X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=5d37de370946553111e0d01a4ca9074574dc8949;hp=62308a8beaab9a5f08eef97a80dcbc7ada3f4e58;hb=4918fe40db262b19093436caca688c75eb632496;hpb=81af631d07084074c6c0346435e7011a6b4582f7 diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 62308a8..5d37de3 100644 --- a/lustre/mdt/mdt_reint.c +++ b/lustre/mdt/mdt_reint.c @@ -47,10 +47,10 @@ #include static inline void mdt_reint_init_ma(struct mdt_thread_info *info, - struct md_attr *ma) + struct md_attr *ma) { ma->ma_need = MA_INODE; - ma->ma_valid = 0; + ma->ma_valid = 0; } /** @@ -59,7 +59,7 @@ static inline void mdt_reint_init_ma(struct mdt_thread_info *info, * Return real version or ENOENT_VERSION if object doesn't exist */ static void mdt_obj_version_get(struct mdt_thread_info *info, - struct mdt_object *o, __u64 *version) + struct mdt_object *o, __u64 *version) { LASSERT(o); @@ -78,20 +78,20 @@ static void mdt_obj_version_get(struct mdt_thread_info *info, * Should be called only during replay. */ static int mdt_version_check(struct ptlrpc_request *req, - __u64 version, int idx) + __u64 version, int idx) { - __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg); - ENTRY; - - if (!exp_connect_vbr(req->rq_export)) - RETURN(0); - - LASSERT(req_is_replay(req)); - /** VBR: version is checked always because costs nothing */ - LASSERT(idx < PTLRPC_NUM_VERSIONS); - /** Sanity check for malformed buffers */ - if (pre_ver == NULL) { - CERROR("No versions in request buffer\n"); + __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg); + + ENTRY; + if (!exp_connect_vbr(req->rq_export)) + RETURN(0); + + LASSERT(req_is_replay(req)); + /** VBR: version is checked always because costs nothing */ + LASSERT(idx < PTLRPC_NUM_VERSIONS); + /** Sanity check for malformed buffers */ + if (pre_ver == NULL) { + CERROR("No versions in request buffer\n"); spin_lock(&req->rq_export->exp_lock); req->rq_export->exp_vbr_failed = 1; spin_unlock(&req->rq_export->exp_lock); @@ -111,18 +111,18 @@ static int mdt_version_check(struct ptlrpc_request *req, * Save pre-versions in reply. */ static void mdt_version_save(struct ptlrpc_request *req, __u64 version, - int idx) + int idx) { - __u64 *reply_ver; + __u64 *reply_ver; - if (!exp_connect_vbr(req->rq_export)) - return; + if (!exp_connect_vbr(req->rq_export)) + return; - LASSERT(!req_is_replay(req)); - LASSERT(req->rq_repmsg != NULL); - reply_ver = lustre_msg_get_versions(req->rq_repmsg); - if (reply_ver) - reply_ver[idx] = version; + LASSERT(!req_is_replay(req)); + LASSERT(req->rq_repmsg != NULL); + reply_ver = lustre_msg_get_versions(req->rq_repmsg); + if (reply_ver) + reply_ver[idx] = version; } /** @@ -131,11 +131,11 @@ static void mdt_version_save(struct ptlrpc_request *req, __u64 version, */ static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx) { - /* save version of file name for replay, it must be ENOENT here */ - if (!req_is_replay(mdt_info_req(info))) { - info->mti_ver[idx] = ENOENT_VERSION; - mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); - } + /* save version of file name for replay, it must be ENOENT here */ + if (!req_is_replay(mdt_info_req(info))) { + info->mti_ver[idx] = ENOENT_VERSION; + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + } } /** @@ -144,44 +144,44 @@ static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx) * Versions are saved in reply only during normal operations not replays. */ void mdt_version_get_save(struct mdt_thread_info *info, - struct mdt_object *mto, int idx) + struct mdt_object *mto, int idx) { - /* don't save versions during replay */ - if (!req_is_replay(mdt_info_req(info))) { - mdt_obj_version_get(info, mto, &info->mti_ver[idx]); - mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); - } + /* don't save versions during replay */ + if (!req_is_replay(mdt_info_req(info))) { + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + } } /** * Get version from disk and check it, no save in reply. */ int mdt_version_get_check(struct mdt_thread_info *info, - struct mdt_object *mto, int idx) + struct mdt_object *mto, int idx) { - /* only check versions during replay */ - if (!req_is_replay(mdt_info_req(info))) - return 0; + /* only check versions during replay */ + if (!req_is_replay(mdt_info_req(info))) + return 0; - mdt_obj_version_get(info, mto, &info->mti_ver[idx]); - return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); } /** * Get version from disk and check if recovery or just save. */ int mdt_version_get_check_save(struct mdt_thread_info *info, - struct mdt_object *mto, int idx) + struct mdt_object *mto, int idx) { - int rc = 0; - - mdt_obj_version_get(info, mto, &info->mti_ver[idx]); - if (req_is_replay(mdt_info_req(info))) - rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], - idx); - else - mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); - return rc; + int rc = 0; + + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + if (req_is_replay(mdt_info_req(info))) + rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], + idx); + else + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + return rc; } /** @@ -190,103 +190,32 @@ int mdt_version_get_check_save(struct mdt_thread_info *info, * This checks version of 'name'. Many reint functions uses 'name' for child not * FID, therefore we need to get object by name and check its version. */ -static int mdt_lookup_version_check(struct mdt_thread_info *info, - struct mdt_object *p, - const struct lu_name *lname, - struct lu_fid *fid, int idx) -{ - int rc, vbrc; - - rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid, - &info->mti_spec); - /* Check version only during replay */ - if (!req_is_replay(mdt_info_req(info))) - return rc; - - info->mti_ver[idx] = ENOENT_VERSION; - if (rc == 0) { - struct mdt_object *child; - child = mdt_object_find(info->mti_env, info->mti_mdt, fid); - if (likely(!IS_ERR(child))) { - mdt_obj_version_get(info, child, &info->mti_ver[idx]); - mdt_object_put(info->mti_env, child); - } - } - vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); - return vbrc ? vbrc : rc; - -} - -static inline int mdt_remote_permission_check(struct mdt_thread_info *info) -{ - struct lu_ucred *uc = mdt_ucred(info); - struct mdt_device *mdt = info->mti_mdt; - - if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) { - if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid && - mdt->mdt_enable_remote_dir_gid != -1) - return -EPERM; - } - - return 0; -} - -/** - * mdt_remote_permission: Check whether the remote operation is permitted, - * - * Only sysadmin can create remote directory / striped directory, - * migrate directory and set default stripedEA on directory, unless - * - * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid. - * - * param[in] info: mdt_thread_info. - * - * retval = 0 remote operation is allowed. - * < 0 remote operation is denied. - */ -static int mdt_remote_permission(struct mdt_thread_info *info) +int mdt_lookup_version_check(struct mdt_thread_info *info, + struct mdt_object *p, + const struct lu_name *lname, + struct lu_fid *fid, int idx) { - struct md_op_spec *spec = &info->mti_spec; - struct lu_attr *attr = &info->mti_attr.ma_attr; - struct obd_export *exp = mdt_info_req(info)->rq_export; - int rc; - - if (info->mti_rr.rr_opcode == REINT_MIGRATE) { - rc = mdt_remote_permission_check(info); - if (rc != 0) - return rc; - } - - if (info->mti_rr.rr_opcode == REINT_CREATE && - (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL && - spec->u.sp_ea.eadatalen != 0)) { - const struct lmv_user_md *lum = spec->u.sp_ea.eadata; - - /* Only new clients can create remote dir( >= 2.4) and - * striped dir(>= 2.6), old client will return -ENOTSUPP */ - if (!mdt_is_dne_client(exp)) - return -ENOTSUPP; + int rc, vbrc; - if (le32_to_cpu(lum->lum_stripe_count) > 1 && - !mdt_is_striped_client(exp)) - return -ENOTSUPP; - - rc = mdt_remote_permission_check(info); - if (rc != 0) - return rc; - } + rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid, + &info->mti_spec); + /* Check version only during replay */ + if (!req_is_replay(mdt_info_req(info))) + return rc; - if (info->mti_rr.rr_opcode == REINT_SETATTR) { - struct md_attr *ma = &info->mti_attr; + info->mti_ver[idx] = ENOENT_VERSION; + if (rc == 0) { + struct mdt_object *child; - if ((ma->ma_valid & MA_LMV)) { - rc = mdt_remote_permission_check(info); - if (rc != 0) - return rc; + child = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (likely(!IS_ERR(child))) { + mdt_obj_version_get(info, child, &info->mti_ver[idx]); + mdt_object_put(info->mti_env, child); } } + vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); + return vbrc ? vbrc : rc; - return 0; } static int mdt_unlock_slaves(struct mdt_thread_info *mti, @@ -322,17 +251,25 @@ static int mdt_unlock_slaves(struct mdt_thread_info *mti, static inline int mdt_object_striped(struct mdt_thread_info *mti, struct mdt_object *obj) { + struct lu_device *bottom_dev; + struct lu_object *bottom_obj; int rc; if (!S_ISDIR(obj->mot_header.loh_attr)) return 0; - rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL, + /* getxattr from bottom obj to avoid reading in shard FIDs */ + bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom); + bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev, + mdt_object_fid(obj), NULL); + if (IS_ERR(bottom_obj)) + return PTR_ERR(bottom_obj); + + rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL, XATTR_NAME_LMV); - if (rc <= 0) - return rc == -ENODATA ? 