X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_open.c;h=456f403799ce63068cfc857a4ca501bfffa45ba2;hp=e225c5b9e349212bfde7e6ef26ea04a8c22e7294;hb=2d5f51b92f6eb45e9fd6ae57f8d9ec349049ea14;hpb=08aa217ce49aba1ded52e0f7adb8a607035123fd diff --git a/lustre/mdt/mdt_open.c b/lustre/mdt/mdt_open.c index e225c5b..456f403 100644 --- a/lustre/mdt/mdt_open.c +++ b/lustre/mdt/mdt_open.c @@ -108,29 +108,34 @@ void mdt_mfd_free(struct mdt_file_data *mfd) static int mdt_create_data(struct mdt_thread_info *info, struct mdt_object *p, struct mdt_object *o) { - struct md_op_spec *spec = &info->mti_spec; - struct md_attr *ma = &info->mti_attr; - int rc = 0; - ENTRY; + struct md_op_spec *spec = &info->mti_spec; + struct md_attr *ma = &info->mti_attr; + int rc = 0; + ENTRY; - if (!md_should_create(spec->sp_cr_flags)) - RETURN(0); + if (!md_should_create(spec->sp_cr_flags)) + RETURN(0); - ma->ma_need = MA_INODE | MA_LOV; - ma->ma_valid = 0; + ma->ma_need = MA_INODE | MA_LOV; + ma->ma_valid = 0; mutex_lock(&o->mot_lov_mutex); - if (!(o->mot_flags & MOF_LOV_CREATED)) { - rc = mdo_create_data(info->mti_env, - p ? mdt_object_child(p) : NULL, - mdt_object_child(o), spec, ma); + if (!(o->mot_flags & MOF_LOV_CREATED)) { + if (p != NULL && (fid_is_obf(mdt_object_fid(p)) || + fid_is_dot_lustre(mdt_object_fid(p)))) + GOTO(unlock, rc = -EPERM); + + rc = mdo_create_data(info->mti_env, + p ? mdt_object_child(p) : NULL, + mdt_object_child(o), spec, ma); if (rc == 0) rc = mdt_attr_get_complex(info, o, ma); - if (rc == 0 && ma->ma_valid & MA_LOV) - o->mot_flags |= MOF_LOV_CREATED; - } + if (rc == 0 && ma->ma_valid & MA_LOV) + o->mot_flags |= MOF_LOV_CREATED; + } +unlock: mutex_unlock(&o->mot_lov_mutex); - RETURN(rc); + RETURN(rc); } static int mdt_ioepoch_opened(struct mdt_object *mo) @@ -577,10 +582,9 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) struct obd_export *exp = req->rq_export; CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n", - mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name, - info->mti_transno, - libcfs_nid2str(exp->exp_connection->c_peer.nid), - rc); + mdt_obd_name(mdt), info->mti_transno, + libcfs_nid2str(exp->exp_connection->c_peer.nid), + rc); RETURN_EXIT; } } else if (info->mti_transno == 0) { @@ -592,18 +596,18 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) } spin_unlock(&mdt->mdt_lut.lut_translock); - CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n", - info->mti_transno, - req->rq_export->exp_obd->obd_last_committed); + CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n", + info->mti_transno, + req->rq_export->exp_obd->obd_last_committed); - req->rq_transno = info->mti_transno; - lustre_msg_set_transno(req->rq_repmsg, info->mti_transno); + req->rq_transno = info->mti_transno; + lustre_msg_set_transno(req->rq_repmsg, info->mti_transno); - /* update lcd in memory only for resent cases */ - ted = &req->rq_export->exp_target_data; - LASSERT(ted); + /* update lcd in memory only for resent cases */ + ted = &req->rq_export->exp_target_data; + LASSERT(ted); mutex_lock(&ted->ted_lcd_lock); - lcd = ted->ted_lcd; + lcd = ted->ted_lcd; if (info->mti_transno < lcd->lcd_last_transno && info->mti_transno != 0) { /* This should happen during replay. Do not update @@ -612,8 +616,8 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) * be checked correctly by xid */ mutex_unlock(&ted->ted_lcd_lock); CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n", - mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name, - info->mti_transno, lcd->lcd_last_transno); + mdt_obd_name(mdt), info->mti_transno, + lcd->lcd_last_transno); RETURN_EXIT; } @@ -704,7 +708,13 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, repbody->ioepoch = o->mot_ioepoch; } } else if (flags & MDS_FMODE_EXEC) { - rc = mdt_write_deny(o); + /* if file is released, we can't deny write because we must + * restore (write) it to access it.