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);
409 lh = &info->mti_lh[MDT_LH_PARENT];
410 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
411 rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
413 GOTO(put_parent, rc);
415 if (!mdt_object_remote(parent)) {
416 rc = mdt_version_get_check_save(info, parent, 0);
418 GOTO(unlock_parent, rc);
422 * Check child name version during replay.
423 * During create replay a file may exist with same name.
425 rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
426 &info->mti_tmp_fid1, 1);
428 GOTO(unlock_parent, rc = -EEXIST);
430 /* -ENOENT is expected here */
432 GOTO(unlock_parent, rc);
434 /* save version of file name for replay, it must be ENOENT here */
435 mdt_enoent_version_save(info, 1);
437 child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
438 if (unlikely(IS_ERR(child)))
439 GOTO(unlock_parent, rc = PTR_ERR(child));
441 ma->ma_need = MA_INODE;
444 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
445 OBD_FAIL_MDS_REINT_CREATE_WRITE);
447 /* Version of child will be updated on disk. */
448 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
449 rc = mdt_version_get_check_save(info, child, 2);
453 /* Let lower layer know current lock mode. */
454 info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
457 * Do not perform lookup sanity check. We know that name does
460 info->mti_spec.sp_cr_lookup = 0;
461 info->mti_spec.sp_feat = &dt_directory_features;
463 rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
464 mdt_object_child(child), &info->mti_spec, ma);
466 rc = mdt_attr_get_complex(info, child, ma);
472 * On DNE, we need to eliminate dependey between 'mkdir a' and
473 * 'mkdir a/b' if b is a striped directory, to achieve this, two
474 * things are done below:
475 * 1. save child and slaves lock.
476 * 2. if the child is a striped directory, relock parent so to
477 * compare against with COS locks to ensure parent was
480 if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
481 struct mdt_lock_handle *lhc;
482 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
485 rc = mdt_object_striped(info, child);
491 if (!mdt_object_remote(parent)) {
492 mdt_object_unlock(info, parent, lh, 1);
493 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
494 rc = mdt_reint_object_lock(info, parent, lh,
495 MDS_INODELOCK_UPDATE,
502 lhc = &info->mti_lh[MDT_LH_CHILD];
503 mdt_lock_handle_init(lhc);
504 mdt_lock_reg_init(lhc, LCK_PW);
505 rc = mdt_reint_striped_lock(info, child, lhc,
506 MDS_INODELOCK_UPDATE, einfo,
511 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
514 /* Return fid & attr to client. */
515 if (ma->ma_valid & MA_INODE)
516 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
517 mdt_object_fid(child));
519 mdt_object_put(info->mti_env, child);
521 mdt_object_unlock(info, parent, lh, rc);
523 mdt_object_put(info->mti_env, parent);
527 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
530 struct mdt_lock_handle *lh;
531 int do_vbr = ma->ma_attr.la_valid &
532 (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
533 __u64 lockpart = MDS_INODELOCK_UPDATE;
534 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
539 rc = mdt_object_striped(info, mo);
545 lh = &info->mti_lh[MDT_LH_PARENT];
546 mdt_lock_reg_init(lh, LCK_PW);
548 /* Even though the new MDT will grant PERM lock to the old
549 * client, but the old client will almost ignore that during
550 * So it needs to revoke both LOOKUP and PERM lock here, so
551 * both new and old client can cancel the dcache */
552 if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
553 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
555 rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
560 /* all attrs are packed into mti_attr in unpack_setattr */
561 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
562 OBD_FAIL_MDS_REINT_SETATTR_WRITE);
564 /* VBR: update version if attr changed are important for recovery */
566 /* update on-disk version of changed object */
567 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
568 rc = mdt_version_get_check_save(info, mo, 0);
570 GOTO(out_unlock, rc);
573 /* Ensure constant striping during chown(). See LU-2789. */
574 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
575 mutex_lock(&mo->mot_lov_mutex);
577 /* all attrs are packed into mti_attr in unpack_setattr */
578 rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
580 if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
581 mutex_unlock(&mo->mot_lov_mutex);
584 GOTO(out_unlock, rc);
585 mdt_dom_obj_lvb_update(info->mti_env, mo, false);
588 mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
593 * Check HSM flags and add HS_DIRTY flag if relevant.
595 * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
596 * and is not RELEASED.
598 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
601 struct lu_ucred *uc = mdt_ucred(info);
606 /* If the file was modified, add the dirty flag */
607 ma->ma_need = MA_HSM;
608 rc = mdt_attr_get_complex(info, mo, ma);
610 CERROR("file attribute read error for "DFID": %d.\n",
611 PFID(mdt_object_fid(mo)), rc);
615 /* If an up2date copy exists in the backend, add dirty flag */
616 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
617 && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
618 ma->ma_hsm.mh_flags |= HS_DIRTY;
620 /* Bump cap so that closes from non-owner writers can
621 * set the HSM state to dirty. */
622 cap_saved = uc->uc_cap;
623 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
624 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
625 uc->uc_cap = cap_saved;
627 CERROR("file attribute change error for "DFID": %d\n",
628 PFID(mdt_object_fid(mo)), rc);
634 static int mdt_reint_setattr(struct mdt_thread_info *info,
635 struct mdt_lock_handle *lhc)
637 struct mdt_device *mdt = info->mti_mdt;
638 struct md_attr *ma = &info->mti_attr;
639 struct mdt_reint_record *rr = &info->mti_rr;
640 struct ptlrpc_request *req = mdt_info_req(info);
641 struct mdt_object *mo;
642 struct mdt_body *repbody;
646 DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
647 (unsigned int)ma->ma_attr.la_valid);
649 if (info->mti_dlm_req)
650 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
652 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
653 mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
655 GOTO(out, rc = PTR_ERR(mo));
657 if (!mdt_object_exists(mo))
658 GOTO(out_put, rc = -ENOENT);
660 if (mdt_object_remote(mo))
661 GOTO(out_put, rc = -EREMOTE);
663 /* revoke lease lock if size is going to be changed */
664 if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
665 !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
666 atomic_read(&mo->mot_lease_count) > 0)) {
667 down_read(&mo->mot_open_sem);
669 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
670 lhc = &info->mti_lh[MDT_LH_LOCAL];
671 mdt_lock_reg_init(lhc, LCK_CW);
673 rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
675 up_read(&mo->mot_open_sem);
679 /* revoke lease lock */
680 mdt_object_unlock(info, mo, lhc, 1);
682 up_read(&mo->mot_open_sem);
685 if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
686 /* Check write access for the O_TRUNC case */
687 if (mdt_write_read(mo) < 0)
688 GOTO(out_put, rc = -ETXTBSY);
690 /* LU-10286: compatibility check for FLR.
691 * Please check the comment in mdt_finish_open() for details */
692 if (!exp_connect_flr(info->mti_exp) ||
693 !exp_connect_overstriping(info->mti_exp)) {
694 rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
695 if (rc < 0 && rc != -ENODATA)
698 if (!exp_connect_flr(info->mti_exp)) {
700 mdt_lmm_is_flr(info->mti_big_lmm))
701 GOTO(out_put, rc = -EOPNOTSUPP);
704 if (!exp_connect_overstriping(info->mti_exp)) {
706 mdt_lmm_is_overstriping(info->mti_big_lmm))
707 GOTO(out_put, rc = -EOPNOTSUPP);
711 /* For truncate, the file size sent from client
712 * is believable, but the blocks are incorrect,
713 * which makes the block size in LSOM attribute
714 * inconsisent with the real block size.
