X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Ftarget%2Ftgt_handler.c;h=24e6936e9472b012a0b2d4febeb57a4a2d6cca71;hp=6b32fca5f107b35e98370547138629ada993e3a4;hb=c1d465de13ccf0eda8020c88661c3cc4d78538ca;hpb=c2d27a0f12688c0d029880919f8b002e557b540c diff --git a/lustre/target/tgt_handler.c b/lustre/target/tgt_handler.c index 6b32fca..24e6936 100644 --- a/lustre/target/tgt_handler.c +++ b/lustre/target/tgt_handler.c @@ -21,7 +21,7 @@ * GPL HEADER END */ /* - * Copyright (c) 2011, 2014, Intel Corporation. + * Copyright (c) 2013, 2017, Intel Corporation. */ /* * lustre/target/tgt_handler.c @@ -42,9 +42,9 @@ #include #include #include -#include #include #include +#include #include "tgt_internal.h" @@ -255,7 +255,9 @@ static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags) if (rc) RETURN(rc); - nodemap = tsi->tsi_exp->exp_target_data.ted_nodemap; + nodemap = nodemap_get_from_exp(tsi->tsi_exp); + if (IS_ERR(nodemap)) + RETURN(PTR_ERR(nodemap)); body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID, NODEMAP_CLIENT_TO_FS, @@ -263,6 +265,7 @@ static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags) body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID, NODEMAP_CLIENT_TO_FS, body->oa.o_gid); + nodemap_putref(nodemap); tsi->tsi_ost_body = body; tsi->tsi_fid = body->oa.o_oi.oi_fid; @@ -340,10 +343,13 @@ static int tgt_request_preprocess(struct tgt_session_info *tsi, dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ); if (dlm_req != NULL) { + union ldlm_wire_policy_data *policy = + &dlm_req->lock_desc.l_policy_data; + if (unlikely(dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS && - dlm_req->lock_desc.l_policy_data.\ - l_inodebits.bits == 0)) { + (policy->l_inodebits.bits | + policy->l_inodebits.try_bits) == 0)) { /* * Lock without inodebits makes no sense and * will oops later in ldlm. If client miss to @@ -389,6 +395,7 @@ static int tgt_handle_request0(struct tgt_session_info *tsi, case SEC_CTX_INIT: case SEC_CTX_INIT_CONT: case SEC_CTX_FINI: + CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val); GOTO(out, rc = 0); } @@ -422,6 +429,23 @@ static int tgt_handle_request0(struct tgt_session_info *tsi, RCL_SERVER)) req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES, RCL_SERVER, 0); + if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER)) + req_capsule_set_size(tsi->tsi_pill, + &RMF_ACL, RCL_SERVER, + LUSTRE_POSIX_ACL_MAX_SIZE_OLD); + + if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO, + RCL_SERVER)) { + struct niobuf_remote *remote_nb = + req_capsule_client_get(tsi->tsi_pill, + &RMF_NIOBUF_REMOTE); + struct ost_body *body = tsi->tsi_ost_body; + + req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO, + RCL_SERVER, + (body->oa.o_flags & OBD_FL_SHORT_IO) ? + remote_nb[0].rnb_len : 0); + } rc = req_capsule_server_pack(tsi->tsi_pill); } @@ -532,7 +556,7 @@ static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id) if (req_can_reconstruct(req, NULL)) { if (!(lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))) { - DEBUG_REQ(D_WARNING, req, "rq_xid "LPU64" matches " + DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches " "saved xid, expected REPLAY or RESENT flag " "(%x)", req->rq_xid, lustre_msg_get_flags(req->rq_reqmsg)); @@ -603,6 +627,65 @@ static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req) RETURN(h); } +static int process_req_last_xid(struct ptlrpc_request *req) +{ + __u64 last_xid; + ENTRY; + + /* check request's xid is consistent with export's last_xid */ + last_xid = lustre_msg_get_last_xid(req->rq_reqmsg); + if (last_xid > req->rq_export->exp_last_xid) + req->rq_export->exp_last_xid = last_xid; + + if (req->rq_xid == 0 || + (req->rq_xid <= req->rq_export->exp_last_xid)) { + DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. " + "last_xid %llx\n", req->rq_xid, + req->rq_export->exp_last_xid); + /* Some request is allowed to be sent during replay, + * such as OUT update requests, FLD requests, so it + * is possible that replay requests has smaller XID + * than the exp_last_xid. + * + * Some non-replay requests may have smaller XID as + * well: + * + * - Client send a no_resend RPC, like statfs; + * - The RPC timedout (or some other error) on client, + * then it's removed from the unreplied list; + * - Client send some other request to bump the + * exp_last_xid on server; + * - The former RPC got chance to be processed; + */ + if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) + RETURN(-EPROTO); + } + + /* The "last_xid" is the minimum xid among unreplied requests, + * if the request is from the previous connection, its xid can + * still be larger than "exp_last_xid", then the above check of + * xid is not enough to determine whether the request is delayed. + * + * For example, if some replay request was delayed and caused + * timeout at client and the replay is restarted, the delayed + * replay request will have the larger xid than "exp_last_xid" + */ + if (req->rq_export->exp_conn_cnt > + lustre_msg_get_conn_cnt(req->rq_reqmsg)) + RETURN(-ESTALE); + + /* try to release in-memory reply data */ + if (tgt_is_multimodrpcs_client(req->rq_export)) { + tgt_handle_received_xid(req->rq_export, + lustre_msg_get_last_xid(req->rq_reqmsg)); + if (!(lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_RESENT | MSG_REPLAY))) + tgt_handle_tag(req->rq_export, + lustre_msg_get_tag(req->rq_reqmsg)); + } + RETURN(0); +} + int tgt_request_handle(struct ptlrpc_request *req) { struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env); @@ -612,10 +695,24 @@ int tgt_request_handle(struct ptlrpc_request *req) struct lu_target *tgt; int request_fail_id = 0; __u32 opc = lustre_msg_get_opc(msg); + struct obd_device *obd; int rc; - + bool is_connect = false; ENTRY; + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) { + if (cfs_fail_val == 0 && + lustre_msg_get_opc(msg) != OBD_PING && + lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) { + struct l_wait_info lwi = { 0 }; + + cfs_fail_val = 1; + cfs_race_state = 0; + l_wait_event(cfs_race_waitq, (cfs_race_state == 1), + &lwi); + } + } + /* Refill the context, to make sure all thread keys are allocated */ lu_env_refill(req->rq_svc_thread->t_env); @@ -627,6 +724,7 @@ int tgt_request_handle(struct ptlrpc_request *req) * target, otherwise that should be connect operation */ if (opc == MDS_CONNECT || opc == OST_CONNECT || opc == MGS_CONNECT) { + is_connect = true; req_capsule_set(&req->rq_pill, &RQF_CONNECT); rc = target_handle_connect(req); if (rc != 0) { @@ -670,20 +768,19 @@ int tgt_request_handle(struct ptlrpc_request *req) GOTO(out, rc); } - /* check request's xid is consistent with export's last_xid */ - if (req->rq_export != NULL) { - __u64 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg); - if (last_xid != 0) - req->rq_export->exp_last_xid = last_xid; - if (req->rq_xid == 0 || - req->rq_xid <= req->rq_export->exp_last_xid) { - DEBUG_REQ(D_ERROR, req, - "Unexpected xid %llx vs. last_xid %llx\n", - req->rq_xid, req->rq_export->exp_last_xid); -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 93, 0) - LBUG(); -#endif - req->rq_status = -EPROTO; + /* Skip last_xid