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1374 LDLM_FL_COS_INCOMPAT;
1376 fid_build_reg_res_name(mdt_object_fid(obj), res);
1377 memset(policy, 0, sizeof(*policy));
1378 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1379 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1380 &lh->mlh_reg_lh, LCK_EX, policy, res,
1383 /* do nothing if both are local */
1391 * TODO, currently we don't save this lock because there is no place to
1392 * hold this lock handle, but to avoid race we need to save this lock.
1394 mdt_object_unlock(info, NULL, lh, 1);
1400 * operation may takes locks of linkea, or directory stripes, group them in
1403 struct mdt_sub_lock {
1404 struct mdt_object *msl_obj;
1405 struct mdt_lock_handle msl_lh;
1406 struct list_head msl_linkage;
1409 static void mdt_unlock_list(struct mdt_thread_info *info,
1410 struct list_head *list, int decref)
1412 struct mdt_sub_lock *msl;
1413 struct mdt_sub_lock *tmp;
1415 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1416 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1417 list_del(&msl->msl_linkage);
1422 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1423 struct mdt_object *obj,
1424 struct mdt_lock_handle *lh,
1425 struct ldlm_enqueue_info *einfo,
1426 struct list_head *slave_locks,
1429 if (mdt_object_remote(obj)) {
1430 mdt_unlock_list(info, slave_locks, decref);
1431 mdt_object_unlock(info, obj, lh, decref);
1433 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1438 * lock parents of links, and also check whether total locks don't exceed
1441 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1442 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1443 * \retval -ev negative errno upon error
1445 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1446 struct mdt_object *pobj,
1447 const struct md_attr *ma,
1448 struct mdt_object *obj,
1449 struct mdt_lock_handle *lhp,
1450 struct ldlm_enqueue_info *peinfo,
1451 struct list_head *parent_slave_locks,
1452 struct list_head *link_locks)
1454 struct mdt_device *mdt = info->mti_mdt;
1455 struct lu_buf *buf = &info->mti_big_buf;
1456 struct lu_name *lname = &info->mti_name;
1457 struct linkea_data ldata = { NULL };
1458 bool blocked = false;
1459 int local_lnkp_cnt = 0;
1464 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1467 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1468 if (buf->lb_buf == NULL)
1472 rc = mdt_links_read(info, obj, &ldata);
1474 if (rc == -ENOENT || rc == -ENODATA)
1479 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1480 linkea_next_entry(&ldata)) {
1481 struct mdt_object *lnkp;
1482 struct mdt_sub_lock *msl;
1486 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1489 /* check if it's also linked to parent */
1490 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1491 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1492 PFID(&fid), PNAME(lname));
1493 /* in case link is remote object, revoke LOOKUP lock */
1494 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1500 /* check if it's linked to a stripe of parent */
1501 if (ma->ma_valid & MA_LMV) {
1502 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1503 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1506 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1507 fid_le_to_cpu(stripe_fid,
1508 &lmv->lmv_stripe_fids[j]);
1509 if (lu_fid_eq(stripe_fid, &fid)) {
1510 CDEBUG(D_INFO, "skip stripe "DFID
1511 ", reovke "DNAME"\n",
1512 PFID(&fid), PNAME(lname));
1513 lnkp = mdt_object_find(info->mti_env,
1516 GOTO(out, rc = PTR_ERR(lnkp));
1522 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1524 mdt_object_put(info->mti_env, lnkp);
1529 /* Check if it's already locked */
1530 list_for_each_entry(msl, link_locks, msl_linkage) {
1531 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1533 DFID" was locked, revoke "DNAME"\n",
1534 PFID(&fid), PNAME(lname));
1535 lnkp = msl->msl_obj;
1541 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1545 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1546 PFID(&fid), PNAME(lname));
1548 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1550 CWARN("%s: cannot find obj "DFID": %ld\n",
1551 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1555 if (!mdt_object_exists(lnkp)) {
1556 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1557 PFID(&fid), PNAME(lname));
1558 mdt_object_put(info->mti_env, lnkp);
1562 if (!mdt_object_remote(lnkp))
1567 GOTO(out, rc = -ENOMEM);
1570 * we can't follow parent-child lock order like other MD
1571 * operations, use lock_try here to avoid deadlock, if the lock
1572 * cannot be taken, drop all locks taken, revoke the blocked
1573 * one, and continue processing the remaining entries, and in
1574 * the end of the loop restart from beginning.