716 rc = mdt_lsom_update(info, mo, true);
721 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
722 if (ma->ma_valid & MA_LOV)
723 GOTO(out_put, rc = -EPROTO);
725 /* MDT supports FMD for regular files due to Data-on-MDT */
726 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
727 ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
728 tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
731 rc = mdt_attr_set(info, mo, ma);
734 } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
735 (ma->ma_valid & MA_INODE)) {
736 struct lu_buf *buf = &info->mti_buf;
737 struct lu_ucred *uc = mdt_ucred(info);
738 struct mdt_lock_handle *lh;
740 __u64 lockpart = MDS_INODELOCK_XATTR;
742 /* reject if either remote or striped dir is disabled */
743 if (ma->ma_valid & MA_LMV) {
744 if (!mdt->mdt_enable_remote_dir ||
745 !mdt->mdt_enable_striped_dir)
746 GOTO(out_put, rc = -EPERM);
748 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
749 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
750 mdt->mdt_enable_remote_dir_gid != -1)
751 GOTO(out_put, rc = -EPERM);
754 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
755 GOTO(out_put, rc = -ENOTDIR);
757 if (ma->ma_attr.la_valid != 0)
758 GOTO(out_put, rc = -EPROTO);
760 if (ma->ma_valid & MA_LOV) {
761 buf->lb_buf = ma->ma_lmm;
762 buf->lb_len = ma->ma_lmm_size;
763 name = XATTR_NAME_LOV;
765 struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
768 buf->lb_len = ma->ma_lmv_size;
770 if (le32_to_cpu(lmu->lum_hash_type) &
771 LMV_HASH_FLAG_SPACE) {
773 * only allow setting "space" hash flag on
776 rc = mdt_object_striped(info, mo);
779 rc = (rc == 1) ? -EPERM : rc);
782 name = XATTR_NAME_DEFAULT_LMV;
783 /* force client to update dir default layout */
784 lockpart |= MDS_INODELOCK_LOOKUP;
787 lh = &info->mti_lh[MDT_LH_PARENT];
788 mdt_lock_reg_init(lh, LCK_PW);
790 rc = mdt_object_lock(info, mo, lh, lockpart);
794 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
797 mdt_object_unlock(info, mo, lh, rc);
801 GOTO(out_put, rc = -EPROTO);
804 /* If file data is modified, add the dirty flag */
805 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
806 rc = mdt_add_dirty_flag(info, mo, ma);
808 ma->ma_need = MA_INODE;
810 rc = mdt_attr_get_complex(info, mo, ma);
814 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
818 mdt_object_put(info->mti_env, mo);
821 mdt_counter_incr(req, LPROC_MDT_SETATTR);
823 mdt_client_compatibility(info);
824 rc2 = mdt_fix_reply(info);
830 static int mdt_reint_create(struct mdt_thread_info *info,
831 struct mdt_lock_handle *lhc)
833 struct ptlrpc_request *req = mdt_info_req(info);
837 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
838 RETURN(err_serious(-ESTALE));
840 if (info->mti_dlm_req)
841 ldlm_request_cancel(mdt_info_req(info),
842 info->mti_dlm_req, 0, LATF_SKIP);
844 if (!lu_name_is_valid(&info->mti_rr.rr_name))
847 switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
849 mdt_counter_incr(req, LPROC_MDT_MKDIR);
857 /* Special file should stay on the same node as parent. */
858 mdt_counter_incr(req, LPROC_MDT_MKNOD);
861 CERROR("%s: Unsupported mode %o\n",
862 mdt_obd_name(info->mti_mdt),
863 info->mti_attr.ma_attr.la_mode);
864 RETURN(err_serious(-EOPNOTSUPP));
867 rc = mdt_create(info);
872 * VBR: save parent version in reply and child version getting by its name.
873 * Version of child is getting and checking during its lookup. If
875 static int mdt_reint_unlink(struct mdt_thread_info *info,
876 struct mdt_lock_handle *lhc)
878 struct mdt_reint_record *rr = &info->mti_rr;
879 struct ptlrpc_request *req = mdt_info_req(info);
880 struct md_attr *ma = &info->mti_attr;
881 struct lu_fid *child_fid = &info->mti_tmp_fid1;
882 struct mdt_object *mp;
883 struct mdt_object *mc;
884 struct mdt_lock_handle *parent_lh;
885 struct mdt_lock_handle *child_lh;
886 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
888 bool cos_incompat = false;
894 DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
895 PNAME(&rr->rr_name));
897 if (info->mti_dlm_req)
898 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
900 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
901 RETURN(err_serious(-ENOENT));
903 if (!fid_is_md_operative(rr->rr_fid1))
906 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
910 if (mdt_object_remote(mp)) {
913 rc = mdt_version_get_check_save(info, mp, 0);
915 GOTO(put_parent, rc);
919 parent_lh = &info->mti_lh[MDT_LH_PARENT];
920 mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
921 rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
924 GOTO(put_parent, rc);
926 /* lookup child object along with version checking */
928 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
930 /* Name might not be able to find during resend of
931 * remote unlink, considering following case.
932 * dir_A is a remote directory, the name entry of
933 * dir_A is on MDT0, the directory is on MDT1,
935 * 1. client sends unlink req to MDT1.
936 * 2. MDT1 sends name delete update to MDT0.
937 * 3. name entry is being deleted in MDT0 synchronously.
938 * 4. MDT1 is restarted.
939 * 5. client resends unlink req to MDT1. So it can not
940 * find the name entry on MDT0 anymore.
941 * In this case, MDT1 only needs to destory the local
944 if (mdt_object_remote(mp) && rc == -ENOENT &&
945 !fid_is_zero(rr->rr_fid2) &&
946 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
948 *child_fid = *rr->rr_fid2;
950 GOTO(unlock_parent, rc);
954 if (!fid_is_md_operative(child_fid))
955 GOTO(unlock_parent, rc = -EPERM);
957 /* We will lock the child regardless it is local or remote. No harm. */
958 mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
960 GOTO(unlock_parent, rc = PTR_ERR(mc));
963 rc = mdt_object_striped(info, mc);
969 mdt_object_put(info->mti_env, mc);
970 mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
975 child_lh = &info->mti_lh[MDT_LH_CHILD];
976 mdt_lock_reg_init(child_lh, LCK_EX);
977 if (info->mti_spec.sp_rm_entry) {
978 struct lu_ucred *uc = mdt_ucred(info);
980 if (!mdt_is_dne_client(req->rq_export))
981 /* Return -ENOTSUPP for old client */
982 GOTO(put_child, rc = -ENOTSUPP);
984 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
985 GOTO(put_child, rc = -EPERM);
987 ma->ma_need = MA_INODE;
989 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
990 NULL, &rr->rr_name, ma, no_name);
994 if (mdt_object_remote(mc)) {
995 struct mdt_body *repbody;
997 if (!fid_is_zero(rr->rr_fid2)) {
998 CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
999 mdt_obd_name(info->mti_mdt),
1000 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1001 GOTO(put_child, rc = -ENOENT);
1003 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1004 mdt_obd_name(info->mti_mdt),
1005 PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1007 if (!mdt_is_dne_client(req->rq_export))
1008 /* Return -ENOTSUPP for old client */
1009 GOTO(put_child, rc = -ENOTSUPP);
1011 /* Revoke the LOOKUP lock of the remote object granted by
1012 * this MDT. Since the unlink will happen on another MDT,
1013 * it will release the LOOKUP lock right away. Then What
1014 * would happen if another client try to grab the LOOKUP
1015 * lock at the same time with unlink XXX */
1016 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1017 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1018 LASSERT(repbody != NULL);
1019 repbody->mbo_fid1 = *mdt_object_fid(mc);
1020 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1021 GOTO(unlock_child, rc = -EREMOTE);
1023 /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1024 * this now because a running HSM restore on the child (unlink
1025 * victim) will hold the layout lock. See LU-4002. */
1026 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1027 if (mdt_object_remote(mp)) {
1028 /* Enqueue lookup lock from parent MDT */
1029 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1030 &child_lh->mlh_rreg_lh,
1031 child_lh->mlh_rreg_mode,
1032 MDS_INODELOCK_LOOKUP, false);
1034 GOTO(put_child, rc);
1036 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1039 rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1042 GOTO(put_child, rc);
1045 * Now we can only make sure we need MA_INODE, in mdd layer, will check
1046 * whether need MA_LOV and MA_COOKIE.