processing for the recovery thread, otherwise, the + * last_xid on same request could be processed twice: first time when + * processing the incoming request, second time when the request is + * being processed by recovery thread. */ + obd = class_exp2obd(req->rq_export); + if (is_connect) { + /* reset the exp_last_xid on each connection. */ + req->rq_export->exp_last_xid = 0; + } else if (obd->obd_recovery_data.trd_processing_task != + current_pid()) { + rc = process_req_last_xid(req); + if (rc) { + req->rq_status = rc; rc = ptlrpc_error(req); GOTO(out, rc); } @@ -692,16 +789,6 @@ int tgt_request_handle(struct ptlrpc_request *req) request_fail_id = tgt->lut_request_fail_id; tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id; - /* try to release in-memory reply data */ - if (tgt_is_multimodrpcs_client(req->rq_export)) { - tgt_handle_received_xid(req->rq_export, - lustre_msg_get_last_xid(req->rq_reqmsg)); - if (!(lustre_msg_get_flags(req->rq_reqmsg) & - (MSG_RESENT | MSG_REPLAY))) - tgt_handle_tag(req->rq_export, - lustre_msg_get_tag(req->rq_reqmsg)); - } - h = tgt_handler_find_check(req); if (IS_ERR(h)) { req->rq_status = PTR_ERR(h); @@ -789,123 +876,6 @@ EXPORT_SYMBOL(tgt_counter_incr); * Unified target generic handlers. */ -/* - * Security functions - */ -static inline void tgt_init_sec_none(struct obd_connect_data *reply) -{ - reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | - OBD_CONNECT_RMT_CLIENT_FORCE); -} - -static int tgt_init_sec_level(struct ptlrpc_request *req) -{ - struct lu_target *tgt = class_exp2tgt(req->rq_export); - char *client; - struct obd_connect_data *data, *reply; - int rc = 0; - bool remote; - ENTRY; - - data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); - reply = req_capsule_server_get(&req->rq_pill, &RMF_CONNECT_DATA); - if (data == NULL || reply == NULL) - RETURN(-EFAULT); - - /* connection from MDT is always trusted */ - if (req->rq_auth_usr_mdt) { - tgt_init_sec_none(reply); - RETURN(0); - } - - if (unlikely(tgt == NULL)) { - DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n", - class_exp2obd(req->rq_export)->obd_name); - RETURN(-EINVAL); - } - - client = libcfs_nid2str(req->rq_peer.nid); - /* no GSS support case */ - if (!req->rq_auth_gss) { - if (tgt->lut_sec_level > LUSTRE_SEC_NONE) { - CWARN("client %s -> target %s does not use GSS, " - "can not run under security level %d.\n", - client, tgt_name(tgt), tgt->lut_sec_level); - RETURN(-EACCES); - } else { - tgt_init_sec_none(reply); - RETURN(0); - } - } - - /* old version case */ - if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) || - !(data->ocd_connect_flags & OBD_CONNECT_MDS_CAPA) || - !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) { - if (tgt->lut_sec_level > LUSTRE_SEC_NONE) { - CWARN("client %s -> target %s uses old version, " - "can not run under security level %d.\n", - client, tgt_name(tgt), tgt->lut_sec_level); - RETURN(-EACCES); - } else { - CWARN("client %s -> target %s uses old version, " - "run under security level %d.\n", - client, tgt_name(tgt), tgt->lut_sec_level); - tgt_init_sec_none(reply); - RETURN(0); - } - } - - remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE; - if (remote) { - if (!req->rq_auth_remote) - CDEBUG(D_SEC, "client (local realm) %s -> target %s " - "asked to be remote.\n", client, tgt_name(tgt)); - } else if (req->rq_auth_remote) { - remote = true; - CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set " - "as remote by default.\n", client, tgt_name(tgt)); - } - - if (remote == 0) { - if (!uid_valid(make_kuid(&init_user_ns, req->rq_auth_uid))) { - CDEBUG(D_SEC, "client %s -> target %s: user is not " - "authenticated!