1576 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1578 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1579 MDS_INODELOCK_UPDATE, true);
1580 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1582 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1583 PFID(&fid), PNAME(lname));
1585 mdt_unlock_list(info, link_locks, 1);
1586 /* also unlock parent locks to avoid deadlock */
1588 mdt_migrate_object_unlock(info, pobj, lhp,
1595 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1596 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1597 MDS_INODELOCK_UPDATE);
1599 mdt_object_put(info->mti_env, lnkp);
1604 if (mdt_object_remote(lnkp)) {
1605 struct ldlm_lock *lock;
1608 * for remote object, set lock cb_atomic,
1609 * so lock can be released in blocking_ast()
1610 * immediately, then the next lock_try will
1611 * have better chance of success.
1613 lock = ldlm_handle2lock(
1614 &msl->msl_lh.mlh_rreg_lh);
1615 LASSERT(lock != NULL);
1616 lock_res_and_lock(lock);
1617 ldlm_set_atomic_cb(lock);
1618 unlock_res_and_lock(lock);
1619 LDLM_LOCK_PUT(lock);
1622 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1627 INIT_LIST_HEAD(&msl->msl_linkage);
1628 msl->msl_obj = lnkp;
1629 list_add_tail(&msl->msl_linkage, link_locks);
1631 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1635 GOTO(out, rc = -EBUSY);
1640 mdt_unlock_list(info, link_locks, rc);
1641 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1643 * parent may have 3 local objects: master object and 2 stripes
1644 * (if it's being migrated too); source may have 2 local
1645 * objects: master and 1 stripe; target has 1 local object.
1651 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1652 struct mdt_object *obj,
1653 const struct md_attr *ma,
1654 struct list_head *slave_locks)
1656 struct mdt_device *mdt = info->mti_mdt;
1657 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1658 struct lu_fid *fid = &info->mti_tmp_fid1;
1659 struct mdt_object *slave;
1660 struct mdt_sub_lock *msl;
1666 LASSERT(mdt_object_remote(obj));
1667 LASSERT(ma->ma_valid & MA_LMV);
1670 if (!lmv_is_sane(lmv))
1673 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1674 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1676 if (!fid_is_sane(fid))
1679 slave = mdt_object_find(info->mti_env, mdt, fid);
1681 GOTO(out, rc = PTR_ERR(slave));
1685 mdt_object_put(info->mti_env, slave);
1686 GOTO(out, rc = -ENOMEM);
1689 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1690 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1691 MDS_INODELOCK_UPDATE, true);
1694 mdt_object_put(info->mti_env, slave);
1698 INIT_LIST_HEAD(&msl->msl_linkage);
1699 msl->msl_obj = slave;
1700 list_add_tail(&msl->msl_linkage, slave_locks);
1706 mdt_unlock_list(info, slave_locks, rc);
1710 /* lock parent and its stripes */
1711 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1712 struct mdt_object *obj,
1713 const struct md_attr *ma,
1714 struct mdt_lock_handle *lh,
1715 struct ldlm_enqueue_info *einfo,
1716 struct list_head *slave_locks)
1720 if (mdt_object_remote(obj)) {
1721 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1722 &lh->mlh_rreg_lh, LCK_PW,
1723 MDS_INODELOCK_UPDATE, false);
1728 * if obj is remote and striped, lock its stripes explicitly
1729 * because it's not striped in LOD layer on this MDT.
1731 if (ma->ma_valid & MA_LMV) {
1732 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1734 mdt_object_unlock(info, obj, lh, rc);
1737 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1745 * in migration, object may be remote, and we need take full lock of it and its
1746 * stripes if it's directory, besides, object may be a remote object on its
1747 * parent, revoke its LOOKUP lock on where its parent is located.
1749 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1750 struct mdt_object *pobj,
1751 struct mdt_object *obj,
1752 struct mdt_lock_handle *lh,
1753 struct ldlm_enqueue_info *einfo,
1754 struct list_head *slave_locks)
1758 if (mdt_object_remote(obj)) {
1759 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1763 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1764 &lh->mlh_rreg_lh, LCK_EX,
1765 MDS_INODELOCK_FULL, false);
1770 * if obj is remote and striped, lock its stripes explicitly
1771 * because it's not striped in LOD layer on this MDT.
