X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=8cb8cb7c19671a55913b717f067a5844d42ab493;hp=80a7aa7857706d3b2272613e8041495429d6cc5b;hb=7936ec229c7f62cfbf5a0e4ff5933149967e7c8f;hpb=0dd85e9ace1e70e60c721eac94ece0cc8a1fccc2 diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 80a7aa7..8cb8cb7 100644 --- a/lustre/mdt/mdt_reint.c +++ b/lustre/mdt/mdt_reint.c @@ -47,6 +47,7 @@ #define DEBUG_SUBSYSTEM S_MDS #include "mdt_internal.h" +#include static inline void mdt_reint_init_ma(struct mdt_thread_info *info, struct md_attr *ma) @@ -66,7 +67,7 @@ static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc, if (repbody->valid & OBD_MD_FLMDSCAPA) RETURN(rc); - if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa && + if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa && exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) { struct lustre_capa *capa; @@ -219,8 +220,8 @@ int mdt_version_get_check_save(struct mdt_thread_info *info, * FID, therefore we need to get object by name and check its version. */ int mdt_lookup_version_check(struct mdt_thread_info *info, - struct mdt_object *p, struct lu_name *lname, - struct lu_fid *fid, int idx) + struct mdt_object *p, const struct lu_name *lname, + struct lu_fid *fid, int idx) { int rc, vbrc; @@ -244,6 +245,79 @@ int mdt_lookup_version_check(struct mdt_thread_info *info, } +/** + * mdt_remote_permission: Check whether the remote operation is permitted, + * + * Before we implement async cross-MDT updates (DNE phase 2). There are a few + * limitations here: + * + * 1.Only sysadmin can create remote directory and striped directory and + * migrate directory now, unless + * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid. + * 2.Remote directory can only be created on MDT0, unless + * lctl set_param mdt.*.enable_remote_dir = 1 + * 3.Only new clients can access remote dir( >= 2.4) and striped dir(>= 2.6), + * old client will return -ENOTSUPP. + * + * XXX these check are only needed for remote synchronization, once async + * update is supported, these check will be removed. + * + * param[in]info: execution environment. + * param[in]parent: the directory of this operation. + * param[in]child: the child of this operation. + * + * retval = 0 remote operation is allowed. + * < 0 remote operation is denied. + */ +static int mdt_remote_permission(struct mdt_thread_info *info, + struct mdt_object *parent, + struct mdt_object *child) +{ + struct mdt_device *mdt = info->mti_mdt; + struct lu_ucred *uc = mdt_ucred(info); + 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; + + /* Only check create remote directory, striped directory and + * migration */ + if (mdt_object_remote(parent) == 0 && mdt_object_remote(child) == 0 && + !(S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL && + spec->u.sp_ea.eadatalen != 0) && + info->mti_rr.rr_opcode != REINT_MIGRATE) + return 0; + + 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; + } + + if (mdt->mdt_enable_remote_dir == 0) { + struct seq_server_site *ss = mdt_seq_site(mdt); + struct lu_seq_range range = { 0 }; + int rc; + + fld_range_set_type(&range, LU_SEQ_RANGE_MDT); + rc = fld_server_lookup(info->mti_env, ss->ss_server_fld, + fid_seq(mdt_object_fid(parent)), &range); + if (rc != 0) + return rc; + + if (range.lsr_index != 0) + return -EPERM; + } + + if (!mdt_is_dne_client(exp)) + return -ENOTSUPP; + + if (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL && + spec->u.sp_ea.eadatalen != 0 && !mdt_is_striped_client(exp)) + return -ENOTSUPP; + + return 0; +} + /* * VBR: we save three versions in reply: * 0 - parent. Check that parent version is the same during replay. @@ -261,43 +335,50 @@ static int mdt_md_create(struct mdt_thread_info *info) struct mdt_body *repbody; struct md_attr *ma = &info->mti_attr; struct mdt_reint_record *rr = &info->mti_rr; - struct lu_name *lname; int rc; ENTRY; - DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID, - rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1)); + 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_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + if (!fid_is_md_operative(rr->rr_fid1)) RETURN(-EPERM); repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen); + parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(parent)) + RETURN(PTR_ERR(parent)); - parent = mdt_object_find_lock(info, rr->rr_fid1, lh, - MDS_INODELOCK_UPDATE); - if (IS_ERR(parent)) - RETURN(PTR_ERR(parent)); + if (!mdt_object_exists(parent)) + GOTO(put_parent, rc = -ENOENT); - rc = mdt_version_get_check_save(info, parent, 0); - if (rc) - GOTO(out_put_parent, rc); + 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, + MDT_CROSS_LOCK); + if (rc) + GOTO(put_parent, rc); - /* - * Check child name version during replay. - * During create replay a file may exist with same name. - */ - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - rc = mdt_lookup_version_check(info, parent, lname, - &info->mti_tmp_fid1, 1); + if (!mdt_object_remote(parent)) { + rc = mdt_version_get_check_save(info, parent, 0); + if (rc) + 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(out_put_parent, rc = -EEXIST); + GOTO(unlock_parent, rc = -EEXIST); /* -ENOENT is expected here */ if (rc != -ENOENT) - GOTO(out_put_parent, rc); + GOTO(unlock_parent, rc); /* save version of file name for replay, it must be ENOENT here */ mdt_enoent_version_save(info, 1); @@ -306,39 +387,11 @@ static int mdt_md_create(struct mdt_thread_info *info) if (likely(!IS_ERR(child))) { struct md_object *next = mdt_object_child(parent); - if (mdt_object_remote(child)) { - struct seq_server_site *ss; - struct lu_ucred *uc = mdt_ucred(info); - - 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) { - CERROR("%s: Creating remote dir is only" - " permitted for administrator or" - " set mdt_enable_remote_dir_gid:" - " rc = %d\n", - mdt_obd_name(mdt), -EPERM); - GOTO(out_put_child, rc = -EPERM); - } - } - - ss = mdt_seq_site(mdt); - if (ss->ss_node_id != 0 && - mdt->mdt_enable_remote_dir == 0) { - CERROR("%s: remote dir is only permitted on" - " MDT0 or set_param" - " mdt.