1048 ma->ma_need = MA_INODE;
1051 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1052 OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1053 /* save version when object is locked */
1054 mdt_version_get_save(info, mc, 1);
1056 mutex_lock(&mc->mot_lov_mutex);
1058 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1059 mdt_object_child(mc), &rr->rr_name, ma, no_name);
1061 mutex_unlock(&mc->mot_lov_mutex);
1063 GOTO(unlock_child, rc);
1065 if (!lu_object_is_dying(&mc->mot_header)) {
1066 rc = mdt_attr_get_complex(info, mc, ma);
1069 } else if (mdt_dom_check_for_discard(info, mc)) {
1070 mdt_dom_discard_data(info, mc);
1072 mdt_handle_last_unlink(info, mc, ma);
1075 if (ma->ma_valid & MA_INODE) {
1076 switch (ma->ma_attr.la_mode & S_IFMT) {
1078 mdt_counter_incr(req, LPROC_MDT_RMDIR);
1086 mdt_counter_incr(req, LPROC_MDT_UNLINK);
1089 LASSERTF(0, "bad file type %o unlinking\n",
1090 ma->ma_attr.la_mode);
1097 mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1099 mdt_object_put(info->mti_env, mc);
1101 mdt_object_unlock(info, mp, parent_lh, rc);
1103 mdt_object_put(info->mti_env, mp);
1104 CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1109 * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1112 static int mdt_reint_link(struct mdt_thread_info *info,
1113 struct mdt_lock_handle *lhc)
1115 struct mdt_reint_record *rr = &info->mti_rr;
1116 struct ptlrpc_request *req = mdt_info_req(info);
1117 struct md_attr *ma = &info->mti_attr;
1118 struct mdt_object *ms;
1119 struct mdt_object *mp;
1120 struct mdt_lock_handle *lhs;
1121 struct mdt_lock_handle *lhp;
1126 DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1127 PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1129 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1130 RETURN(err_serious(-ENOENT));
1132 if (info->mti_dlm_req)
1133 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1135 /* Invalid case so return error immediately instead of
1137 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1140 if (!fid_is_md_operative(rr->rr_fid1) ||
1141 !fid_is_md_operative(rr->rr_fid2))
1144 /* step 1: find target parent dir */
1145 mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1147 RETURN(PTR_ERR(mp));
1149 rc = mdt_version_get_check_save(info, mp, 0);
1151 GOTO(put_parent, rc);
1153 /* step 2: find source */
1154 ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1156 GOTO(put_parent, rc = PTR_ERR(ms));
1158 if (!mdt_object_exists(ms)) {
1159 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1160 mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1161 GOTO(put_source, rc = -ENOENT);
1164 cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1166 lhp = &info->mti_lh[MDT_LH_PARENT];
1167 mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1168 rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1171 GOTO(put_source, rc);
1173 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1175 lhs = &info->mti_lh[MDT_LH_CHILD];
1176 mdt_lock_reg_init(lhs, LCK_EX);
1177 rc = mdt_reint_object_lock(info, ms, lhs,
1178 MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1181 GOTO(unlock_parent, rc);
1183 /* step 3: link it */
1184 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1185 OBD_FAIL_MDS_REINT_LINK_WRITE);
1187 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1188 rc = mdt_version_get_check_save(info, ms, 1);
1190 GOTO(unlock_source, rc);
1192 /** check target version by name during replay */
1193 rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1194 &info->mti_tmp_fid1, 2);
1195 if (rc != 0 && rc != -ENOENT)
1196 GOTO(unlock_source, rc);
1197 /* save version of file name for replay, it must be ENOENT here */
1198 if (!req_is_replay(mdt_info_req(info))) {
1199 if (rc != -ENOENT) {
1200 CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1201 PNAME(&rr->rr_name));
1202 GOTO(unlock_source, rc = -EEXIST);
1204 info->mti_ver[2] = ENOENT_VERSION;
1205 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1208 rc = mdo_link(info->mti_env, mdt_object_child(mp),
1209 mdt_object_child(ms), &rr->rr_name, ma);
1212 mdt_counter_incr(req, LPROC_MDT_LINK);
1216 mdt_object_unlock(info, ms, lhs, rc);
1218 mdt_object_unlock(info, mp, lhp, rc);
1220 mdt_object_put(info->mti_env, ms);
1222 mdt_object_put(info->mti_env, mp);
1226 * lock the part of the directory according to the hash of the name
1227 * (lh->mlh_pdo_hash) in parallel directory lock.
1229 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1230 struct mdt_lock_handle *lh,
1231 struct mdt_object *obj, __u64 ibits,
1234 struct ldlm_res_id *res = &info->mti_res_id;
1235 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1236 union ldlm_policy_data *policy = &info->mti_policy;
1237 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1241 * Finish res_id initializing by name hash marking part of
1242 * directory which is taking modification.
1244 LASSERT(lh->mlh_pdo_hash != 0);
1245 fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1246 memset(policy, 0, sizeof(*policy));
1247 policy->l_inodebits.bits = ibits;
1249 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1250 dlmflags |= LDLM_FL_COS_INCOMPAT;
1252 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1253 * going to be sent to client. If it is - mdt_intent_policy() path will
1254 * fix it up and turn FL_LOCAL flag off.
1256 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1257 policy, res, dlmflags,
1258 &info->mti_exp->exp_handle.h_cookie);
1263 * Get BFL lock for rename or migrate process.