\n", client, tgt_name(tgt)); - RETURN(-EACCES); - } - } - - - switch (tgt->lut_sec_level) { - case LUSTRE_SEC_NONE: - if (remote) { - CDEBUG(D_SEC, - "client %s -> target %s is set as remote, " - "can not run under security level %d.\n", - client, tgt_name(tgt), tgt->lut_sec_level); - RETURN(-EACCES); - } - tgt_init_sec_none(reply); - break; - case LUSTRE_SEC_REMOTE: - if (!remote) - tgt_init_sec_none(reply); - break; - case LUSTRE_SEC_ALL: - if (remote) - break; - reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | - OBD_CONNECT_RMT_CLIENT_FORCE); - reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA; - reply->ocd_connect_flags &= ~OBD_CONNECT_MDS_CAPA; - break; - default: - RETURN(-EINVAL); - } - - RETURN(rc); -} - int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp) { struct lu_target *tgt = class_exp2tgt(exp); @@ -969,7 +939,7 @@ int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp return rc; } -int tgt_adapt_sptlrpc_conf(struct lu_target *tgt, int initial) +int tgt_adapt_sptlrpc_conf(struct lu_target *tgt) { struct sptlrpc_rule_set tmp_rset; int rc; @@ -980,7 +950,7 @@ int tgt_adapt_sptlrpc_conf(struct lu_target *tgt, int initial) } sptlrpc_rule_set_init(&tmp_rset); - rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset, initial); + rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset); if (rc) { CERROR("%s: failed get sptlrpc rules: rc = %d\n", tgt_name(tgt), rc); @@ -1006,10 +976,6 @@ int tgt_connect(struct tgt_session_info *tsi) ENTRY; - rc = tgt_init_sec_level(req); - if (rc != 0) - GOTO(out, rc); - /* XXX: better to call this check right after getting new export but * before last_rcvd slot allocation to avoid server load upon insecure * connects. This is to be fixed after unifiyng all targets. @@ -1071,6 +1037,35 @@ int tgt_obd_log_cancel(struct tgt_session_info *tsi) return err_serious(-EOPNOTSUPP); } +int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf) +{ + struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env); + struct ptlrpc_request *req = tgt_ses_req(tsi); + struct obd_export *exp = req->rq_export; + struct ptlrpc_bulk_desc *desc; + struct l_wait_info *lwi = &tti->tti_u.update.tti_wait_info; + int i; + int rc; + + ENTRY; + + desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1, + PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC, + MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops); + if (desc == NULL) + RETURN(-ENOMEM); + + for (i = 0; i < rdbuf->rb_nbufs; i++) + desc->bd_frag_ops->add_iov_frag(desc, + rdbuf->rb_bufs[i].lb_buf, + rdbuf->rb_bufs[i].lb_len); + + rc = target_bulk_io(exp, desc, lwi); + ptlrpc_free_bulk(desc); + RETURN(rc); +} +EXPORT_SYMBOL(tgt_send_buffer); + int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob) { struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env); @@ -1094,13 +1089,13 @@ int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob) RETURN(-ENOMEM); if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE)) - /* old client requires reply size in it's PAGE_CACHE_SIZE, + /* old client requires reply size in it's PAGE_SIZE, * which is rdpg->rp_count */ nob = rdpg->rp_count; for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0; i++, tmpcount -= tmpsize) { - tmpsize = min_t(int, tmpcount, PAGE_CACHE_SIZE); + tmpsize = min_t(int, tmpcount, PAGE_SIZE); desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0, tmpsize); } @@ -1161,14 +1156,14 @@ static int tgt_obd_idx_read(struct tgt_session_info *tsi) GOTO(out, rc = -EFAULT); rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT, exp_max_brw_size(tsi->tsi_exp)); - rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE -1) >> PAGE_CACHE_SHIFT; + rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT; /* allocate pages to store the containers */ OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0])); if (rdpg->rp_pages == NULL) GOTO(out, rc = -ENOMEM); for (i = 0; i < rdpg->rp_npages; i++) { - rdpg->rp_pages[i] = alloc_page(GFP_IOFS); + rdpg->rp_pages[i] = alloc_page(GFP_NOFS); if (rdpg->rp_pages[i] == NULL) GOTO(out, rc = -ENOMEM); } @@ -1218,6 +1213,7 @@ int tgt_sync(const struct lu_env *env, struct lu_target *tgt, tgt->lut_obd->obd_last_committed) { rc = dt_object_sync(env, obj, start, end); } + atomic_inc(&tgt->lut_sync_count); RETURN(rc); } @@ -1226,14 +1222,27 @@ EXPORT_SYMBOL(tgt_sync); * Unified target DLM handlers. */ -/* Ensure that data and metadata are synced to the disk when lock is cancelled - * (if requested) */ +/** + * Unified target BAST + * + * Ensure data and metadata are synced to disk when lock is canceled if Sync on + * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough, + * but if it's cross-MDT lock, because remote object version is not set, a + * filesystem sync is needed. + * + * \param lock server side lock + * \param desc lock desc + * \param data ldlm_cb_set_arg + * \param flag indicates whether this cancelling or blocking callback + * \retval 0 on success + * \retval negative number on error + */ static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, void *data, int flag) { struct lu_env env; struct lu_target *tgt; - struct dt_object *obj; + struct dt_object *obj = NULL; struct lu_fid fid; int rc = 0; @@ -1248,10 +1257,12 @@ static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, } if (flag == LDLM_CB_CANCELING && - (lock->l_granted_mode & (LCK_PW | LCK_GROUP)) && + (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) && (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL || (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL && - lock->l_flags & LDLM_FL_CBPENDING))) { + ldlm_is_cbpending(lock))) && + ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) || + lock->l_resource->lr_type == LDLM_EXTENT)) { __u64 start = 0; __u64 end = OBD_OBJECT_EOF; @@ -1261,28 +1272,30 @@ static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, ost_fid_from_resid(&fid, &lock->l_resource->lr_name, tgt->lut_lsd.lsd_osd_index); - obj = dt_locate(&env, tgt->lut_bottom, &fid); - if (IS_ERR(obj)) - GOTO(err_env, rc = PTR_ERR(obj)); - - if (!dt_object_exists(obj)) - GOTO(err_put, rc = -ENOENT); if (lock->l_resource->lr_type == LDLM_EXTENT) { + obj = dt_locate(&env, tgt->lut_bottom, &fid); + if (IS_ERR(obj)) + GOTO(err_env, rc = PTR_ERR(obj)); + + if (!dt_object_exists(obj)) + GOTO(err_put, rc = -ENOENT); + start = lock->l_policy_data.l_extent.start; end = lock->l_policy_data.l_extent.end; } rc = tgt_sync(&env, tgt, obj, start, end); if (rc < 0) { - CERROR("%s: syncing "DFID" ("LPU64"-"LPU64") on lock " + CERROR("%s: syncing "DFID" (%llu-%llu) on lock " "cancel: rc = %d\n", tgt_name(tgt), PFID(&fid), lock->l_policy_data.l_extent.start, lock->l_policy_data.l_extent.end, rc); } err_put: - lu_object_put(&env, &obj->do_lu); + if (obj != NULL) + dt_object_put(&env, obj); err_env: lu_env_fini(&env); } @@ -1450,15 +1463,43 @@ TGT_LLOG_HDL_VAR(0, LLOG_ORIGIN_HANDLE_CLOSE, tgt_llog_close), }; EXPORT_SYMBOL(tgt_llog_handlers); +/* + * sec context handlers + */ +/* XXX: Implement based on mdt_sec_ctx_handle()? */ +static int