*.enable_remote_dir=1\n", - mdt_obd_name(mdt)); - GOTO(out_put_child, rc = -EPERM); - } - if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) { - /* Return -EIO for old client */ - GOTO(out_put_child, rc = -EIO); - } + rc = mdt_remote_permission(info, parent, child); + if (rc != 0) + GOTO(out_put_child, rc); - } - ma->ma_need = MA_INODE; + ma->ma_need = MA_INODE; ma->ma_valid = 0; /* capa for cross-ref will be stored here */ ma->ma_capa = req_capsule_server_get(info->mti_pill, @@ -349,7 +402,7 @@ static int mdt_md_create(struct mdt_thread_info *info) OBD_FAIL_MDS_REINT_CREATE_WRITE); /* Version of child will be updated on disk. */ - info->mti_mos = child; + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child)); rc = mdt_version_get_check_save(info, child, 2); if (rc) GOTO(out_put_child, rc); @@ -365,9 +418,8 @@ static int mdt_md_create(struct mdt_thread_info *info) info->mti_spec.sp_cr_lookup = 0; info->mti_spec.sp_feat = &dt_directory_features; - rc = mdo_create(info->mti_env, next, lname, - mdt_object_child(child), - &info->mti_spec, ma); + rc = mdo_create(info->mti_env, next, &rr->rr_name, + mdt_object_child(child), &info->mti_spec, ma); if (rc == 0) rc = mdt_attr_get_complex(info, child, ma); @@ -383,19 +435,116 @@ out_put_child: rc = PTR_ERR(child); } mdt_create_pack_capa(info, rc, child, repbody); -out_put_parent: - mdt_object_unlock_put(info, parent, lh, rc); - RETURN(rc); +unlock_parent: + mdt_object_unlock(info, parent, lh, rc); +put_parent: + mdt_object_put(info->mti_env, parent); + RETURN(rc); +} + +static int mdt_unlock_slaves(struct mdt_thread_info *mti, + struct mdt_object *obj, __u64 ibits, + struct mdt_lock_handle *s0_lh, + struct mdt_object *s0_obj, + struct ldlm_enqueue_info *einfo) +{ + ldlm_policy_data_t *policy = &mti->mti_policy; + int rc; + ENTRY; + + if (!S_ISDIR(obj->mot_header.loh_attr)) + RETURN(0); + + /* Unlock stripe 0 */ + if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) { + LASSERT(s0_obj != NULL); + mdt_object_unlock_put(mti, s0_obj, s0_lh, 1); + } + + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = ibits; + + rc = mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo, + policy); + RETURN(rc); +} + +/** + * Lock slave stripes if necessary, the lock handles of slave stripes + * will be stored in einfo->ei_cbdata. + **/ +static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj, + ldlm_mode_t mode, __u64 ibits, + struct mdt_lock_handle *s0_lh, + struct mdt_object **s0_objp, + struct ldlm_enqueue_info *einfo) +{ + ldlm_policy_data_t *policy = &mti->mti_policy; + struct lu_buf *buf = &mti->mti_buf; + struct lmv_mds_md_v1 *lmv; + struct lu_fid *fid = &mti->mti_tmp_fid1; + int rc; + ENTRY; + + if (!S_ISDIR(obj->mot_header.loh_attr)) + RETURN(0); + + buf->lb_buf = mti->mti_xattr_buf; + buf->lb_len = sizeof(mti->mti_xattr_buf); + rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf, + XATTR_NAME_LMV); + if (rc == -ERANGE) { + rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV); + if (rc > 0) { + buf->lb_buf = mti->mti_big_lmm; + buf->lb_len = mti->mti_big_lmmsize; + } + } + + if (rc == -ENODATA || rc == -ENOENT) + RETURN(0); + + if (rc <= 0) + RETURN(rc); + + lmv = buf->lb_buf; + if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1) + RETURN(-EINVAL); + + /* Sigh, 0_stripe and master object are different + * object, though they are in the same MDT, to avoid + * adding osd_object_lock here, so we will enqueue the + * stripe0 lock in MDT0 for now */ + fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]); + *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits); + if (IS_ERR(*s0_objp)) + RETURN(PTR_ERR(*s0_objp)); + + memset(einfo, 0, sizeof(*einfo)); + einfo->ei_type = LDLM_IBITS; + einfo->ei_mode = mode; + einfo->ei_cb_bl = mdt_remote_blocking_ast; + einfo->ei_cb_cp = ldlm_completion_ast; + einfo->ei_enq_slave = 1; + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = ibits; + + rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo, + policy); + RETURN(rc); } int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, struct md_attr *ma, int flags) { - struct mdt_lock_handle *lh; - int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS); - __u64 lockpart = MDS_INODELOCK_UPDATE; - int rc; - ENTRY; + struct mdt_lock_handle *lh; + int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS); + __u64 lockpart = MDS_INODELOCK_UPDATE; + struct ldlm_enqueue_info *einfo = &info->mti_einfo; + struct mdt_lock_handle *s0_lh; + struct mdt_object *s0_obj = NULL; + int rc; + ENTRY; /* attr shouldn't be set on remote object */ LASSERT(!mdt_object_remote(mo)); @@ -414,6 +563,12 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, if (rc != 0) RETURN(rc); + s0_lh = &info->mti_lh[MDT_LH_LOCAL]; + mdt_lock_reg_init(s0_lh, LCK_EX); + rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo); + if (rc != 0) + GOTO(out_unlock, rc); + if (mdt_object_exists(mo) == 0) GOTO(out_unlock, rc = -ENOENT); @@ -428,19 +583,28 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, /* VBR: update version if attr changed are important for recovery */ if (do_vbr) { /* update on-disk version of changed object */ - info->mti_mos = mo; + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo)); rc = mdt_version_get_check_save(info, mo, 0); if (rc) GOTO(out_unlock, rc); } + /* Ensure constant striping during chown(). See LU-2789. */ + if (ma->ma_attr.la_valid & (LA_UID|LA_GID)) + mutex_lock(&mo->mot_lov_mutex); + /* all attrs are packed into mti_attr in unpack_setattr */ rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma); + + if (ma->ma_attr.la_valid & (LA_UID|LA_GID)) + mutex_unlock(&mo->mot_lov_mutex); + if (rc != 