1773 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1774 struct md_attr *ma = &info->mti_attr;
1776 ma->ma_lmv = info->mti_big_lmm;
1777 ma->ma_lmv_size = info->mti_big_lmmsize;
1779 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1781 mdt_object_unlock(info, obj, lh, rc);
1785 if (ma->ma_valid & MA_LMV) {
1786 rc = mdt_lock_remote_slaves(info, obj, ma,
1789 mdt_object_unlock(info, obj, lh, rc);
1793 if (mdt_object_remote(pobj)) {
1794 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1799 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1807 * lookup source by name, if parent is striped directory, we need to find the
1808 * corresponding stripe where source is located, and then lookup there.
1810 * besides, if parent is migrating too, and file is already in target stripe,
1811 * this should be a redo of 'lfs migrate' on client side.
1813 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1814 struct mdt_object *pobj,
1815 const struct md_attr *ma,
1816 const struct lu_name *lname,
1817 struct mdt_object **spobj,
1818 struct mdt_object **sobj)
1820 const struct lu_env *env = info->mti_env;
1821 struct lu_fid *fid = &info->mti_tmp_fid1;
1822 struct mdt_object *stripe;
1825 if (ma->ma_valid & MA_LMV) {
1826 /* if parent is striped, lookup on corresponding stripe */
1827 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1828 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1829 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1830 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1831 LMV_HASH_FLAG_MIGRATION;
1834 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1835 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1838 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1844 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1845 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1847 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1849 stripe = mdt_object_find(env, info->mti_mdt, fid);
1851 return PTR_ERR(stripe);
1854 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1856 if (rc == -ENOENT && is_migrating) {
1858 * if parent is migrating, and lookup child failed on
1859 * source stripe, lookup again on target stripe, if it
1860 * exists, it means previous migration was interrupted,
1861 * and current file was migrated already.
1863 mdt_object_put(env, stripe);
1865 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1866 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1868 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1874 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1876 stripe = mdt_object_find(env, info->mti_mdt, fid);
1878 return PTR_ERR(stripe);
1881 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1882 fid, &info->mti_spec);
1883 mdt_object_put(env, stripe);
1884 return rc ?: -EALREADY;
1886 mdt_object_put(env, stripe);
1891 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1897 mdt_object_get(env, stripe);
1902 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1903 if (IS_ERR(*sobj)) {
1904 mdt_object_put(env, stripe);
1905 rc = PTR_ERR(*sobj);
1913 /* end lease and close file for regular file */
1914 static int mdd_migrate_close(struct mdt_thread_info *info,
1915 struct mdt_object *obj)
1917 struct close_data *data;
1918 struct mdt_body *repbody;
1919 struct ldlm_lock *lease;
1924 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1926 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1930 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1935 lease = ldlm_handle2lock(&data->cd_handle);
1939 /* check if the lease was already canceled */
1940 lock_res_and_lock(lease);
1941 rc = ldlm_is_cancel(lease);
1942 unlock_res_and_lock(lease);
1946 LDLM_DEBUG(lease, DFID" lease broken",
1947 PFID(mdt_object_fid(obj)));
1951 * cancel server side lease, client side counterpart should have been
1952 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1954 ldlm_lock_cancel(lease);
1955 ldlm_reprocess_all(lease->l_resource, lease);
1956 LDLM_LOCK_PUT(lease);
1959 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1960 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1961 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1967 * migrate file in below steps:
1968 * 1. lock parent and its stripes
1969 * 2. lookup source by name
1970 * 3. lock parents of source links if source is not directory
1971 * 4. reject if source is in HSM
1972 * 5. take source open_sem and close file if source is regular file
1973 * 6. lock source and its stripes if it's directory
1974 * 7. lock target so subsequent change to it can trigger COS
1976 * 9. unlock above locks
1977 * 10. sync device if source has links
1979 static int mdt_reint_migrate(struct mdt_thread_info *info,
1980 struct mdt_lock_handle *unused)
1982 const struct lu_env *env = info->mti_env;
1983 struct mdt_device *mdt = info->mti_mdt;
1984 struct ptlrpc_request *req = mdt_info_req(info);
1985 struct mdt_reint_record *rr = &info->mti_rr;
1986 struct lu_ucred *uc = mdt_ucred(info);
1987 struct md_attr *ma = &info->mti_attr;
1988 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1989 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1990 struct mdt_object *pobj;
1991 struct mdt_object *spobj = NULL;
1992 struct mdt_object *sobj = NULL;
1993 struct mdt_object *tobj;
1994 struct lustre_handle rename_lh = { 0 };
1995 struct mdt_lock_handle *lhp;
1996 struct mdt_lock_handle *lhs;
1997 struct mdt_lock_handle *lht;
1998 LIST_HEAD(parent_slave_locks);
1999 LIST_HEAD(child_slave_locks);
2000 LIST_HEAD(link_locks);
2001 int lock_retries = 5;
2002 bool open_sem_locked = false;
2003 bool do_sync = false;
2007 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2008 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2010 if (info->mti_dlm_req)
2011 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2013 if (!fid_is_md_operative(rr->rr_fid1) ||
2014 !fid_is_md_operative(rr->rr_fid2))
2017 /* don't allow migrate . or .. */
2018 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2021 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2024 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2025 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2026 mdt->mdt_enable_remote_dir_gid != -1)
2030 * Note: do not enqueue rename lock for replay request, because
2031 * if other MDT holds rename lock, but being blocked to wait for
2032 * this MDT to finish its recovery, and the failover MDT can not
2033 * get rename lock, which will cause deadlock.