*/ + if ((ma->ma_valid & MA_HSM) && + (ma->ma_hsm.mh_flags & HS_RELEASED)) + rc = 0; + else + rc = mdt_write_deny(o); } if (rc) RETURN(rc); @@ -827,9 +837,13 @@ int mdt_finish_open(struct mdt_thread_info *info, * that looks like it was actually almost succesful and a failure at the * same time */ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) { - mdt_set_disposition(info, rep, DISP_OPEN_LOCK | \ - DISP_OPEN_OPEN | DISP_LOOKUP_NEG | \ - DISP_LOOKUP_POS); + mdt_set_disposition(info, rep, DISP_OPEN_OPEN | + DISP_LOOKUP_NEG | + DISP_LOOKUP_POS); + + if (flags & MDS_OPEN_LOCK) + mdt_set_disposition(info, rep, DISP_OPEN_LOCK); + RETURN(-ENOENT); } @@ -846,7 +860,7 @@ int mdt_finish_open(struct mdt_thread_info *info, } } #ifdef CONFIG_FS_POSIX_ACL - else if (exp->exp_connect_flags & OBD_CONNECT_ACL) { + else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) { const struct lu_env *env = info->mti_env; struct md_object *next = mdt_object_child(o); struct lu_buf *buf = &info->mti_buf; @@ -876,8 +890,8 @@ int mdt_finish_open(struct mdt_thread_info *info, } #endif - if (info->mti_mdt->mdt_opts.mo_mds_capa && - exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) { + if (info->mti_mdt->mdt_opts.mo_mds_capa && + exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1); @@ -888,9 +902,9 @@ int mdt_finish_open(struct mdt_thread_info *info, RETURN(rc); repbody->valid |= OBD_MD_FLMDSCAPA; } - if (info->mti_mdt->mdt_opts.mo_oss_capa && - exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA && - S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) { + if (info->mti_mdt->mdt_opts.mo_oss_capa && + exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA && + S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2); @@ -906,13 +920,11 @@ int mdt_finish_open(struct mdt_thread_info *info, * If we are following a symlink, don't open; and do not return open * handle for special nodes as client required. */ - if (islnk || (!isreg && !isdir && - (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) { - lustre_msg_set_transno(req->rq_repmsg, 0); - RETURN(0); - } - - mdt_set_disposition(info, rep, DISP_OPEN_OPEN); + if (islnk || (!isreg && !isdir && + (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) { + lustre_msg_set_transno(req->rq_repmsg, 0); + RETURN(0); + } /* * We need to return the existing object's fid back, so it is done here, @@ -957,11 +969,15 @@ int mdt_finish_open(struct mdt_thread_info *info, else repbody->valid |= OBD_MD_FLEASIZE; } + mdt_set_disposition(info, rep, DISP_OPEN_OPEN); RETURN(0); } } rc = mdt_mfd_open(info, p, o, flags, created); + if (!rc) + mdt_set_disposition(info, rep, DISP_OPEN_OPEN); + RETURN(rc); } @@ -994,7 +1010,7 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD); ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD, RCL_SERVER); - ma->ma_need = MA_INODE; + ma->ma_need = MA_INODE | MA_HSM; if (ma->ma_lmm_size > 0) ma->ma_need |= MA_LOV; @@ -1040,25 +1056,35 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, mdt_export_evict(exp); RETURN_EXIT; } - rc = mdt_object_exists(child); - if (rc > 0) { - - mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA); - rc = mdt_attr_get_complex(info, child, ma); - if (rc == 0) - rc = mdt_finish_open(info, parent, child, - flags, 1, ldlm_rep); - } else if (rc < 0) { - /* the child object was created on remote server */ - repbody->fid1 = *rr->rr_fid2; - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); - rc = 0; - } else if (rc == 0) { - /* the child does not exist, we should do regular open */ - mdt_object_put(env, parent); - mdt_object_put(env, child); - GOTO(regular_open, 0); - } + + if (unlikely(mdt_object_remote(child))) { + /* the child object was created on remote server */ + if (!mdt_is_dne_client(exp)) { + /* Return -EIO for old client */ + mdt_object_put(env, parent); + mdt_object_put(env, child); + GOTO(out, rc = -EIO); + } + repbody->fid1 = *rr->rr_fid2; + repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); + rc = 0; + } else { + if (mdt_object_exists(child)) { + mdt_set_capainfo(info, 1, rr->rr_fid2, + BYPASS_CAPA); + rc = mdt_attr_get_complex(info, child, ma); + if (rc == 0) + rc = mdt_finish_open(info, parent, + child, flags, + 1, ldlm_rep); + } else { + /* the child does not exist, we should do + * regular open */ + mdt_object_put(env, parent); + mdt_object_put(env, child); + GOTO(regular_open, 0); + } + } mdt_object_put(env, parent); mdt_object_put(env, child); GOTO(out, rc); @@ -1089,30 +1115,156 @@ int mdt_open_by_fid(struct mdt_thread_info* info, if (IS_ERR(o)) RETURN(rc = PTR_ERR(o)); - rc = mdt_object_exists(o); - if (rc > 0) { - mdt_set_disposition(info, rep, (DISP_IT_EXECD | - DISP_LOOKUP_EXECD | - DISP_LOOKUP_POS)); - - rc = mdt_attr_get_complex(info, o, ma); - if (rc == 0) - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); - } else if (rc == 0) { - rc = -ENOENT; - } else { + if (unlikely(mdt_object_remote(o))) { /* the child object was created on remote server */ struct mdt_body *repbody; + + mdt_set_disposition(info, rep, (DISP_IT_EXECD | + DISP_LOOKUP_EXECD | + DISP_LOOKUP_POS)); repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); repbody->fid1 = *rr->rr_fid2; repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); rc = 0; - } + } else { + if (mdt_object_exists(o)) { + mdt_set_disposition(info, rep, (DISP_IT_EXECD | + DISP_LOOKUP_EXECD | + DISP_LOOKUP_POS)); + + rc = mdt_attr_get_complex(info, o, ma); + if (rc == 0) + rc = mdt_finish_open(info, NULL, o, flags, 0, + rep); + } else { + rc = -ENOENT; + } + } mdt_object_put(info->mti_env, o); RETURN(rc); } +/* lock object for open */ +static int mdt_object_open_lock(struct mdt_thread_info *info, + struct mdt_object *obj, + struct mdt_lock_handle *lhc, + __u64 *ibits) +{ + struct md_attr *ma = &info->mti_attr; + __u64 open_flags = info->mti_spec.sp_cr_flags; + ldlm_mode_t lm = LCK_CR; + bool try_layout = false; + bool create_layout = false; + int rc = 0; + ENTRY; + + *ibits = 0; + if (open_flags & MDS_OPEN_LOCK) { + if (open_flags & FMODE_WRITE) + lm = LCK_CW; + /* if file is released, we can't deny write because we must + * restore (write) it to access it. */ + else if ((open_flags & MDS_FMODE_EXEC) && + !((ma->ma_valid & MA_HSM) && + (ma->ma_hsm.mh_flags & HS_RELEASED))) + lm = LCK_PR; + else + lm = LCK_CR; + + *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN; + } + + if (S_ISREG(lu_object_attr(&obj->mot_obj.mo_lu))) { + if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) && + md_should_create(open_flags)) + create_layout = true; + if (exp_connect_layout(info->mti_exp) && !create_layout && + ma->ma_need & MA_LOV) + try_layout = true; + } + + mdt_lock_handle_init(lhc); + mdt_lock_reg_init(lhc, lm); + + /* one problem to return layout lock on open is that it may result + * in too many layout locks cached on the client side. */ + if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) { + /* return lookup lock to validate inode at the client side, + * this is pretty important otherwise mdt will return layout + * lock