0) GOTO(out_unlock, rc); EXIT; out_unlock: + mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo); mdt_object_unlock(info, mo, lh, rc); return rc; } @@ -588,6 +752,16 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, buf, XATTR_NAME_LOV, 0); if (rc) GOTO(out_put, rc); + } else if ((ma->ma_valid & MA_LMV) && (ma->ma_valid & MA_INODE)) { + struct lu_buf *buf = &info->mti_buf; + + LASSERT(ma->ma_attr.la_valid == 0); + buf->lb_buf = ma->ma_lmv; + buf->lb_len = ma->ma_lmv_size; + rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), + buf, XATTR_NAME_DEFAULT_LMV, 0); + if (rc) + GOTO(out_put, rc); } else LBUG(); @@ -598,12 +772,12 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, ma->ma_need = MA_INODE; ma->ma_valid = 0; rc = mdt_attr_get_complex(info, mo, ma); - if (rc != 0) - GOTO(out_put, rc); + if (rc != 0) + GOTO(out_put, rc); - mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); + mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); - if (info->mti_mdt->mdt_opts.mo_oss_capa && + if (info->mti_mdt->mdt_lut.lut_oss_capa && exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA && S_ISREG(lu_object_attr(&mo->mot_obj)) && (ma->ma_attr.la_valid & LA_SIZE) && !som_au) { @@ -645,8 +819,10 @@ static int mdt_reint_create(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0); - LASSERT(info->mti_rr.rr_namelen > 0); - switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { + if (!lu_name_is_valid(&info->mti_rr.rr_name)) + RETURN(-EPROTO); + + switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { case S_IFDIR: mdt_counter_incr(req, LPROC_MDT_MKDIR); break; @@ -685,13 +861,15 @@ 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 lu_name *lname; - int rc; - int no_name = 0; + struct ldlm_enqueue_info *einfo = &info->mti_einfo; + struct mdt_lock_handle *s0_lh = NULL; + struct mdt_object *s0_obj = NULL; + int rc; + int no_name = 0; ENTRY; - DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1), - rr->rr_name); + DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1), + PNAME(&rr->rr_name)); if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0); @@ -699,8 +877,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) RETURN(err_serious(-ENOENT)); - if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + if (!fid_is_md_operative(rr->rr_fid1)) RETURN(-EPERM); + /* * step 1: Found the parent. */ @@ -711,23 +890,13 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, } parent_lh = &info->mti_lh[MDT_LH_PARENT]; - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - if (mdt_object_remote(mp)) { - mdt_lock_reg_init(parent_lh, LCK_EX); - rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh, - parent_lh->mlh_rreg_mode, - MDS_INODELOCK_UPDATE); - if (rc != ELDLM_OK) - GOTO(put_parent, rc); - - } else { - mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name, - rr->rr_namelen); - rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE, - MDT_LOCAL_LOCK); - if (rc) - GOTO(put_parent, rc); + mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name); + rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE, + MDT_CROSS_LOCK); + if (rc != 0) + GOTO(put_parent, rc); + if (!mdt_object_remote(mp)) { rc = mdt_version_get_check_save(info, mp, 0); if (rc) GOTO(unlock_parent, rc); @@ -736,7 +905,7 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, /* step 2: find & lock the child */ /* lookup child object along with version checking */ fid_zero(child_fid); - rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1); + rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1); if (rc != 0) { /* Name might not be able to find during resend of * remote unlink, considering following case. @@ -762,11 +931,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, } } - if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid)) + if (!fid_is_md_operative(child_fid)) GOTO(unlock_parent, rc = -EPERM); - mdt_reint_init_ma(info, ma); - /* We will lock the child regardless it is local or remote. No harm. */ mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid); if (IS_ERR(mc)) @@ -778,18 +945,18 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, struct mdt_body *repbody; if (!fid_is_zero(rr->rr_fid2)) { - CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n", + CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n", mdt_obd_name(info->mti_mdt), - (char *)rr->rr_name, PFID(mdt_object_fid(mc))); + PNAME(&rr->rr_name), PFID(mdt_object_fid(mc))); GOTO(put_child, rc = -ENOENT); } - CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n", + CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n", mdt_obd_name(info->mti_mdt), - (char *)rr->rr_name, PFID(mdt_object_fid(mc))); + PNAME(&rr->rr_name), PFID(mdt_object_fid(mc))); if (!mdt_is_dne_client(req->rq_export)) - /* Return -EIO for old client */ - GOTO(put_child, rc = -EIO); + /* Return -ENOTSUPP for old client */ + GOTO(put_child, rc = -ENOTSUPP); if (info->mti_spec.sp_rm_entry) { struct lu_ucred *uc = mdt_ucred(info); @@ -806,7 +973,7 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, ma->ma_valid = 0; mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); rc = mdo_unlink(info->mti_env, mdt_object_child(mp), - NULL, lname, ma, no_name); + NULL, &rr->rr_name, ma, no_name); GOTO(put_child, rc); } /* Revoke the LOOKUP lock of the remote object granted by @@ -823,8 +990,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, GOTO(unlock_child, rc = -EREMOTE); } else if (info->mti_spec.sp_rm_entry) { rc = -EPERM; - CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n", - mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc); + CDEBUG(D_INFO, "%s: no rm_entry on local dir '"DNAME"': " + "rc = %d\n", + mdt_obd_name(info->mti_mdt), PNAME(&rr->rr_name), rc); GOTO(put_child, rc); } @@ -836,21 +1004,34 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, MDT_CROSS_LOCK); if (rc != 0) GOTO(put_child, rc); + /* + * Now