2035 if (!req_is_replay(req)) {
2036 rc = mdt_rename_lock(info, &rename_lh);
2038 CERROR("%s: can't lock FS for rename: rc = %d\n",
2039 mdt_obd_name(info->mti_mdt), rc);
2044 /* pobj is master object of parent */
2045 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2047 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2049 if (unlikely(!info->mti_big_lmm)) {
2050 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2051 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2052 if (!info->mti_big_lmm)
2053 GOTO(put_parent, rc = -ENOMEM);
2056 ma->ma_lmv = info->mti_big_lmm;
2057 ma->ma_lmv_size = info->mti_big_lmmsize;
2059 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2061 GOTO(put_parent, rc);
2064 /* lock parent object */
2065 lhp = &info->mti_lh[MDT_LH_PARENT];
2066 mdt_lock_reg_init(lhp, LCK_PW);
2067 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2068 &parent_slave_locks);
2070 GOTO(put_parent, rc);
2073 * spobj is the corresponding stripe against name if pobj is striped
2074 * directory, which is the real parent, and no need to lock, because
2075 * we've taken full lock of pobj.
2077 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2079 GOTO(unlock_parent, rc);
2081 /* lock parents of source links, and revoke LOOKUP lock of links */
2082 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2083 &parent_slave_locks, &link_locks);
2084 if (rc == -EBUSY && lock_retries-- > 0) {
2085 mdt_object_put(env, sobj);
2086 mdt_object_put(env, spobj);
2091 GOTO(put_source, rc);
2094 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2095 * with one request, if total lock count exceeds this limit, we will
2096 * drop all locks after migration, and synchronous device in the end.
2100 /* TODO: DoM migration is not supported yet */
2101 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2102 ma->ma_lmm = info->mti_big_lmm;
2103 ma->ma_lmm_size = info->mti_big_lmmsize;
2105 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2107 GOTO(put_source, rc);
2109 if (ma->ma_valid & MA_LOV &&
2110 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2111 GOTO(put_source, rc = -EOPNOTSUPP);
2114 /* if migration HSM is allowed */
2115 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2116 ma->ma_need = MA_HSM;
2118 rc = mdt_attr_get_complex(info, sobj, ma);
2120 GOTO(unlock_links, rc);
2122 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2123 GOTO(unlock_links, rc = -EOPNOTSUPP);
2126 /* end lease and close file for regular file */
2127 if (info->mti_spec.sp_migrate_close) {
2128 /* try to hold open_sem so that nobody else can open the file */
2129 if (!down_write_trylock(&sobj->mot_open_sem)) {
2131 mdd_migrate_close(info, sobj);
2132 GOTO(unlock_links, rc = -EBUSY);
2134 open_sem_locked = true;
2135 rc = mdd_migrate_close(info, sobj);
2137 GOTO(unlock_open_sem, rc);
2142 lhs = &info->mti_lh[MDT_LH_OLD];
2143 mdt_lock_reg_init(lhs, LCK_EX);
2144 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2145 &child_slave_locks);
2147 GOTO(unlock_open_sem, rc);
2150 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2152 GOTO(unlock_source, rc = PTR_ERR(tobj));
2154 lht = &info->mti_lh[MDT_LH_NEW];
2155 mdt_lock_reg_init(lht, LCK_EX);
2156 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2158 GOTO(put_target, rc);
2160 /* Don't do lookup sanity check. We know name doesn't exist. */
2161 info->mti_spec.sp_cr_lookup = 0;
2162 info->mti_spec.sp_feat = &dt_directory_features;
2164 rc = mdo_migrate(env, mdt_object_child(pobj),
2165 mdt_object_child(sobj), &rr->rr_name,
2166 mdt_object_child(tobj), &info->mti_spec, ma);
2169 mdt_object_unlock(info, tobj, lht, rc);
2171 mdt_object_put(env, tobj);
2173 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2174 &child_slave_locks, rc);
2176 if (open_sem_locked)
2177 up_write(&sobj->mot_open_sem);
2179 /* if we've got too many locks to save into RPC,
2180 * then just commit before the locks are released */
2182 mdt_device_sync(env, mdt);
2183 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2185 mdt_object_put(env, sobj);
2186 mdt_object_put(env, spobj);
2188 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2189 &parent_slave_locks, rc);
2191 mdt_object_put(env, pobj);
2193 if (lustre_handle_is_used(&rename_lh))
2194 mdt_rename_unlock(&rename_lh);