0 : rc; + lu_object_put(mti->mti_env, bottom_obj); - return 1; + return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc; } /** @@ -362,12 +299,12 @@ static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj, policy); } -static inline int mdt_reint_striped_lock(struct mdt_thread_info *info, - struct mdt_object *o, - struct mdt_lock_handle *lh, - __u64 ibits, - struct ldlm_enqueue_info *einfo, - bool cos_incompat) +int mdt_reint_striped_lock(struct mdt_thread_info *info, + struct mdt_object *o, + struct mdt_lock_handle *lh, + __u64 ibits, + struct ldlm_enqueue_info *einfo, + bool cos_incompat) { int rc; @@ -396,16 +333,143 @@ static inline int mdt_reint_striped_lock(struct mdt_thread_info *info, return rc; } -static inline void -mdt_reint_striped_unlock(struct mdt_thread_info *info, struct mdt_object *o, - struct mdt_lock_handle *lh, - struct ldlm_enqueue_info *einfo, int decref) +void mdt_reint_striped_unlock(struct mdt_thread_info *info, + struct mdt_object *o, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, int decref) { if (einfo->ei_cbdata) mdt_unlock_slaves(info, o, einfo, decref); mdt_object_unlock(info, o, lh, decref); } +static int mdt_restripe(struct mdt_thread_info *info, + struct mdt_object *parent, + const struct lu_name *lname, + const struct lu_fid *tfid, + struct md_op_spec *spec, + struct md_attr *ma) +{ + struct mdt_device *mdt = info->mti_mdt; + struct lu_fid *fid = &info->mti_tmp_fid2; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; + struct lmv_mds_md_v1 *lmv; + struct mdt_object *child; + struct mdt_lock_handle *lhp; + struct mdt_lock_handle *lhc; + struct mdt_body *repbody; + int rc; + + ENTRY; + if (!mdt->mdt_enable_dir_restripe) + RETURN(-EPERM); + + rc = mdt_version_get_check_save(info, parent, 0); + if (rc) + RETURN(rc); + + lhp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lhp, LCK_PW, lname); + rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE, + true); + if (rc) + RETURN(rc); + + rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV); + if (rc) + GOTO(unlock_parent, rc); + + if (ma->ma_valid & MA_LMV) { + /* don't allow restripe if parent dir layout is changing */ + lmv = &ma->ma_lmv->lmv_md_v1; + if (!lmv_is_sane(lmv)) + GOTO(unlock_parent, rc = -EBADF); + + if (lmv_is_layout_changing(lmv)) + GOTO(unlock_parent, rc = -EBUSY); + } + + fid_zero(fid); + rc = mdt_lookup_version_check(info, parent, lname, fid, 1); + if (rc) + GOTO(unlock_parent, rc); + + child = mdt_object_find(info->mti_env, mdt, fid); + if (IS_ERR(child)) + GOTO(unlock_parent, rc = PTR_ERR(child)); + + if (!mdt_object_exists(child)) + GOTO(out_child, rc = -ENOENT); + + if (mdt_object_remote(child)) { + struct mdt_body *repbody; + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + if (!repbody) + GOTO(out_child, rc = -EPROTO); + + repbody->mbo_fid1 = *fid; + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); + GOTO(out_child, rc = -EREMOTE); + } + + /* lock object */ + lhc = &info->mti_lh[MDT_LH_CHILD]; + mdt_lock_reg_init(lhc, LCK_EX); + + /* enqueue object remote LOOKUP lock */ + if (mdt_object_remote(parent)) { + rc = mdt_remote_object_lock(info, parent, fid, + &lhc->mlh_rreg_lh, + lhc->mlh_rreg_mode, + MDS_INODELOCK_LOOKUP, false); + if (rc != ELDLM_OK) + GOTO(out_child, rc); + } + + rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo, + true); + if (rc) + GOTO(unlock_child, rc); + + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child)); + rc = mdt_version_get_check_save(info, child, 1); + if (rc) + GOTO(unlock_child, rc); + + spin_lock(&mdt->mdt_restriper.mdr_lock); + if (child->mot_restriping) { + /* race? */ + spin_unlock(&mdt->mdt_restriper.mdr_lock); + GOTO(unlock_child, rc = -EBUSY); + } + child->mot_restriping = 1; + spin_unlock(&mdt->mdt_restriper.mdr_lock); + + *fid = *tfid; + rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma); + if (rc) + GOTO(restriping_clear, rc); + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + if (!repbody) + GOTO(restriping_clear, rc = -EPROTO); + + mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid); + EXIT; + +restriping_clear: + child->mot_restriping = 0; +unlock_child: + mdt_reint_striped_unlock(info, child, lhc, einfo, rc); +out_child: + mdt_object_put(info->mti_env, child); +unlock_parent: + mdt_object_unlock(info, parent, lhp, rc); + + return rc; +} + /* * VBR: we save three versions in reply: * 0 - parent. Check that parent version is the same during replay. @@ -416,23 +480,62 @@ mdt_reint_striped_unlock(struct mdt_thread_info *info, struct mdt_object *o, */ static int mdt_create(struct mdt_thread_info *info) { - struct mdt_device *mdt = info->mti_mdt; - struct mdt_object *parent; - struct mdt_object *child; - struct mdt_lock_handle *lh; - struct mdt_body *repbody; - struct md_attr *ma = &info->mti_attr; + struct mdt_device *mdt = info->mti_mdt; + struct mdt_object *parent; + struct mdt_object *child; + struct mdt_lock_handle *lh; + struct mdt_body *repbody; + struct md_attr *ma = &info->mti_attr; struct mdt_reint_record *rr = &info->mti_rr; + struct md_op_spec *spec = &info->mti_spec; + bool restripe = false; int rc; - ENTRY; - DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") " - "in "DFID, + ENTRY; + DEBUG_REQ(D_INODE, mdt_info_req(info), + "Create ("DNAME"->"DFID") in "DFID, PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1)); if (!fid_is_md_operative(rr->rr_fid1)) RETURN(-EPERM); + if (S_ISDIR(ma->ma_attr.la_mode) && + spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) { + const struct lmv_user_md *lum = spec->u.sp_ea.eadata; + struct lu_ucred *uc = mdt_ucred(info); + struct obd_export *exp = mdt_info_req(info)->rq_export; + + /* Only new clients can create remote dir( >= 2.4) and + * striped dir(>= 2.6), old client will return -ENOTSUPP + */ + if (!mdt_is_dne_client(exp)) + RETURN(-ENOTSUPP); + + if (le32_to_cpu(lum->lum_stripe_count) > 1) { + if (!mdt_is_striped_client(exp)) + RETURN(-ENOTSUPP); + + if (!mdt->mdt_enable_striped_dir) + RETURN(-EPERM); + } else if (!mdt->mdt_enable_remote_dir) { + RETURN(-EPERM); + } + + if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) && + (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) == + LMV_HASH_TYPE_CRUSH) + RETURN(-EPROTO); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN) && + uc->uc_gid != mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) + RETURN(-EPERM); + + /* restripe if later found dir exists */ + if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT) + restripe = true; + } + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); @@ -442,6 +545,29 @@ static int mdt_create(struct mdt_thread_info *info) if (!mdt_object_exists(parent)) GOTO(put_parent, rc = -ENOENT); + /* + * LU-10235: check if name exists locklessly first to avoid massive + * lock recalls on existing directories. + */ + rc = mdt_lookup_version_check(info, parent, &rr->rr_name, + &info->mti_tmp_fid1, 1); + if (rc == 0) { + if (!restripe) + GOTO(put_parent, rc = -EEXIST); + + rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec, + ma); + } + + /* -ENOENT is expected here */ + if (rc != -ENOENT) + GOTO(put_parent, rc); + + /* save version of file name for replay, it must be ENOENT here */ + mdt_enoent_version_save(info, 1); + + OBD_RACE(OBD_FAIL_MDS_CREATE_RACE); + lh = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name); rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE); @@ -454,30 +580,10 @@ static int mdt_create(struct mdt_thread_info *info) GOTO(unlock_parent, rc); } - /* - * Check child name version during replay. - * During create replay a file may exist with same name. - */ - rc = mdt_lookup_version_check(info, parent, &rr->rr_name, - &info->mti_tmp_fid1, 1); - if (rc == 0) - GOTO(unlock_parent, rc = -EEXIST); - - /* -ENOENT is expected here */ - if (rc != -ENOENT) - GOTO(unlock_parent, rc); - - /* save version of file name for replay, it must be ENOENT here */ - mdt_enoent_version_save(info, 1); - child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2); if (unlikely(IS_ERR(child))) GOTO(unlock_parent, rc = PTR_ERR(child)); - rc = mdt_remote_permission(info); - if (rc != 0) - GOTO(put_child, rc); - ma->ma_need = MA_INODE; ma->ma_valid = 0; @@ -519,7 +625,7 @@ static int mdt_create(struct mdt_thread_info *info) */ if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) { struct mdt_lock_handle *lhc; - struct ldlm_enqueue_info *einfo = &info->mti_einfo; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; bool cos_incompat; rc = mdt_object_striped(info, child); @@ -555,13 +661,14 @@ static int mdt_create(struct mdt_thread_info *info) if (ma->ma_valid & MA_INODE) mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(child)); + EXIT; put_child: mdt_object_put(info->mti_env, child); unlock_parent: mdt_object_unlock(info, parent, lh, rc); put_parent: mdt_object_put(info->mti_env, parent); - RETURN(rc); + return rc; } static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, @@ -571,11 +678,11 @@ static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, int do_vbr = ma->ma_attr.la_valid & (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS); __u64 lockpart = MDS_INODELOCK_UPDATE; - struct ldlm_enqueue_info *einfo = &info->mti_einfo; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; bool cos_incompat; int rc; - ENTRY; + ENTRY; rc = mdt_object_striped(info, mo); if (rc < 0) RETURN(rc); @@ -588,7 +695,8 @@ static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, /* Even though the new MDT will grant PERM lock to the old * client, but the old client will almost ignore that during * So it needs to revoke both LOOKUP and PERM lock here, so - * both new and old client can cancel the dcache */ + * both new and old client can cancel the dcache + */ if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM; @@ -638,9 +746,11 @@ out_unlock: int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, struct md_attr *ma) { + struct lu_ucred *uc = mdt_ucred(info); + cfs_cap_t cap_saved; int rc; - ENTRY; + ENTRY; /* If the file was modified, add the dirty flag */ ma->ma_need = MA_HSM; rc = mdt_attr_get_complex(info, mo, ma); @@ -655,7 +765,13 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) { ma->ma_hsm.mh_flags |= HS_DIRTY; + /* Bump cap so that closes from non-owner writers can + * set the HSM state to dirty. + */ + cap_saved = uc->uc_cap; + uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER); rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm); + uc->uc_cap = cap_saved; if (rc) CERROR("file attribute change error for "DFID": %d\n", PFID(mdt_object_fid(mo)), rc); @@ -665,26 +781,30 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, } static int mdt_reint_setattr(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + struct mdt_lock_handle *lhc) { - struct md_attr *ma = &info->mti_attr; - struct mdt_reint_record *rr = &info->mti_rr; - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_object *mo; - struct mdt_body *repbody; - int rc, rc2; - ENTRY; + struct mdt_device *mdt = info->mti_mdt; + struct md_attr *ma = &info->mti_attr; + struct mdt_reint_record *rr = &info->mti_rr; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_object *mo; + struct mdt_body *repbody; + ktime_t kstart = ktime_get(); + int rc, rc2; - DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1), - (unsigned int)ma->ma_attr.la_valid); + ENTRY; + DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1), + (unsigned int)ma->ma_attr.la_valid); if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); + OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); - if (IS_ERR(mo)) - GOTO(out, rc = PTR_ERR(mo)); + mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1); + if (IS_ERR(mo)) + GOTO(out, rc = PTR_ERR(mo)); if (!mdt_object_exists(mo)) GOTO(out_put, rc = -ENOENT); @@ -692,20 +812,54 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (mdt_object_remote(mo)) GOTO(out_put, rc = -EREMOTE); + ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid; + /* revoke lease lock if size is going to be changed */ + if (unlikely(ma->ma_attr.la_valid & LA_SIZE && + !