for each open. + * However this is a double-edged sword because changing + * permission will revoke huge # of LOOKUP locks. */ + *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP; + if (!mdt_object_lock_try(info, obj, lhc, *ibits, + MDT_CROSS_LOCK)) { + *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP); + if (*ibits != 0) + rc = mdt_object_lock(info, obj, lhc, *ibits, + MDT_CROSS_LOCK); + } + } else if (*ibits != 0) { + rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK); + } + + CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64 + ", open_flags = "LPO64", try_layout = %d, rc = %d\n", + PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc); + + /* will change layout, revoke layout locks by enqueuing EX lock. */ + if (rc == 0 && create_layout) { + struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT]; + + CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID + ", open_flags = "LPO64"\n", + PFID(mdt_object_fid(obj)), open_flags); + + LASSERT(!try_layout); + mdt_lock_handle_init(ll); + mdt_lock_reg_init(ll, LCK_EX); + rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT, + MDT_LOCAL_LOCK); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2); + } + + RETURN(rc); +} + +static void mdt_object_open_unlock(struct mdt_thread_info *info, + struct mdt_object *obj, + struct mdt_lock_handle *lhc, + __u64 ibits, int rc) +{ + __u64 open_flags = info->mti_spec.sp_cr_flags; + struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT]; + + /* Release local layout lock - the layout lock put in MDT_LH_LAYOUT + * will never return to client side. */ + if (lustre_handle_is_used(&ll->mlh_reg_lh)) { + LASSERT(!(ibits & MDS_INODELOCK_LAYOUT)); + mdt_object_unlock(info, obj, ll, 1); + } + + /* Cross-ref case, the lock should be returned to the client */ + if (ibits == 0 || rc == -EREMOTE) + return; + + if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) { + /* for the open request, the lock will only return to client + * if open or layout lock is granted. */ + rc = 1; + } + + if (rc != 0) { + struct ldlm_reply *ldlm_rep; + + ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP); + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK); + mdt_object_unlock(info, obj, lhc, 1); + } +} + int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, struct mdt_lock_handle *lhc) { @@ -1124,7 +1276,7 @@ int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, struct mdt_object *parent= NULL; struct mdt_object *o; int rc; - ldlm_mode_t lm; + __u64 ibits = 0; ENTRY; if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) { @@ -1146,35 +1298,26 @@ int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, if (IS_ERR(o)) RETURN(rc = PTR_ERR(o)); - rc = mdt_object_exists(o); - if (rc == 0) { - mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD | - DISP_LOOKUP_NEG)); - GOTO(out, rc = -ENOENT); - } else if (rc < 0) { - CERROR("NFS remote open shouldn't happen.\n"); - GOTO(out, rc); - } - mdt_set_disposition(info, rep, (DISP_IT_EXECD | - DISP_LOOKUP_EXECD | - DISP_LOOKUP_POS)); + if (mdt_object_remote(o)) { + CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n", + mdt_obd_name(info->mti_mdt), + PFID(rr->rr_fid2)); + GOTO(out, rc = -EREMOTE); + } else if (!mdt_object_exists(o)) { + mdt_set_disposition(info, rep, + DISP_IT_EXECD | + DISP_LOOKUP_EXECD | + DISP_LOOKUP_NEG); + GOTO(out, rc = -ENOENT); + } - if (flags & FMODE_WRITE) - lm = LCK_CW; - else if (flags & MDS_FMODE_EXEC) - lm = LCK_PR; - else - lm = LCK_CR; + mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); - mdt_lock_handle_init(lhc); - mdt_lock_reg_init(lhc, lm); - rc = mdt_object_lock(info, o, lhc, - MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN, - MDT_CROSS_LOCK); + rc = mdt_attr_get_complex(info, o, ma); if (rc) GOTO(out, rc); - rc = mdt_attr_get_complex(info, o, ma); + rc = mdt_object_open_lock(info, o, lhc, &ibits); if (rc) GOTO(out, rc); @@ -1189,15 +1332,16 @@ int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, } } - if (flags & MDS_OPEN_LOCK) - mdt_set_disposition(info, rep, DISP_OPEN_LOCK); rc = mdt_finish_open(info, parent, o, flags, 0, rep); - - if (!