2199 static int mdt_object_lock_save(struct mdt_thread_info *info,
2200 struct mdt_object *dir,
2201 struct mdt_lock_handle *lh,
2202 int idx, bool cos_incompat)
2206 /* we lock the target dir if it is local */
2207 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2212 /* get and save correct version after locking */
2213 mdt_version_get_save(info, dir, idx);
2218 * determine lock order of sobj and tobj
2220 * there are two situations we need to lock tobj before sobj:
2221 * 1. sobj is child of tobj
2222 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2223 * larger than that of tobj
2225 * \retval 1 lock tobj before sobj
2226 * \retval 0 lock sobj before tobj
2227 * \retval -ev negative errno upon error
2229 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2230 struct mdt_object *sobj,
2231 struct mdt_object *tobj)
2233 struct md_attr *ma = &info->mti_attr;
2234 struct lu_fid *spfid = &info->mti_tmp_fid1;
2235 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2236 struct lmv_mds_md_v1 *lmv;
2241 /* sobj and tobj are the same */
2245 if (fid_is_root(mdt_object_fid(sobj)))
2248 if (fid_is_root(mdt_object_fid(tobj)))
2251 /* check whether sobj is child of tobj */
2252 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2253 mdt_object_fid(tobj));
2260 /* check whether sobj and tobj are children of the same parent */
2261 rc = mdt_attr_get_pfid(info, sobj, spfid);
2265 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2269 if (!lu_fid_eq(spfid, tpfid))
2272 /* check whether sobj and tobj are sibling stripes */
2273 ma->ma_need = MA_LMV;
2275 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2276 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2277 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2281 if (!(ma->ma_valid & MA_LMV))
2284 lmv = &ma->ma_lmv->lmv_md_v1;
2285 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2287 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2290 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2294 if (!(ma->ma_valid & MA_LMV))
2297 lmv = &ma->ma_lmv->lmv_md_v1;
2298 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2300 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2302 /* check stripe index of sobj and tobj */
2303 if (sindex == tindex)
2306 return sindex < tindex ? 0 : 1;
2310 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2311 * 2 - srcdir child; 3 - tgtdir child.
2312 * Update on disk version of srcdir child.
2314 static int mdt_reint_rename(struct mdt_thread_info *info,
2315 struct mdt_lock_handle *unused)
2317 struct mdt_device *mdt = info->mti_mdt;
2318 struct mdt_reint_record *rr = &info->mti_rr;
2319 struct md_attr *ma = &info->mti_attr;
2320 struct ptlrpc_request *req = mdt_info_req(info);
2321 struct mdt_object *msrcdir = NULL;
2322 struct mdt_object *mtgtdir = NULL;
2323 struct mdt_object *mold;
2324 struct mdt_object *mnew = NULL;
2325 struct lustre_handle rename_lh = { 0 };
2326 struct mdt_lock_handle *lh_srcdirp;
2327 struct mdt_lock_handle *lh_tgtdirp;
2328 struct mdt_lock_handle *lh_oldp = NULL;
2329 struct mdt_lock_handle *lh_newp = NULL;
2330 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2331 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2333 bool reverse = false, discard = false;
2338 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2339 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2340 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2342 if (info->mti_dlm_req)
2343 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2345 if (!fid_is_md_operative(rr->rr_fid1) ||
2346 !fid_is_md_operative(rr->rr_fid2))
2349 /* find both parents. */
2350 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2351 if (IS_ERR(msrcdir))
2352 RETURN(PTR_ERR(msrcdir));
2354 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2356 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2358 mdt_object_get(info->mti_env, mtgtdir);
2360 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2361 if (IS_ERR(mtgtdir))
2362 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2366 * Note: do not enqueue rename lock for replay request, because
2367 * if other MDT holds rename lock, but being blocked to wait for
2368 * this MDT to finish its recovery, and the failover MDT can not
2369 * get rename lock, which will cause deadlock.
