X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=a4ea444abafd36b3554db1b8ea7bf6c97f1e37c0;hp=909091c7280f373eac6982800b4262f2c888efa0;hb=640ed6104453e912a5c7766d265a36a30a31761d;hpb=e5a5abca5637d59187475bfc2e0ef45e2af41420 diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 909091c..a4ea444 100644 --- a/lustre/mdt/mdt_reint.c +++ b/lustre/mdt/mdt_reint.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2013, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -46,7 +42,9 @@ #define DEBUG_SUBSYSTEM S_MDS +#include #include "mdt_internal.h" +#include static inline void mdt_reint_init_ma(struct mdt_thread_info *info, struct md_attr *ma) @@ -55,33 +53,6 @@ static inline void mdt_reint_init_ma(struct mdt_thread_info *info, ma->ma_valid = 0; } -static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc, - struct mdt_object *object, - struct mdt_body *repbody) -{ - ENTRY; - - /* for cross-ref mkdir, mds capa has been fetched from remote obj, then - * we won't go to below*/ - if (repbody->valid & OBD_MD_FLMDSCAPA) - RETURN(rc); - - if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa && - exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) { - struct lustre_capa *capa; - - capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1); - LASSERT(capa); - capa->lc_opc = CAPA_OPC_MDS_DEFAULT; - rc = mo_capa_get(info->mti_env, mdt_object_child(object), capa, - 0); - if (rc == 0) - repbody->valid |= OBD_MD_FLMDSCAPA; - } - - RETURN(rc); -} - /** * Get version of object by fid. * @@ -90,14 +61,15 @@ static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc, static void mdt_obj_version_get(struct mdt_thread_info *info, struct mdt_object *o, __u64 *version) { - LASSERT(o); + LASSERT(o); + if (mdt_object_exists(o) && !mdt_object_remote(o) && !fid_is_obf(mdt_object_fid(o))) - *version = dt_version_get(info->mti_env, mdt_obj2dt(o)); - else - *version = ENOENT_VERSION; - CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n", - PFID(mdt_object_fid(o)), *version); + *version = dt_version_get(info->mti_env, mdt_obj2dt(o)); + else + *version = ENOENT_VERSION; + CDEBUG(D_INODE, "FID "DFID" version is %#llx\n", + PFID(mdt_object_fid(o)), *version); } /** @@ -125,7 +97,7 @@ static int mdt_version_check(struct ptlrpc_request *req, spin_unlock(&req->rq_export->exp_lock); RETURN(-EOVERFLOW); } else if (pre_ver[idx] != version) { - CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n", + CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n", pre_ver[idx], version); spin_lock(&req->rq_export->exp_lock); req->rq_export->exp_vbr_failed = 1; @@ -218,9 +190,10 @@ 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. */ -int mdt_lookup_version_check(struct mdt_thread_info *info, - struct mdt_object *p, struct lu_name *lname, - struct lu_fid *fid, int idx) +static int mdt_lookup_version_check(struct mdt_thread_info *info, + struct mdt_object *p, + const struct lu_name *lname, + struct lu_fid *fid, int idx) { int rc, vbrc; @@ -244,6 +217,123 @@ int mdt_lookup_version_check(struct mdt_thread_info *info, } +static int mdt_unlock_slaves(struct mdt_thread_info *mti, + struct mdt_object *obj, + struct ldlm_enqueue_info *einfo, + int decref) +{ + union ldlm_policy_data *policy = &mti->mti_policy; + struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL]; + struct lustre_handle_array *slave_locks = einfo->ei_cbdata; + int i; + + LASSERT(S_ISDIR(obj->mot_header.loh_attr)); + LASSERT(slave_locks); + + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = einfo->ei_inodebits; + mdt_lock_handle_init(lh); + mdt_lock_reg_init(lh, einfo->ei_mode); + for (i = 0; i < slave_locks->ha_count; i++) { + if (test_bit(i, (void *)slave_locks->ha_map)) + lh->mlh_rreg_lh = slave_locks->ha_handles[i]; + else + lh->mlh_reg_lh = slave_locks->ha_handles[i]; + mdt_object_unlock(mti, NULL, lh, decref); + slave_locks->ha_handles[i].cookie = 0ull; + } + + return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo, + policy); +} + +static inline int mdt_object_striped(struct mdt_thread_info *mti, + struct mdt_object *obj) +{ + int rc; + + if (!S_ISDIR(obj->mot_header.loh_attr)) + return 0; + + rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL, + XATTR_NAME_LMV); + if (rc <= 0) + return rc == -ENODATA ? 0 : rc; + + return 1; +} + +/** + * 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, + enum ldlm_mode mode, __u64 ibits, + struct ldlm_enqueue_info *einfo) +{ + union ldlm_policy_data *policy = &mti->mti_policy; + + LASSERT(S_ISDIR(obj->mot_header.loh_attr)); + + einfo->ei_type = LDLM_IBITS; + einfo->ei_mode = mode; + einfo->ei_cb_bl = mdt_remote_blocking_ast; + einfo->ei_cb_local_bl = mdt_blocking_ast; + einfo->ei_cb_cp = ldlm_completion_ast; + einfo->ei_enq_slave = 1; + einfo->ei_namespace = mti->mti_mdt->mdt_namespace; + einfo->ei_inodebits = ibits; + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = ibits; + + return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo, + policy); +} + +int mdt_reint_striped_lock(struct mdt_thread_info *info, + struct mdt_object *o, + struct mdt_lock_handle *lh, + __u64 ibits, + struct ldlm_enqueue_info *einfo, + bool cos_incompat) +{ + int rc; + + LASSERT(!mdt_object_remote(o)); + + memset(einfo, 0, sizeof(*einfo)); + + rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat); + if (rc) + return rc; + + rc = mdt_object_striped(info, o); + if (rc != 1) { + if (rc < 0) + mdt_object_unlock(info, o, lh, rc); + return rc; + } + + rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo); + if (rc) { + mdt_object_unlock(info, o, lh, rc); + if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME)) + rc = 0; + } + + return rc; +} + +void mdt_reint_striped_unlock(struct mdt_thread_info *info, + struct mdt_object *o, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, int decref) +{ + if (einfo->ei_cbdata) + mdt_unlock_slaves(info, o, einfo, decref); + mdt_object_unlock(info, o, lh, decref); +} + /* * VBR: we save three versions in reply: * 0 - parent. Check that parent version is the same during replay. @@ -252,156 +342,200 @@ int mdt_lookup_version_check(struct mdt_thread_info *info, * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity * check. */ -static int mdt_md_create(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_reint_record *rr = &info->mti_rr; - struct lu_name *lname; - int rc; - ENTRY; + 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; + 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); + if (S_ISDIR(ma->ma_attr.la_mode) && + spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) { + const struct lmv_user_md *lum = spec->u.sp_ea.eadata; + struct lu_ucred *uc = mdt_ucred(info); + struct obd_export *exp = mdt_info_req(info)->rq_export; + + /* Only new clients can create remote dir( >= 2.4) and + * striped dir(>= 2.6), old client will return -ENOTSUPP */ + if (!mdt_is_dne_client(exp)) + RETURN(-ENOTSUPP); + + if (le32_to_cpu(lum->lum_stripe_count) > 1) { + if (!mdt_is_striped_client(exp)) + RETURN(-ENOTSUPP); + + if (!mdt->mdt_enable_striped_dir) + RETURN(-EPERM); + } else if (!mdt->mdt_enable_remote_dir) { + RETURN(-EPERM); + } + + if (!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); + } + 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_lock(info, rr->rr_fid1, lh, - MDS_INODELOCK_UPDATE); - if (IS_ERR(parent)) - RETURN(PTR_ERR(parent)); - - rc = mdt_version_get_check_save(info, parent, 0); - if (rc) - GOTO(out_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); + parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(parent)) + RETURN(PTR_ERR(parent)); + + if (!mdt_object_exists(parent)) + GOTO(put_parent, rc = -ENOENT); + + 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); + if (rc) + GOTO(put_parent, rc); + + 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); child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2); - 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); - } - } + if (unlikely(IS_ERR(child))) + GOTO(unlock_parent, rc = PTR_ERR(child)); - 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); - } + ma->ma_need = MA_INODE; + ma->ma_valid = 0; + + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_CREATE_WRITE); + + /* Version of child will be updated on disk. */ + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child)); + rc = mdt_version_get_check_save(info, child, 2); + if (rc) + GOTO(put_child, rc); + + /* Let lower layer know current lock mode. */ + info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode); + + /* + * Do not perform lookup sanity check. We know that name does + * not exist. + */ + info->mti_spec.sp_cr_lookup = 0; + info->mti_spec.sp_feat = &dt_directory_features; + + rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name, + mdt_object_child(child), &info->mti_spec, ma); + if (rc == 0) + rc = mdt_attr_get_complex(info, child, ma); + if (rc < 0) + GOTO(put_child, rc); + + /* + * On DNE, we need to eliminate dependey between 'mkdir a' and + * 'mkdir a/b' if