(flags & MDS_OPEN_LOCK) || rc) - mdt_object_unlock(info, o, lhc, 1); - + if (!rc) { + mdt_set_disposition(info, rep, DISP_LOOKUP_POS); + if (flags & MDS_OPEN_LOCK) + mdt_set_disposition(info, rep, DISP_OPEN_LOCK); + } GOTO(out, rc); + out: + mdt_object_open_unlock(info, o, lhc, ibits, rc); mdt_object_put(env, o); if (parent != NULL) mdt_object_put(env, parent); @@ -1211,9 +1355,10 @@ int mdt_pin(struct mdt_thread_info* info) } /* Cross-ref request. Currently it can only be a pure open (w/o create) */ -int mdt_cross_open(struct mdt_thread_info* info, - const struct lu_fid *fid, - struct ldlm_reply *rep, __u32 flags) +static int mdt_cross_open(struct mdt_thread_info *info, + const struct lu_fid *parent_fid, + const struct lu_fid *fid, + struct ldlm_reply *rep, __u32 flags) { struct md_attr *ma = &info->mti_attr; struct mdt_object *o; @@ -1224,35 +1369,46 @@ int mdt_cross_open(struct mdt_thread_info* info, if (IS_ERR(o)) RETURN(rc = PTR_ERR(o)); - rc = mdt_object_exists(o); - if (rc > 0) { - /* Do permission check for cross-open. */ - rc = mo_permission(info->mti_env, NULL, mdt_object_child(o), - NULL, flags | MDS_OPEN_CROSS); - if (rc) - goto out; - - mdt_set_capainfo(info, 0, fid, BYPASS_CAPA); - rc = mdt_attr_get_complex(info, o, ma); - if (rc == 0) - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); - } else if (rc == 0) { - /* - * Something is wrong here. lookup was positive but there is - * no object! - */ - CERROR("Cross-ref object doesn't exist!\n"); + if (mdt_object_remote(o)) { + /* Something is wrong here, the object is on another MDS! */ + CERROR("%s: "DFID" isn't on this server!: rc = %d\n", + mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); + LU_OBJECT_DEBUG(D_WARNING, info->mti_env, + &o->mot_obj.mo_lu, + "Object isn't on this server! FLD error?\n"); rc = -EFAULT; - } else { - /* Something is wrong here, the object is on another MDS! */ - CERROR("The object isn't on this server! FLD error?\n"); - LU_OBJECT_DEBUG(D_WARNING, info->mti_env, - &o->mot_obj.mo_lu, - "Object isn't on this server! FLD error?\n"); + } else { + if (mdt_object_exists(o)) { + /* Do permission check for cross-open. */ + rc = mo_permission(info->mti_env, NULL, + mdt_object_child(o), + NULL, flags | MDS_OPEN_CROSS); + if (rc) + goto out; + + mdt_set_capainfo(info, 0, fid, BYPASS_CAPA); + rc = mdt_attr_get_complex(info, o, ma); + if (rc != 0) + GOTO(out, rc); - rc = -EFAULT; - } + /* Do not create lov object if the fid is opened + * under OBF */ + if (S_ISREG(ma->ma_attr.la_mode) && + !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) && + fid_is_obf(parent_fid)) + GOTO(out, rc = -EPERM); + rc = mdt_finish_open(info, NULL, o, flags, 0, rep); + } else { + /* + * Something is wrong here. lookup was positive but + * there is no object! + */ + CERROR("%s: "DFID" doesn't exist!: rc = %d\n", + mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); + rc = -EFAULT; + } + } out: mdt_object_put(info->mti_env, o); RETURN(rc); @@ -1270,6 +1426,7 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) struct lu_fid *child_fid = &info->mti_tmp_fid1; struct md_attr *ma = &info->mti_attr; __u64 create_flags = info->mti_spec.sp_cr_flags; + __u64 ibits; struct mdt_reint_record *rr = &info->mti_rr; struct lu_name *lname; int result, rc; @@ -1310,8 +1467,15 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) PFID(rr->rr_fid1), rr->rr_name, PFID(rr->rr_fid2), create_flags, ma->ma_attr.la_mode, msg_flags); - - if (req_is_replay(req) || + if (info->mti_cross_ref) { + /* This is cross-ref open */ + mdt_set_disposition(info, ldlm_rep, + (DISP_IT_EXECD | DISP_LOOKUP_EXECD | + DISP_LOOKUP_POS)); + result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1, + ldlm_rep, create_flags); + GOTO(out, result); + } else if (req_is_replay(req) || (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) { /* This is a replay request or from liblustre with ea. */ result = mdt_open_by_fid(info, ldlm_rep); @@ -1330,14 +1494,24 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) GOTO(out, result = -EFAULT); } CDEBUG(D_INFO, "No object(1), continue as regular open.\n"); - } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref && - create_flags & MDS_OPEN_LOCK) || + } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) || (create_flags & MDS_OPEN_BY_FID)) { result = mdt_open_by_fid_lock(info, ldlm_rep, lhc); - if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT)) + /* If result is 0 then open by FID has found the file + * and there is nothing left for us to do here. More + * generally if it is anything other than -ENOENT or + * -EREMOTE then we return that now. If -ENOENT and + * MDS_OPEN_CREAT is set then we must create the file + * below. If -EREMOTE then we need to return a LOOKUP + * lock to the client, which we do below. Hence this + * odd looking condition. See LU-2523. */ + if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) && + result != -EREMOTE) GOTO(out, result); + if (unlikely(rr->rr_namelen == 0)) GOTO(out, result = -EINVAL); + CDEBUG(D_INFO, "No object(2), continue as regular open.\n"); } @@ -1347,14 +1521,6 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) mdt_set_disposition(info, ldlm_rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); - if (info->mti_cross_ref) { - /* This is cross-ref open */ - mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS); - result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep, - create_flags); - GOTO(out, result); - } - lh = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen); @@ -1418,8 +1584,9 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); if (result == -ENOENT) { - if (mdt_object_obf(parent)) - GOTO(out_child, result = -EPERM); + /* Create under OBF and .lustre is not permitted */ + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + GOTO(out_child, result = -EPERM); /* save versions in reply */ mdt_version_get_save(info, parent, 0); @@ -1462,12 +1629,10 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) } created = 1; } else { - /* We have to get attr & lov ea for this object */ - result = mdt_attr_get_complex(info, child, ma); /* * The object is on remote node, return its FID for remote open. */ - if (result == -EREMOTE) { + if (mdt_object_remote(child)) { /* * Check if this lock already was sent to client and * this is resent case. For resent case do not take lock @@ -1502,67 +1667,66 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); if (rc != 0) result = rc; + else + result = -EREMOTE; GOTO(out_child, result); - } + } else { + if (mdt_object_exists(child)) { + /* We have to get attr & LOV EA & HSM for this + * object */ + ma->ma_need |= MA_HSM; + result = mdt_attr_get_complex(info, child, ma); + } else { + /*object non-exist!!!