2371 if (!req_is_replay(req)) {
2373 * Normally rename RPC is handled on the MDT with the target
2374 * directory (if target exists, it's on the MDT with the
2375 * target), if the source directory is remote, it's a hint that
2376 * source is remote too (this may not be true, but it won't
2377 * cause any issue), return -EXDEV early to avoid taking
2380 if (!mdt->mdt_enable_remote_rename &&
2381 mdt_object_remote(msrcdir))
2382 GOTO(out_put_tgtdir, rc = -EXDEV);
2384 rc = mdt_rename_lock(info, &rename_lh);
2386 CERROR("%s: can't lock FS for rename: rc = %d\n",
2387 mdt_obd_name(mdt), rc);
2388 GOTO(out_put_tgtdir, rc);
2392 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2394 GOTO(out_unlock_rename, rc);
2398 /* source needs to be looked up after locking source parent, otherwise
2399 * this rename may race with unlink source, and cause rename hang, see
2400 * sanityn.sh 55b, so check parents first, if later we found source is
2401 * remote, relock parents. */
2402 cos_incompat = (mdt_object_remote(msrcdir) ||
2403 mdt_object_remote(mtgtdir));
2405 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2407 /* lock parents in the proper order. */
2408 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2409 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2412 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2413 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2416 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2419 GOTO(out_unlock_rename, rc);
2421 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2423 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2426 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2427 GOTO(out_unlock_rename, rc);
2430 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2433 GOTO(out_unlock_rename, rc);
2435 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2437 if (mtgtdir != msrcdir) {
2438 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2440 } else if (!mdt_object_remote(mtgtdir) &&
2441 lh_srcdirp->mlh_pdo_hash !=
2442 lh_tgtdirp->mlh_pdo_hash) {
2443 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2444 MDS_INODELOCK_UPDATE,
2446 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2449 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2450 GOTO(out_unlock_rename, rc);
2454 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2455 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2457 /* find mold object. */
2459 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2461 GOTO(out_unlock_parents, rc);
2463 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2464 GOTO(out_unlock_parents, rc = -EINVAL);
2466 if (!fid_is_md_operative(old_fid))
2467 GOTO(out_unlock_parents, rc = -EPERM);
2469 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2471 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2473 if (!mdt_object_exists(mold)) {
2474 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2476 "object does not exist");
2477 GOTO(out_put_old, rc = -ENOENT);
2480 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2481 GOTO(out_put_old, rc = -EXDEV);
2483 /* Check if @mtgtdir is subdir of @mold, before locking child
2484 * to avoid reverse locking. */
2485 if (mtgtdir != msrcdir) {
2486 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2491 GOTO(out_put_old, rc);
2495 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2496 /* save version after locking */
2497 mdt_version_get_save(info, mold, 2);
2499 if (!cos_incompat && mdt_object_remote(mold)) {
2500 cos_incompat = true;
2501 mdt_object_put(info->mti_env, mold);
2502 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2503 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2507 /* find mnew object:
2508 * mnew target object may not exist now
2509 * lookup with version checking */
2511 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2514 /* the new_fid should have been filled at this moment */
2515 if (lu_fid_eq(old_fid, new_fid))
2516 GOTO(out_put_old, rc);
2518 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2519 lu_fid_eq(new_fid, rr->rr_fid2))
2520 GOTO(out_put_old, rc = -EINVAL);
2522 if (!fid_is_md_operative(new_fid))
2523 GOTO(out_put_old, rc = -EPERM);
2525 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2527 GOTO(out_put_old, rc = PTR_ERR(mnew));
2529 if (!mdt_object_exists(mnew)) {