1265 static int mdt_rename_lock(struct mdt_thread_info *info,
1266 struct lustre_handle *lh)
1271 if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1272 struct lu_fid *fid = &info->mti_tmp_fid1;
1273 struct mdt_object *obj;
1275 /* XXX, right now, it has to use object API to
1276 * enqueue lock cross MDT, so it will enqueue
1277 * rename lock(with LUSTRE_BFL_FID) by root object */
1279 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1281 RETURN(PTR_ERR(obj));
1283 rc = mdt_remote_object_lock(info, obj,
1284 &LUSTRE_BFL_FID, lh,
1286 MDS_INODELOCK_UPDATE, false);
1287 mdt_object_put(info->mti_env, obj);
1289 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1290 union ldlm_policy_data *policy = &info->mti_policy;
1291 struct ldlm_res_id *res_id = &info->mti_res_id;
1294 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1295 memset(policy, 0, sizeof *policy);
1296 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1297 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1298 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1299 LDLM_IBITS, policy, LCK_EX, &flags,
1301 ldlm_completion_ast, NULL, NULL, 0,
1303 &info->mti_exp->exp_handle.h_cookie,
1310 static void mdt_rename_unlock(struct lustre_handle *lh)
1313 LASSERT(lustre_handle_is_used(lh));
1314 /* Cancel the single rename lock right away */
1315 ldlm_lock_decref_and_cancel(lh, LCK_EX);
1319 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1320 const struct lu_fid *fid,
1323 struct mdt_object *dir;
1328 dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1332 /* check early, the real version will be saved after locking */
1333 rc = mdt_version_get_check(info, dir, idx);
1337 if (!mdt_object_exists(dir))
1338 GOTO(out_put, rc = -ENOENT);
1340 if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1341 GOTO(out_put, rc = -ENOTDIR);
1345 mdt_object_put(info->mti_env, dir);
1350 * in case obj is remote obj on its parent, revoke LOOKUP lock,
1351 * herein we don't really check it, just do revoke.
1353 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1354 struct mdt_object *pobj,
1355 struct mdt_object *obj)
1357 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1360 mdt_lock_handle_init(lh);
1361 mdt_lock_reg_init(lh, LCK_EX);
1363 if (mdt_object_remote(pobj)) {
1364 /* don't bother to check if pobj and obj are on the same MDT. */
1365 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1366 &lh->mlh_rreg_lh, LCK_EX,
1367 MDS_INODELOCK_LOOKUP, false);
1368 } else if (mdt_object_remote(obj)) {
1369 struct ldlm_res_id *res = &info->mti_res_id;
1370 union ldlm_policy_data *policy = &info->mti_policy;
1371 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1372 LDLM_FL_COS_INCOMPAT;
1374 fid_build_reg_res_name(mdt_object_fid(obj), res);
1375 memset(policy, 0, sizeof(*policy));
1376 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1377 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1378 &lh->mlh_reg_lh, LCK_EX, policy, res,
1381 /* do nothing if both are local */
1389 * TODO, currently we don't save this lock because there is no place to
1390 * hold this lock handle, but to avoid race we need to save this lock.
1392 mdt_object_unlock(info, NULL, lh, 1);
1398 * operation may takes locks of linkea, or directory stripes, group them in
1401 struct mdt_sub_lock {
1402 struct mdt_object *msl_obj;
1403 struct mdt_lock_handle msl_lh;
1404 struct list_head msl_linkage;
1407 static void mdt_unlock_list(struct mdt_thread_info *info,
1408 struct list_head *list, int decref)
1410 struct mdt_sub_lock *msl;
1411 struct mdt_sub_lock *tmp;
1413 list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1414 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1415 list_del(&msl->msl_linkage);
1420 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1421 struct mdt_object *obj,
1422 struct mdt_lock_handle *lh,
1423 struct ldlm_enqueue_info *einfo,
1424 struct list_head *slave_locks,
1427 if (mdt_object_remote(obj)) {
1428 mdt_unlock_list(info, slave_locks, decref);
1429 mdt_object_unlock(info, obj, lh, decref);
1431 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1436 * lock parents of links, and also check whether total locks don't exceed
1439 * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat
1440 * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS
1441 * \retval -ev negative errno upon error
1443 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1444 struct mdt_object *pobj,
1445 const struct md_attr *ma,
1446 struct mdt_object *obj,
1447 struct mdt_lock_handle *lhp,
1448 struct ldlm_enqueue_info *peinfo,
1449 struct list_head *parent_slave_locks,
1450 struct list_head *link_locks)
1452 struct mdt_device *mdt = info->mti_mdt;
1453 struct lu_buf *buf = &info->mti_big_buf;
1454 struct lu_name *lname = &info->mti_name;
1455 struct linkea_data ldata = { NULL };
1456 bool blocked = false;
1457 int local_lnkp_cnt = 0;
1462 if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1465 buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1466 if (buf->lb_buf == NULL)
1470 rc = mdt_links_read(info, obj, &ldata);
1472 if (rc == -ENOENT || rc == -ENODATA)
1477 for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1478 linkea_next_entry(&ldata)) {
1479 struct mdt_object *lnkp;
1480 struct mdt_sub_lock *msl;
1484 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1487 /* check if it's also linked to parent */
1488 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1489 CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1490 PFID(&fid), PNAME(lname));
1491 /* in case link is remote object, revoke LOOKUP lock */
1492 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1498 /* check if it's linked to a stripe of parent */
1499 if (ma->ma_valid & MA_LMV) {
1500 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1501 struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1504 for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1505 fid_le_to_cpu(stripe_fid,
1506 &lmv->lmv_stripe_fids[j]);
1507 if (lu_fid_eq(stripe_fid, &fid)) {
1508 CDEBUG(D_INFO, "skip stripe "DFID
1509 ", reovke "DNAME"\n",
1510 PFID(&fid), PNAME(lname));
1511 lnkp = mdt_object_find(info->mti_env,
1514 GOTO(out, rc = PTR_ERR(lnkp));
1520 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1522 mdt_object_put(info->mti_env, lnkp);
1527 /* Check if it's already locked */
1528 list_for_each_entry(msl, link_locks, msl_linkage) {
1529 if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1531 DFID" was locked, revoke "DNAME"\n",
1532 PFID(&fid), PNAME(lname));
1533 lnkp = msl->msl_obj;
1539 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1543 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1544 PFID(&fid), PNAME(lname));
1546 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1548 CWARN("%s: cannot find obj "DFID": %ld\n",
1549 mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1553 if (!mdt_object_exists(lnkp)) {
1554 CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1555 PFID(&fid), PNAME(lname));
1556 mdt_object_put(info->mti_env, lnkp);
1560 if (!mdt_object_remote(lnkp))
1565 GOTO(out, rc = -ENOMEM);
1568 * we can't follow parent-child lock order like other MD
1569 * operations, use lock_try here to avoid deadlock, if the lock
1570 * cannot be taken, drop all locks taken, revoke the blocked
1571 * one, and continue processing the remaining entries, and in
1572 * the end of the loop restart from beginning.
1574 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1576 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1577 MDS_INODELOCK_UPDATE, true);
1578 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1580 CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1581 PFID(&fid), PNAME(lname));
1583 mdt_unlock_list(info, link_locks, 1);
1584 /* also unlock parent locks to avoid deadlock */
1586 mdt_migrate_object_unlock(info, pobj, lhp,
1593 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1594 rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1595 MDS_INODELOCK_UPDATE);
1597 mdt_object_put(info->mti_env, lnkp);
1602 if (mdt_object_remote(lnkp)) {
1603 struct ldlm_lock *lock;
1606 * for remote object, set lock cb_atomic,
1607 * so lock can be released in blocking_ast()
1608 * immediately, then the next lock_try will
1609 * have better chance of success.