b is a striped directory, to achieve this, two + * things are done below: + * 1. save child and slaves lock. + * 2. if the child is a striped directory, relock parent so to + * compare against with COS locks to ensure parent was + * committed to disk. + */ + if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) { + struct mdt_lock_handle *lhc; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; + bool cos_incompat; + + rc = mdt_object_striped(info, child); + if (rc < 0) + GOTO(put_child, rc); + + cos_incompat = rc; + if (cos_incompat) { + if (!mdt_object_remote(parent)) { + mdt_object_unlock(info, parent, lh, 1); + mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name); + rc = mdt_reint_object_lock(info, parent, lh, + MDS_INODELOCK_UPDATE, + true); + if (rc) + GOTO(put_child, rc); + } } - 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, - &RMF_CAPA1); - LASSERT(ma->ma_capa); - - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_CREATE_WRITE); - - /* Version of child will be updated on disk. */ - info->mti_mos = child; - rc = mdt_version_get_check_save(info, child, 2); - if (rc) - GOTO(out_put_child, rc); - - /* Let lower layer know current lock mode. */ - info->mti_spec.sp_cr_mode = - mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode); - /* - * Do not perform lookup sanity check. We know that name does - * not exist. - */ - 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); - if (rc == 0) - rc = mdt_attr_get_complex(info, child, ma); - - if (rc == 0) { - /* Return fid & attr to client. */ - if (ma->ma_valid & MA_INODE) - mdt_pack_attr2body(info, repbody, &ma->ma_attr, - mdt_object_fid(child)); - } -out_put_child: - mdt_object_put(info->mti_env, child); - } else { - 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); + lhc = &info->mti_lh[MDT_LH_CHILD]; + mdt_lock_handle_init(lhc); + mdt_lock_reg_init(lhc, LCK_PW); + rc = mdt_reint_striped_lock(info, child, lhc, + MDS_INODELOCK_UPDATE, einfo, + cos_incompat); + if (rc) + GOTO(put_child, rc); + + mdt_reint_striped_unlock(info, child, lhc, einfo, rc); + } + + /* Return fid & attr to client. */ + if (ma->ma_valid & MA_INODE) + mdt_pack_attr2body(info, repbody, &ma->ma_attr, + mdt_object_fid(child)); +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); } -int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, - struct md_attr *ma, int flags) +static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, + struct md_attr *ma) { - 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_PROJID | LA_FLAGS); + __u64 lockpart = MDS_INODELOCK_UPDATE; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; + bool cos_incompat; + int rc; + ENTRY; + + rc = mdt_object_striped(info, mo); + if (rc < 0) + RETURN(rc); - /* attr shouldn't be set on remote object */ - LASSERT(!mdt_object_remote(mo)); + cos_incompat = rc; - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_reg_init(lh, LCK_PW); + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lh, LCK_PW); /* Even though the new MDT will grant PERM lock to the old * client, but the old client will almost ignore that during @@ -410,39 +544,41 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM; - rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK); - if (rc != 0) - RETURN(rc); + rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo, + cos_incompat); + if (rc != 0) + RETURN(rc); - if (mdt_object_exists(mo) == 0) - GOTO(out_unlock, rc = -ENOENT); + /* all attrs are packed into mti_attr in unpack_setattr */ + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_SETATTR_WRITE); - /* all attrs are packed into mti_attr in unpack_setattr */ - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_SETATTR_WRITE); + /* VBR: update version if attr changed are important for recovery */ + if (do_vbr) { + /* update on-disk version of changed object */ + 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); + } - /* This is only for set ctime when rename's source is on remote MDS. */ - if (unlikely(ma->ma_attr.la_valid == LA_CTIME)) - ma->ma_attr_flags |= MDS_VTX_BYPASS; + /* Ensure constant striping during chown(). See LU-2789. */ + if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID)) + mutex_lock(&mo->mot_lov_mutex); - /* VBR: update version if attr changed are important for recovery */ - if (do_vbr) { - /* update on-disk version of changed object */ - info->mti_mos = mo; - rc = mdt_version_get_check_save(info, mo, 0); - if (rc) - GOTO(out_unlock, rc); - } + /* all attrs are packed into mti_attr in unpack_setattr */ + rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma); - /* all attrs are packed into mti_attr in unpack_setattr */ - rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma); - if (rc != 0) - GOTO(out_unlock, rc); + if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID)) + mutex_unlock(&mo->mot_lov_mutex); - EXIT; + if (rc != 0) + GOTO(out_unlock, rc); + mdt_dom_obj_lvb_update(info->mti_env, mo, false); + EXIT; out_unlock: - mdt_object_unlock(info, mo, lh, rc); - return rc; + mdt_reint_striped_unlock(info, mo, lh, einfo, rc); + return rc; } /** @@ -454,6 +590,8 @@ out_unlock: int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, struct md_attr *ma) { + struct lu_ucred *uc = mdt_ucred(info); + cfs_cap_t cap_saved; int rc; ENTRY; @@ -469,127 +607,152 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, /* If an up2date copy exists in the backend, add dirty flag */ if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS) && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) { - struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD]; - ma->ma_hsm.mh_flags |= HS_DIRTY; - mdt_lock_reg_init(lh, LCK_PW); - rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR, - MDT_LOCAL_LOCK); - if (rc != 0) - RETURN(rc); - + /* Bump cap so that closes from non-owner writers can + * set the HSM state to dirty. */ + cap_saved = uc->uc_cap; + uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER); rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm); + uc->uc_cap = cap_saved; if (rc) CERROR("file attribute change error for "DFID": %d\n", PFID(mdt_object_fid(mo)), rc); - mdt_object_unlock(info, mo, lh, rc); } RETURN(rc); } static int mdt_reint_setattr(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + struct mdt_lock_handle *lhc) { - struct md_attr *ma = &info->mti_attr; - struct mdt_reint_record *rr = &info->mti_rr; - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_export_data *med = &req->rq_export->exp_mdt_data; - struct mdt_file_data *mfd; - struct mdt_object *mo; - struct mdt_body *repbody; - int som_au, rc, rc2; - ENTRY; + struct mdt_device *mdt = info->mti_mdt; + struct md_attr *ma = &info->mti_attr; + struct mdt_reint_record *rr = &info->mti_rr; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_object *mo; + struct mdt_body *repbody; + int rc, rc2; + ENTRY; - DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1), - (unsigned int)ma->ma_attr.la_valid); + 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); + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); - if (IS_ERR(mo)) - GOTO(out, rc = PTR_ERR(mo)); - - /* start a log jounal handle if needed */ - if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) { - if ((ma->ma_attr.la_valid & LA_SIZE) || - (rr->rr_flags & MRF_OPEN_TRUNC)) { - /* Check write access for the O_TRUNC case */ - if (mdt_write_read(mo) < 0) - GOTO(out_put, rc = -ETXTBSY); - } - } else if (info->mti_ioepoch && - (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) { - /* Truncate case. IOEpoch is opened. */ - rc = mdt_write_get(mo); - if (rc) - GOTO(out_put, rc); - - mfd = mdt_mfd_new(med); - if (mfd == NULL) { - mdt_write_put(mo); - GOTO(out_put, rc = -ENOMEM); - } + mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1); + if (IS_ERR(mo)) + GOTO(out, rc = PTR_ERR(mo)); + + if (!mdt_object_exists(mo)) + GOTO(out_put, rc = -ENOENT); + + if (mdt_object_remote(mo)) + GOTO(out_put, rc = -EREMOTE); + + /* revoke lease lock if size is going to be changed */ + if (unlikely(ma->ma_attr.la_valid & LA_SIZE && + !