2530 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2532 "object does not exist");
2533 GOTO(out_put_new, rc = -ENOENT);
2536 if (mdt_object_remote(mnew)) {
2537 struct mdt_body *repbody;
2539 /* Always send rename req to the target child MDT */
2540 repbody = req_capsule_server_get(info->mti_pill,
2542 LASSERT(repbody != NULL);
2543 repbody->mbo_fid1 = *new_fid;
2544 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2545 GOTO(out_put_new, rc = -EXDEV);
2547 /* Before locking the target dir, check we do not replace
2548 * a dir with a non-dir, otherwise it may deadlock with
2549 * link op which tries to create a link in this dir
2550 * back to this non-dir. */
2551 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2552 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2553 GOTO(out_put_new, rc = -EISDIR);
2555 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2556 mdt_lock_reg_init(lh_oldp, LCK_EX);
2557 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2558 if (mdt_object_remote(msrcdir)) {
2559 /* Enqueue lookup lock from the parent MDT */
2560 rc = mdt_remote_object_lock(info, msrcdir,
2561 mdt_object_fid(mold),
2562 &lh_oldp->mlh_rreg_lh,
2563 lh_oldp->mlh_rreg_mode,
2564 MDS_INODELOCK_LOOKUP,
2567 GOTO(out_put_new, rc);
2569 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2572 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2575 GOTO(out_unlock_old, rc);
2577 /* Check if @msrcdir is subdir of @mnew, before locking child
2578 * to avoid reverse locking. */
2579 if (mtgtdir != msrcdir) {
2580 rc = mdo_is_subdir(info->mti_env,
2581 mdt_object_child(msrcdir), new_fid);
2585 GOTO(out_unlock_old, rc);
2589 /* We used to acquire MDS_INODELOCK_FULL here but we
2590 * can't do this now because a running HSM restore on
2591 * the rename onto victim will hold the layout
2592 * lock. See LU-4002. */
2594 lh_newp = &info->mti_lh[MDT_LH_NEW];
2595 mdt_lock_reg_init(lh_newp, LCK_EX);
2596 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2597 MDS_INODELOCK_LOOKUP |
2598 MDS_INODELOCK_UPDATE,
2601 GOTO(out_unlock_old, rc);
2603 /* get and save version after locking */
2604 mdt_version_get_save(info, mnew, 3);
2605 } else if (rc != -EREMOTE && rc != -ENOENT) {
2606 GOTO(out_put_old, rc);
2608 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2609 mdt_lock_reg_init(lh_oldp, LCK_EX);
2610 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2611 if (mdt_object_remote(msrcdir)) {
2612 /* Enqueue lookup lock from the parent MDT */
2613 rc = mdt_remote_object_lock(info, msrcdir,
2614 mdt_object_fid(mold),
2615 &lh_oldp->mlh_rreg_lh,
2616 lh_oldp->mlh_rreg_mode,
2617 MDS_INODELOCK_LOOKUP,
2620 GOTO(out_put_old, rc);
2622 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2625 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2628 GOTO(out_unlock_old, rc);
2630 mdt_enoent_version_save(info, 3);
2633 /* step 5: rename it */
2634 mdt_reint_init_ma(info, ma);
2636 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2637 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2640 mutex_lock(&mnew->mot_lov_mutex);
2642 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2643 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2644 mnew != NULL ? mdt_object_child(mnew) : NULL,
2645 &rr->rr_tgt_name, ma);
2648 mutex_unlock(&mnew->mot_lov_mutex);
2650 /* handle last link of tgt object */
2652 mdt_counter_incr(req, LPROC_MDT_RENAME);
2654 mdt_handle_last_unlink(info, mnew, ma);
2655 discard = mdt_dom_check_for_discard(info, mnew);
2657 mdt_rename_counter_tally(info, info->mti_mdt, req,
2663 mdt_object_unlock(info, mnew, lh_newp, rc);
2665 mdt_object_unlock(info, mold, lh_oldp, rc);
2667 if (mnew && !discard)
2668 mdt_object_put(info->mti_env, mnew);
2670 mdt_object_put(info->mti_env, mold);
2672 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2673 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2675 if (lustre_handle_is_used(&rename_lh))
2676 mdt_rename_unlock(&rename_lh);
2678 mdt_object_put(info->mti_env, mtgtdir);
2680 mdt_object_put(info->mti_env, msrcdir);
2682 /* The DoM discard can be done right in the place above where it is
2683 * assigned, meanwhile it is done here after rename unlock due to
2684 * compatibility with old clients, for them the discard blocks
2685 * the main thread until completion. Check LU-11359 for details.