we can only make sure we need MA_INODE, in mdd layer, will check + * whether need MA_LOV and MA_COOKIE. + */ + ma->ma_need = MA_INODE; + ma->ma_valid = 0; - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_UNLINK_WRITE); - /* save version when object is locked */ - mdt_version_get_save(info, mc, 1); - /* - * Now we can only make sure we need MA_INODE, in mdd layer, will check - * whether need MA_LOV and MA_COOKIE. - */ - ma->ma_need = MA_INODE; - ma->ma_valid = 0; - mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); + s0_lh = &info->mti_lh[MDT_LH_LOCAL]; + mdt_lock_reg_init(s0_lh, LCK_EX); + rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh, + &s0_obj, einfo); + if (rc != 0) + GOTO(unlock_child, rc); + + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_UNLINK_WRITE); + /* save version when object is locked */ + mdt_version_get_save(info, mc, 1); + + mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); + + mutex_lock(&mc->mot_lov_mutex); rc = mdo_unlink(info->mti_env, mdt_object_child(mp), - mdt_object_child(mc), lname, ma, no_name); + mdt_object_child(mc), &rr->rr_name, ma, no_name); + + mutex_unlock(&mc->mot_lov_mutex); + if (rc == 0 && !lu_object_is_dying(&mc->mot_header)) rc = mdt_attr_get_complex(info, mc, ma); if (rc == 0) @@ -876,7 +1057,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, } EXIT; + unlock_child: + mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo); mdt_object_unlock(info, mc, child_lh, rc); put_child: mdt_object_put(info->mti_env, mc); @@ -902,12 +1085,11 @@ static int mdt_reint_link(struct mdt_thread_info *info, struct mdt_object *mp; struct mdt_lock_handle *lhs; struct mdt_lock_handle *lhp; - struct lu_name *lname; int rc; ENTRY; - DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s", - PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name); + 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)); @@ -920,14 +1102,13 @@ static int mdt_reint_link(struct mdt_thread_info *info, if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) RETURN(-EPERM); - if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) || - fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2)) + if (!fid_is_md_operative(rr->rr_fid1) || + !fid_is_md_operative(rr->rr_fid2)) RETURN(-EPERM); /* step 1: find & lock the target parent dir */ lhp = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name, - rr->rr_namelen); + mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name); mp = mdt_object_find_lock(info, rr->rr_fid2, lhp, MDS_INODELOCK_UPDATE); if (IS_ERR(mp)) @@ -945,6 +1126,13 @@ static int mdt_reint_link(struct mdt_thread_info *info, if (IS_ERR(ms)) GOTO(out_unlock_parent, rc = PTR_ERR(ms)); + if (!mdt_object_exists(ms)) { + mdt_object_put(info->mti_env, ms); + CDEBUG(D_INFO, "%s: "DFID" does not exist.\n", + mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1)); + GOTO(out_unlock_parent, rc = -ENOENT); + } + if (mdt_object_remote(ms)) { mdt_object_put(info->mti_env, ms); CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n", @@ -954,7 +1142,7 @@ static int mdt_reint_link(struct mdt_thread_info *info, } rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE | - MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); + MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK); if (rc != 0) { mdt_object_put(info->mti_env, ms); GOTO(out_unlock_parent, rc); @@ -964,29 +1152,29 @@ static int mdt_reint_link(struct mdt_thread_info *info, mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_LINK_WRITE); - info->mti_mos = ms; + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms)); rc = mdt_version_get_check_save(info, ms, 1); if (rc) GOTO(out_unlock_child, rc); - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); /** check target version by name during replay */ - rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2); + rc = mdt_lookup_version_check(info, mp, &rr->rr_name, + &info->mti_tmp_fid1, 2); if (rc != 0 && rc != -ENOENT) GOTO(out_unlock_child, rc); /* save version of file name for replay, it must be ENOENT here */ if (!req_is_replay(mdt_info_req(info))) { if (rc != -ENOENT) { - CDEBUG(D_INFO, "link target %.*s existed!\n", - rr->rr_namelen, (char *)rr->rr_name); + CDEBUG(D_INFO, "link target "DNAME" existed!\n", + PNAME(&rr->rr_name)); GOTO(out_unlock_child, rc = -EEXIST); } info->mti_ver[2] = ENOENT_VERSION; mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2); } - rc = mdo_link(info->mti_env, mdt_object_child(mp), - mdt_object_child(ms), lname, ma); + rc = mdo_link(info->mti_env, mdt_object_child(mp), + mdt_object_child(ms), &rr->rr_name, ma); if (rc == 0) mdt_counter_incr(req, LPROC_MDT_LINK); @@ -1030,47 +1218,69 @@ static int mdt_pdir_hash_lock(struct mdt_thread_info *info, return rc; } +enum mdt_rename_lock { + MRL_RENAME, + MRL_MIGRATE, +}; + +/** + * Get BFL lock for rename or migrate process, right now, it does not support + * cross-MDT rename, so we only need global rename lock during migration. + **/ static int mdt_rename_lock(struct mdt_thread_info *info, - struct lustre_handle *lh) + struct lustre_handle *lh, + enum mdt_rename_lock rename_lock) { struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; ldlm_policy_data_t *policy = &info->mti_policy; struct ldlm_res_id *res_id = &info->mti_res_id; __u64 flags = 0; - int rc; + int rc; ENTRY; - fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id); + /* XXX only do global rename lock for migration */ + if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0 && + rename_lock == MRL_MIGRATE) { + 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 */ + lu_root_fid(fid); + obj = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (IS_ERR(obj)) + RETURN(PTR_ERR(obj)); + + LASSERT(mdt_object_remote(obj)); + rc = mdt_remote_object_lock(info, obj, + &LUSTRE_BFL_FID, lh, + LCK_EX, + MDS_INODELOCK_UPDATE); + mdt_object_put(info->mti_env, obj); + } else { + fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id); + 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); + } - memset(policy, 0, sizeof *policy); - policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 5, 53, 0) - /* In phase I, we will not do cross-rename, so local BFL lock would - * be enough - */ - flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; - /* - * Current node is controller, that is mdt0, where we should - * take BFL lock. - */ - 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); -#else -#warning "Local rename lock is invalid for DNE phase II." -#endif - RETURN(rc); + RETURN(rc); } static void mdt_rename_unlock(struct lustre_handle *lh) { - ENTRY; - LASSERT(lustre_handle_is_used(lh)); - ldlm_lock_decref(lh, LCK_EX); - EXIT; + ENTRY; + LASSERT(lustre_handle_is_used(lh)); + /* Cancel the single rename lock right away */ + ldlm_lock_decref_and_cancel(lh, LCK_EX); + EXIT; } /* @@ -1078,37 +1288,478 @@ static void mdt_rename_unlock(struct lustre_handle *lh) * target. Source should not be ancestor of target dir. May be other rename * checks can be moved here later. */ -static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid) +static int mdt_rename_sanity(struct mdt_thread_info *info, + const struct lu_fid *dir_fid, + const struct lu_fid *fid) { - struct mdt_reint_record *rr = &info->mti_rr; - struct lu_fid dst_fid = *rr->rr_fid2; struct mdt_object *dst; + struct lu_fid dst_fid = *dir_fid; int rc = 0; ENTRY; - do { - LASSERT(fid_is_sane(&dst_fid)); - dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid); - if (!IS_ERR(dst)) { - rc = mdo_is_subdir(info->mti_env, - mdt_object_child(dst), fid, - &dst_fid); - mdt_object_put(info->mti_env, dst); - if (rc != -EREMOTE && rc < 0) { - CERROR("Failed mdo_is_subdir(), rc %d\n", rc); - } else { - /* check the found fid */ - if (lu_fid_eq(&dst_fid, fid)) - rc = -EINVAL; - } - } else { - rc = PTR_ERR(dst); - } - } while (rc == -EREMOTE); + /* If the source and target are in the same directory, they can not + * be parent/child relationship, so subdir check is not needed */ + if (lu_fid_eq(dir_fid, fid)) + return 0; + + do { + LASSERT(fid_is_sane(&dst_fid)); + dst = mdt_object_find(info->mti_env, info->mti_mdt, &dst_fid); + if (!IS_ERR(dst)) { + rc = mdo_is_subdir(info->mti_env, + mdt_object_child(dst), fid, + &dst_fid); + mdt_object_put(info->mti_env, dst); + if (rc != -EREMOTE && rc < 0) { + CERROR("%s: failed subdir check in "DFID" for " + DFID": rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(dir_fid), PFID(fid), rc); + /* Return EINVAL only if a parent is the @fid */ + if (rc == -EINVAL) + rc = -EIO; + } else { + /* check the found fid */ + if (lu_fid_eq(&dst_fid, fid)) + rc = -EINVAL; + } + } else { + rc = PTR_ERR(dst); + } + } while (rc == -EREMOTE); RETURN(rc); } +/* 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) +{ + 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); + } +} + +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) +{ + struct lu_buf *buf = &info->mti_big_buf; + struct linkea_data ldata = { 0 }; + int count; + int rc; + ENTRY; + + if (S_ISDIR(lu_object_attr(&obj->mot_obj))) + RETURN(0); + + buf = lu_buf_check_and_alloc(buf, PATH_MAX); + if (buf->lb_buf == NULL) + RETURN(-ENOMEM); + + ldata.ld_buf = buf; + rc = mdt_links_read(info, obj, &ldata); + if (rc != 0) { + if (rc == -ENOENT || rc == -ENODATA) + rc = 0; + RETURN(rc); + } + + 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; + + linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, + &name, &fid); + ldata.ld_lee = (struct link_ea_entry *)((char *)ldata.ld_lee + + ldata.ld_reclen); + 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)); + 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); + continue; + } + + 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); + continue; + } + + OBD_ALLOC_PTR(mll); + if (mll == NULL) { + mdt_object_put(info->mti_env, mdt_pobj); + GOTO(out, rc = -ENOMEM); + } + + mdt_lock_pdo_init(&mll->mll_lh, LCK_PW, &name); + rc = mdt_object_lock(info, mdt_pobj, &mll->mll_lh, + MDS_INODELOCK_UPDATE, + MDT_CROSS_LOCK); + if (rc != 0) { + CERROR("%s: cannot lock "DFID": rc =%d\n", + mdt_obd_name(mdt), PFID(&fid), rc); + mdt_object_put(info->mti_env, mdt_pobj); + OBD_FREE_PTR(mll); + GOTO(out, rc); + } + + CFS_INIT_LIST_HEAD(&mll->mll_list); + mll->mll_obj = mdt_pobj; + list_add_tail(&mll->mll_list, lock_list); + } +out: + if (rc != 0) + mdt_unlock_list(info, lock_list, rc); + 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) +{ + 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; + int rc; + ENTRY; + + CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1), + PNAME(&rr->rr_name), PFID(rr->rr_fid2)); + + /* 1: lock the source dir. */ + msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(msrcdir)) { + CERROR("%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)); + } + + lh_dirp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lh_dirp, LCK_PW, &rr->rr_name); + rc = mdt_object_lock(info, msrcdir, lh_dirp, + MDS_INODELOCK_UPDATE, + MDT_CROSS_LOCK); + if (rc) + GOTO(out_put_parent, rc); + + if (!mdt_object_remote(msrcdir)) { + rc = mdt_version_get_check_save(info, msrcdir, 0); + if (rc) + GOTO(out_unlock_parent, rc); + } + + /* 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); + + if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2)) + GOTO(out_unlock_parent, rc = -EINVAL); + + if (!fid_is_md_operative(old_fid)) + GOTO(out_unlock_parent, rc = -EPERM); + + 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 (mdt_object_remote(mold)) { + CERROR("%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 (S_ISREG(lu_object_attr(&mold->mot_obj)) && + !mdt_object_remote(msrcdir)) { + CERROR("%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); + } + + rc = mdt_remote_permission(info, msrcdir, mold); + if (rc != 0) + GOTO(out_put_child, rc); + + /* 3: iterate the linkea of the object and lock all of the objects */ + CFS_INIT_LIST_HEAD(&lock_list); + rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list); + if (rc != 0) + GOTO(out_put_child, rc); + + /* 4: lock of the object migrated object */ + lh_childp = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lh_childp, LCK_EX); + rc = mdt_object_lock(info, mold, lh_childp, + MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE | + MDS_INODELOCK_LAYOUT, MDT_CROSS_LOCK); + if (rc != 0) + GOTO(out_unlock_list, rc); + + 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_list, 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)) { + CERROR("%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); + } + + if (!fid_is_sane(&lmm1->lmv_stripe_fids[1])) + GOTO(out_unlock_child, rc = -EINVAL); + + 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)) { + CERROR("%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); + } + + 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); + 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)) { + CERROR("%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); + } + } + + /* 5: migrate it */ + mdt_reint_init_ma(info, ma); + + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_RENAME_WRITE); + + 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); +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: + mdt_unlock_list(info, &lock_list, rc); +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); + + RETURN(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; + + 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); + + RETURN(dir); +out_put: + mdt_object_put(info->mti_env, dir); + return ERR_PTR(rc); +} + +static int mdt_object_lock_save(struct mdt_thread_info *info, + struct mdt_object *dir, + struct mdt_lock_handle *lh, + int idx) +{ + int rc; + + /* we lock the target dir if it is local */ + rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE, + MDT_LOCAL_LOCK); + if (rc != 0) + return rc; + + /* get and save correct version after locking */ + mdt_version_get_save(info, dir, idx); + return 0; +} + + +static int mdt_rename_parents_lock(struct mdt_thread_info *info, + struct mdt_object **srcp, + struct mdt_object **tgtp) +{ + struct mdt_reint_record *rr = &info->mti_rr; + const struct lu_fid *fid_src = rr->rr_fid1; + const struct lu_fid *fid_tgt = rr->rr_fid2; + struct mdt_lock_handle *lh_src = &info->mti_lh[MDT_LH_PARENT]; + struct mdt_lock_handle *lh_tgt = &info->mti_lh[MDT_LH_CHILD]; + struct mdt_object *src; + struct mdt_object *tgt; + int reverse = 0; + int rc; + ENTRY; + + /* Check if the @src is not a child of the @tgt, otherwise a + * reverse locking must take place. + * + * Note: cannot be called after object_find, because if the object + * is destroyed in between it gets stuck in lu_object_find_at(), + * waiting for the last ref. */ + rc = mdt_rename_sanity(info, fid_src, fid_tgt); + if (rc == -EINVAL) + reverse = 1; + else if (rc) + RETURN(rc); + + /* find both parents. */ + src = mdt_object_find_check(info, fid_src, 0); + if (IS_ERR(src)) + RETURN(PTR_ERR(src)); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5); + + if (lu_fid_eq(fid_src, fid_tgt)) { + tgt = src; + mdt_object_get(info->mti_env, tgt); + } else { + tgt = mdt_object_find_check(info, fid_tgt, 1); + if (IS_ERR(tgt)) + GOTO(err_src_put, rc = PTR_ERR(tgt)); + + if (unlikely(mdt_object_remote(tgt))) { + CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID + "on different MDTs\n", PFID(fid_src), + PFID(fid_tgt)); + GOTO(err_tgt_put, rc = -EXDEV); + } + } + + /* lock parents in the proper order. */ + if (reverse) { + rc = mdt_object_lock_save(info, tgt, lh_tgt, 1); + if (rc) + GOTO(err_tgt_put, rc); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); + + rc = mdt_object_lock_save(info, src, lh_src, 0); + } else { + rc = mdt_object_lock_save(info, src, lh_src, 0); + if (rc) + GOTO(err_tgt_put, rc); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); + + if (tgt != src) + rc = mdt_object_lock_save(info, tgt, lh_tgt, 1); + else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) { + rc = mdt_pdir_hash_lock(info, lh_tgt, tgt, + MDS_INODELOCK_UPDATE); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10); + } + } + if (rc) + GOTO(err_unlock, rc); + + if (lu_object_is_dying(&tgt->mot_header)) + GOTO(err_unlock, rc = -ENOENT); + + *srcp = src; + *tgtp = tgt; + RETURN(0); + +err_unlock: + /* The order does not matter as the handle is checked inside, + * as well as not used handle. */ + mdt_object_unlock(info, src, lh_src, rc); + mdt_object_unlock(info, tgt, lh_tgt, rc); +err_tgt_put: + mdt_object_put(info->mti_env, tgt); +err_src_put: + mdt_object_put(info->mti_env, src); + RETURN(rc); +} + /* * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent; * 2 - src child; 3 - tgt child. @@ -1125,245 +1776,241 @@ static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid) * And tgt_c will be still in the same MDT as the original * src_c. */ -static int mdt_reint_rename(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) +static int mdt_reint_rename_internal(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc) { - struct mdt_reint_record *rr = &info->mti_rr; - struct md_attr *ma = &info->mti_attr; - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_object *msrcdir; - struct mdt_object *mtgtdir; - struct mdt_object *mold; - struct mdt_object *mnew = NULL; - struct mdt_lock_handle *lh_srcdirp; - struct mdt_lock_handle *lh_tgtdirp; - struct mdt_lock_handle *lh_oldp; - struct mdt_lock_handle *lh_newp; - struct lu_fid *old_fid = &info->mti_tmp_fid1; - struct lu_fid *new_fid = &info->mti_tmp_fid2; - struct lustre_handle rename_lh = { 0 }; - struct lu_name slname = { 0 }; - struct lu_name *lname; - int rc; - ENTRY; - - if (info->mti_dlm_req) - ldlm_request_cancel(req, info->mti_dlm_req, 