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) && + atomic_read(&mo->mot_lease_count) > 0)) { + down_read(&mo->mot_open_sem); + + if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */ + lhc = &info->mti_lh[MDT_LH_LOCAL]; + mdt_lock_reg_init(lhc, LCK_CW); + + rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN); + if (rc != 0) { + up_read(&mo->mot_open_sem); + GOTO(out_put, rc); + } + + /* revoke lease lock */ + mdt_object_unlock(info, mo, lhc, 1); + } + up_read(&mo->mot_open_sem); + } + if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) { /* Check write access for the O_TRUNC case */ if (mdt_write_read(mo) < 0) GOTO(out_put, rc = -ETXTBSY); /* LU-10286: compatibility check for FLR. - * Please check the comment in mdt_finish_open() for details */ - if (!exp_connect_flr(info->mti_exp)) { + * Please check the comment in mdt_finish_open() for details + */ + if (!exp_connect_flr(info->mti_exp) || + !exp_connect_overstriping(info->mti_exp)) { rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV); if (rc < 0 && rc != -ENODATA) GOTO(out_put, rc); - if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm)) - GOTO(out_put, rc = -EOPNOTSUPP); + if (!exp_connect_flr(info->mti_exp)) { + if (rc > 0 && + mdt_lmm_is_flr(info->mti_big_lmm)) + GOTO(out_put, rc = -EOPNOTSUPP); + } + + if (!exp_connect_overstriping(info->mti_exp)) { + if (rc > 0 && + mdt_lmm_is_overstriping(info->mti_big_lmm)) + GOTO(out_put, rc = -EOPNOTSUPP); + } } /* For truncate, the file size sent from client @@ -722,39 +876,64 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (ma->ma_valid & MA_LOV) GOTO(out_put, rc = -EPROTO); + /* MDT supports FMD for regular files due to Data-on-MDT */ + if (S_ISREG(lu_object_attr(&mo->mot_obj)) && + ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) + tgt_fmd_update(info->mti_exp, mdt_object_fid(mo), + req->rq_xid); + rc = mdt_attr_set(info, mo, ma); if (rc) GOTO(out_put, rc); } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) && (ma->ma_valid & MA_INODE)) { - struct lu_buf *buf = &info->mti_buf; - struct mdt_lock_handle *lh; + struct lu_buf *buf = &info->mti_buf; + struct lu_ucred *uc = mdt_ucred(info); + struct mdt_lock_handle *lh; + const char *name; + __u64 lockpart = MDS_INODELOCK_XATTR; + + /* reject if either remote or striped dir is disabled */ + if (ma->ma_valid & MA_LMV) { + if (!mdt->mdt_enable_remote_dir || + !mdt->mdt_enable_striped_dir) + GOTO(out_put, rc = -EPERM); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN) && + uc->uc_gid != mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) + GOTO(out_put, rc = -EPERM); + } - rc = mdt_remote_permission(info); - if (rc < 0) - GOTO(out_put, rc); + if (!S_ISDIR(lu_object_attr(&mo->mot_obj))) + GOTO(out_put, rc = -ENOTDIR); if (ma->ma_attr.la_valid != 0) GOTO(out_put, rc = -EPROTO); - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_reg_init(lh, LCK_PW); - - rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR); - if (rc != 0) - GOTO(out_put, rc); - if (ma->ma_valid & MA_LOV) { buf->lb_buf = ma->ma_lmm; buf->lb_len = ma->ma_lmm_size; + name = XATTR_NAME_LOV; } else { - buf->lb_buf = ma->ma_lmv; + struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md; + + buf->lb_buf = lmu; buf->lb_len = ma->ma_lmv_size; + name = XATTR_NAME_DEFAULT_LMV; + /* force client to update dir default layout */ + lockpart |= MDS_INODELOCK_LOOKUP; } + + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lh, LCK_PW); + + rc = mdt_object_lock(info, mo, lh, lockpart); + if (rc != 0) + GOTO(out_put, rc); + rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf, - (ma->ma_valid & MA_LOV) ? - XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV, - 0); + name, 0); mdt_object_unlock(info, mo, lh, rc); if (rc) @@ -767,8 +946,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (ma->ma_attr_flags & MDS_DATA_MODIFIED) rc = mdt_add_dirty_flag(info, mo, ma); - ma->ma_need = MA_INODE; - ma->ma_valid = 0; + ma->ma_need = MA_INODE; + ma->ma_valid = 0; rc = mdt_attr_get_complex(info, mo, ma); if (rc != 0) GOTO(out_put, rc); @@ -780,24 +959,26 @@ out_put: mdt_object_put(info->mti_env, mo); out: if (rc == 0) - mdt_counter_incr(req, LPROC_MDT_SETATTR); + mdt_counter_incr(req, LPROC_MDT_SETATTR, + ktime_us_delta(ktime_get(), kstart)); - mdt_client_compatibility(info); - rc2 = mdt_fix_reply(info); - if (rc == 0) - rc = rc2; - return rc; + mdt_client_compatibility(info); + rc2 = mdt_fix_reply(info); + if (rc == 0) + rc = rc2; + return rc; } static int mdt_reint_create(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + struct mdt_lock_handle *lhc) { - struct ptlrpc_request *req = mdt_info_req(info); - int rc; - ENTRY; + struct ptlrpc_request *req = mdt_info_req(info); + ktime_t kstart = ktime_get(); + int rc; - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) - RETURN(err_serious(-ESTALE)); + ENTRY; + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) + RETURN(err_serious(-ESTALE)); if (info->mti_dlm_req) ldlm_request_cancel(mdt_info_req(info), @@ -808,16 +989,12 @@ static int mdt_reint_create(struct mdt_thread_info *info, switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { case S_IFDIR: - mdt_counter_incr(req, LPROC_MDT_MKDIR); - break; - case S_IFREG: - case S_IFLNK: - case S_IFCHR: - case S_IFBLK: - case S_IFIFO: + case S_IFREG: + case S_IFLNK: + case S_IFCHR: + case S_IFBLK: + case S_IFIFO: case S_IFSOCK: - /* Special file should stay on the same node as parent. */ - mdt_counter_incr(req, LPROC_MDT_MKNOD); break; default: CERROR("%s: Unsupported mode %o\n", @@ -827,6 +1004,16 @@ static int mdt_reint_create(struct mdt_thread_info *info, } rc = mdt_create(info); + if (rc == 0) { + if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR) + mdt_counter_incr(req, LPROC_MDT_MKDIR, + ktime_us_delta(ktime_get(), kstart)); + else + /* Special file should stay on the same node as parent*/ + mdt_counter_incr(req, LPROC_MDT_MKNOD, + ktime_us_delta(ktime_get(), kstart)); + } + RETURN(rc); } @@ -845,14 +1032,14 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, struct mdt_object *mc; struct mdt_lock_handle *parent_lh; struct mdt_lock_handle *child_lh; - struct ldlm_enqueue_info *einfo = &info->mti_einfo; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; __u64 lock_ibits; - bool cos_incompat = false, discard = false; + bool cos_incompat = false; int no_name = 0; + ktime_t kstart = ktime_get(); int rc; ENTRY; - DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1), PNAME(&rr->rr_name)); @@ -902,15 +1089,15 @@ relock: * find the name entry on MDT0 anymore. * In this case, MDT1 only needs to destory the local * directory. - * */ + */ if (mdt_object_remote(mp) && rc == -ENOENT && !fid_is_zero(rr->rr_fid2) && lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) { no_name = 1; *child_fid = *rr->rr_fid2; - } else { + } else { GOTO(unlock_parent, rc); - } + } } if (!fid_is_md_operative(child_fid)) @@ -924,7 +1111,7 @@ relock: if (!cos_incompat) { rc = mdt_object_striped(info, mc); if (rc < 0) - GOTO(unlock_parent, rc = PTR_ERR(mc)); + GOTO(put_child, rc); cos_incompat = rc; if (cos_incompat) { @@ -974,7 +1161,8 @@ relock: * this MDT. Since the unlink will happen on another MDT, * it will release the LOOKUP lock right away. Then What * would happen if another client try to grab the LOOKUP - * lock at the same time with unlink XXX */ + * lock at the same time with unlink XXX + */ mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP); repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); LASSERT(repbody != NULL); @@ -984,7 +1172,8 @@ relock: } /* We used to acquire MDS_INODELOCK_FULL here but we can't do * this now because a running HSM restore on the child (unlink - * victim) will hold the layout lock. See LU-4002. */ + * victim) will hold the layout lock. See LU-4002. + */ lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; if (mdt_object_remote(mp)) { /* Enqueue lookup lock from parent MDT */ @@ -1028,8 +1217,8 @@ relock: rc = mdt_attr_get_complex(info, mc, ma); if (rc) GOTO(out_stat, rc); - } else { - discard = true; + } else if (mdt_dom_check_for_discard(info, mc)) { + mdt_dom_discard_data(info, mc); } mdt_handle_last_unlink(info, mc, ma); @@ -1037,7 +1226,8 @@ out_stat: if (ma->ma_valid & MA_INODE) { switch (ma->ma_attr.la_mode & S_IFMT) { case S_IFDIR: - mdt_counter_incr(req, LPROC_MDT_RMDIR); + mdt_counter_incr(req, LPROC_MDT_RMDIR, + ktime_us_delta(ktime_get(), kstart)); break; case S_IFREG: case S_IFLNK: @@ -1045,7 +1235,8 @@ out_stat: case S_IFBLK: case S_IFIFO: case S_IFSOCK: - mdt_counter_incr(req, LPROC_MDT_UNLINK); + mdt_counter_incr(req, LPROC_MDT_UNLINK, + ktime_us_delta(ktime_get(), kstart)); break; default: LASSERTF(0, "bad file type %o unlinking\n", @@ -1058,14 +1249,13 @@ out_stat: unlock_child: mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc); put_child: - if (discard) - mdt_dom_check_and_discard(info, mc); mdt_object_put(info->mti_env, mc); unlock_parent: mdt_object_unlock(info, mp, parent_lh, rc); put_parent: mdt_object_put(info->mti_env, mp); - return rc; + CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE); + return rc; } /* @@ -1082,21 +1272,28 @@ static int mdt_reint_link(struct mdt_thread_info *info, struct mdt_object *mp; struct mdt_lock_handle *lhs; struct mdt_lock_handle *lhp; + ktime_t kstart = ktime_get(); bool cos_incompat; int rc; - ENTRY; + ENTRY; DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME, PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name)); if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) RETURN(err_serious(-ENOENT)); + if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) { + req->rq_no_reply = 1; + RETURN(err_serious(-ENOENT)); + } + if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); /* Invalid case so return error immediately instead of - * processing it */ + * processing it + */ if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) RETURN(-EPERM); @@ -1172,7 +1369,8 @@ static int mdt_reint_link(struct mdt_thread_info *info, mdt_object_child(ms), &rr->rr_name, ma); if (rc == 0) - mdt_counter_incr(req, LPROC_MDT_LINK); + mdt_counter_incr(req, LPROC_MDT_LINK, + ktime_us_delta(ktime_get(), kstart)); EXIT; unlock_source: @@ -1216,8 +1414,9 @@ static int mdt_pdir_hash_lock(struct mdt_thread_info *info, * going to be sent to client. If it is - mdt_intent_policy() path will * fix it up and turn FL_LOCAL flag off. */ - rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy, - res, dlmflags, &info->mti_exp->exp_handle.h_cookie); + rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, + policy, res, dlmflags, + &info->mti_exp->exp_handle.h_cookie); return rc; } @@ -1228,15 +1427,16 @@ static int mdt_rename_lock(struct mdt_thread_info *info, struct lustre_handle *lh) { int rc; - ENTRY; + ENTRY; if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) { struct lu_fid *fid = &info->mti_tmp_fid1; struct mdt_object *obj; /* XXX, right now, it has to use object API to * enqueue lock cross MDT, so it will enqueue - * rename lock(with LUSTRE_BFL_FID) by root object */ + * rename lock(with LUSTRE_BFL_FID) by root object + */ lu_root_fid(fid); obj = mdt_object_find(info->mti_env, info->mti_mdt, fid); if (IS_ERR(obj)) @@ -1254,15 +1454,16 @@ static int mdt_rename_lock(struct mdt_thread_info *info, __u64 flags = 0; fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id); - memset(policy, 0, sizeof *policy); + memset(policy, 0, sizeof(*policy)); policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; - rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy, - LCK_EX, &flags, ldlm_blocking_ast, - ldlm_completion_ast, NULL, NULL, 0, - LVB_T_NONE, - &info->mti_exp->exp_handle.h_cookie, - lh); + rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id, + LDLM_IBITS, policy, LCK_EX, &flags, + ldlm_blocking_ast, + ldlm_completion_ast, NULL, NULL, 0, + LVB_T_NONE, + &info->mti_exp->exp_handle.h_cookie, + lh); RETURN(rc); } RETURN(rc); @@ -1277,473 +1478,868 @@ static void mdt_rename_unlock(struct lustre_handle *lh) EXIT; } -/* Update object linkEA */ -struct mdt_lock_list { - struct mdt_object *mll_obj; - struct mdt_lock_handle mll_lh; - struct list_head mll_list; -}; - -static void mdt_unlock_list(struct mdt_thread_info *info, - struct list_head *list, int rc) +static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info, + const struct lu_fid *fid, + int idx) { - struct mdt_lock_list *mll; - struct mdt_lock_list *mll2; - - list_for_each_entry_safe(mll, mll2, list, mll_list) { - mdt_object_unlock_put(info, mll->mll_obj, &mll->mll_lh, rc); - list_del(&mll->mll_list); - OBD_FREE_PTR(mll); + struct mdt_object *dir; + int rc; + + ENTRY; + dir = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (IS_ERR(dir)) + RETURN(dir); + + /* check early, the real version will be saved after locking */ + rc = mdt_version_get_check(info, dir, idx); + if (rc) + GOTO(out_put, rc); + + if (!mdt_object_exists(dir)) + GOTO(out_put, rc = -ENOENT); + + if (!S_ISDIR(lu_object_attr(&dir->mot_obj))) + GOTO(out_put, rc = -ENOTDIR); + + RETURN(dir); +out_put: + mdt_object_put(info->mti_env, dir); + return ERR_PTR(rc); +} + +/* + * in case obj is remote obj on its parent, revoke LOOKUP lock, + * herein we don't really check it, just do revoke. + */ +int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info, + struct mdt_object *pobj, + struct mdt_object *obj) +{ + struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL]; + int rc; + + mdt_lock_handle_init(lh); + mdt_lock_reg_init(lh, LCK_EX); + + if (mdt_object_remote(pobj)) { + /* don't bother to check if pobj and obj are on the same MDT. */ + rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_EX, + MDS_INODELOCK_LOOKUP, false); + } else if (mdt_object_remote(obj)) { + struct ldlm_res_id *res = &info->mti_res_id; + union ldlm_policy_data *policy = &info->mti_policy; + __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB | + LDLM_FL_COS_INCOMPAT; + + fid_build_reg_res_name(mdt_object_fid(obj), res); + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP; + rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace, + &lh->mlh_reg_lh, LCK_EX, policy, res, + dlmflags, NULL); + } else { + /* do nothing if both are local */ + return 0; + } + + if (rc != ELDLM_OK) + return rc; + + /* + * TODO, currently we don't save this lock because there is no place to + * hold this lock handle, but to avoid race we need to save this lock. + */ + mdt_object_unlock(info, NULL, lh, 1); + + return 0; +} + +/* + * operation may takes locks of linkea, or directory stripes, group them in + * different list. + */ +struct mdt_sub_lock { + struct mdt_object *msl_obj; + struct mdt_lock_handle msl_lh; + struct list_head msl_linkage; +}; + +static void mdt_unlock_list(struct mdt_thread_info *info, + struct list_head *list, int decref) +{ + struct mdt_sub_lock *msl; + struct mdt_sub_lock *tmp; + + list_for_each_entry_safe(msl, tmp, list, msl_linkage) { + mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref); + list_del(&msl->msl_linkage); + OBD_FREE_PTR(msl); } } -static int mdt_lock_objects_in_linkea(struct mdt_thread_info *info, - struct mdt_object *obj, - struct mdt_object *pobj, - struct list_head *lock_list) +static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info, + struct mdt_object *obj, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks, + int decref) { - struct lu_buf *buf = &info->mti_big_buf; - struct linkea_data ldata = { NULL }; - int count; - int retry_count; - int rc; - ENTRY; + if (mdt_object_remote(obj)) { + mdt_unlock_list(info, slave_locks, decref); + mdt_object_unlock(info, obj, lh, decref); + } else { + mdt_reint_striped_unlock(info, obj, lh, einfo, decref); + } +} + +/* + * lock parents of links, and also check whether total locks don't exceed + * RS_MAX_LOCKS. + * + * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat + * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS + * \retval -ev negative errno upon error + */ +static int mdt_link_parents_lock(struct mdt_thread_info *info, + struct mdt_object *pobj, + const struct md_attr *ma, + struct mdt_object *obj, + struct mdt_lock_handle *lhp, + struct ldlm_enqueue_info *peinfo, + struct list_head *parent_slave_locks, + struct list_head *link_locks) +{ + struct mdt_device *mdt = info->mti_mdt; + struct lu_buf *buf = &info->mti_big_buf; + struct lu_name *lname = &info->mti_name; + struct linkea_data ldata = { NULL }; + bool blocked = false; + int local_lnkp_cnt = 0; + int rc; + ENTRY; if (S_ISDIR(lu_object_attr(&obj->mot_obj))) RETURN(0); - buf = lu_buf_check_and_alloc(buf, PATH_MAX); + buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE); if (buf->lb_buf == NULL) RETURN(-ENOMEM); ldata.ld_buf = buf; rc = mdt_links_read(info, obj, &ldata); - if (rc != 0) { + if (rc) { if (rc == -ENOENT || rc == -ENODATA) rc = 0; RETURN(rc); } - /* ignore the migrating parent(@pobj) */ - retry_count = ldata.ld_leh->leh_reccount - 1; - -again: - LASSERT(ldata.ld_leh != NULL); - ldata.ld_lee = (struct link_ea_entry *)(ldata.ld_leh + 1); - for (count = 0; count < ldata.ld_leh->leh_reccount; count++) { - struct mdt_device *mdt = info->mti_mdt; - struct mdt_object *mdt_pobj; - struct mdt_lock_list *mll; - struct lu_name name; - struct lu_fid fid; + for (linkea_first_entry(&ldata); ldata.ld_lee && !rc; + linkea_next_entry(&ldata)) { + struct mdt_object *lnkp; + struct mdt_sub_lock *msl; + struct lu_fid fid; __u64 ibits; - linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, - &name, &fid); - mdt_pobj = mdt_object_find(info->mti_env, mdt, &fid); - if (IS_ERR(mdt_pobj)) { - CWARN("%s: cannot find obj "DFID": rc = %ld\n", - mdt_obd_name(mdt), PFID(&fid), PTR_ERR(mdt_pobj)); - goto next; + linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname, + &fid); + + /* check if it's also linked to parent */ + if (lu_fid_eq(mdt_object_fid(pobj), &fid)) { + CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n", + PFID(&fid), PNAME(lname)); + /* in case link is remote object, revoke LOOKUP lock */ + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + continue; } - if (!mdt_object_exists(mdt_pobj)) { - CDEBUG(D_INFO, "%s: obj "DFID" does not exist\n", - mdt_obd_name(mdt), PFID(&fid)); - mdt_object_put(info->mti_env, mdt_pobj); - goto next; + lnkp = NULL; + + /* check if it's linked to a stripe of parent */ + if (ma->ma_valid & MA_LMV) { + struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + struct lu_fid *stripe_fid = &info->mti_tmp_fid1; + int j = 0; + + for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) { + fid_le_to_cpu(stripe_fid, + &lmv->lmv_stripe_fids[j]); + if (lu_fid_eq(stripe_fid, &fid)) { + CDEBUG(D_INFO, "skip stripe "DFID + ", reovke "DNAME"\n", + PFID(&fid), PNAME(lname)); + lnkp = mdt_object_find(info->mti_env, + mdt, &fid); + if (IS_ERR(lnkp)) + GOTO(out, rc = PTR_ERR(lnkp)); + break; + } + } + + if (lnkp) { + rc = mdt_revoke_remote_lookup_lock(info, lnkp, + obj); + mdt_object_put(info->mti_env, lnkp); + continue; + } } - /* Check if the object already exists in the list */ - list_for_each_entry(mll, lock_list, mll_list) { - if (mll->mll_obj == mdt_pobj) { - mdt_object_put(info->mti_env, mdt_pobj); - goto next; + /* Check if it's already locked */ + list_for_each_entry(msl, link_locks, msl_linkage) { + if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) { + CDEBUG(D_INFO, + DFID" was locked, revoke "DNAME"\n", + PFID(&fid), PNAME(lname)); + lnkp = msl->msl_obj; + break; } } - if (mdt_pobj == pobj) { - CDEBUG(D_INFO, "%s: skipping parent obj "DFID"\n", - mdt_obd_name(mdt), PFID(&fid)); - mdt_object_put(info->mti_env, mdt_pobj); - goto next; + if (lnkp) { + rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj); + continue; } - OBD_ALLOC_PTR(mll); - if (mll == NULL) { - mdt_object_put(info->mti_env, mdt_pobj); - GOTO(out, rc = -ENOMEM); + CDEBUG(D_INFO, "lock "DFID":"DNAME"\n", + PFID(&fid), PNAME(lname)); + + lnkp = mdt_object_find(info->mti_env, mdt, &fid); + if (IS_ERR(lnkp)) { + CWARN("%s: cannot find obj "DFID": %ld\n", + mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp)); + continue; } - /* Since this needs to lock all of objects in linkea, to avoid - * deadlocks, because it does not follow parent-child order as - * other MDT operation, let's use try_lock here and if the lock - * cannot be gotten because of conflicting locks, then drop all - * current locks, send an AST to the client, and start again. */ - mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name); + if (!mdt_object_exists(lnkp)) { + CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n", + PFID(&fid), PNAME(lname)); + mdt_object_put(info->mti_env, lnkp); + continue; + } + + if (!mdt_object_remote(lnkp)) + local_lnkp_cnt++; + + OBD_ALLOC_PTR(msl); + if (msl == NULL) + GOTO(out, rc = -ENOMEM); + + /* + * we can't follow parent-child lock order like other MD + * operations, use lock_try here to avoid deadlock, if the lock + * cannot be taken, drop all locks taken, revoke the blocked + * one, and continue processing the remaining entries, and in + * the end of the loop restart from beginning. + */ + mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname); ibits = 0; - rc = mdt_object_lock_try(info, mdt_pobj, &mll->mll_lh, &ibits, + rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits, MDS_INODELOCK_UPDATE, true); if (!