2688 mdt_dom_discard_data(info, mnew);
2689 mdt_object_put(info->mti_env, mnew);
2694 static int mdt_reint_resync(struct mdt_thread_info *info,
2695 struct mdt_lock_handle *lhc)
2697 struct mdt_reint_record *rr = &info->mti_rr;
2698 struct ptlrpc_request *req = mdt_info_req(info);
2699 struct md_attr *ma = &info->mti_attr;
2700 struct mdt_object *mo;
2701 struct ldlm_lock *lease;
2702 struct mdt_body *repbody;
2703 struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2709 DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2711 if (info->mti_dlm_req)
2712 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2714 mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2716 GOTO(out, rc = PTR_ERR(mo));
2718 if (!mdt_object_exists(mo))
2719 GOTO(out_obj, rc = -ENOENT);
2721 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2722 GOTO(out_obj, rc = -EINVAL);
2724 if (mdt_object_remote(mo))
2725 GOTO(out_obj, rc = -EREMOTE);
2727 lease = ldlm_handle2lock(rr->rr_lease_handle);
2729 GOTO(out_obj, rc = -ESTALE);
2731 /* It's really necessary to grab open_sem and check if the lease lock
2732 * has been lost. There would exist a concurrent writer coming in and
2733 * generating some dirty data in memory cache, the writeback would fail
2734 * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2735 if (!down_write_trylock(&mo->mot_open_sem))
2736 GOTO(out_put_lease, rc = -EBUSY);
2738 lock_res_and_lock(lease);
2739 lease_broken = ldlm_is_cancel(lease);
2740 unlock_res_and_lock(lease);
2742 GOTO(out_unlock, rc = -EBUSY);
2744 /* the file has yet opened by anyone else after we took the lease. */
2745 layout.mlc_opc = MD_LAYOUT_RESYNC;
2746 lhc = &info->mti_lh[MDT_LH_LOCAL];
2747 rc = mdt_layout_change(info, mo, lhc, &layout);
2749 GOTO(out_unlock, rc);
2751 mdt_object_unlock(info, mo, lhc, 0);
2753 ma->ma_need = MA_INODE;
2755 rc = mdt_attr_get_complex(info, mo, ma);
2757 GOTO(out_unlock, rc);
2759 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2760 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2764 up_write(&mo->mot_open_sem);
2766 LDLM_LOCK_PUT(lease);
2768 mdt_object_put(info->mti_env, mo);
2770 mdt_client_compatibility(info);
2771 rc2 = mdt_fix_reply(info);
2777 struct mdt_reinter {
2778 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2779 enum lprocfs_extra_opc mr_extra_opc;
2782 static const struct mdt_reinter mdt_reinters[] = {
2784 .mr_handler = &mdt_reint_setattr,
2785 .mr_extra_opc = MDS_REINT_SETATTR,
2788 .mr_handler = &mdt_reint_create,
2789 .mr_extra_opc = MDS_REINT_CREATE,
2792 .mr_handler = &mdt_reint_link,
2793 .mr_extra_opc = MDS_REINT_LINK,
2796 .mr_handler = &mdt_reint_unlink,
2797 .mr_extra_opc = MDS_REINT_UNLINK,
2800 .mr_handler = &mdt_reint_rename,
2801 .mr_extra_opc = MDS_REINT_RENAME,
2804 .mr_handler = &mdt_reint_open,
2805 .mr_extra_opc = MDS_REINT_OPEN,
2807 [REINT_SETXATTR] = {
2808 .mr_handler = &mdt_reint_setxattr,
2809 .mr_extra_opc = MDS_REINT_SETXATTR,
2812 .mr_handler = &mdt_reint_unlink,
2813 .mr_extra_opc = MDS_REINT_UNLINK,
2816 .mr_handler = &mdt_reint_migrate,
2817 .mr_extra_opc = MDS_REINT_RENAME,
2820 .mr_handler = &mdt_reint_resync,
2821 .mr_extra_opc = MDS_REINT_RESYNC,
2825 int mdt_reint_rec(struct mdt_thread_info *info,
2826 struct mdt_lock_handle *lhc)
2828 const struct mdt_reinter *mr;
2832 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2835 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2836 if (mr->mr_handler == NULL)
2839 rc = (*mr->mr_handler)(info, lhc);
2841 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2842 PTLRPC_LAST_CNTR + mr->mr_extra_opc);