1611 lock = ldlm_handle2lock(
1612 &msl->msl_lh.mlh_rreg_lh);
1613 LASSERT(lock != NULL);
1614 lock_res_and_lock(lock);
1615 ldlm_set_atomic_cb(lock);
1616 unlock_res_and_lock(lock);
1617 LDLM_LOCK_PUT(lock);
1620 mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1625 INIT_LIST_HEAD(&msl->msl_linkage);
1626 msl->msl_obj = lnkp;
1627 list_add_tail(&msl->msl_linkage, link_locks);
1629 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1633 GOTO(out, rc = -EBUSY);
1638 mdt_unlock_list(info, link_locks, rc);
1639 else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1641 * parent may have 3 local objects: master object and 2 stripes
1642 * (if it's being migrated too); source may have 2 local
1643 * objects: master and 1 stripe; target has 1 local object.
1649 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1650 struct mdt_object *obj,
1651 const struct md_attr *ma,
1652 struct list_head *slave_locks)
1654 struct mdt_device *mdt = info->mti_mdt;
1655 const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1656 struct lu_fid *fid = &info->mti_tmp_fid1;
1657 struct mdt_object *slave;
1658 struct mdt_sub_lock *msl;
1664 LASSERT(mdt_object_remote(obj));
1665 LASSERT(ma->ma_valid & MA_LMV);
1668 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1671 if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1674 for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1675 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1677 if (!fid_is_sane(fid))
1680 slave = mdt_object_find(info->mti_env, mdt, fid);
1682 GOTO(out, rc = PTR_ERR(slave));
1686 mdt_object_put(info->mti_env, slave);
1687 GOTO(out, rc = -ENOMEM);
1690 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1691 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1692 MDS_INODELOCK_UPDATE, true);
1695 mdt_object_put(info->mti_env, slave);
1699 INIT_LIST_HEAD(&msl->msl_linkage);
1700 msl->msl_obj = slave;
1701 list_add_tail(&msl->msl_linkage, slave_locks);
1707 mdt_unlock_list(info, slave_locks, rc);
1711 /* lock parent and its stripes */
1712 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1713 struct mdt_object *obj,
1714 const struct md_attr *ma,
1715 struct mdt_lock_handle *lh,
1716 struct ldlm_enqueue_info *einfo,
1717 struct list_head *slave_locks)
1721 if (mdt_object_remote(obj)) {
1722 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1723 &lh->mlh_rreg_lh, LCK_PW,
1724 MDS_INODELOCK_UPDATE, false);
1729 * if obj is remote and striped, lock its stripes explicitly
1730 * because it's not striped in LOD layer on this MDT.
1732 if (ma->ma_valid & MA_LMV) {
1733 rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1735 mdt_object_unlock(info, obj, lh, rc);
1738 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1746 * in migration, object may be remote, and we need take full lock of it and its
1747 * stripes if it's directory, besides, object may be a remote object on its
1748 * parent, revoke its LOOKUP lock on where its parent is located.
1750 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1751 struct mdt_object *pobj,
1752 struct mdt_object *obj,
1753 struct mdt_lock_handle *lh,
1754 struct ldlm_enqueue_info *einfo,
1755 struct list_head *slave_locks)
1759 if (mdt_object_remote(obj)) {
1760 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1764 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1765 &lh->mlh_rreg_lh, LCK_EX,
1766 MDS_INODELOCK_FULL, false);
1771 * if obj is remote and striped, lock its stripes explicitly
1772 * because it's not striped in LOD layer on this MDT.
1774 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1775 struct md_attr *ma = &info->mti_attr;
1777 ma->ma_lmv = info->mti_big_lmm;
1778 ma->ma_lmv_size = info->mti_big_lmmsize;
1780 rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1782 mdt_object_unlock(info, obj, lh, rc);
1786 if (ma->ma_valid & MA_LMV) {
1787 rc = mdt_lock_remote_slaves(info, obj, ma,
1790 mdt_object_unlock(info, obj, lh, rc);
1794 if (mdt_object_remote(pobj)) {
1795 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1800 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1808 * lookup source by name, if parent is striped directory, we need to find the
1809 * corresponding stripe where source is located, and then lookup there.
1811 * besides, if parent is migrating too, and file is already in target stripe,
1812 * this should be a redo of 'lfs migrate' on client side.
1814 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1815 struct mdt_object *pobj,
1816 const struct md_attr *ma,
1817 const struct lu_name *lname,
1818 struct mdt_object **spobj,
1819 struct mdt_object **sobj)
1821 const struct lu_env *env = info->mti_env;
1822 struct lu_fid *fid = &info->mti_tmp_fid1;
1823 struct mdt_object *stripe;
1826 if (ma->ma_valid & MA_LMV) {
1827 /* if parent is striped, lookup on corresponding stripe */
1828 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1829 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1830 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1831 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1832 LMV_HASH_FLAG_MIGRATION;
1835 hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1836 stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1839 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1845 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1846 rc += le32_to_cpu(lmv->lmv_migrate_offset);
1848 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1850 stripe = mdt_object_find(env, info->mti_mdt, fid);
1852 return PTR_ERR(stripe);
1855 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1857 if (rc == -ENOENT && is_migrating) {
1859 * if parent is migrating, and lookup child failed on
1860 * source stripe, lookup again on target stripe, if it
1861 * exists, it means previous migration was interrupted,
1862 * and current file was migrated already.
