X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=5d37de370946553111e0d01a4ca9074574dc8949;hp=241bed15f54d4d2b5926b2901e51b28183a8a28a;hb=4918fe40db262b19093436caca688c75eb632496;hpb=c0fa0ba4a8efcd774f1fe27986a0217c76dedf6d diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 241bed1..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; } /** @@ -190,30 +190,31 @@ 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 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; + 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; } @@ -342,6 +343,133 @@ void mdt_reint_striped_unlock(struct mdt_thread_info *info, 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. @@ -352,17 +480,18 @@ void mdt_reint_striped_unlock(struct mdt_thread_info *info, */ 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; + struct md_op_spec *spec = &info->mti_spec; + bool restripe = false; int rc; - ENTRY; + 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)); @@ -377,7 +506,8 @@ static int mdt_create(struct mdt_thread_info *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 */ + * striped dir(>= 2.6), old client will return -ENOTSUPP + */ if (!mdt_is_dne_client(exp)) RETURN(-ENOTSUPP); @@ -391,10 +521,19 @@ static int mdt_create(struct mdt_thread_info *info) 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); @@ -406,6 +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); @@ -418,22 +580,6 @@ 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)); @@ -515,13 +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, @@ -534,8 +681,8 @@ static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, 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); @@ -548,7 +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; @@ -601,8 +749,8 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, 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); @@ -618,7 +766,8 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, ma->ma_hsm.mh_flags |= HS_DIRTY; /* Bump cap so that closes from non-owner writers can - * set the HSM state to dirty. */ + * 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); @@ -640,15 +789,18 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, struct ptlrpc_request *req = mdt_info_req(info); struct mdt_object *mo; struct mdt_body *repbody; + ktime_t kstart = ktime_get(); int rc, rc2; - ENTRY; + 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, mdt, rr->rr_fid1); if (IS_ERR(mo)) @@ -660,6 +812,7 @@ 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) && @@ -688,7 +841,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, GOTO(out_put, rc = -ETXTBSY); /* LU-10286: compatibility check for FLR. - * Please check the comment in mdt_finish_open() for details */ + * 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); @@ -766,19 +920,6 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, buf->lb_buf = lmu; buf->lb_len = ma->ma_lmv_size; - - if (le32_to_cpu(lmu->lum_hash_type) & - LMV_HASH_FLAG_SPACE) { - /* - * only allow setting "space" hash flag on - * plain directory. - */ - rc = mdt_object_striped(info, mo); - if (rc) - GOTO(out_put, - rc = (rc == 1) ? -EPERM : rc); - } - name = XATTR_NAME_DEFAULT_LMV; /* force client to update dir default layout */ lockpart |= MDS_INODELOCK_LOOKUP; @@ -805,8 +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); @@ -818,24 +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), @@ -846,16 +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", @@ -865,6 +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); } @@ -887,10 +1036,10 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, __u64 lock_ibits; 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)); @@ -940,15 +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)) @@ -1012,7 +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); @@ -1022,7 +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 */ @@ -1075,7 +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: @@ -1083,7 +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", @@ -1102,7 +1255,7 @@ unlock_parent: put_parent: mdt_object_put(info->mti_env, mp); CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE); - return rc; + return rc; } /* @@ -1119,21 +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); @@ -1209,7 +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: @@ -1266,15 +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)) @@ -1292,7 +1454,7 @@ 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(info->mti_env, ns, res_id, @@ -1324,7 +1486,6 @@ static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info, int rc; ENTRY; - dir = mdt_object_find(info->mti_env, info->mti_mdt, fid); if (IS_ERR(dir)) RETURN(dir); @@ -1399,9 +1560,9 @@ int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info, * different list. */ struct mdt_sub_lock { - struct mdt_object *msl_obj; - struct mdt_lock_handle msl_lh; - struct list_head msl_linkage; + 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, @@ -1458,7 +1619,6 @@ static int mdt_link_parents_lock(struct mdt_thread_info *info, int rc; ENTRY; - if (S_ISDIR(lu_object_attr(&obj->mot_obj))) RETURN(0); @@ -1660,17 +1820,13 @@ static int mdt_lock_remote_slaves(struct mdt_thread_info *info, int rc; ENTRY; - LASSERT(mdt_object_remote(obj)); LASSERT(ma->ma_valid & MA_LMV); LASSERT(lmv); - if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1) + if (!lmv_is_sane(lmv)) RETURN(-EINVAL); - if (le32_to_cpu(lmv->lmv_stripe_count) < 1) - RETURN(0); - for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) { fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]); @@ -1774,9 +1930,6 @@ static int mdt_migrate_object_lock(struct mdt_thread_info *info, if (S_ISDIR(lu_object_attr(&obj->mot_obj))) { struct md_attr *ma = &info->mti_attr; - ma->ma_lmv = info->mti_big_lmm; - ma->ma_lmv_size = info->mti_big_lmmsize; - ma->ma_valid = 0; rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV); if (rc) { mdt_object_unlock(info, obj, lh, rc); @@ -1826,25 +1979,15 @@ static int mdt_migrate_lookup(struct mdt_thread_info *info, 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; - __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type); - __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count); - bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) & - LMV_HASH_FLAG_MIGRATION; - - if (is_migrating) { - hash_type = le32_to_cpu(lmv->lmv_migrate_hash); - stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset); - } - rc = lmv_name_to_stripe_index(hash_type, stripe_count, - lname->ln_name, - lname->ln_namelen); + 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; - if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION) - rc += le32_to_cpu(lmv->lmv_migrate_offset); - fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]); stripe = mdt_object_find(env, info->mti_mdt, fid); @@ -1854,20 +1997,17 @@ static int mdt_migrate_lookup(struct mdt_thread_info *info, fid_zero(fid); rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid, &info->mti_spec); - if (rc == -ENOENT && is_migrating) { + if (rc == -ENOENT && lmv_is_layout_changing(lmv)) { /* - * if parent is migrating, 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. + * 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); - hash_type = le32_to_cpu(lmv->lmv_hash_type); - stripe_count = le32_to_cpu(lmv->lmv_migrate_offset); - - rc = lmv_name_to_stripe_index(hash_type, stripe_count, - lname->ln_name, + rc = lmv_name_to_stripe_index(lmv, lname->ln_name, lname->ln_namelen); if (rc < 0) return rc; @@ -1977,8 +2117,8 @@ close: * 9. unlock above locks * 10. sync device if source has links */ -static int mdt_reint_migrate(struct mdt_thread_info *info, - struct mdt_lock_handle *unused) +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; @@ -2003,8 +2143,8 @@ static int mdt_reint_migrate(struct mdt_thread_info *info, bool open_sem_locked = false; bool do_sync = false; int rc; - ENTRY; + ENTRY; CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2)); @@ -2022,7 +2162,7 @@ static int mdt_reint_migrate(struct mdt_thread_info *info, if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration) RETURN(-EPERM); - if (!md_capable(uc, CFS_CAP_SYS_ADMIN) && + 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); @@ -2032,8 +2172,10 @@ static int mdt_reint_migrate(struct mdt_thread_info *info, * 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_is_replay(req)) { + if (req && !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", @@ -2043,20 +2185,22 @@ static int mdt_reint_migrate(struct mdt_thread_info *info, } /* pobj is master object of parent */ - pobj = mdt_parent_find_check(info, rr->rr_fid1, 0); + pobj = mdt_object_find(env, mdt, rr->rr_fid1); if (IS_ERR(pobj)) GOTO(unlock_rename, rc = PTR_ERR(pobj)); - if (unlikely(!info->mti_big_lmm)) { - info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC); - OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize); - if (!info->mti_big_lmm) - GOTO(put_parent, rc = -ENOMEM); + if (req) { + rc = mdt_version_get_check(info, pobj, 0); + if (rc) + GOTO(put_parent, rc); } - ma->ma_lmv = info->mti_big_lmm; - ma->ma_lmv_size = info->mti_big_lmmsize; - ma->ma_valid = 0; + if (!mdt_object_exists(pobj)) + GOTO(put_parent, rc = -ENOENT); + + 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); @@ -2098,18 +2242,14 @@ lock_parent: */ do_sync = rc; - /* TODO: DoM migration is not supported yet */ + /* TODO: DoM migration is not supported, migrate dirent only */ if (S_ISREG(lu_object_attr(&sobj->mot_obj))) { - ma->ma_lmm = info->mti_big_lmm; - ma->ma_lmm_size = info->mti_big_lmmsize; - ma->ma_valid = 0; rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV); if (rc) - GOTO(put_source, rc); + GOTO(unlock_links, rc); - if (ma->ma_valid & MA_LOV && - mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM) - GOTO(put_source, rc = -EOPNOTSUPP); + if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm)) + info->mti_spec.sp_migrate_nsonly = 1; } /* if migration HSM is allowed */ @@ -2164,7 +2304,11 @@ lock_parent: rc = mdo_migrate(env, mdt_object_child(pobj), mdt_object_child(sobj), &rr->rr_name, - mdt_object_child(tobj), &info->mti_spec, ma); + 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); @@ -2178,7 +2322,8 @@ unlock_open_sem: 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 */ + * 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); @@ -2271,10 +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; @@ -2308,6 +2449,71 @@ 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. @@ -2327,15 +2533,17 @@ static int mdt_reint_rename(struct mdt_thread_info *info, 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, 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)); @@ -2399,7 +2607,8 @@ static int mdt_reint_rename(struct mdt_thread_info *info, /* 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)); @@ -2482,7 +2691,8 @@ relock: 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); @@ -2507,7 +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); @@ -2548,35 +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); @@ -2590,43 +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); } @@ -2650,19 +2851,23 @@ 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 = 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 && !discard) @@ -2696,16 +2901,16 @@ 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 = { .mlc_mirror_id = rr->rr_mirror_id }; - 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; + ENTRY; DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1)); if (info->mti_dlm_req) @@ -2731,7 +2936,8 @@ static int mdt_reint_resync(struct mdt_thread_info *info, /* 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); @@ -2743,10 +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); @@ -2824,8 +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);