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) && + atomic_read(&mo->mot_lease_count) > 0)) { + down_read(&mo->mot_open_sem); + + if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */ + lhc = &info->mti_lh[MDT_LH_LOCAL]; + mdt_lock_reg_init(lhc, LCK_CW); + + rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN); + if (rc != 0) { + up_read(&mo->mot_open_sem); + GOTO(out_put, rc); + } - mdt_ioepoch_open(info, mo, 0); - repbody->ioepoch = mo->mot_ioepoch; + /* revoke lease lock */ + mdt_object_unlock(info, mo, lhc, 1); + } + up_read(&mo->mot_open_sem); + } - mdt_object_get(info->mti_env, mo); - mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC); - mfd->mfd_object = mo; - mfd->mfd_xid = req->rq_xid; + if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) { + /* Check write access for the O_TRUNC case */ + if (mdt_write_read(mo) < 0) + GOTO(out_put, rc = -ETXTBSY); - spin_lock(&med->med_open_lock); - cfs_list_add(&mfd->mfd_list, &med->med_open_head); - spin_unlock(&med->med_open_lock); - repbody->handle.cookie = mfd->mfd_handle.h_cookie; - } + /* LU-10286: compatibility check for FLR. + * Please check the comment in mdt_finish_open() for details */ + if (!exp_connect_flr(info->mti_exp)) { + rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV); + if (rc < 0 && rc != -ENODATA) + GOTO(out_put, rc); - som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE; - if (som_au) { - /* SOM Attribute update case. Find the proper mfd and update - * SOM attributes on the proper object. */ - LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM); - LASSERT(info->mti_ioepoch); - - spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle, - req_is_replay(req)); - if (mfd == NULL) { - spin_unlock(&med->med_open_lock); - CDEBUG(D_INODE, "no handle for file close: " - "fid = "DFID": cookie = "LPX64"\n", - PFID(info->mti_rr.rr_fid1), - info->mti_ioepoch->handle.cookie); - GOTO(out_put, rc = -ESTALE); - } - LASSERT(mfd->mfd_mode == MDS_FMODE_SOM); - LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE)); - - class_handle_unhash(&mfd->mfd_handle); - cfs_list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); - - mdt_mfd_close(info, mfd); - } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) { - LASSERT((ma->ma_valid & MA_LOV) == 0); - rc = mdt_attr_set(info, mo, ma, rr->rr_flags); - if (rc) - GOTO(out_put, rc); - } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) { - struct lu_buf *buf = &info->mti_buf; - LASSERT(ma->ma_attr.la_valid == 0); - buf->lb_buf = ma->ma_lmm; - buf->lb_len = ma->ma_lmm_size; - rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), - buf, XATTR_NAME_LOV, 0); + if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm)) + GOTO(out_put, rc = -EOPNOTSUPP); + } + + /* For truncate, the file size sent from client + * is believable, but the blocks are incorrect, + * which makes the block size in LSOM attribute + * inconsisent with the real block size. + */ + rc = mdt_lsom_update(info, mo, true); + if (rc) + GOTO(out_put, rc); + } + + if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) { + if (ma->ma_valid & MA_LOV) + GOTO(out_put, rc = -EPROTO); + + rc = mdt_attr_set(info, mo, ma); + if (rc) + GOTO(out_put, rc); + } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) && + (ma->ma_valid & MA_INODE)) { + struct lu_buf *buf = &info->mti_buf; + struct lu_ucred *uc = mdt_ucred(info); + struct mdt_lock_handle *lh; + + /* reject if either remote or striped dir is disabled */ + if (ma->ma_valid & MA_LMV) { + if (!mdt->mdt_enable_remote_dir || + !mdt->mdt_enable_striped_dir) + GOTO(out_put, rc = -EPERM); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN) && + uc->uc_gid != mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) + GOTO(out_put, rc = -EPERM); + } + + if (ma->ma_attr.la_valid != 0) + GOTO(out_put, rc = -EPROTO); + + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lh, LCK_PW); + + rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR); + if (rc != 0) + GOTO(out_put, rc); + + if (ma->ma_valid & MA_LOV) { + buf->lb_buf = ma->ma_lmm; + buf->lb_len = ma->ma_lmm_size; + } else { + 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, + (ma->ma_valid & MA_LOV) ? + XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV, + 0); + + mdt_object_unlock(info, mo, lh, rc); if (rc) GOTO(out_put, rc); - } else - LBUG(); + } else { + GOTO(out_put, rc = -EPROTO); + } /* If file data is modified, add the dirty flag */ if (ma->ma_attr_flags & MDS_DATA_MODIFIED) @@ -598,31 +761,16 @@ 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); - - mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); - - if (info->mti_mdt->mdt_opts.mo_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) { - struct lustre_capa *capa; - - capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2); - LASSERT(capa); - capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC; - rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0); - if (rc) - GOTO(out_put, rc); - repbody->valid |= OBD_MD_FLOSSCAPA; - } + if (rc != 0) + GOTO(out_put, rc); - EXIT; + mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); + + EXIT; out_put: - mdt_object_put(info->mti_env, mo); + mdt_object_put(info->mti_env, mo); out: - if (rc == 0) + if (rc == 0) mdt_counter_incr(req, LPROC_MDT_SETATTR); mdt_client_compatibility(info); @@ -642,11 +790,14 @@ static int mdt_reint_create(struct mdt_thread_info *info, 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), info->mti_dlm_req, 0); + if (info->mti_dlm_req) + ldlm_request_cancel(mdt_info_req(info), + info->mti_dlm_req, 0, LATF_SKIP); + + if (!lu_name_is_valid(&info->mti_rr.rr_name)) + RETURN(-EPROTO); - LASSERT(info->mti_rr.rr_namelen > 0); - switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { + switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { case S_IFDIR: mdt_counter_incr(req, LPROC_MDT_MKDIR); break; @@ -666,7 +817,7 @@ static int mdt_reint_create(struct mdt_thread_info *info, RETURN(err_serious(-EOPNOTSUPP)); } - rc = mdt_md_create(info); + rc = mdt_create(info); RETURN(rc); } @@ -675,68 +826,59 @@ static int mdt_reint_create(struct mdt_thread_info *info, * Version of child is getting and checking during its lookup. If */ static int mdt_reint_unlink(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + 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 lu_fid *child_fid = &info->mti_tmp_fid1; - struct mdt_object *mp; - 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 mdt_reint_record *rr = &info->mti_rr; + struct ptlrpc_request *req = mdt_info_req(info); + struct md_attr *ma = &info->mti_attr; + struct lu_fid *child_fid = &info->mti_tmp_fid1; + struct mdt_object *mp; + struct mdt_object *mc; + struct mdt_lock_handle *parent_lh; + struct mdt_lock_handle *child_lh; + struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; + __u64 lock_ibits; + bool cos_incompat = false, discard = false; + int no_name = 0; + int rc; + 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); + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) - RETURN(err_serious(-ENOENT)); + 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. - */ + mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); - if (IS_ERR(mp)) { - rc = PTR_ERR(mp); - GOTO(out, rc); - } + if (IS_ERR(mp)) + RETURN(PTR_ERR(mp)); - 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); - + cos_incompat = true; } 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); - rc = mdt_version_get_check_save(info, mp, 0); if (rc) - GOTO(unlock_parent, rc); + GOTO(put_parent, rc); } - /* step 2: find & lock the child */ +relock: + parent_lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name); + rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE, + cos_incompat); + if (rc != 0) + GOTO(put_parent, rc); + /* 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. @@ -762,126 +904,163 @@ 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)) GOTO(unlock_parent, rc = PTR_ERR(mc)); - child_lh = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(child_lh, LCK_EX); + if (!cos_incompat) { + rc = mdt_object_striped(info, mc); + if (rc < 0) + GOTO(unlock_parent, rc); + + cos_incompat = rc; + if (cos_incompat) { + mdt_object_put(info->mti_env, mc); + mdt_object_unlock(info, mp, parent_lh, -EAGAIN); + goto relock; + } + } + + child_lh = &info->mti_lh[MDT_LH_CHILD]; + mdt_lock_reg_init(child_lh, LCK_EX); + if (info->mti_spec.sp_rm_entry) { + struct lu_ucred *uc = mdt_ucred(info); + + if (!mdt_is_dne_client(req->rq_export)) + /* Return -ENOTSUPP for old client */ + GOTO(put_child, rc = -ENOTSUPP); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) + GOTO(put_child, rc = -EPERM); + + ma->ma_need = MA_INODE; + ma->ma_valid = 0; + rc = mdo_unlink(info->mti_env, mdt_object_child(mp), + NULL, &rr->rr_name, ma, no_name); + GOTO(put_child, rc); + } + if (mdt_object_remote(mc)) { 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); - - if (info->mti_spec.sp_rm_entry) { - struct lu_ucred *uc = mdt_ucred(info); - - if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) { - CERROR("%s: unlink remote entry is only " - "permitted for administrator: rc = %d\n", - mdt_obd_name(info->mti_mdt), - -EPERM); - GOTO(put_child, rc = -EPERM); - } + /* Return -ENOTSUPP for old client */ + GOTO(put_child, rc = -ENOTSUPP); - ma->ma_need = MA_INODE; - 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); - GOTO(put_child, rc); - } /* Revoke the LOOKUP lock of the remote object granted by * 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 */ - mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP, - MDT_CROSS_LOCK); + mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP); repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); LASSERT(repbody != NULL); - repbody->fid1 = *mdt_object_fid(mc); - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); + repbody->mbo_fid1 = *mdt_object_fid(mc); + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); 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); - GOTO(put_child, rc); } + /* 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. */ + lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE; + if (mdt_object_remote(mp)) { + /* Enqueue lookup lock from parent MDT */ + rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc), + &child_lh->mlh_rreg_lh, + child_lh->mlh_rreg_mode, + MDS_INODELOCK_LOOKUP, false); + if (rc != ELDLM_OK) + GOTO(put_child, rc); - rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL, - MDT_CROSS_LOCK); - if (rc != 0) { - GOTO(put_child, rc); + lock_ibits &= ~MDS_INODELOCK_LOOKUP; } - 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); + rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo, + cos_incompat); + 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); + + mutex_lock(&mc->mot_lov_mutex); rc = mdo_unlink(info->mti_env, mdt_object_child(mp), - mdt_object_child(mc), lname, ma, no_name); - if (rc == 0 && !lu_object_is_dying(&mc->mot_header)) + mdt_object_child(mc), &rr->rr_name, ma, no_name); + + mutex_unlock(&mc->mot_lov_mutex); + if (rc != 0) + GOTO(unlock_child, rc); + + if (!lu_object_is_dying(&mc->mot_header)) { rc = mdt_attr_get_complex(info, mc, ma); - if (rc == 0) - mdt_handle_last_unlink(info, mc, ma); + if (rc) + GOTO(out_stat, rc); + } else { + discard = mdt_dom_check_for_discard(info, mc); + } + mdt_handle_last_unlink(info, mc, ma); - if (ma->ma_valid & MA_INODE) { - switch (ma->ma_attr.la_mode & S_IFMT) { - case S_IFDIR: +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); - break; - case S_IFREG: - case S_IFLNK: - case S_IFCHR: - case S_IFBLK: - case S_IFIFO: - case S_IFSOCK: + break; + case S_IFREG: + case S_IFLNK: + case S_IFCHR: + case S_IFBLK: + case S_IFIFO: + case S_IFSOCK: mdt_counter_incr(req, LPROC_MDT_UNLINK); - break; - default: - LASSERTF(0, "bad file type %o unlinking\n", - ma->ma_attr.la_mode); - } - } + break; + default: + LASSERTF(0, "bad file type %o unlinking\n", + ma->ma_attr.la_mode); + } + } + + EXIT; - EXIT; unlock_child: - mdt_object_unlock(info, mc, child_lh, rc); + mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc); put_child: mdt_object_put(info->mti_env, mc); unlock_parent: mdt_object_unlock(info, mp, parent_lh, rc); put_parent: mdt_object_put(info->mti_env, mp); -out: + + /* discard is just a PW DOM lock to drop the data on a client + * no need to keep objects being get and locked, do that after all. + */ + if (discard) + mdt_dom_discard_data(info, child_fid); + return rc; } @@ -890,110 +1069,117 @@ out: * name. */ static int mdt_reint_link(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + 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 *ms; - struct mdt_object *mp; - struct mdt_lock_handle *lhs; - struct mdt_lock_handle *lhp; - struct lu_name *lname; - int rc; - ENTRY; + 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 *ms; + struct mdt_object *mp; + struct mdt_lock_handle *lhs; + struct mdt_lock_handle *lhp; + bool cos_incompat; + 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)); + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) + RETURN(err_serious(-ENOENT)); - if (info->mti_dlm_req) - ldlm_request_cancel(req, info->mti_dlm_req, 0); + 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 */ - if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) - RETURN(-EPERM); + /* Invalid case so return error immediately instead of + * processing it */ + 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); - mp = mdt_object_find_lock(info, rr->rr_fid2, lhp, - MDS_INODELOCK_UPDATE); - if (IS_ERR(mp)) - RETURN(PTR_ERR(mp)); - - rc = mdt_version_get_check_save(info, mp, 0); - if (rc) - GOTO(out_unlock_parent, rc); - - /* step 2: find & lock the source */ - lhs = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(lhs, LCK_EX); - - ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); - if (IS_ERR(ms)) - GOTO(out_unlock_parent, rc = PTR_ERR(ms)); - - if (mdt_object_remote(ms)) { - mdt_object_put(info->mti_env, ms); - CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n", - mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1), - PFID(rr->rr_fid2)); - GOTO(out_unlock_parent, rc = -EXDEV); - } + /* step 1: find target parent dir */ + mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2); + if (IS_ERR(mp)) + RETURN(PTR_ERR(mp)); - rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE | - MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); - if (rc != 0) { - mdt_object_put(info->mti_env, ms); - GOTO(out_unlock_parent, rc); - } + rc = mdt_version_get_check_save(info, mp, 0); + if (rc) + GOTO(put_parent, rc); - /* step 3: link it */ - mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, - OBD_FAIL_MDS_REINT_LINK_WRITE); + /* step 2: find source */ + ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(ms)) + GOTO(put_parent, rc = PTR_ERR(ms)); - info->mti_mos = ms; - rc = mdt_version_get_check_save(info, ms, 1); - if (rc) - GOTO(out_unlock_child, rc); + if (!mdt_object_exists(ms)) { + CDEBUG(D_INFO, "%s: "DFID" does not exist.\n", + mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1)); + GOTO(put_source, rc = -ENOENT); + } - 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); - 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))) { + cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms)); + + lhp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name); + rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE, + cos_incompat); + if (rc != 0) + GOTO(put_source, rc); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5); + + lhs = &info->mti_lh[MDT_LH_CHILD]; + mdt_lock_reg_init(lhs, LCK_EX); + rc = mdt_reint_object_lock(info, ms, lhs, + MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR, + cos_incompat); + if (rc != 0) + GOTO(unlock_parent, rc); + + /* step 3: link it */ + mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, + OBD_FAIL_MDS_REINT_LINK_WRITE); + + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms)); + rc = mdt_version_get_check_save(info, ms, 1); + if (rc) + GOTO(unlock_source, rc); + + /** check target version by name during replay */ + rc = mdt_lookup_version_check(info, mp, &rr->rr_name, + &info->mti_tmp_fid1, 2); + if (rc != 0 && rc != -ENOENT) + GOTO(unlock_source, 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); - GOTO(out_unlock_child, rc = -EEXIST); + CDEBUG(D_INFO, "link target "DNAME" existed!\n", + PNAME(&rr->rr_name)); + GOTO(unlock_source, rc = -EEXIST); } - info->mti_ver[2] = ENOENT_VERSION; - mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2); - } + 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) + if (rc == 0) mdt_counter_incr(req, LPROC_MDT_LINK); - EXIT; -out_unlock_child: - mdt_object_unlock_put(info, ms, lhs, rc); -out_unlock_parent: - mdt_object_unlock_put(info, mp, lhp, rc); - return rc; + EXIT; +unlock_source: + mdt_object_unlock(info, ms, lhs, rc); +unlock_parent: + mdt_object_unlock(info, mp, lhp, rc); +put_source: + mdt_object_put(info->mti_env, ms); +put_parent: + mdt_object_put(info->mti_env, mp); + return rc; } /** * lock the part of the directory according to the hash of the name @@ -1001,11 +1187,13 @@ out_unlock_parent: */ static int mdt_pdir_hash_lock(struct mdt_thread_info *info, struct mdt_lock_handle *lh, - struct mdt_object *obj, __u64 ibits) + struct mdt_object *obj, __u64 ibits, + bool cos_incompat) { struct ldlm_res_id *res = &info->mti_res_id; struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; - ldlm_policy_data_t *policy = &info->mti_policy; + union ldlm_policy_data *policy = &info->mti_policy; + __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; int rc; /* @@ -1016,100 +1204,1024 @@ static int mdt_pdir_hash_lock(struct mdt_thread_info *info, fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res); memset(policy, 0, sizeof(*policy)); policy->l_inodebits.bits = ibits; + if (cos_incompat && + (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX)) + dlmflags |= LDLM_FL_COS_INCOMPAT; /* * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is * going to be sent to client. If it is - mdt_intent_policy() path will * fix it up and turn FL_LOCAL flag off. */ - rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy, - res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB, + rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, + policy, res, dlmflags, &info->mti_exp->exp_handle.h_cookie); return rc; } +/** + * Get BFL lock for rename or migrate process. + **/ static int mdt_rename_lock(struct mdt_thread_info *info, - struct lustre_handle *lh) + struct lustre_handle *lh) { - 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); - - 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); + 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 */ + lu_root_fid(fid); + obj = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (IS_ERR(obj)) + RETURN(PTR_ERR(obj)); + + rc = mdt_remote_object_lock(info, obj, + &LUSTRE_BFL_FID, lh, + LCK_EX, + MDS_INODELOCK_UPDATE, false); + mdt_object_put(info->mti_env, obj); + } else { + struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; + union ldlm_policy_data *policy = &info->mti_policy; + struct ldlm_res_id *res_id = &info->mti_res_id; + __u64 flags = 0; + + 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(info->mti_env, ns, res_id, + LDLM_IBITS, policy, LCK_EX, &flags, + ldlm_blocking_ast, + ldlm_completion_ast, NULL, NULL, 0, + LVB_T_NONE, + &info->mti_exp->exp_handle.h_cookie, + lh); + RETURN(rc); + } + RETURN(rc); } 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; } -/* - * This is is_subdir() variant, it is CMD if cmm forwards it to correct - * 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 struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info, + const struct lu_fid *fid, + int idx) { - struct mdt_reint_record *rr = &info->mti_rr; - struct lu_fid dst_fid = *rr->rr_fid2; - struct mdt_object *dst; - int rc = 0; - ENTRY; + struct mdt_object *dir; + int rc; - 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); + 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); +} + +/* + * in case obj is remote obj on its parent, revoke LOOKUP lock, + * herein we don't really check it, just do revoke. + */ +int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info, + struct mdt_object *pobj, + struct mdt_object *obj) +{ + struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL]; + int rc; + + mdt_lock_handle_init(lh); + mdt_lock_reg_init(lh, LCK_EX); + + if (mdt_object_remote(pobj)) { + /* don't bother to check if pobj and obj are on the same MDT. */ + rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_EX, + MDS_INODELOCK_LOOKUP, false); + } else if (mdt_object_remote(obj)) { + struct ldlm_res_id *res = &info->mti_res_id; + union ldlm_policy_data *policy = &info->mti_policy; + __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB | + LDLM_FL_COS_INCOMPAT; + + fid_build_reg_res_name(mdt_object_fid(obj), res); + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP; + rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace, + &lh->mlh_reg_lh, LCK_EX, policy, res, + dlmflags, NULL); + } else { + /* do nothing if both are local */ + return 0; + } + + if (rc != ELDLM_OK) + return rc; + + /* + * TODO, currently we don't save this lock because there is no place to + * hold this lock handle, but to avoid race we need to save this lock. + */ + mdt_object_unlock(info, NULL, lh, 1); + + return 0; +} + +/* + * operation may takes locks of linkea, or directory stripes, group them in + * different list. + */ +struct mdt_sub_lock { + struct mdt_object *msl_obj; + struct mdt_lock_handle msl_lh; + struct list_head msl_linkage; +}; + +static void mdt_unlock_list(struct mdt_thread_info *info, + struct list_head *list, int decref) +{ + struct mdt_sub_lock *msl; + struct mdt_sub_lock *tmp; + + list_for_each_entry_safe(msl, tmp, list, msl_linkage) { + mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref); + list_del(&msl->msl_linkage); + OBD_FREE_PTR(msl); + } +} + +/* + * lock parents of links, and also check whether total locks don't exceed + * RS_MAX_LOCKS. + * + * \retval 0 on success, and locks can be saved in ptlrpc_reply_stat + * \retval 1 on success, but total lock count may exceed RS_MAX_LOCKS + * \retval -ev negative errno upon error + */ +static int mdt_lock_links(struct mdt_thread_info *info, + struct mdt_object *pobj, + const struct md_attr *ma, + struct mdt_object *obj, + struct list_head *link_locks) +{ + struct mdt_device *mdt = info->mti_mdt; + struct lu_buf *buf = &info->mti_big_buf; + struct lu_name *lname = &info->mti_name; + struct linkea_data ldata = { NULL }; + bool blocked = false; + int retries = 5; + int local_lnkp_cnt = 0; + int rc; + + ENTRY; + + if (S_ISDIR(lu_object_attr(&obj->mot_obj))) + RETURN(0); + + buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE); + if (buf->lb_buf == NULL) + RETURN(-ENOMEM); + + ldata.ld_buf = buf; + rc = mdt_links_read(info, obj, &ldata); + if (rc) { + if (rc == -ENOENT || rc == -ENODATA) + rc = 0; + RETURN(rc); + } + +repeat: + for (linkea_first_entry(&ldata); ldata.ld_lee && !rc; + linkea_next_entry(&ldata)) { + struct mdt_object *lnkp; + struct mdt_sub_lock *msl; + struct lu_fid fid; + __u64 ibits; + + linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname, + &fid); + + /* check if it's also linked to parent */ + if (lu_fid_eq(mdt_object_fid(pobj), &fid)) { + CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n", + PFID(&fid), PNAME(lname)); + /* in case link is remote object, revoke LOOKUP lock */ + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + continue; + } + + lnkp = NULL; + + /* check if it's linked to a stripe of parent */ + if (ma->ma_valid & MA_LMV) { + struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + struct lu_fid *stripe_fid = &info->mti_tmp_fid1; + int j = 0; + + for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) { + fid_le_to_cpu(stripe_fid, + &lmv->lmv_stripe_fids[j]); + if (lu_fid_eq(stripe_fid, &fid)) { + CDEBUG(D_INFO, "skip stripe "DFID + ", reovke "DNAME"\n", + PFID(&fid), PNAME(lname)); + lnkp = mdt_object_find(info->mti_env, + mdt, &fid); + if (IS_ERR(lnkp)) + GOTO(out, rc = PTR_ERR(lnkp)); + break; + } + } + + if (lnkp) { + rc = mdt_revoke_remote_lookup_lock(info, lnkp, + obj); + mdt_object_put(info->mti_env, lnkp); + continue; + } + } + + /* Check if it's already locked */ + list_for_each_entry(msl, link_locks, msl_linkage) { + if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) { + CDEBUG(D_INFO, + DFID" was locked, revoke "DNAME"\n", + PFID(&fid), PNAME(lname)); + lnkp = msl->msl_obj; + break; + } + } + + if (lnkp) { + rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj); + continue; + } + + CDEBUG(D_INFO, "lock "DFID":"DNAME"\n", + PFID(&fid), PNAME(lname)); + + lnkp = mdt_object_find(info->mti_env, mdt, &fid); + if (IS_ERR(lnkp)) { + CWARN("%s: cannot find obj "DFID": %ld\n", + mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp)); + continue; + } + + if (!mdt_object_exists(lnkp)) { + CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n", + PFID(&fid), PNAME(lname)); + mdt_object_put(info->mti_env, lnkp); + continue; + } + + if (!mdt_object_remote(lnkp)) + local_lnkp_cnt++; + + OBD_ALLOC_PTR(msl); + if (msl == NULL) + GOTO(out, rc = -ENOMEM); + + /* + * we can't follow parent-child lock order like other MD + * operations, use lock_try here to avoid deadlock, if the lock + * cannot be taken, drop all locks taken, revoke the blocked + * one, and continue processing the remaining entries, and in + * the end of the loop restart from beginning. + */ + mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname); + ibits = 0; + rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits, + MDS_INODELOCK_UPDATE, true); + if (!(ibits & MDS_INODELOCK_UPDATE)) { + blocked = true; + + CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n", + PFID(&fid), PNAME(lname), retries); + + mdt_unlock_list(info, link_locks, 1); + + mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname); + rc = mdt_object_lock(info, lnkp, &msl->msl_lh, + MDS_INODELOCK_UPDATE); + if (rc) { + mdt_object_put(info->mti_env, lnkp); + OBD_FREE_PTR(msl); + GOTO(out, rc); + } + + if (mdt_object_remote(lnkp)) { + struct ldlm_lock *lock; + + /* + * for remote object, set lock cb_atomic, + * so lock can be released in blocking_ast() + * immediately, then the next lock_try will + * have better chance of success. + */ + lock = ldlm_handle2lock( + &msl->msl_lh.mlh_rreg_lh); + LASSERT(lock != NULL); + lock_res_and_lock(lock); + ldlm_set_atomic_cb(lock); + unlock_res_and_lock(lock); + LDLM_LOCK_PUT(lock); + } + + mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1); + OBD_FREE_PTR(msl); + continue; + } + + INIT_LIST_HEAD(&msl->msl_linkage); + msl->msl_obj = lnkp; + list_add_tail(&msl->msl_linkage, link_locks); + + rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj); + } + + if (blocked) { + rc = -EBUSY; + if (--retries > 0) { + mdt_unlock_list(info, link_locks, rc); + blocked = false; + local_lnkp_cnt = 0; + goto repeat; + } + } + + EXIT; +out: + if (rc) + mdt_unlock_list(info, link_locks, rc); + else if (local_lnkp_cnt > RS_MAX_LOCKS - 6) + /* + * parent may have 3 local objects: master object and 2 stripes + * (if it's being migrated too); source may have 2 local + * objects: master and 1 stripe; target has 1 local object. + */ + rc = 1; + return rc; +} + +static int mdt_lock_remote_slaves(struct mdt_thread_info *info, + struct mdt_object *obj, + const struct md_attr *ma, + struct list_head *slave_locks) +{ + struct mdt_device *mdt = info->mti_mdt; + const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + struct lu_fid *fid = &info->mti_tmp_fid1; + struct mdt_object *slave; + struct mdt_sub_lock *msl; + int i; + int rc; + + ENTRY; - RETURN(rc); + LASSERT(mdt_object_remote(obj)); + LASSERT(ma->ma_valid & MA_LMV); + LASSERT(lmv); + + if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1) + 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]); + + slave = mdt_object_find(info->mti_env, mdt, fid); + if (IS_ERR(slave)) + GOTO(out, rc = PTR_ERR(slave)); + + OBD_ALLOC_PTR(msl); + if (!msl) { + mdt_object_put(info->mti_env, slave); + GOTO(out, rc = -ENOMEM); + } + + mdt_lock_reg_init(&msl->msl_lh, LCK_EX); + rc = mdt_reint_object_lock(info, slave, &msl->msl_lh, + MDS_INODELOCK_UPDATE, true); + if (rc) { + OBD_FREE_PTR(msl); + mdt_object_put(info->mti_env, slave); + GOTO(out, rc); + } + + INIT_LIST_HEAD(&msl->msl_linkage); + msl->msl_obj = slave; + list_add_tail(&msl->msl_linkage, slave_locks); + } + EXIT; + +out: + if (rc) + mdt_unlock_list(info, slave_locks, rc); + return rc; +} + +static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info, + struct mdt_object *obj, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks, + int decref) +{ + if (mdt_object_remote(obj)) { + mdt_unlock_list(info, slave_locks, decref); + mdt_object_unlock(info, obj, lh, decref); + } else { + mdt_reint_striped_unlock(info, obj, lh, einfo, decref); + } +} + +/* lock parent and its stripes */ +static int mdt_migrate_parent_lock(struct mdt_thread_info *info, + struct mdt_object *obj, + const struct md_attr *ma, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks) +{ + int rc; + + if (mdt_object_remote(obj)) { + rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_PW, + MDS_INODELOCK_UPDATE, false); + if (rc != ELDLM_OK) + return rc; + + /* + * if obj is remote and striped, lock its stripes explicitly + * because it's not striped in LOD layer on this MDT. + */ + if (ma->ma_valid & MA_LMV) { + rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks); + if (rc) + mdt_object_unlock(info, obj, lh, rc); + } + } else { + rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE, + einfo, true); + } + + return rc; +} + +/* + * in migration, object may be remote, and we need take full lock of it and its + * stripes if it's directory, besides, object may be a remote object on its + * parent, revoke its LOOKUP lock on where its parent is located. + */ +static int mdt_migrate_object_lock(struct mdt_thread_info *info, + struct mdt_object *pobj, + struct mdt_object *obj, + struct mdt_lock_handle *lh, + struct ldlm_enqueue_info *einfo, + struct list_head *slave_locks) +{ + int rc; + + if (mdt_object_remote(obj)) { + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + if (rc) + return rc; + + rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj), + &lh->mlh_rreg_lh, LCK_EX, + MDS_INODELOCK_FULL, false); + if (rc != ELDLM_OK) + return rc; + + /* + * if obj is remote and striped, lock its stripes explicitly + * because it's not striped in LOD layer on this MDT. + */ + if (S_ISDIR(lu_object_attr(&obj->mot_obj))) { + struct md_attr *ma = &info->mti_attr; + + 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); + return rc; + } + + if (ma->ma_valid & MA_LMV) { + rc = mdt_lock_remote_slaves(info, obj, ma, + slave_locks); + if (rc) + mdt_object_unlock(info, obj, lh, rc); + } + } + } else { + if (mdt_object_remote(pobj)) { + rc = mdt_revoke_remote_lookup_lock(info, pobj, obj); + if (rc) + return rc; + } + + rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL, + einfo, true); + } + + return rc; } /* - * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent; - * 2 - src child; 3 - tgt child. - * Update on disk version of src child. + * lookup source by name, if parent is striped directory, we need to find the + * corresponding stripe where source is located, and then lookup there. + * + * besides, if parent is migrating too, and file is already in target stripe, + * this should be a redo of 'lfs migrate' on client side. + */ +static int mdt_migrate_lookup(struct mdt_thread_info *info, + struct mdt_object *pobj, + const struct md_attr *ma, + const struct lu_name *lname, + struct mdt_object **spobj, + struct mdt_object **sobj) +{ + const struct lu_env *env = info->mti_env; + struct lu_fid *fid = &info->mti_tmp_fid1; + struct mdt_object *stripe; + int rc; + + if (ma->ma_valid & MA_LMV) { + /* if parent is striped, lookup on corresponding stripe */ + struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1; + __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 (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); + if (IS_ERR(stripe)) + return PTR_ERR(stripe); + + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid, + &info->mti_spec); + if (rc == -ENOENT && is_migrating) { + /* + * 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. + */ + 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, + lname->ln_namelen); + if (rc < 0) + return rc; + + fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]); + + stripe = mdt_object_find(env, info->mti_mdt, fid); + if (IS_ERR(stripe)) + return PTR_ERR(stripe); + + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(stripe), lname, + fid, &info->mti_spec); + mdt_object_put(env, stripe); + return rc ?: -EALREADY; + } else if (rc) { + mdt_object_put(env, stripe); + return rc; + } + } else { + fid_zero(fid); + rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid, + &info->mti_spec); + if (rc) + return rc; + + stripe = pobj; + mdt_object_get(env, stripe); + } + + *spobj = stripe; + + *sobj = mdt_object_find(env, info->mti_mdt, fid); + if (IS_ERR(*sobj)) { + mdt_object_put(env, stripe); + rc = PTR_ERR(*sobj); + *spobj = NULL; + *sobj = NULL; + } + + return rc; +} + +/* end lease and close file for regular file */ +static int mdd_migrate_close(struct mdt_thread_info *info, + struct mdt_object *obj) +{ + struct close_data *data; + struct mdt_body *repbody; + struct ldlm_lock *lease; + int rc; + int rc2; + + rc = -EPROTO; + if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH, + RCL_CLIENT) || + !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA, + RCL_CLIENT)) + goto close; + + data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); + if (!data) + goto close; + + rc = -ESTALE; + lease = ldlm_handle2lock(&data->cd_handle); + if (!lease) + goto close; + + /* check if the lease was already canceled */ + lock_res_and_lock(lease); + rc = ldlm_is_cancel(lease); + unlock_res_and_lock(lease); + + if (rc) { + rc = -EAGAIN; + LDLM_DEBUG(lease, DFID" lease broken", + PFID(mdt_object_fid(obj))); + } + + /* + * cancel server side lease, client side counterpart should have been + * cancelled, it's okay to cancel it now as we've held mot_open_sem. + */ + ldlm_lock_cancel(lease); + ldlm_reprocess_all(lease->l_resource); + LDLM_LOCK_PUT(lease); + +close: + rc2 = mdt_close_internal(info, mdt_info_req(info), NULL); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; + + return rc ?: rc2; +} + +/* + * migrate file in below steps: + * 1. lock parent and its stripes + * 2. lookup source by name + * 3. lock parents of source links if source is not directory + * 4. reject if source is in HSM + * 5. take source open_sem and close file if source is regular file + * 6. lock source and its stripes if it's directory + * 7. lock target so subsequent change to it can trigger COS + * 8. migrate file + * 9. unlock above locks + * 10. sync device if source has links + */ +static int mdt_reint_migrate_internal(struct mdt_thread_info *info) +{ + const struct lu_env *env = info->mti_env; + struct mdt_device *mdt = info->mti_mdt; + struct mdt_reint_record *rr = &info->mti_rr; + struct lu_ucred *uc = mdt_ucred(info); + struct md_attr *ma = &info->mti_attr; + struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0]; + struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1]; + struct mdt_object *pobj; + struct mdt_object *spobj = NULL; + struct mdt_object *sobj = NULL; + struct mdt_object *tobj; + struct mdt_lock_handle *lhp; + struct mdt_lock_handle *lhs; + struct mdt_lock_handle *lht; + LIST_HEAD(parent_slave_locks); + LIST_HEAD(child_slave_locks); + LIST_HEAD(link_locks); + bool open_sem_locked = false; + bool do_sync = false; + int rc; + ENTRY; + + CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1), + PNAME(&rr->rr_name), PFID(rr->rr_fid2)); + + /* don't allow migrate . or .. */ + if (lu_name_is_dot_or_dotdot(&rr->rr_name)) + RETURN(-EBUSY); + + if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration) + RETURN(-EPERM); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN) && + uc->uc_gid != mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) + RETURN(-EPERM); + + /* pobj is master object of parent */ + pobj = mdt_object_find_check(info, rr->rr_fid1, 0); + if (IS_ERR(pobj)) + RETURN(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); + } + + ma->ma_lmv = info->mti_big_lmm; + ma->ma_lmv_size = info->mti_big_lmmsize; + ma->ma_valid = 0; + rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV); + if (rc) + GOTO(put_parent, rc); + + /* lock parent object */ + lhp = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lhp, LCK_PW); + rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo, + &parent_slave_locks); + if (rc) + GOTO(put_parent, rc); + + /* + * spobj is the corresponding stripe against name if pobj is striped + * directory, which is the real parent, and no need to lock, because + * we've taken full lock of pobj. + */ + rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj); + if (rc) + GOTO(unlock_parent, rc); + + /* lock parents of source links, and revoke LOOKUP lock of links */ + rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks); + if (rc < 0) + GOTO(put_source, rc); + + /* + * RS_MAX_LOCKS is the limit of number of locks that can be saved along + * with one request, if total lock count exceeds this limit, we will + * drop all locks after migration, and synchronous device in the end. + */ + do_sync = rc; + + /* TODO: DoM migration is not supported yet */ + 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); + + if (ma->ma_valid & MA_LOV && + mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM) + GOTO(put_source, rc = -EOPNOTSUPP); + } + + /* if migration HSM is allowed */ + if (!mdt->mdt_opts.mo_migrate_hsm_allowed) { + ma->ma_need = MA_HSM; + ma->ma_valid = 0; + rc = mdt_attr_get_complex(info, sobj, ma); + if (rc) + GOTO(unlock_links, rc); + + if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0) + GOTO(unlock_links, rc = -EOPNOTSUPP); + } + + /* end lease and close file for regular file */ + if (info->mti_spec.sp_migrate_close) { + /* try to hold open_sem so that nobody else can open the file */ + if (!down_write_trylock(&sobj->mot_open_sem)) { + /* close anyway */ + mdd_migrate_close(info, sobj); + GOTO(unlock_links, rc = -EBUSY); + } else { + open_sem_locked = true; + rc = mdd_migrate_close(info, sobj); + if (rc) + GOTO(unlock_open_sem, rc); + } + } + + /* lock source */ + lhs = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lhs, LCK_EX); + rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo, + &child_slave_locks); + if (rc) + GOTO(unlock_open_sem, rc); + + /* lock target */ + tobj = mdt_object_find(env, mdt, rr->rr_fid2); + if (IS_ERR(tobj)) + GOTO(unlock_source, rc = PTR_ERR(tobj)); + + lht = &info->mti_lh[MDT_LH_NEW]; + mdt_lock_reg_init(lht, LCK_EX); + rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true); + if (rc) + GOTO(put_target, rc); + + /* Don't do lookup sanity check. We know name doesn't exist. */ + info->mti_spec.sp_cr_lookup = 0; + info->mti_spec.sp_feat = &dt_directory_features; + + rc = mdo_migrate(env, mdt_object_child(pobj), + mdt_object_child(sobj), &rr->rr_name, + mdt_object_child(tobj), &info->mti_spec, ma); + EXIT; + + mdt_object_unlock(info, tobj, lht, rc); +put_target: + mdt_object_put(env, tobj); +unlock_source: + mdt_migrate_object_unlock(info, sobj, lhs, seinfo, + &child_slave_locks, rc); +unlock_open_sem: + if (open_sem_locked) + up_write(&sobj->mot_open_sem); +unlock_links: + mdt_unlock_list(info, &link_locks, rc); +put_source: + mdt_object_put(env, sobj); + mdt_object_put(env, spobj); +unlock_parent: + mdt_migrate_object_unlock(info, pobj, lhp, peinfo, + &parent_slave_locks, rc); +put_parent: + mdt_object_put(env, pobj); + + if (!rc && do_sync) + mdt_device_sync(env, mdt); + + return rc; +} + +static int mdt_object_lock_save(struct mdt_thread_info *info, + struct mdt_object *dir, + struct mdt_lock_handle *lh, + int idx, bool cos_incompat) +{ + int rc; + + /* we lock the target dir if it is local */ + rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE, + cos_incompat); + if (rc != 0) + return rc; + + /* get and save correct version after locking */ + mdt_version_get_save(info, dir, idx); + return 0; +} + +/* + * determine lock order of sobj and tobj + * + * there are two situations we need to lock tobj before sobj: + * 1. sobj is child of tobj + * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is + * larger than that of tobj + * + * \retval 1 lock tobj before sobj + * \retval 0 lock sobj before tobj + * \retval -ev negative errno upon error + */ +static int mdt_rename_determine_lock_order(struct mdt_thread_info *info, + struct mdt_object *sobj, + struct mdt_object *tobj) +{ + struct md_attr *ma = &info->mti_attr; + struct lu_fid *spfid = &info->mti_tmp_fid1; + struct lu_fid *tpfid = &info->mti_tmp_fid2; + struct lmv_mds_md_v1 *lmv; + __u32 sindex; + __u32 tindex; + int rc; + + /* sobj and tobj are the same */ + if (sobj == tobj) + return 0; + + if (fid_is_root(mdt_object_fid(sobj))) + return 0; + + if (fid_is_root(mdt_object_fid(tobj))) + return 1; + + /* check whether sobj is child of tobj */ + rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj), + mdt_object_fid(tobj)); + if (rc < 0) + return rc; + + if (rc == 1) + return 1; + + /* check whether sobj and tobj are children of the same parent */ + rc = mdt_attr_get_pfid(info, sobj, spfid); + if (rc) + return rc; + + rc = mdt_attr_get_pfid(info, tobj, tpfid); + if (rc) + return rc; + + if (!lu_fid_eq(spfid, tpfid)) + 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; + + if (!(ma->ma_valid & MA_LMV)) + return 0; + + lmv = &ma->ma_lmv->lmv_md_v1; + if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE)) + return 0; + sindex = le32_to_cpu(lmv->lmv_master_mdt_index); + + ma->ma_valid = 0; + rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV); + if (rc) + return rc; + + if (!(ma->ma_valid & MA_LMV)) + return -ENODATA; + + lmv = &ma->ma_lmv->lmv_md_v1; + if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE)) + return -EINVAL; + tindex = le32_to_cpu(lmv->lmv_master_mdt_index); + + /* check stripe index of sobj and tobj */ + if (sindex == tindex) + return -EINVAL; + + return sindex < tindex ? 0 : 1; +} + +/* + * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent; + * 2 - srcdir child; 3 - tgtdir child. + * Update on disk version of srcdir child. */ /** * For DNE phase I, only these renames are allowed @@ -1122,262 +2234,539 @@ 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; + 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; + __u64 lock_ibits; + bool reverse = false; + bool cos_incompat, discard = false; + int rc; + ENTRY; - if (info->mti_dlm_req) - ldlm_request_cancel(req, info->mti_dlm_req, 0); + 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)); - 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); + /* find both parents. */ + msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0); + if (IS_ERR(msrcdir)) + RETURN(PTR_ERR(msrcdir)); - 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); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5); - rc = mdt_rename_lock(info, &rename_lh); - if (rc) { - CERROR("Can't lock FS for rename, rc %d\n", rc); - RETURN(rc); + if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) { + mtgtdir = msrcdir; + mdt_object_get(info->mti_env, mtgtdir); + } else { + mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1); + if (IS_ERR(mtgtdir)) + GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir)); } - 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); - } + rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir); + if (rc < 0) + GOTO(out_put_tgtdir, rc); + + reverse = rc; + + /* source needs to be looked up after locking source parent, otherwise + * this rename may race with unlink source, and cause rename hang, see + * sanityn.sh 55b, so check parents first, if later we found source is + * remote, relock parents. */ + cos_incompat = (mdt_object_remote(msrcdir) || + mdt_object_remote(mtgtdir)); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5); + + /* lock parents in the proper order. */ + lh_srcdirp = &info->mti_lh[MDT_LH_PARENT]; + lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD]; + +relock: + mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name); + mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name); + + if (reverse) { + rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1, + cos_incompat); + if (rc) + GOTO(out_put_tgtdir, rc); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); + + rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0, + cos_incompat); + if (rc != 0) { + mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); + GOTO(out_put_tgtdir, rc); + } + } else { + rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0, + cos_incompat); + if (rc) + GOTO(out_put_tgtdir, rc); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); + + if (mtgtdir != msrcdir) { + rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1, + cos_incompat); + } else if (lh_srcdirp->mlh_pdo_hash != + lh_tgtdirp->mlh_pdo_hash) { + rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir, + MDS_INODELOCK_UPDATE, + cos_incompat); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10); + } + if (rc != 0) { + mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); + GOTO(out_put_tgtdir, rc); } } - /* 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); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5); + + /* find mold 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); - } + GOTO(out_unlock_parents, rc = PTR_ERR(mold)); - 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 (!mdt_object_exists(mold)) { + LU_OBJECT_DEBUG(D_INODE, info->mti_env, + &mold->mot_obj, + "object does not exist"); + GOTO(out_put_old, rc = -ENOENT); + } - if (lu_fid_eq(new_fid, rr->rr_fid1) || - lu_fid_eq(new_fid, rr->rr_fid2)) - GOTO(out_unlock_old, rc = -EINVAL); + /* Check if @mtgtdir is subdir of @mold, before locking child + * to avoid reverse locking. */ + if (mtgtdir != msrcdir) { + rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir), + old_fid); + if (rc) { + if (rc == 1) + rc = -EINVAL; + GOTO(out_put_old, rc); + } + } - if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid)) - GOTO(out_unlock_old, rc = -EPERM); + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold)); + /* save version after locking */ + mdt_version_get_save(info, mold, 2); - 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); - } + if (!cos_incompat && mdt_object_remote(mold)) { + cos_incompat = true; + mdt_object_put(info->mti_env, mold); + mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN); + mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN); + goto relock; + } - 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)); + /* find mnew object: + * mnew 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); + + 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_exists(mnew)) { + LU_OBJECT_DEBUG(D_INODE, info->mti_env, + &mnew->mot_obj, + "object does not exist"); + GOTO(out_put_new, rc = -ENOENT); + } 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); + struct mdt_body *repbody; + + /* Always send rename req to the target child MDT */ + repbody = req_capsule_server_get(info->mti_pill, + &RMF_MDT_BODY); + LASSERT(repbody != NULL); + repbody->mbo_fid1 = *new_fid; + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); + GOTO(out_put_new, rc = -EXDEV); + } + /* 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. */ + 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]; + mdt_lock_reg_init(lh_oldp, 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_object_lock(info, mnew, lh_newp, - MDS_INODELOCK_FULL, 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 { - /* 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); + rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits, + cos_incompat); + if (rc != 0) + GOTO(out_unlock_old, rc); + + /* Check if @msrcdir is subdir of @mnew, before locking child + * to avoid reverse locking. */ + if (mtgtdir != msrcdir) { + rc = mdo_is_subdir(info->mti_env, + mdt_object_child(msrcdir), new_fid); + if (rc) { + if (rc == 1) + rc = -EINVAL; + 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_reint_object_lock(info, mnew, lh_newp, + MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_UPDATE, + cos_incompat); + if (rc != 0) + GOTO(out_unlock_old, rc); + + /* get and save version after locking */ + mdt_version_get_save(info, mnew, 3); + } else if (rc != -EREMOTE && rc != -ENOENT) { + GOTO(out_put_old, rc); + } else { + lh_oldp = &info->mti_lh[MDT_LH_OLD]; + mdt_lock_reg_init(lh_oldp, 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); + if (rc != 0) + GOTO(out_unlock_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); + discard = mdt_dom_check_for_discard(info, mnew); + } 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(info, mtgtdir, lh_tgtdirp, rc); + mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); +out_put_tgtdir: + mdt_object_put(info->mti_env, mtgtdir); +out_put_srcdir: + mdt_object_put(info->mti_env, msrcdir); + + /* If 'discard' is set then new_fid must exits. + * DOM data discard need neither object nor lock, + * so do this at the end. + */ + if (discard) + mdt_dom_discard_data(info, new_fid); + + return rc; +} + +static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc, bool rename) +{ + struct mdt_reint_record *rr = &info->mti_rr; + struct ptlrpc_request *req = mdt_info_req(info); + struct lustre_handle rename_lh = { 0 }; + int rc; + ENTRY; + + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); + + if (!fid_is_md_operative(rr->rr_fid1) || + !fid_is_md_operative(rr->rr_fid2)) + RETURN(-EPERM); + + /* Note: do not enqueue rename lock for replay request, because + * if other MDT holds rename lock, but being blocked to wait for + * this MDT to finish its recovery, and the failover MDT can not + * get rename lock, which will cause deadlock. */ + if (!req_is_replay(req)) { + rc = mdt_rename_lock(info, &rename_lh); + if (rc != 0) { + CERROR("%s: can't lock FS for rename: rc = %d\n", + mdt_obd_name(info->mti_mdt), rc); + RETURN(rc); + } + } + + if (rename) + rc = mdt_reint_rename_internal(info, lhc); + else + rc = mdt_reint_migrate_internal(info); + if (lustre_handle_is_used(&rename_lh)) mdt_rename_unlock(&rename_lh); + + RETURN(rc); +} + +static int mdt_reint_rename(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc) +{ + return mdt_reint_rename_or_migrate(info, lhc, true); +} + +static int mdt_reint_migrate(struct mdt_thread_info *info, + struct mdt_lock_handle *lhc) +{ + return mdt_reint_rename_or_migrate(info, lhc, false); +} + +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; + + DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1)); + + if (info->mti_dlm_req) + ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); + + mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(mo)) + GOTO(out, rc = PTR_ERR(mo)); + + if (!mdt_object_exists(mo)) + GOTO(out_obj, rc = -ENOENT); + + if (!S_ISREG(lu_object_attr(&mo->mot_obj))) + GOTO(out_obj, rc = -EINVAL); + + if (mdt_object_remote(mo)) + GOTO(out_obj, rc = -EREMOTE); + + lease = ldlm_handle2lock(rr->rr_lease_handle); + if (lease == NULL) + GOTO(out_obj, rc = -ESTALE); + + /* It's really necessary to grab open_sem and check if the lease lock + * has been lost. There would exist a concurrent writer coming in and + * generating some dirty data in memory cache, the writeback would fail + * after the layout version is increased by MDS_REINT_RESYNC RPC. */ + if (!down_write_trylock(&mo->mot_open_sem)) + GOTO(out_put_lease, rc = -EBUSY); + + lock_res_and_lock(lease); + lease_broken = ldlm_is_cancel(lease); + unlock_res_and_lock(lease); + if (lease_broken) + GOTO(out_unlock, rc = -EBUSY); + + /* 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); + if (rc) + GOTO(out_unlock, rc); + + ma->ma_need = MA_INODE; + ma->ma_valid = 0; + rc = mdt_attr_get_complex(info, mo, ma); + if (rc != 0) + GOTO(out_unlock, rc); + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); + + EXIT; +out_unlock: + up_write(&mo->mot_open_sem); +out_put_lease: + LDLM_LOCK_PUT(lease); +out_obj: + mdt_object_put(info->mti_env, mo); +out: + mdt_client_compatibility(info); + rc2 = mdt_fix_reply(info); + if (rc == 0) + rc = rc2; return rc; } -typedef int (*mdt_reinter)(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc); - -static mdt_reinter reinters[REINT_MAX] = { - [REINT_SETATTR] = mdt_reint_setattr, - [REINT_CREATE] = mdt_reint_create, - [REINT_LINK] = mdt_reint_link, - [REINT_UNLINK] = mdt_reint_unlink, - [REINT_RENAME] = mdt_reint_rename, - [REINT_OPEN] = mdt_reint_open, - [REINT_SETXATTR] = mdt_reint_setxattr, - [REINT_RMENTRY] = mdt_reint_unlink +struct mdt_reinter { + int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *); + enum lprocfs_extra_opc mr_extra_opc; +}; + +static const struct mdt_reinter mdt_reinters[] = { + [REINT_SETATTR] = { + .mr_handler = &mdt_reint_setattr, + .mr_extra_opc = MDS_REINT_SETATTR, + }, + [REINT_CREATE] = { + .mr_handler = &mdt_reint_create, + .mr_extra_opc = MDS_REINT_CREATE, + }, + [REINT_LINK] = { + .mr_handler = &mdt_reint_link, + .mr_extra_opc = MDS_REINT_LINK, + }, + [REINT_UNLINK] = { + .mr_handler = &mdt_reint_unlink, + .mr_extra_opc = MDS_REINT_UNLINK, + }, + [REINT_RENAME] = { + .mr_handler = &mdt_reint_rename, + .mr_extra_opc = MDS_REINT_RENAME, + }, + [REINT_OPEN] = { + .mr_handler = &mdt_reint_open, + .mr_extra_opc = MDS_REINT_OPEN, + }, + [REINT_SETXATTR] = { + .mr_handler = &mdt_reint_setxattr, + .mr_extra_opc = MDS_REINT_SETXATTR, + }, + [REINT_RMENTRY] = { + .mr_handler = &mdt_reint_unlink, + .mr_extra_opc = MDS_REINT_UNLINK, + }, + [REINT_MIGRATE] = { + .mr_handler = &mdt_reint_migrate, + .mr_extra_opc = MDS_REINT_RENAME, + }, + [REINT_RESYNC] = { + .mr_handler = &mdt_reint_resync, + .mr_extra_opc = MDS_REINT_RESYNC, + }, }; int mdt_reint_rec(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + struct mdt_lock_handle *lhc) { - int rc; - ENTRY; + const struct mdt_reinter *mr; + int rc; + ENTRY; + + if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters))) + RETURN(-EPROTO); + + mr = &mdt_reinters[info->mti_rr.rr_opcode]; + if (mr->mr_handler == NULL) + RETURN(-EPROTO); - rc = reinters[info->mti_rr.rr_opcode](info, lhc); + rc = (*mr->mr_handler)(info, lhc); - RETURN(rc); + lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats, + PTLRPC_LAST_CNTR + mr->mr_extra_opc); + + RETURN(rc); }