1864 mdt_object_put(env, stripe);
1866 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1867 stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1869 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1875 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1877 stripe = mdt_object_find(env, info->mti_mdt, fid);
1879 return PTR_ERR(stripe);
1882 rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1883 fid, &info->mti_spec);
1884 mdt_object_put(env, stripe);
1885 return rc ?: -EALREADY;
1887 mdt_object_put(env, stripe);
1892 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1898 mdt_object_get(env, stripe);
1903 *sobj = mdt_object_find(env, info->mti_mdt, fid);
1904 if (IS_ERR(*sobj)) {
1905 mdt_object_put(env, stripe);
1906 rc = PTR_ERR(*sobj);
1914 /* end lease and close file for regular file */
1915 static int mdd_migrate_close(struct mdt_thread_info *info,
1916 struct mdt_object *obj)
1918 struct close_data *data;
1919 struct mdt_body *repbody;
1920 struct ldlm_lock *lease;
1925 if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1927 !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1931 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1936 lease = ldlm_handle2lock(&data->cd_handle);
1940 /* check if the lease was already canceled */
1941 lock_res_and_lock(lease);
1942 rc = ldlm_is_cancel(lease);
1943 unlock_res_and_lock(lease);
1947 LDLM_DEBUG(lease, DFID" lease broken",
1948 PFID(mdt_object_fid(obj)));
1952 * cancel server side lease, client side counterpart should have been
1953 * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1955 ldlm_lock_cancel(lease);
1956 ldlm_reprocess_all(lease->l_resource, lease);
1957 LDLM_LOCK_PUT(lease);
1960 rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1961 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1962 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1968 * migrate file in below steps:
1969 * 1. lock parent and its stripes
1970 * 2. lookup source by name
1971 * 3. lock parents of source links if source is not directory
1972 * 4. reject if source is in HSM
1973 * 5. take source open_sem and close file if source is regular file
1974 * 6. lock source and its stripes if it's directory
1975 * 7. lock target so subsequent change to it can trigger COS
1977 * 9. unlock above locks
1978 * 10. sync device if source has links
1980 static int mdt_reint_migrate(struct mdt_thread_info *info,
1981 struct mdt_lock_handle *unused)
1983 const struct lu_env *env = info->mti_env;
1984 struct mdt_device *mdt = info->mti_mdt;
1985 struct ptlrpc_request *req = mdt_info_req(info);
1986 struct mdt_reint_record *rr = &info->mti_rr;
1987 struct lu_ucred *uc = mdt_ucred(info);
1988 struct md_attr *ma = &info->mti_attr;
1989 struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1990 struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1991 struct mdt_object *pobj;
1992 struct mdt_object *spobj = NULL;
1993 struct mdt_object *sobj = NULL;
1994 struct mdt_object *tobj;
1995 struct lustre_handle rename_lh = { 0 };
1996 struct mdt_lock_handle *lhp;
1997 struct mdt_lock_handle *lhs;
1998 struct mdt_lock_handle *lht;
1999 LIST_HEAD(parent_slave_locks);
2000 LIST_HEAD(child_slave_locks);
2001 LIST_HEAD(link_locks);
2002 int lock_retries = 5;
2003 bool open_sem_locked = false;
2004 bool do_sync = false;
2008 CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2009 PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2011 if (info->mti_dlm_req)
2012 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2014 if (!fid_is_md_operative(rr->rr_fid1) ||
2015 !fid_is_md_operative(rr->rr_fid2))
2018 /* don't allow migrate . or .. */
2019 if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2022 if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2025 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2026 uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2027 mdt->mdt_enable_remote_dir_gid != -1)
2031 * Note: do not enqueue rename lock for replay request, because
2032 * if other MDT holds rename lock, but being blocked to wait for
2033 * this MDT to finish its recovery, and the failover MDT can not
2034 * get rename lock, which will cause deadlock.
2036 if (!req_is_replay(req)) {
2037 rc = mdt_rename_lock(info, &rename_lh);
2039 CERROR("%s: can't lock FS for rename: rc = %d\n",
2040 mdt_obd_name(info->mti_mdt), rc);
2045 /* pobj is master object of parent */
2046 pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2048 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2050 if (unlikely(!info->mti_big_lmm)) {
2051 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2052 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2053 if (!info->mti_big_lmm)
2054 GOTO(put_parent, rc = -ENOMEM);
2057 ma->ma_lmv = info->mti_big_lmm;
2058 ma->ma_lmv_size = info->mti_big_lmmsize;
2060 rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2062 GOTO(put_parent, rc);
2065 /* lock parent object */
2066 lhp = &info->mti_lh[MDT_LH_PARENT];
2067 mdt_lock_reg_init(lhp, LCK_PW);
2068 rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2069 &parent_slave_locks);
2071 GOTO(put_parent, rc);
2074 * spobj is the corresponding stripe against name if pobj is striped
2075 * directory, which is the real parent, and no need to lock, because
2076 * we've taken full lock of pobj.
2078 rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2080 GOTO(unlock_parent, rc);
2082 /* lock parents of source links, and revoke LOOKUP lock of links */
2083 rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2084 &parent_slave_locks, &link_locks);
2085 if (rc == -EBUSY && lock_retries-- > 0) {
2086 mdt_object_put(env, sobj);
2087 mdt_object_put(env, spobj);
2092 GOTO(put_source, rc);
2095 * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2096 * with one request, if total lock count exceeds this limit, we will
2097 * drop all locks after migration, and synchronous device in the end.
2101 /* TODO: DoM migration is not supported yet */
2102 if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2103 ma->ma_lmm = info->mti_big_lmm;
2104 ma->ma_lmm_size = info->mti_big_lmmsize;
2106 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2108 GOTO(put_source, rc);
2110 if (ma->ma_valid & MA_LOV &&
2111 mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2112 GOTO(put_source, rc = -EOPNOTSUPP);
2115 /* if migration HSM is allowed */
2116 if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2117 ma->ma_need = MA_HSM;
2119 rc = mdt_attr_get_complex(info, sobj, ma);
2121 GOTO(unlock_links, rc);
2123 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2124 GOTO(unlock_links, rc = -EOPNOTSUPP);
2127 /* end lease and close file for regular file */
2128 if (info->mti_spec.sp_migrate_close) {
2129 /* try to hold open_sem so that nobody else can open the file */
2130 if (!down_write_trylock(&sobj->mot_open_sem)) {
2132 mdd_migrate_close(info, sobj);
2133 GOTO(unlock_links, rc = -EBUSY);
2135 open_sem_locked = true;
2136 rc = mdd_migrate_close(info, sobj);
2138 GOTO(unlock_open_sem, rc);
2143 lhs = &info->mti_lh[MDT_LH_OLD];
2144 mdt_lock_reg_init(lhs, LCK_EX);
2145 rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2146 &child_slave_locks);
2148 GOTO(unlock_open_sem, rc);
2151 tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2153 GOTO(unlock_source, rc = PTR_ERR(tobj));
2155 lht = &info->mti_lh[MDT_LH_NEW];
2156 mdt_lock_reg_init(lht, LCK_EX);
2157 rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2159 GOTO(put_target, rc);
2161 /* Don't do lookup sanity check. We know name doesn't exist. */
2162 info->mti_spec.sp_cr_lookup = 0;
2163 info->mti_spec.sp_feat = &dt_directory_features;
2165 rc = mdo_migrate(env, mdt_object_child(pobj),
2166 mdt_object_child(sobj), &rr->rr_name,
2167 mdt_object_child(tobj), &info->mti_spec, ma);
2170 mdt_object_unlock(info, tobj, lht, rc);
2172 mdt_object_put(env, tobj);
2174 mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2175 &child_slave_locks, rc);
2177 if (open_sem_locked)
2178 up_write(&sobj->mot_open_sem);
2180 /* if we've got too many locks to save into RPC,
2181 * then just commit before the locks are released */
2183 mdt_device_sync(env, mdt);
2184 mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2186 mdt_object_put(env, sobj);
2187 mdt_object_put(env, spobj);
2189 mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2190 &parent_slave_locks, rc);
2192 mdt_object_put(env, pobj);
2194 if (lustre_handle_is_used(&rename_lh))
2195 mdt_rename_unlock(&rename_lh);
2200 static int mdt_object_lock_save(struct mdt_thread_info *info,
2201 struct mdt_object *dir,
2202 struct mdt_lock_handle *lh,
2203 int idx, bool cos_incompat)
2207 /* we lock the target dir if it is local */
2208 rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2213 /* get and save correct version after locking */
2214 mdt_version_get_save(info, dir, idx);
2219 * determine lock order of sobj and tobj
2221 * there are two situations we need to lock tobj before sobj:
2222 * 1. sobj is child of tobj
2223 * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2224 * larger than that of tobj
2226 * \retval 1 lock tobj before sobj
2227 * \retval 0 lock sobj before tobj
2228 * \retval -ev negative errno upon error
2230 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2231 struct mdt_object *sobj,
2232 struct mdt_object *tobj)
2234 struct md_attr *ma = &info->mti_attr;
2235 struct lu_fid *spfid = &info->mti_tmp_fid1;
2236 struct lu_fid *tpfid = &info->mti_tmp_fid2;
2237 struct lmv_mds_md_v1 *lmv;
2242 /* sobj and tobj are the same */
2246 if (fid_is_root(mdt_object_fid(sobj)))
2249 if (fid_is_root(mdt_object_fid(tobj)))
2252 /* check whether sobj is child of tobj */
2253 rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2254 mdt_object_fid(tobj));
2261 /* check whether sobj and tobj are children of the same parent */
2262 rc = mdt_attr_get_pfid(info, sobj, spfid);
2266 rc = mdt_attr_get_pfid(info, tobj, tpfid);
2270 if (!lu_fid_eq(spfid, tpfid))
2273 /* check whether sobj and tobj are sibling stripes */
2274 ma->ma_need = MA_LMV;
2276 ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2277 ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2278 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2282 if (!(ma->ma_valid & MA_LMV))
2285 lmv = &ma->ma_lmv->lmv_md_v1;
2286 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2288 sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2291 rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2295 if (!(ma->ma_valid & MA_LMV))
2298 lmv = &ma->ma_lmv->lmv_md_v1;
2299 if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2301 tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2303 /* check stripe index of sobj and tobj */
2304 if (sindex == tindex)
2307 return sindex < tindex ? 0 : 1;
2311 * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2312 * 2 - srcdir child; 3 - tgtdir child.