0); + struct mdt_reint_record *rr = &info->mti_rr; + struct md_attr *ma = &info->mti_attr; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_object *msrcdir = NULL; + struct mdt_object *mtgtdir = NULL; + struct mdt_object *mold; + struct mdt_object *mnew = NULL; + struct mdt_lock_handle *lh_srcdirp; + struct mdt_lock_handle *lh_tgtdirp; + struct mdt_lock_handle *lh_oldp = 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; + int rc; + ENTRY; - DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s", - PFID(rr->rr_fid1), rr->rr_name, - PFID(rr->rr_fid2), rr->rr_tgt); + 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 (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) || - fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2)) - RETURN(-EPERM); + lh_srcdirp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name); + lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD]; + mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name); - rc = mdt_rename_lock(info, &rename_lh); - if (rc) { - CERROR("Can't lock FS for rename, rc %d\n", rc); + /* step 1&2: lock the source and target dirs. */ + rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir); + if (rc) RETURN(rc); - } - - lh_newp = &info->mti_lh[MDT_LH_NEW]; - - /* step 1: lock the source dir. */ - lh_srcdirp = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name, - rr->rr_namelen); - msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp, - MDS_INODELOCK_UPDATE); - if (IS_ERR(msrcdir)) - GOTO(out_rename_lock, rc = PTR_ERR(msrcdir)); - - rc = mdt_version_get_check_save(info, msrcdir, 0); - if (rc) - GOTO(out_unlock_source, rc); - - /* step 2: find & lock the target dir. */ - lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt, - rr->rr_tgtlen); - if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) { - mdt_object_get(info->mti_env, msrcdir); - mtgtdir = msrcdir; - if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) { - rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir, - MDS_INODELOCK_UPDATE); - if (rc) - GOTO(out_unlock_source, rc); - OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10); - } - } else { - mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt, - rr->rr_fid2); - if (IS_ERR(mtgtdir)) - GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir)); - - /* check early, the real version will be saved after locking */ - rc = mdt_version_get_check(info, mtgtdir, 1); - if (rc) - GOTO(out_put_target, rc); - if (unlikely(mdt_object_remote(mtgtdir))) { - CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID - "on different MDTs\n", PFID(rr->rr_fid1), - PFID(rr->rr_fid2)); - GOTO(out_put_target, rc = -EXDEV); - } else { - if (likely(mdt_object_exists(mtgtdir))) { - /* we lock the target dir if it is local */ - rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp, - MDS_INODELOCK_UPDATE, - MDT_LOCAL_LOCK); - if (rc != 0) - GOTO(out_put_target, rc); - /* get and save correct version after locking */ - mdt_version_get_save(info, mtgtdir, 1); - } else { - GOTO(out_put_target, rc = -ESTALE); - } - } - } + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5); - /* step 3: find & lock the old object. */ - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - mdt_name_copy(&slname, lname); - fid_zero(old_fid); - rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2); - if (rc != 0) - GOTO(out_unlock_target, rc); + /* step 3: find & lock the old object. */ + fid_zero(old_fid); + rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2); + if (rc != 0) + GOTO(out_unlock_parents, rc); - if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2)) - GOTO(out_unlock_target, rc = -EINVAL); + if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2)) + GOTO(out_unlock_parents, rc = -EINVAL); - if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid)) - GOTO(out_unlock_target, rc = -EPERM); + if (!fid_is_md_operative(old_fid)) + GOTO(out_unlock_parents, rc = -EPERM); mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid); if (IS_ERR(mold)) - GOTO(out_unlock_target, rc = PTR_ERR(mold)); - - lh_oldp = &info->mti_lh[MDT_LH_OLD]; - mdt_lock_reg_init(lh_oldp, LCK_EX); - rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP | - MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); - if (rc != 0) { - mdt_object_put(info->mti_env, mold); - GOTO(out_unlock_target, rc); - } - - info->mti_mos = mold; - /* save version after locking */ - mdt_version_get_save(info, mold, 2); - mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA); - - /* step 4: find & lock the new object. */ - /* new target object may not exist now */ - lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen); - /* lookup with version checking */ - fid_zero(new_fid); - rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3); - if (rc == 0) { - /* the new_fid should have been filled at this moment */ - if (lu_fid_eq(old_fid, new_fid)) - GOTO(out_unlock_old, rc); - - if (lu_fid_eq(new_fid, rr->rr_fid1) || - lu_fid_eq(new_fid, rr->rr_fid2)) - GOTO(out_unlock_old, rc = -EINVAL); - - if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid)) - GOTO(out_unlock_old, rc = -EPERM); + GOTO(out_unlock_parents, rc = PTR_ERR(mold)); + + /* Check if @mtgtdir is subdir of @mold, before locking child + * to avoid reverse locking. */ + rc = mdt_rename_sanity(info, rr->rr_fid2, old_fid); + if (rc) + GOTO(out_put_old, rc); + + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold)); + /* save version after locking */ + mdt_version_get_save(info, mold, 2); + mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA); + + /* step 4: find & lock the new object. */ + /* new target object may not exist now */ + /* lookup with version checking */ + fid_zero(new_fid); + rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid, + 3); + if (rc == 0) { + /* the new_fid should have been filled at this moment */ + if (lu_fid_eq(old_fid, new_fid)) + GOTO(out_put_old, rc); + + if (lu_fid_eq(new_fid, rr->rr_fid1) || + lu_fid_eq(new_fid, rr->rr_fid2)) + GOTO(out_put_old, rc = -EINVAL); + + if (!fid_is_md_operative(new_fid)) + GOTO(out_put_old, rc = -EPERM); if (mdt_object_remote(mold)) { CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n", PFID(old_fid)); - GOTO(out_unlock_old, rc = -EXDEV); + GOTO(out_put_old, rc = -EXDEV); } - mdt_lock_reg_init(lh_newp, LCK_EX); - mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid); - if (IS_ERR(mnew)) - GOTO(out_unlock_old, rc = PTR_ERR(mnew)); + mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid); + if (IS_ERR(mnew)) + GOTO(out_put_old, rc = PTR_ERR(mnew)); if (mdt_object_remote(mnew)) { - mdt_object_put(info->mti_env, mnew); CDEBUG(D_INFO, "src child "DFID" is on another MDT\n", PFID(new_fid)); - GOTO(out_unlock_old, rc = -EXDEV); + GOTO(out_put_new, rc = -EXDEV); } + lh_oldp = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lh_oldp, LCK_EX); + rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); + if (rc != 0) + GOTO(out_put_new, rc); + + /* Check if @msrcdir is subdir of @mnew, before locking child + * to avoid reverse locking. */ + rc = mdt_rename_sanity(info, rr->rr_fid1, new_fid); + if (rc) + GOTO(out_unlock_old, rc); + /* 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. */ + + lh_newp = &info->mti_lh[MDT_LH_NEW]; + mdt_lock_reg_init(lh_newp, LCK_EX); rc = mdt_object_lock(info, mnew, lh_newp, MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE, - MDT_CROSS_LOCK); - if (rc != 0) { - mdt_object_put(info->mti_env, mnew); - GOTO(out_unlock_old, rc); - } - /* get and save version after locking */ - mdt_version_get_save(info, mnew, 3); - mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA); - } else if (rc != -EREMOTE && rc != -ENOENT) { - GOTO(out_unlock_old, rc); - } else { + MDT_LOCAL_LOCK); + if (rc != 0) + GOTO(out_unlock_old, rc); + + /* get and save version after locking */ + mdt_version_get_save(info, mnew, 3); + mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA); + } else if (rc != -EREMOTE && rc != -ENOENT) { + GOTO(out_put_old, rc); + } else { /* If mnew does not exist and mold are remote directory, * it only allows rename if they are under same directory */ if (mtgtdir != msrcdir && mdt_object_remote(mold)) { CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n", PFID(old_fid)); - GOTO(out_unlock_old, rc = -EXDEV); + GOTO(out_put_old, rc = -EXDEV); } + + lh_oldp = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lh_oldp, LCK_EX); + rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); + if (rc != 0) + GOTO(out_put_old, rc); + mdt_enoent_version_save(info, 3); - } + } - /* step 5: rename it */ - mdt_reint_init_ma(info, ma); + /* step 5: rename it */ + mdt_reint_init_ma(info, ma); - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_RENAME_WRITE); + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_RENAME_WRITE); + if (mnew != NULL) + mutex_lock(&mnew->mot_lov_mutex); - /* Check if @dst is subdir of @src. */ - rc = mdt_rename_sanity(info, old_fid); - if (rc) - GOTO(out_unlock_new, rc); + rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir), + mdt_object_child(mtgtdir), old_fid, &rr->rr_name, + mnew != NULL ? mdt_object_child(mnew) : NULL, + &rr->rr_tgt_name, ma); - rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir), - mdt_object_child(mtgtdir), old_fid, &slname, - (mnew ? mdt_object_child(mnew) : NULL), - lname, ma); + if (mnew != NULL) + mutex_unlock(&mnew->mot_lov_mutex); - /* handle last link of tgt object */ - if (rc == 0) { + /* handle last link of tgt object */ + if (rc == 0) { mdt_counter_incr(req, LPROC_MDT_RENAME); - if (mnew) - mdt_handle_last_unlink(info, mnew, ma); + if (mnew) + mdt_handle_last_unlink(info, mnew, ma); mdt_rename_counter_tally(info, info->mti_mdt, req, - msrcdir, mtgtdir); - } + msrcdir, mtgtdir); + } - EXIT; -out_unlock_new: - if (mnew) - mdt_object_unlock_put(info, mnew, lh_newp, rc); + EXIT; + if (mnew != NULL) + mdt_object_unlock(info, mnew, lh_newp, rc); out_unlock_old: - mdt_object_unlock_put(info, mold, lh_oldp, rc); -out_unlock_target: - mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); -out_put_target: - mdt_object_put(info->mti_env, mtgtdir); -out_unlock_source: - mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc); -out_rename_lock: + mdt_object_unlock(info, mold, lh_oldp, rc); +out_put_new: + if (mnew != NULL) + mdt_object_put(info->mti_env, mnew); +out_put_old: + mdt_object_put(info->mti_env, mold); +out_unlock_parents: + mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc); + mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc); + return rc; +} + +static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc, + enum mdt_rename_lock rename_lock) +{ + 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); + + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) || + fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2)) + RETURN(-EPERM); + + rc = mdt_rename_lock(info, &rename_lh, rename_lock); + if (rc != 0) { + CERROR("%s: can't lock FS for rename: rc = %d\n", + mdt_obd_name(info->mti_mdt), rc); + RETURN(rc); + } + + if (rename_lock == MRL_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; + + 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, MRL_RENAME); +} + +static int mdt_reint_migrate(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc) +{ + return mdt_reint_rename_or_migrate(info, lhc, MRL_MIGRATE); } typedef int (*mdt_reinter)(struct mdt_thread_info *info, @@ -1377,7 +2024,8 @@ static mdt_reinter reinters[REINT_MAX] = { [REINT_RENAME] = mdt_reint_rename, [REINT_OPEN] = mdt_reint_open, [REINT_SETXATTR] = mdt_reint_setxattr, - [REINT_RMENTRY] = mdt_reint_unlink + [REINT_RMENTRY] = mdt_reint_unlink, + [REINT_MIGRATE] = mdt_reint_migrate, }; int mdt_reint_rec(struct mdt_thread_info *info,