*/ + LBUG(); + } + } } LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh)); - /* get openlock if this is not replay and if a client requested it */ - if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) { - ldlm_mode_t lm; - - if (create_flags & FMODE_WRITE) - lm = LCK_CW; - else if (create_flags & MDS_FMODE_EXEC) - lm = LCK_PR; - else - lm = LCK_CR; - mdt_lock_handle_init(lhc); - mdt_lock_reg_init(lhc, lm); - rc = mdt_object_lock(info, child, lhc, - MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN, - MDT_CROSS_LOCK); - if (rc) { - result = rc; - GOTO(out_child, result); - } else { - result = -EREMOTE; - mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK); - } - } + /* get openlock if this is not replay and if a client requested it */ + if (!req_is_replay(req)) { + rc = mdt_object_open_lock(info, child, lhc, &ibits); + if (rc != 0) + GOTO(out_child, result = rc); + else if (create_flags & MDS_OPEN_LOCK) + mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK); + } - /* Try to open it now. */ - rc = mdt_finish_open(info, parent, child, create_flags, - created, ldlm_rep); - if (rc) { - result = rc; - if (lustre_handle_is_used(&lhc->mlh_reg_lh)) - /* openlock was acquired and mdt_finish_open failed - - drop the openlock */ - mdt_object_unlock(info, child, lhc, 1); - if (created) { - ma->ma_need = 0; - ma->ma_valid = 0; - ma->ma_cookie_size = 0; - rc = mdo_unlink(info->mti_env, - mdt_object_child(parent), - mdt_object_child(child), - lname, - &info->mti_attr); - if (rc != 0) - CERROR("Error in cleanup of open\n"); - } - } + /* Try to open it now. */ + rc = mdt_finish_open(info, parent, child, create_flags, + created, ldlm_rep); + if (rc) { + result = rc; + /* openlock will be released if mdt_finish_open failed */ + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK); + if (created) { + ma->ma_need = 0; + ma->ma_valid = 0; + ma->ma_cookie_size = 0; + rc = mdo_unlink(info->mti_env, + mdt_object_child(parent), + mdt_object_child(child), + lname, + &info->mti_attr, 0); + if (rc != 0) + CERROR("%s: "DFID" cleanup of open: rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(mdt_object_fid(child)), rc); + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); + } + } EXIT; out_child: + mdt_object_open_unlock(info, child, lhc, ibits, result); mdt_object_put(info->mti_env, child); out_parent: mdt_object_unlock_put(info, parent, lh, result || !created); out: - if (result && result != -EREMOTE) - lustre_msg_set_transno(req->rq_repmsg, 0); - return result; + if (result) + lustre_msg_set_transno(req->rq_repmsg, 0); + return result; } #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \ @@ -1605,6 +1769,24 @@ int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd) rc = mo_attr_set(info->mti_env, next, ma); } + /* If file data is modified, add the dirty flag. + * + * If MDS_CLOSE_CLEANUP is set, this file is being closed due to an + * eviction, file could have been modified and now dirty + * regarding to HSM archive, check this! + * The logic here is to mark a file dirty if there's a chance it was + * dirtied before the client was evicted, so that we don't have to wait + * for a release attempt before finding out the file was actually dirty + * and fail the release. Aggressively marking it dirty here will cause + * the policy engine to attempt to re-archive it; when rearchiving, we + * can compare the current version to the LMA data_version and make the + * archive request into a noop if it's not actually dirty. + */ + if ((ma->ma_attr_flags & MDS_DATA_MODIFIED) || + ((ma->ma_attr_flags & MDS_CLOSE_CLEANUP) && + (mode & (FMODE_WRITE|MDS_FMODE_TRUNC)))) + rc = mdt_add_dirty_flag(info, o, ma); + ma->ma_need |= MA_INODE; ma->ma_valid &= ~MA_INODE;