2313 * Update on disk version of srcdir child.
2315 static int mdt_reint_rename(struct mdt_thread_info *info,
2316 struct mdt_lock_handle *unused)
2318 struct mdt_device *mdt = info->mti_mdt;
2319 struct mdt_reint_record *rr = &info->mti_rr;
2320 struct md_attr *ma = &info->mti_attr;
2321 struct ptlrpc_request *req = mdt_info_req(info);
2322 struct mdt_object *msrcdir = NULL;
2323 struct mdt_object *mtgtdir = NULL;
2324 struct mdt_object *mold;
2325 struct mdt_object *mnew = NULL;
2326 struct lustre_handle rename_lh = { 0 };
2327 struct mdt_lock_handle *lh_srcdirp;
2328 struct mdt_lock_handle *lh_tgtdirp;
2329 struct mdt_lock_handle *lh_oldp = NULL;
2330 struct mdt_lock_handle *lh_newp = NULL;
2331 struct lu_fid *old_fid = &info->mti_tmp_fid1;
2332 struct lu_fid *new_fid = &info->mti_tmp_fid2;
2334 bool reverse = false, discard = false;
2339 DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2340 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2341 PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2343 if (info->mti_dlm_req)
2344 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2346 if (!fid_is_md_operative(rr->rr_fid1) ||
2347 !fid_is_md_operative(rr->rr_fid2))
2350 /* find both parents. */
2351 msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2352 if (IS_ERR(msrcdir))
2353 RETURN(PTR_ERR(msrcdir));
2355 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2357 if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2359 mdt_object_get(info->mti_env, mtgtdir);
2361 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2362 if (IS_ERR(mtgtdir))
2363 GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2367 * Note: do not enqueue rename lock for replay request, because
2368 * if other MDT holds rename lock, but being blocked to wait for
2369 * this MDT to finish its recovery, and the failover MDT can not
2370 * get rename lock, which will cause deadlock.
2372 if (!req_is_replay(req)) {
2374 * Normally rename RPC is handled on the MDT with the target
2375 * directory (if target exists, it's on the MDT with the
2376 * target), if the source directory is remote, it's a hint that
2377 * source is remote too (this may not be true, but it won't
2378 * cause any issue), return -EXDEV early to avoid taking
2381 if (!mdt->mdt_enable_remote_rename &&
2382 mdt_object_remote(msrcdir))
2383 GOTO(out_put_tgtdir, rc = -EXDEV);
2385 rc = mdt_rename_lock(info, &rename_lh);
2387 CERROR("%s: can't lock FS for rename: rc = %d\n",
2388 mdt_obd_name(mdt), rc);
2389 GOTO(out_put_tgtdir, rc);
2393 rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2395 GOTO(out_unlock_rename, rc);
2399 /* source needs to be looked up after locking source parent, otherwise
2400 * this rename may race with unlink source, and cause rename hang, see
2401 * sanityn.sh 55b, so check parents first, if later we found source is
2402 * remote, relock parents. */
2403 cos_incompat = (mdt_object_remote(msrcdir) ||
2404 mdt_object_remote(mtgtdir));
2406 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2408 /* lock parents in the proper order. */
2409 lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2410 lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2413 mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2414 mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2417 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2420 GOTO(out_unlock_rename, rc);
2422 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2424 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2427 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2428 GOTO(out_unlock_rename, rc);
2431 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2434 GOTO(out_unlock_rename, rc);
2436 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2438 if (mtgtdir != msrcdir) {
2439 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2441 } else if (!mdt_object_remote(mtgtdir) &&
2442 lh_srcdirp->mlh_pdo_hash !=
2443 lh_tgtdirp->mlh_pdo_hash) {
2444 rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2445 MDS_INODELOCK_UPDATE,
2447 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2450 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2451 GOTO(out_unlock_rename, rc);
2455 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2456 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2458 /* find mold object. */
2460 rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2462 GOTO(out_unlock_parents, rc);
2464 if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2465 GOTO(out_unlock_parents, rc = -EINVAL);
2467 if (!fid_is_md_operative(old_fid))
2468 GOTO(out_unlock_parents, rc = -EPERM);
2470 mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2472 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2474 if (!mdt_object_exists(mold)) {
2475 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2477 "object does not exist");
2478 GOTO(out_put_old, rc = -ENOENT);
2481 if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2482 GOTO(out_put_old, rc = -EXDEV);
2484 /* Check if @mtgtdir is subdir of @mold, before locking child
2485 * to avoid reverse locking. */
2486 if (mtgtdir != msrcdir) {
2487 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2492 GOTO(out_put_old, rc);
2496 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2497 /* save version after locking */
2498 mdt_version_get_save(info, mold, 2);
2500 if (!cos_incompat && mdt_object_remote(mold)) {
2501 cos_incompat = true;
2502 mdt_object_put(info->mti_env, mold);
2503 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2504 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2508 /* find mnew object:
2509 * mnew target object may not exist now
2510 * lookup with version checking */
2512 rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2515 /* the new_fid should have been filled at this moment */
2516 if (lu_fid_eq(old_fid, new_fid))
2517 GOTO(out_put_old, rc);
2519 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2520 lu_fid_eq(new_fid, rr->rr_fid2))
2521 GOTO(out_put_old, rc = -EINVAL);
2523 if (!fid_is_md_operative(new_fid))
2524 GOTO(out_put_old, rc = -EPERM);
2526 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2528 GOTO(out_put_old, rc = PTR_ERR(mnew));
2530 if (!mdt_object_exists(mnew)) {
2531 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2533 "object does not exist");
2534 GOTO(out_put_new, rc = -ENOENT);
2537 if (mdt_object_remote(mnew)) {
2538 struct mdt_body *repbody;