(ibits & MDS_INODELOCK_UPDATE)) { - mdt_unlock_list(info, lock_list, 0); - CDEBUG(D_INFO, "%s: busy lock on "DFID" %s retry %d\n", - mdt_obd_name(mdt), PFID(&fid), name.ln_name, - retry_count); + CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n", + PFID(&fid), PNAME(lname)); - if (retry_count == 0) { - mdt_object_put(info->mti_env, mdt_pobj); - OBD_FREE_PTR(mll); - GOTO(out, rc = -EBUSY); - } + mdt_unlock_list(info, link_locks, 1); + /* also unlock parent locks to avoid deadlock */ + if (!blocked) + mdt_migrate_object_unlock(info, pobj, lhp, + peinfo, + parent_slave_locks, + 1); + + blocked = true; - mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name); - rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh, + mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname); + rc = mdt_object_lock(info, lnkp, &msl->msl_lh, MDS_INODELOCK_UPDATE); - if (rc != 0) { - mdt_object_put(info->mti_env, mdt_pobj); - OBD_FREE_PTR(mll); + if (rc) { + mdt_object_put(info->mti_env, lnkp); + OBD_FREE_PTR(msl); GOTO(out, rc); } - if (mdt_object_remote(mdt_pobj)) { + if (mdt_object_remote(lnkp)) { struct ldlm_lock *lock; - /* For remote object, Set lock to cb_atomic, + /* + * for remote object, set lock cb_atomic, * so lock can be released in blocking_ast() - * immediately, then the next try_lock will - * have better chance to succeds */ - lock = - ldlm_handle2lock(&mll->mll_lh.mlh_rreg_lh); + * immediately, then the next lock_try will + * have better chance of success. + */ + lock = ldlm_handle2lock( + &msl->msl_lh.mlh_rreg_lh); LASSERT(lock != NULL); lock_res_and_lock(lock); ldlm_set_atomic_cb(lock); unlock_res_and_lock(lock); LDLM_LOCK_PUT(lock); } - mdt_object_unlock_put(info, mdt_pobj, &mll->mll_lh, rc); - OBD_FREE_PTR(mll); - retry_count--; - goto again; + + mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1); + OBD_FREE_PTR(msl); + continue; } - rc = 0; - INIT_LIST_HEAD(&mll->mll_list); - mll->mll_obj = mdt_pobj; - list_add_tail(&mll->mll_list, lock_list); -next: - ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee + - ldata.ld_reclen); + + INIT_LIST_HEAD(&msl->msl_linkage); + msl->msl_obj = lnkp; + list_add_tail(&msl->msl_linkage, link_locks); + + rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj); } + + if (blocked) + GOTO(out, rc = -EBUSY); + + EXIT; out: - if (rc != 0) - mdt_unlock_list(info, lock_list, rc); - RETURN(rc); + if (rc) + mdt_unlock_list(info, link_locks, rc); + else if (local_lnkp_cnt > RS_MAX_LOCKS - 6) + /* + * parent may have 3 local objects: master object and 2 stripes + * (if it's being migrated too); source may have 2 local + * objects: master and 1 stripe; target has 1 local object. + */ + rc = 1; + return rc; } -/* migrate files from one MDT to another MDT */ -static int mdt_reint_migrate_internal(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) +static int mdt_lock_remote_slaves(struct mdt_thread_info *info, + struct mdt_object *obj, + const struct md_attr *ma, + struct list_head *slave_locks) { - struct mdt_reint_record *rr = &info->mti_rr; - struct md_attr *ma = &info->mti_attr; - struct mdt_object *msrcdir; - struct mdt_object *mold; - struct mdt_object *mnew = NULL; - struct mdt_lock_handle *lh_dirp; - struct mdt_lock_handle *lh_childp; - struct mdt_lock_handle *lh_tgtp = NULL; - struct lu_fid *old_fid = &info->mti_tmp_fid1; - struct list_head lock_list; - __u64 lock_ibits; - struct ldlm_lock *lease = NULL; - bool lock_open_sem = false; - int rc; + struct mdt_device *mdt = info->mti_mdt; + const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + struct lu_fid *fid = &info->mti_tmp_fid1; + struct mdt_object *slave; + struct mdt_sub_lock *msl; + int i; + int rc; + ENTRY; + LASSERT(mdt_object_remote(obj)); + LASSERT(ma->ma_valid & MA_LMV); + LASSERT(lmv); - CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1), - PNAME(&rr->rr_name), PFID(rr->rr_fid2)); + if (!lmv_is_sane(lmv)) + RETURN(-EINVAL); - /* 1: lock the source dir. */ - msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); - if (IS_ERR(msrcdir)) { - CDEBUG(D_OTHER, "%s: cannot find source dir "DFID" : rc = %d\n", - mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1), - (int)PTR_ERR(msrcdir)); - RETURN(PTR_ERR(msrcdir)); + for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) { + fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]); + + if (!fid_is_sane(fid)) + continue; + + slave = mdt_object_find(info->mti_env, mdt, fid); + if (IS_ERR(slave)) + GOTO(out, rc = PTR_ERR(slave)); + + OBD_ALLOC_PTR(msl); + if (!msl) { + mdt_object_put(info->mti_env, slave); + GOTO(out, rc = -ENOMEM); + } + + mdt_lock_reg_init(&msl->msl_lh, LCK_EX); + rc = mdt_reint_object_lock(info, slave, &msl->msl_lh, + MDS_INODELOCK_UPDATE, true); + if (rc) { + OBD_FREE_PTR(msl); + mdt_object_put(info->mti_env, slave); + GOTO(out, rc); + } + + INIT_LIST_HEAD(&msl->msl_linkage); + msl->msl_obj = slave; + list_add_tail(&msl->msl_linkage, slave_locks); } + EXIT; - lh_dirp = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name); - rc = mdt_reint_object_lock(info, msrcdir, lh_dirp, MDS_INODELOCK_UPDATE, - true); +out: if (rc) - GOTO(out_put_parent, rc); + mdt_unlock_list(info, slave_locks, rc); + return rc; +} - if (!mdt_object_remote(msrcdir)) { - rc = mdt_version_get_check_save(info, msrcdir, 0); - if (rc) - GOTO(out_unlock_parent, rc); +/* lock parent and its stripes */ +static int mdt_migrate_parent_lock(struct mdt_thread_info *info, + struct mdt_object *obj, + const struct md_attr *ma, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks) +{ + int rc; + + if (mdt_object_remote(obj)) { + rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_PW, + MDS_INODELOCK_UPDATE, false); + if (rc != ELDLM_OK) + return rc; + + /* + * if obj is remote and striped, lock its stripes explicitly + * because it's not striped in LOD layer on this MDT. + */ + if (ma->ma_valid & MA_LMV) { + rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks); + if (rc) + mdt_object_unlock(info, obj, lh, rc); + } + } else { + rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE, + einfo, true); } - /* 2: sanity check and find the object to be migrated. */ - fid_zero(old_fid); - rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2); - if (rc != 0) - GOTO(out_unlock_parent, rc); + return rc; +} - if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2)) - GOTO(out_unlock_parent, rc = -EINVAL); +/* + * in migration, object may be remote, and we need take full lock of it and its + * stripes if it's directory, besides, object may be a remote object on its + * parent, revoke its LOOKUP lock on where its parent is located. + */ +static int mdt_migrate_object_lock(struct mdt_thread_info *info, + struct mdt_object *pobj, + struct mdt_object *obj, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks) +{ + int rc; - if (!fid_is_md_operative(old_fid)) - GOTO(out_unlock_parent, rc = -EPERM); + if (mdt_object_remote(obj)) { + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + if (rc) + return rc; - if (lu_fid_eq(old_fid, &info->mti_mdt->mdt_md_root_fid)) - GOTO(out_unlock_parent, rc = -EPERM); + rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_EX, + MDS_INODELOCK_FULL, false); + if (rc != ELDLM_OK) + return rc; - mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid); - if (IS_ERR(mold)) - GOTO(out_unlock_parent, rc = PTR_ERR(mold)); + /* + * if obj is remote and striped, lock its stripes explicitly + * because it's not striped in LOD layer on this MDT. + */ + if (S_ISDIR(lu_object_attr(&obj->mot_obj))) { + struct md_attr *ma = &info->mti_attr; - if (!mdt_object_exists(mold)) { - LU_OBJECT_DEBUG(D_INODE, info->mti_env, - &mold->mot_obj, - "object does not exist"); - GOTO(out_put_child, rc = -ENOENT); - } + rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV); + if (rc) { + mdt_object_unlock(info, obj, lh, rc); + return rc; + } - if (mdt_object_remote(mold)) { - CDEBUG(D_OTHER, "%s: source "DFID" is on the remote MDT\n", - mdt_obd_name(info->mti_mdt), PFID(old_fid)); - GOTO(out_put_child, rc = -EREMOTE); + if (ma->ma_valid & MA_LMV) { + rc = mdt_lock_remote_slaves(info, obj, ma, + slave_locks); + if (rc) + mdt_object_unlock(info, obj, lh, rc); + } + } + } else { + if (mdt_object_remote(pobj)) { + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + if (rc) + return rc; + } + + rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL, + einfo, true); } - if (S_ISREG(lu_object_attr(&mold->mot_obj)) && - !mdt_object_remote(msrcdir)) { - CDEBUG(D_OTHER, "%s: parent "DFID" is still on the same" - " MDT, which should be migrated first:" - " rc = %d\n", mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(msrcdir)), -EPERM); - GOTO(out_put_child, rc = -EPERM); + return rc; +} + +/* + * lookup source by name, if parent is striped directory, we need to find the + * corresponding stripe where source is located, and then lookup there. + * + * besides, if parent is migrating too, and file is already in target stripe, + * this should be a redo of 'lfs migrate' on client side. + */ +static int mdt_migrate_lookup(struct mdt_thread_info *info, + struct mdt_object *pobj, + const struct md_attr *ma, + const struct lu_name *lname, + struct mdt_object **spobj, + struct mdt_object **sobj) +{ + const struct lu_env *env = info->mti_env; + struct lu_fid *fid = &info->mti_tmp_fid1; + struct mdt_object *stripe; + int rc; + + if (ma->ma_valid & MA_LMV) { + /* if parent is striped, lookup on corresponding stripe */ + struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + + if (!lmv_is_sane(lmv)) + return -EBADF; + + rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name, + lname->ln_namelen); + if (rc < 0) + return rc; + + fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]); + + stripe = mdt_object_find(env, info->mti_mdt, fid); + if (IS_ERR(stripe)) + return PTR_ERR(stripe); + + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid, + &info->mti_spec); + if (rc == -ENOENT && lmv_is_layout_changing(lmv)) { + /* + * if parent layout is changeing, and lookup child + * failed on source stripe, lookup again on target + * stripe, if it exists, it means previous migration + * was interrupted, and current file was migrated + * already. + */ + mdt_object_put(env, stripe); + + rc = lmv_name_to_stripe_index(lmv, lname->ln_name, + lname->ln_namelen); + if (rc < 0) + return rc; + + fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]); + + stripe = mdt_object_find(env, info->mti_mdt, fid); + if (IS_ERR(stripe)) + return PTR_ERR(stripe); + + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(stripe), lname, + fid, &info->mti_spec); + mdt_object_put(env, stripe); + return rc ?: -EALREADY; + } else if (rc) { + mdt_object_put(env, stripe); + return rc; + } + } else { + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid, + &info->mti_spec); + if (rc) + return rc; + + stripe = pobj; + mdt_object_get(env, stripe); } - rc = mdt_remote_permission(info); - if (rc != 0) - GOTO(out_put_child, rc); + *spobj = stripe; - /* 3: iterate the linkea of the object and lock all of the objects */ - INIT_LIST_HEAD(&lock_list); - rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list); - if (rc != 0) - GOTO(out_put_child, rc); + *sobj = mdt_object_find(env, info->mti_mdt, fid); + if (IS_ERR(*sobj)) { + mdt_object_put(env, stripe); + rc = PTR_ERR(*sobj); + *spobj = NULL; + *sobj = NULL; + } - if (info->mti_spec.sp_migrate_close) { - struct close_data *data; - struct mdt_body *repbody; - bool lease_broken = false; + return rc; +} - if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH, +/* end lease and close file for regular file */ +static int mdd_migrate_close(struct mdt_thread_info *info, + struct mdt_object *obj) +{ + struct close_data *data; + struct mdt_body *repbody; + struct ldlm_lock *lease; + int rc; + int rc2; + + rc = -EPROTO; + if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH, RCL_CLIENT) || - !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA, + !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA, RCL_CLIENT)) - GOTO(out_lease, rc = -EPROTO); + goto close; - data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); - if (data == NULL) - GOTO(out_lease, rc = -EPROTO); + data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); + if (!data) + goto close; - lease = ldlm_handle2lock(&data->cd_handle); - if (lease == NULL) - GOTO(out_lease, rc = -ESTALE); + rc = -ESTALE; + lease = ldlm_handle2lock(&data->cd_handle); + if (!lease) + goto close; - /* try to hold open_sem so that nobody else can open the file */ - if (!down_write_trylock(&mold->mot_open_sem)) { - ldlm_lock_cancel(lease); - GOTO(out_lease, rc = -EBUSY); - } + /* check if the lease was already canceled */ + lock_res_and_lock(lease); + rc = ldlm_is_cancel(lease); + unlock_res_and_lock(lease); - lock_open_sem = true; - /* Check if the lease open lease has already canceled */ - lock_res_and_lock(lease); - lease_broken = ldlm_is_cancel(lease); - unlock_res_and_lock(lease); + if (rc) { + rc = -EAGAIN; + LDLM_DEBUG(lease, DFID" lease broken", + PFID(mdt_object_fid(obj))); + } - LDLM_DEBUG(lease, DFID " lease broken? %d", - PFID(mdt_object_fid(mold)), lease_broken); + /* + * cancel server side lease, client side counterpart should have been + * cancelled, it's okay to cancel it now as we've held mot_open_sem. + */ + ldlm_lock_cancel(lease); + ldlm_reprocess_all(lease->l_resource, lease); + LDLM_LOCK_PUT(lease); - /* Cancel server side lease. Client side counterpart should - * have been cancelled. It's okay to cancel it now as we've - * held mot_open_sem. */ - ldlm_lock_cancel(lease); +close: + rc2 = mdt_close_internal(info, mdt_info_req(info), NULL); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; - if (lease_broken) - GOTO(out_lease, rc = -EAGAIN); -out_lease: - rc = mdt_close_internal(info, mdt_info_req(info), NULL); - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; - if (rc != 0) - GOTO(out_unlock_list, rc); - } + return rc ?: rc2; +} - /* 4: lock of the object migrated object */ - lh_childp = &info->mti_lh[MDT_LH_OLD]; - mdt_lock_reg_init(lh_childp, LCK_EX); - lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE | - MDS_INODELOCK_LAYOUT; - if (mdt_object_remote(msrcdir)) { - /* Enqueue lookup lock from the parent MDT */ - rc = mdt_remote_object_lock(info, msrcdir, mdt_object_fid(mold), - &lh_childp->mlh_rreg_lh, - lh_childp->mlh_rreg_mode, - MDS_INODELOCK_LOOKUP, false); - if (rc != ELDLM_OK) - GOTO(out_unlock_list, rc); +/* + * migrate file in below steps: + * 1. lock parent and its stripes + * 2. lookup source by name + * 3. lock parents of source links if source is not directory + * 4. reject if source is in HSM + * 5. take source open_sem and close file if source is regular file + * 6. lock source and its stripes if it's directory + * 7. lock target so subsequent change to it can trigger COS + * 8. migrate file + * 9. unlock above locks + * 10. sync device if source has links + */ +int mdt_reint_migrate(struct mdt_thread_info *info, + struct mdt_lock_handle *unused) +{ + const struct lu_env *env = info->mti_env; + struct mdt_device *mdt = info->mti_mdt; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_reint_record *rr = &info->mti_rr; + struct lu_ucred *uc = mdt_ucred(info); + struct md_attr *ma = &info->mti_attr; + struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0]; + struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1]; + struct mdt_object *pobj; + struct mdt_object *spobj = NULL; + struct mdt_object *sobj = NULL; + struct mdt_object *tobj; + struct lustre_handle rename_lh = { 0 }; + struct mdt_lock_handle *lhp; + struct mdt_lock_handle *lhs; + struct mdt_lock_handle *lht; + LIST_HEAD(parent_slave_locks); + LIST_HEAD(child_slave_locks); + LIST_HEAD(link_locks); + int lock_retries = 5; + bool open_sem_locked = false; + bool do_sync = false; + int rc; - lock_ibits &= ~MDS_INODELOCK_LOOKUP; - } + ENTRY; + CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1), + PNAME(&rr->rr_name), PFID(rr->rr_fid2)); - rc = mdt_reint_object_lock(info, mold, lh_childp, lock_ibits, true); - if (rc != 0) - GOTO(out_unlock_child, rc); + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); - /* Migration is incompatible with HSM. */ - ma->ma_need = MA_HSM; - ma->ma_valid = 0; - rc = mdt_attr_get_complex(info, mold, ma); - if (rc != 0) - GOTO(out_unlock_child, rc); - - if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) { - rc = -ENOSYS; - CDEBUG(D_OTHER, - "%s: cannot migrate HSM archived file "DFID": rc = %d\n", - mdt_obd_name(info->mti_mdt), PFID(old_fid), rc); - GOTO(out_unlock_child, rc); - } + if (!fid_is_md_operative(rr->rr_fid1) || + !fid_is_md_operative(rr->rr_fid2)) + RETURN(-EPERM); - ma->ma_need = MA_LMV; - ma->ma_valid = 0; - ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf; - ma->ma_lmv_size = sizeof(info->mti_xattr_buf); - rc = mdt_stripe_get(info, mold, ma, XATTR_NAME_LMV); - if (rc != 0) - GOTO(out_unlock_child, rc); - - if ((ma->ma_valid & MA_LMV)) { - struct lmv_mds_md_v1 *lmm1; - - lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv); - lmm1 = &ma->ma_lmv->lmv_md_v1; - if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) { - CDEBUG(D_OTHER, "%s: can not migrate striped dir "DFID - ": rc = %d\n", mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(mold)), -EPERM); - GOTO(out_unlock_child, rc = -EPERM); - } + /* don't allow migrate . or .. */ + if (lu_name_is_dot_or_dotdot(&rr->rr_name)) + RETURN(-EBUSY); - if (!fid_is_sane(&lmm1->lmv_stripe_fids[1])) - GOTO(out_unlock_child, rc = -EINVAL); + if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration) + RETURN(-EPERM); - mnew = mdt_object_find(info->mti_env, info->mti_mdt, - &lmm1->lmv_stripe_fids[1]); - if (IS_ERR(mnew)) - GOTO(out_unlock_child, rc = PTR_ERR(mnew)); - - if (!mdt_object_remote(mnew)) { - CDEBUG(D_OTHER, - "%s: "DFID" being migrated is on this MDT:" - " rc = %d\n", mdt_obd_name(info->mti_mdt), - PFID(rr->rr_fid2), -EPERM); - GOTO(out_put_new, rc = -EPERM); - } + if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) && + uc->uc_gid != mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) + RETURN(-EPERM); - lh_tgtp = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(lh_tgtp, LCK_EX); - rc = mdt_remote_object_lock(info, mnew, - mdt_object_fid(mnew), - &lh_tgtp->mlh_rreg_lh, - lh_tgtp->mlh_rreg_mode, - MDS_INODELOCK_UPDATE, false); + /* + * Note: do not enqueue rename lock for replay request, because + * if other MDT holds rename lock, but being blocked to wait for + * this MDT to finish its recovery, and the failover MDT can not + * get rename lock, which will cause deadlock. + * + * req is NULL if this is called by directory auto-split. + */ + if (req && !req_is_replay(req)) { + rc = mdt_rename_lock(info, &rename_lh); if (rc != 0) { - lh_tgtp = NULL; - GOTO(out_put_new, rc); - } - } else { - mnew = mdt_object_find(info->mti_env, info->mti_mdt, - rr->rr_fid2); - if (IS_ERR(mnew)) - GOTO(out_unlock_child, rc = PTR_ERR(mnew)); - if (!mdt_object_remote(mnew)) { - CDEBUG(D_OTHER, "%s: Migration "DFID" is on this MDT:" - " rc = %d\n", mdt_obd_name(info->mti_mdt), - PFID(rr->rr_fid2), -EXDEV); - GOTO(out_put_new, rc = -EXDEV); + CERROR("%s: can't lock FS for rename: rc = %d\n", + mdt_obd_name(info->mti_mdt), rc); + RETURN(rc); } } - /* 5: migrate it */ - mdt_reint_init_ma(info, ma); + /* pobj is master object of parent */ + pobj = mdt_object_find(env, mdt, rr->rr_fid1); + if (IS_ERR(pobj)) + GOTO(unlock_rename, rc = PTR_ERR(pobj)); - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_RENAME_WRITE); + if (req) { + rc = mdt_version_get_check(info, pobj, 0); + if (rc) + GOTO(put_parent, rc); + } - rc = mdo_migrate(info->mti_env, mdt_object_child(msrcdir), - mdt_object_child(mold), &rr->rr_name, - mdt_object_child(mnew), ma); - if (rc != 0) - GOTO(out_unlock_new, rc); + if (!mdt_object_exists(pobj)) + GOTO(put_parent, rc = -ENOENT); -out_unlock_new: - if (lh_tgtp != NULL) - mdt_object_unlock(info, mnew, lh_tgtp, rc); -out_put_new: - if (mnew) - mdt_object_put(info->mti_env, mnew); -out_unlock_child: - mdt_object_unlock(info, mold, lh_childp, rc); -out_unlock_list: - /* we don't really modify linkea objects, so we can safely decref these - * locks, and this can avoid saving them as COS locks, which may prevent - * subsequent migrate. */ - mdt_unlock_list(info, &lock_list, 1); - if (lease != NULL) { - ldlm_reprocess_all(lease->l_resource); - LDLM_LOCK_PUT(lease); + if (!S_ISDIR(lu_object_attr(&pobj->mot_obj))) + GOTO(put_parent, rc = -ENOTDIR); + + rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV); + if (rc) + GOTO(put_parent, rc); + +lock_parent: + /* lock parent object */ + lhp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lhp, LCK_PW); + rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo, + &parent_slave_locks); + if (rc) + GOTO(put_parent, rc); + + /* + * spobj is the corresponding stripe against name if pobj is striped + * directory, which is the real parent, and no need to lock, because + * we've taken full lock of pobj. + */ + rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj); + if (rc) + GOTO(unlock_parent, rc); + + /* lock parents of source links, and revoke LOOKUP lock of links */ + rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo, + &parent_slave_locks, &link_locks); + if (rc == -EBUSY && lock_retries-- > 0) { + mdt_object_put(env, sobj); + mdt_object_put(env, spobj); + goto lock_parent; } - if (lock_open_sem) - up_write(&mold->mot_open_sem); -out_put_child: - mdt_object_put(info->mti_env, mold); -out_unlock_parent: - mdt_object_unlock(info, msrcdir, lh_dirp, rc); -out_put_parent: - mdt_object_put(info->mti_env, msrcdir); + if (rc < 0) + GOTO(put_source, rc); - RETURN(rc); -} + /* + * RS_MAX_LOCKS is the limit of number of locks that can be saved along + * with one request, if total lock count exceeds this limit, we will + * drop all locks after migration, and synchronous device in the end. + */ + do_sync = rc; -static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info, - const struct lu_fid *fid, - int idx) -{ - struct mdt_object *dir; - int rc; - ENTRY; + /* TODO: DoM migration is not supported, migrate dirent only */ + if (S_ISREG(lu_object_attr(&sobj->mot_obj))) { + rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV); + if (rc) + GOTO(unlock_links, rc); - dir = mdt_object_find(info->mti_env, info->mti_mdt, fid); - if (IS_ERR(dir)) - RETURN(dir); + if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm)) + info->mti_spec.sp_migrate_nsonly = 1; + } - /* check early, the real version will be saved after locking */ - rc = mdt_version_get_check(info, dir, idx); + /* if migration HSM is allowed */ + if (!mdt->mdt_opts.mo_migrate_hsm_allowed) { + ma->ma_need = MA_HSM; + ma->ma_valid = 0; + rc = mdt_attr_get_complex(info, sobj, ma); + if (rc) + GOTO(unlock_links, rc); + + if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) + GOTO(unlock_links, rc = -EOPNOTSUPP); + } + + /* end lease and close file for regular file */ + if (info->mti_spec.sp_migrate_close) { + /* try to hold open_sem so that nobody else can open the file */ + if (!down_write_trylock(&sobj->mot_open_sem)) { + /* close anyway */ + mdd_migrate_close(info, sobj); + GOTO(unlock_links, rc = -EBUSY); + } else { + open_sem_locked = true; + rc = mdd_migrate_close(info, sobj); + if (rc) + GOTO(unlock_open_sem, rc); + } + } + + /* lock source */ + lhs = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lhs, LCK_EX); + rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo, + &child_slave_locks); if (rc) - GOTO(out_put, rc); + GOTO(unlock_open_sem, rc); - RETURN(dir); -out_put: - mdt_object_put(info->mti_env, dir); - return ERR_PTR(rc); + /* lock target */ + tobj = mdt_object_find(env, mdt, rr->rr_fid2); + if (IS_ERR(tobj)) + GOTO(unlock_source, rc = PTR_ERR(tobj)); + + lht = &info->mti_lh[MDT_LH_NEW]; + mdt_lock_reg_init(lht, LCK_EX); + rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true); + if (rc) + GOTO(put_target, rc); + + /* Don't do lookup sanity check. We know name doesn't exist. */ + info->mti_spec.sp_cr_lookup = 0; + info->mti_spec.sp_feat = &dt_directory_features; + + rc = mdo_migrate(env, mdt_object_child(pobj), + mdt_object_child(sobj), &rr->rr_name, + mdt_object_child(tobj), + &info->mti_spec, ma); + if (!rc) + lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats, + LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC); + EXIT; + + mdt_object_unlock(info, tobj, lht, rc); +put_target: + mdt_object_put(env, tobj); +unlock_source: + mdt_migrate_object_unlock(info, sobj, lhs, seinfo, + &child_slave_locks, rc); +unlock_open_sem: + if (open_sem_locked) + up_write(&sobj->mot_open_sem); +unlock_links: + /* if we've got too many locks to save into RPC, + * then just commit before the locks are released + */ + if (!rc && do_sync) + mdt_device_sync(env, mdt); + mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc); +put_source: + mdt_object_put(env, sobj); + mdt_object_put(env, spobj); +unlock_parent: + mdt_migrate_object_unlock(info, pobj, lhp, peinfo, + &parent_slave_locks, rc); +put_parent: + mdt_object_put(env, pobj); +unlock_rename: + if (lustre_handle_is_used(&rename_lh)) + mdt_rename_unlock(&rename_lh); + + return rc; } static int mdt_object_lock_save(struct mdt_thread_info *info, @@ -1820,10 +2416,6 @@ static int mdt_rename_determine_lock_order(struct mdt_thread_info *info, return 0; /* check whether sobj and tobj are sibling stripes */ - ma->ma_need = MA_LMV; - ma->ma_valid = 0; - ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf; - ma->ma_lmv_size = sizeof(info->mti_xattr_buf); rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV); if (rc) return rc; @@ -1857,24 +2449,79 @@ static int mdt_rename_determine_lock_order(struct mdt_thread_info *info, } /* + * lock rename source object. + * + * Both source and source parent may be remote, and source may be a remote + * object on source parent, to avoid overriding lock handle, store remote + * LOOKUP lock separately in @lhr. + * + * \retval 0 on success + * \retval -ev negative errno upon error + */ +static int mdt_rename_source_lock(struct mdt_thread_info *info, + struct mdt_object *parent, + struct mdt_object *child, + struct mdt_lock_handle *lhc, + struct mdt_lock_handle *lhr, + __u64 ibits, + bool cos_incompat) +{ + int rc; + + rc = mdt_is_remote_object(info, parent, child); + if (rc < 0) + return rc; + + if (rc) { + /* enqueue remote LOOKUP lock from the parent MDT */ + __u64 rmt_ibits = MDS_INODELOCK_LOOKUP; + + if (mdt_object_remote(parent)) { + rc = mdt_remote_object_lock(info, parent, + mdt_object_fid(child), + &lhr->mlh_rreg_lh, + lhr->mlh_rreg_mode, + rmt_ibits, false); + if (rc != ELDLM_OK) + return rc; + } else { + LASSERT(mdt_object_remote(child)); + rc = mdt_object_local_lock(info, child, lhr, + &rmt_ibits, 0, true); + if (rc < 0) + return rc; + } + + ibits &= ~MDS_INODELOCK_LOOKUP; + } + + if (mdt_object_remote(child)) { + rc = mdt_remote_object_lock(info, child, mdt_object_fid(child), + &lhc->mlh_rreg_lh, + lhc->mlh_rreg_mode, + ibits, false); + if (rc == ELDLM_OK) + rc = 0; + } else { + rc = mdt_reint_object_lock(info, child, lhc, ibits, + cos_incompat); + } + + if (!rc) + mdt_object_unlock(info, child, lhr, rc); + + return rc; +} + +/* * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent; * 2 - srcdir child; 3 - tgtdir child. * Update on disk version of srcdir child. */ -/** - * For DNE phase I, only these renames are allowed - * mv src_p/src_c tgt_p/tgt_c - * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT. - * 2. src_p and tgt_p are same directory, and tgt_c does not - * exists. In this case, all of modification will happen - * in the MDT where ithesource parent is, only one remote - * update is needed, i.e. set c_time/m_time on the child. - * And tgt_c will be still in the same MDT as the original - * src_c. - */ -static int mdt_reint_rename_internal(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) +static int mdt_reint_rename(struct mdt_thread_info *info, + struct mdt_lock_handle *unused) { + struct mdt_device *mdt = info->mti_mdt; struct mdt_reint_record *rr = &info->mti_rr; struct md_attr *ma = &info->mti_attr; struct ptlrpc_request *req = mdt_info_req(info); @@ -1882,24 +2529,34 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info, struct mdt_object *mtgtdir = NULL; struct mdt_object *mold; struct mdt_object *mnew = NULL; + struct lustre_handle rename_lh = { 0 }; struct mdt_lock_handle *lh_srcdirp; struct mdt_lock_handle *lh_tgtdirp; struct mdt_lock_handle *lh_oldp = NULL; + struct mdt_lock_handle *lh_rmt = NULL; struct mdt_lock_handle *lh_newp = NULL; struct lu_fid *old_fid = &info->mti_tmp_fid1; struct lu_fid *new_fid = &info->mti_tmp_fid2; __u64 lock_ibits; - bool reverse = false; - bool cos_incompat, discard = false; + bool reverse = false, discard = false; + bool cos_incompat; + ktime_t kstart = ktime_get(); int rc; - ENTRY; + ENTRY; DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME, PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name)); + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); + + if (!fid_is_md_operative(rr->rr_fid1) || + !fid_is_md_operative(rr->rr_fid2)) + RETURN(-EPERM); + /* find both parents. */ - msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0); + msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0); if (IS_ERR(msrcdir)) RETURN(PTR_ERR(msrcdir)); @@ -1909,21 +2566,49 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info, mtgtdir = msrcdir; mdt_object_get(info->mti_env, mtgtdir); } else { - mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1); + mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1); if (IS_ERR(mtgtdir)) GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir)); } + /* + * Note: do not enqueue rename lock for replay request, because + * if other MDT holds rename lock, but being blocked to wait for + * this MDT to finish its recovery, and the failover MDT can not + * get rename lock, which will cause deadlock. + */ + if (!req_is_replay(req)) { + /* + * Normally rename RPC is handled on the MDT with the target + * directory (if target exists, it's on the MDT with the + * target), if the source directory is remote, it's a hint that + * source is remote too (this may not be true, but it won't + * cause any issue), return -EXDEV early to avoid taking + * rename_lock. + */ + if (!mdt->mdt_enable_remote_rename && + mdt_object_remote(msrcdir)) + GOTO(out_put_tgtdir, rc = -EXDEV); + + rc = mdt_rename_lock(info, &rename_lh); + if (rc != 0) { + CERROR("%s: can't lock FS for rename: rc = %d\n", + mdt_obd_name(mdt), rc); + GOTO(out_put_tgtdir, rc); + } + } + rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir); if (rc < 0) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); reverse = rc; /* source needs to be looked up after locking source parent, otherwise * this rename may race with unlink source, and cause rename hang, see * sanityn.sh 55b, so check parents first, if later we found source is - * remote, relock parents. */ + * remote, relock parents. + */ cos_incompat = (mdt_object_remote(msrcdir) || mdt_object_remote(mtgtdir)); @@ -1941,7 +2626,7 @@ relock: rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1, cos_incompat); if (rc) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); @@ -1949,20 +2634,21 @@ relock: cos_incompat); if (rc != 0) { mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); } } else { rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0, cos_incompat); if (rc) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); if (mtgtdir != msrcdir) { rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1, cos_incompat); - } else if (lh_srcdirp->mlh_pdo_hash != + } else if (!mdt_object_remote(mtgtdir) && + lh_srcdirp->mlh_pdo_hash != lh_tgtdirp->mlh_pdo_hash) { rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir, MDS_INODELOCK_UPDATE, @@ -1971,7 +2657,7 @@ relock: } if (rc != 0) { mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); } } @@ -2001,8 +2687,12 @@ relock: GOTO(out_put_old, rc = -ENOENT); } + if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename) + GOTO(out_put_old, rc = -EXDEV); + /* Check if @mtgtdir is subdir of @mold, before locking child - * to avoid reverse locking. */ + * to avoid reverse locking. + */ if (mtgtdir != msrcdir) { rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir), old_fid); @@ -2027,7 +2717,8 @@ relock: /* find mnew object: * mnew target object may not exist now - * lookup with version checking */ + * lookup with version checking + */ fid_zero(new_fid); rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid, 3); @@ -2068,35 +2759,25 @@ relock: /* Before locking the target dir, check we do not replace * a dir with a non-dir, otherwise it may deadlock with * link op which tries to create a link in this dir - * back to this non-dir. */ + * back to this non-dir. + */ if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) && !S_ISDIR(lu_object_attr(&mold->mot_obj))) GOTO(out_put_new, rc = -EISDIR); lh_oldp = &info->mti_lh[MDT_LH_OLD]; + lh_rmt = &info->mti_lh[MDT_LH_RMT]; mdt_lock_reg_init(lh_oldp, LCK_EX); + mdt_lock_reg_init(lh_rmt, LCK_EX); lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR; - if (mdt_object_remote(msrcdir)) { - /* Enqueue lookup lock from the parent MDT */ - rc = mdt_remote_object_lock(info, msrcdir, - mdt_object_fid(mold), - &lh_oldp->mlh_rreg_lh, - lh_oldp->mlh_rreg_mode, - MDS_INODELOCK_LOOKUP, - false); - if (rc != ELDLM_OK) - GOTO(out_put_new, rc); - - lock_ibits &= ~MDS_INODELOCK_LOOKUP; - } - - rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits, - cos_incompat); - if (rc != 0) - GOTO(out_unlock_old, rc); + rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp, + lh_rmt, lock_ibits, cos_incompat); + if (rc < 0) + GOTO(out_put_new, rc); /* Check if @msrcdir is subdir of @mnew, before locking child - * to avoid reverse locking. */ + * to avoid reverse locking. + */ if (mtgtdir != msrcdir) { rc = mdo_is_subdir(info->mti_env, mdt_object_child(msrcdir), new_fid); @@ -2110,43 +2791,43 @@ relock: /* We used to acquire MDS_INODELOCK_FULL here but we * can't do this now because a running HSM restore on * the rename onto victim will hold the layout - * lock. See LU-4002. */ + * lock. See LU-4002. + */ lh_newp = &info->mti_lh[MDT_LH_NEW]; mdt_lock_reg_init(lh_newp, LCK_EX); - rc = mdt_reint_object_lock(info, mnew, lh_newp, - MDS_INODELOCK_LOOKUP | - MDS_INODELOCK_UPDATE, + lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; + if (mdt_object_remote(mtgtdir)) { + rc = mdt_remote_object_lock(info, mtgtdir, + mdt_object_fid(mnew), + &lh_newp->mlh_rreg_lh, + lh_newp->mlh_rreg_mode, + MDS_INODELOCK_LOOKUP, + false); + if (rc != ELDLM_OK) + GOTO(out_unlock_old, rc); + + lock_ibits &= ~MDS_INODELOCK_LOOKUP; + } + rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits, cos_incompat); if (rc != 0) - GOTO(out_unlock_old, rc); + GOTO(out_unlock_new, rc); /* get and save version after locking */ mdt_version_get_save(info, mnew, 3); - } else if (rc != -EREMOTE && rc != -ENOENT) { + } else if (rc != -ENOENT) { GOTO(out_put_old, rc); } else { lh_oldp = &info->mti_lh[MDT_LH_OLD]; + lh_rmt = &info->mti_lh[MDT_LH_RMT]; mdt_lock_reg_init(lh_oldp, LCK_EX); + mdt_lock_reg_init(lh_rmt, LCK_EX); lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR; - if (mdt_object_remote(msrcdir)) { - /* Enqueue lookup lock from the parent MDT */ - rc = mdt_remote_object_lock(info, msrcdir, - mdt_object_fid(mold), - &lh_oldp->mlh_rreg_lh, - lh_oldp->mlh_rreg_mode, - MDS_INODELOCK_LOOKUP, - false); - if (rc != ELDLM_OK) - GOTO(out_put_old, rc); - - lock_ibits &= ~MDS_INODELOCK_LOOKUP; - } - - rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits, - cos_incompat); + rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp, + lh_rmt, lock_ibits, cos_incompat); if (rc != 0) - GOTO(out_unlock_old, rc); + GOTO(out_put_old, rc); mdt_enoent_version_save(info, 3); } @@ -2170,106 +2851,67 @@ relock: /* handle last link of tgt object */ if (rc == 0) { - mdt_counter_incr(req, LPROC_MDT_RENAME); + mdt_counter_incr(req, LPROC_MDT_RENAME, + ktime_us_delta(ktime_get(), kstart)); if (mnew) { mdt_handle_last_unlink(info, mnew, ma); - discard = true; + discard = mdt_dom_check_for_discard(info, mnew); } - mdt_rename_counter_tally(info, info->mti_mdt, req, - msrcdir, mtgtdir); + msrcdir, mtgtdir, + ktime_us_delta(ktime_get(), kstart)); } EXIT; +out_unlock_new: if (mnew != NULL) mdt_object_unlock(info, mnew, lh_newp, rc); out_unlock_old: + mdt_object_unlock(info, NULL, lh_rmt, rc); mdt_object_unlock(info, mold, lh_oldp, rc); out_put_new: - if (mnew != NULL) { - if (discard) - mdt_dom_check_and_discard(info, mnew); + if (mnew && !discard) mdt_object_put(info->mti_env, mnew); - } out_put_old: mdt_object_put(info->mti_env, mold); out_unlock_parents: mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); +out_unlock_rename: + if (lustre_handle_is_used(&rename_lh)) + mdt_rename_unlock(&rename_lh); out_put_tgtdir: mdt_object_put(info->mti_env, mtgtdir); out_put_srcdir: mdt_object_put(info->mti_env, msrcdir); - return rc; -} - -static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc, bool rename) -{ - struct mdt_reint_record *rr = &info->mti_rr; - struct ptlrpc_request *req = mdt_info_req(info); - struct lustre_handle rename_lh = { 0 }; - int rc; - ENTRY; - - if (info->mti_dlm_req) - ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); - if (!fid_is_md_operative(rr->rr_fid1) || - !fid_is_md_operative(rr->rr_fid2)) - RETURN(-EPERM); - - /* Note: do not enqueue rename lock for replay request, because - * if other MDT holds rename lock, but being blocked to wait for - * this MDT to finish its recovery, and the failover MDT can not - * get rename lock, which will cause deadlock. */ - if (!req_is_replay(req)) { - rc = mdt_rename_lock(info, &rename_lh); - if (rc != 0) { - CERROR("%s: can't lock FS for rename: rc = %d\n", - mdt_obd_name(info->mti_mdt), rc); - RETURN(rc); - } + /* The DoM discard can be done right in the place above where it is + * assigned, meanwhile it is done here after rename unlock due to + * compatibility with old clients, for them the discard blocks + * the main thread until completion. Check LU-11359 for details. + */ + if (discard) { + mdt_dom_discard_data(info, mnew); + mdt_object_put(info->mti_env, mnew); } - - if (rename) - rc = mdt_reint_rename_internal(info, lhc); - else - rc = mdt_reint_migrate_internal(info, lhc); - - if (lustre_handle_is_used(&rename_lh)) - mdt_rename_unlock(&rename_lh); - - RETURN(rc); -} - -static int mdt_reint_rename(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) -{ - return mdt_reint_rename_or_migrate(info, lhc, true); -} - -static int mdt_reint_migrate(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) -{ - return mdt_reint_rename_or_migrate(info, lhc, false); + return rc; } static int mdt_reint_resync(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { struct mdt_reint_record *rr = &info->mti_rr; - struct ptlrpc_request *req = mdt_info_req(info); - struct md_attr *ma = &info->mti_attr; - struct mdt_object *mo; - struct ldlm_lock *lease; - struct mdt_body *repbody; - struct md_layout_change layout = { 0 }; - bool lease_broken; - int rc, rc2; - ENTRY; + struct ptlrpc_request *req = mdt_info_req(info); + struct md_attr *ma = &info->mti_attr; + struct mdt_object *mo; + struct ldlm_lock *lease; + struct mdt_body *repbody; + struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id }; + bool lease_broken; + int rc, rc2; - DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1)); + ENTRY; + DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1)); if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); @@ -2287,14 +2929,15 @@ static int mdt_reint_resync(struct mdt_thread_info *info, if (mdt_object_remote(mo)) GOTO(out_obj, rc = -EREMOTE); - lease = ldlm_handle2lock(rr->rr_handle); + lease = ldlm_handle2lock(rr->rr_lease_handle); if (lease == NULL) GOTO(out_obj, rc = -ESTALE); /* It's really necessary to grab open_sem and check if the lease lock * has been lost. There would exist a concurrent writer coming in and * generating some dirty data in memory cache, the writeback would fail - * after the layout version is increased by MDS_REINT_RESYNC RPC. */ + * after the layout version is increased by MDS_REINT_RESYNC RPC. + */ if (!down_write_trylock(&mo->mot_open_sem)) GOTO(out_put_lease, rc = -EBUSY); @@ -2306,10 +2949,13 @@ static int mdt_reint_resync(struct mdt_thread_info *info, /* the file has yet opened by anyone else after we took the lease. */ layout.mlc_opc = MD_LAYOUT_RESYNC; - rc = mdt_layout_change(info, mo, &layout); + lhc = &info->mti_lh[MDT_LH_LOCAL]; + rc = mdt_layout_change(info, mo, lhc, &layout); if (rc) GOTO(out_unlock, rc); + mdt_object_unlock(info, mo, lhc, 0); + ma->ma_need = MA_INODE; ma->ma_valid = 0; rc = mdt_attr_get_complex(info, mo, ma); @@ -2387,8 +3033,8 @@ int mdt_reint_rec(struct mdt_thread_info *info, { const struct mdt_reinter *mr; int rc; - ENTRY; + ENTRY; if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters))) RETURN(-EPROTO);