2540 /* Always send rename req to the target child MDT */
2541 repbody = req_capsule_server_get(info->mti_pill,
2543 LASSERT(repbody != NULL);
2544 repbody->mbo_fid1 = *new_fid;
2545 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2546 GOTO(out_put_new, rc = -EXDEV);
2548 /* Before locking the target dir, check we do not replace
2549 * a dir with a non-dir, otherwise it may deadlock with
2550 * link op which tries to create a link in this dir
2551 * back to this non-dir. */
2552 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2553 !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2554 GOTO(out_put_new, rc = -EISDIR);
2556 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2557 mdt_lock_reg_init(lh_oldp, LCK_EX);
2558 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2559 if (mdt_object_remote(msrcdir)) {
2560 /* Enqueue lookup lock from the parent MDT */
2561 rc = mdt_remote_object_lock(info, msrcdir,
2562 mdt_object_fid(mold),
2563 &lh_oldp->mlh_rreg_lh,
2564 lh_oldp->mlh_rreg_mode,
2565 MDS_INODELOCK_LOOKUP,
2568 GOTO(out_put_new, rc);
2570 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2573 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2576 GOTO(out_unlock_old, rc);
2578 /* Check if @msrcdir is subdir of @mnew, before locking child
2579 * to avoid reverse locking. */
2580 if (mtgtdir != msrcdir) {
2581 rc = mdo_is_subdir(info->mti_env,
2582 mdt_object_child(msrcdir), new_fid);
2586 GOTO(out_unlock_old, rc);
2590 /* We used to acquire MDS_INODELOCK_FULL here but we
2591 * can't do this now because a running HSM restore on
2592 * the rename onto victim will hold the layout
2593 * lock. See LU-4002. */
2595 lh_newp = &info->mti_lh[MDT_LH_NEW];
2596 mdt_lock_reg_init(lh_newp, LCK_EX);
2597 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2598 MDS_INODELOCK_LOOKUP |
2599 MDS_INODELOCK_UPDATE,
2602 GOTO(out_unlock_old, rc);
2604 /* get and save version after locking */
2605 mdt_version_get_save(info, mnew, 3);
2606 } else if (rc != -EREMOTE && rc != -ENOENT) {
2607 GOTO(out_put_old, rc);
2609 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2610 mdt_lock_reg_init(lh_oldp, LCK_EX);
2611 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2612 if (mdt_object_remote(msrcdir)) {
2613 /* Enqueue lookup lock from the parent MDT */
2614 rc = mdt_remote_object_lock(info, msrcdir,
2615 mdt_object_fid(mold),
2616 &lh_oldp->mlh_rreg_lh,
2617 lh_oldp->mlh_rreg_mode,
2618 MDS_INODELOCK_LOOKUP,
2621 GOTO(out_put_old, rc);
2623 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2626 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2629 GOTO(out_unlock_old, rc);
2631 mdt_enoent_version_save(info, 3);
2634 /* step 5: rename it */
2635 mdt_reint_init_ma(info, ma);
2637 mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2638 OBD_FAIL_MDS_REINT_RENAME_WRITE);
2641 mutex_lock(&mnew->mot_lov_mutex);
2643 rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2644 mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2645 mnew != NULL ? mdt_object_child(mnew) : NULL,
2646 &rr->rr_tgt_name, ma);
2649 mutex_unlock(&mnew->mot_lov_mutex);
2651 /* handle last link of tgt object */
2653 mdt_counter_incr(req, LPROC_MDT_RENAME);
2655 mdt_handle_last_unlink(info, mnew, ma);
2656 discard = mdt_dom_check_for_discard(info, mnew);
2658 mdt_rename_counter_tally(info, info->mti_mdt, req,
2664 mdt_object_unlock(info, mnew, lh_newp, rc);
2666 mdt_object_unlock(info, mold, lh_oldp, rc);
2668 if (mnew && !discard)
2669 mdt_object_put(info->mti_env, mnew);
2671 mdt_object_put(info->mti_env, mold);
2673 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2674 mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2676 if (lustre_handle_is_used(&rename_lh))
2677 mdt_rename_unlock(&rename_lh);
2679 mdt_object_put(info->mti_env, mtgtdir);
2681 mdt_object_put(info->mti_env, msrcdir);
2683 /* The DoM discard can be done right in the place above where it is
2684 * assigned, meanwhile it is done here after rename unlock due to
2685 * compatibility with old clients, for them the discard blocks
2686 * the main thread until completion. Check LU-11359 for details.
2689 mdt_dom_discard_data(info, mnew);
2690 mdt_object_put(info->mti_env, mnew);
2695 static int mdt_reint_resync(struct mdt_thread_info *info,
2696 struct mdt_lock_handle *lhc)
2698 struct mdt_reint_record *rr = &info->mti_rr;
2699 struct ptlrpc_request *req = mdt_info_req(info);
2700 struct md_attr *ma = &info->mti_attr;
2701 struct mdt_object *mo;
2702 struct ldlm_lock *lease;
2703 struct mdt_body *repbody;
2704 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 rc = mdt_layout_change(info, mo, &layout);
2748 GOTO(out_unlock, rc);
2750 ma->ma_need = MA_INODE;
2752 rc = mdt_attr_get_complex(info, mo, ma);
2754 GOTO(out_unlock, rc);
2756 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2757 mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2761 up_write(&mo->mot_open_sem);
2763 LDLM_LOCK_PUT(lease);
2765 mdt_object_put(info->mti_env, mo);
2767 mdt_client_compatibility(info);
2768 rc2 = mdt_fix_reply(info);
2774 struct mdt_reinter {
2775 int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2776 enum lprocfs_extra_opc mr_extra_opc;
2779 static const struct mdt_reinter mdt_reinters[] = {
2781 .mr_handler = &mdt_reint_setattr,
2782 .mr_extra_opc = MDS_REINT_SETATTR,
2785 .mr_handler = &mdt_reint_create,
2786 .mr_extra_opc = MDS_REINT_CREATE,
2789 .mr_handler = &mdt_reint_link,
2790 .mr_extra_opc = MDS_REINT_LINK,
2793 .mr_handler = &mdt_reint_unlink,
2794 .mr_extra_opc = MDS_REINT_UNLINK,
2797 .mr_handler = &mdt_reint_rename,
2798 .mr_extra_opc = MDS_REINT_RENAME,
2801 .mr_handler = &mdt_reint_open,
2802 .mr_extra_opc = MDS_REINT_OPEN,
2804 [REINT_SETXATTR] = {
2805 .mr_handler = &mdt_reint_setxattr,
2806 .mr_extra_opc = MDS_REINT_SETXATTR,
2809 .mr_handler = &mdt_reint_unlink,
2810 .mr_extra_opc = MDS_REINT_UNLINK,
2813 .mr_handler = &mdt_reint_migrate,
2814 .mr_extra_opc = MDS_REINT_RENAME,
2817 .mr_handler = &mdt_reint_resync,
2818 .mr_extra_opc = MDS_REINT_RESYNC,
2822 int mdt_reint_rec(struct mdt_thread_info *info,
2823 struct mdt_lock_handle *lhc)
2825 const struct mdt_reinter *mr;
2829 if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2832 mr = &mdt_reinters[info->mti_rr.rr_opcode];
2833 if (mr->mr_handler == NULL)
2836 rc = (*mr->mr_handler)(info, lhc);
2838 lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2839 PTLRPC_LAST_CNTR + mr->mr_extra_opc);