tgt_sec_ctx_handle(struct tgt_session_info *tsi) +{ + return 0; +} + +struct tgt_handler tgt_sec_ctx_handlers[] = { +TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle), +TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle), +TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle), +}; +EXPORT_SYMBOL(tgt_sec_ctx_handlers); + +int (*tgt_lfsck_in_notify_local)(const struct lu_env *env, + struct dt_device *key, + struct lfsck_req_local *lrl, + struct thandle *th) = NULL; + +void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *, + struct dt_device *, + struct lfsck_req_local *, + struct thandle *)) +{ + tgt_lfsck_in_notify_local = notify; +} +EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local); + int (*tgt_lfsck_in_notify)(const struct lu_env *env, struct dt_device *key, - struct lfsck_request *lr, - struct thandle *th) = NULL; + struct lfsck_request *lr) = NULL; void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *, struct dt_device *, - struct lfsck_request *, - struct thandle *)) + struct lfsck_request *)) { tgt_lfsck_in_notify = notify; } @@ -1466,11 +1507,15 @@ EXPORT_SYMBOL(tgt_register_lfsck_in_notify); static int (*tgt_lfsck_query)(const struct lu_env *env, struct dt_device *key, - struct lfsck_request *lr) = NULL; + struct lfsck_request *req, + struct lfsck_reply *rep, + struct lfsck_query *que) = NULL; void tgt_register_lfsck_query(int (*query)(const struct lu_env *, struct dt_device *, - struct lfsck_request *)) + struct lfsck_request *, + struct lfsck_reply *, + struct lfsck_query *)) { tgt_lfsck_query = query; } @@ -1489,7 +1534,7 @@ static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi) if (lr == NULL) RETURN(-EPROTO); - rc = tgt_lfsck_in_notify(env, key, lr, NULL); + rc = tgt_lfsck_in_notify(env, key, lr); RETURN(rc); } @@ -1509,8 +1554,8 @@ static int tgt_handle_lfsck_query(struct tgt_session_info *tsi) if (reply == NULL) RETURN(-ENOMEM); - rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, request); - reply->lr_status = rc; + rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, + request, reply, NULL); RETURN(rc < 0 ? rc : 0); } @@ -1564,6 +1609,41 @@ void tgt_io_thread_done(struct ptlrpc_thread *thread) EXIT; } EXPORT_SYMBOL(tgt_io_thread_done); + +/** + * Helper function for getting Data-on-MDT file server DLM lock + * if asked by client. + */ +int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, + struct lustre_handle *lh, int mode, __u64 *flags) +{ + union ldlm_policy_data policy = { + .l_inodebits.bits = MDS_INODELOCK_DOM, + }; + int rc; + + ENTRY; + + LASSERT(lh != NULL); + LASSERT(ns != NULL); + LASSERT(!lustre_handle_is_used(lh)); + + rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, &policy, mode, + flags, ldlm_blocking_ast, + ldlm_completion_ast, ldlm_glimpse_ast, + NULL, 0, LVB_T_NONE, NULL, lh); + + RETURN(rc == ELDLM_OK ? 0 : -EIO); +} +EXPORT_SYMBOL(tgt_mdt_data_lock); + +void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode) +{ + LASSERT(lustre_handle_is_used(lh)); + ldlm_lock_decref(lh, mode); +} +EXPORT_SYMBOL(tgt_mdt_data_unlock); + /** * Helper function for getting server side [start, start+count] DLM lock * if asked by client. @@ -1572,8 +1652,8 @@ int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, __u64 start, __u64 end, struct lustre_handle *lh, int mode, __u64 *flags) { - ldlm_policy_data_t policy; - int rc; + union ldlm_policy_data policy; + int rc; ENTRY; @@ -1601,26 +1681,31 @@ int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, } EXPORT_SYMBOL(tgt_extent_lock); -void tgt_extent_unlock(struct lustre_handle *lh, ldlm_mode_t mode) +void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode) { LASSERT(lustre_handle_is_used(lh)); ldlm_lock_decref(lh, mode); } EXPORT_SYMBOL(tgt_extent_unlock); -int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, - struct obd_ioobj *obj, struct niobuf_remote *nb, - struct lustre_handle *lh, int mode) +static int tgt_brw_lock(struct obd_export *exp, struct ldlm_res_id *res_id, + struct obd_ioobj *obj, struct niobuf_remote *nb, + struct lustre_handle *lh, enum ldlm_mode mode) { + struct ldlm_namespace *ns = exp->exp_obd->obd_namespace; __u64 flags = 0; int nrbufs = obj->ioo_bufcnt; int i; + int rc; ENTRY; LASSERT(mode == LCK_PR || mode == LCK_PW); LASSERT(!lustre_handle_is_used(lh)); + if (ns->ns_obd->obd_recovering) + RETURN(0); + if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK)) RETURN(0); @@ -1628,14 +1713,19 @@ int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK)) RETURN(-EFAULT); - RETURN(tgt_extent_lock(ns, res_id, nb[0].rnb_offset, - nb[nrbufs - 1].rnb_offset + - nb[nrbufs - 1].rnb_len - 1, - lh, mode, &flags)); + /* MDT IO for data-on-mdt */ + if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS) + rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags); + else + rc = tgt_extent_lock(ns, res_id, nb[0].rnb_offset, + nb[nrbufs - 1].rnb_offset + + nb[nrbufs - 1].rnb_len - 1, + lh, mode, &flags); + RETURN(rc); } -void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob, - struct lustre_handle *lh, int mode) +static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob, + struct lustre_handle *lh, enum ldlm_mode mode) { ENTRY; @@ -1648,18 +1738,15 @@ void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob, tgt_extent_unlock(lh, mode); EXIT; } - -static __u32 tgt_checksum_bulk(struct lu_target *tgt, - struct ptlrpc_bulk_desc *desc, int opc, - cksum_type_t cksum_type) +static int tgt_checksum_niobuf(struct lu_target *tgt, + struct niobuf_local *local_nb, int npages, + int opc, enum cksum_types cksum_type, + __u32 *cksum) { struct cfs_crypto_hash_desc *hdesc; unsigned int bufsize; int i, err; unsigned char cfs_alg = cksum_obd2cfs(cksum_type); - __u32 cksum; - - LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type)); hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0); if (IS_ERR(hdesc)) { @@ -1669,53 +1756,64 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, } CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg)); - for (i = 0; i < desc->bd_iov_count; i++) { + for (i = 0; i < npages; i++) { /* corrupt the data before we compute the checksum, to * simulate a client->OST data error */ if (i == 0 && opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) { - int off = BD_GET_KIOV(desc, i).kiov_offset & - ~PAGE_MASK; - int len = BD_GET_KIOV(desc, i).kiov_len; + int off = local_nb[i].lnb_page_offset & ~PAGE_MASK; + int len = local_nb[i].lnb_len; struct page *np = tgt_page_to_corrupt; - char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off; if (np) { - char *ptr2 = kmap(np) + off; + char *ptr = ll_kmap_atomic(local_nb[i].lnb_page, + KM_USER0); + char *ptr2 = page_address(np); + + memcpy(ptr2 + off, ptr + off, len); + memcpy(ptr2 + off, "bad3", min(4, len)); + ll_kunmap_atomic(ptr, KM_USER0); + + /* LU-8376 to preserve original index for + * display in dump_all_bulk_pages() */ + np->index = i; - memcpy(ptr2, ptr, len); - memcpy(ptr2, "bad3", min(4, len)); - kunmap(np); - BD_GET_KIOV(desc, i).kiov_page = np; + cfs_crypto_hash_update_page(hdesc, np, off, + len); + continue; } else { CERROR("%s: can't alloc page for corruption\n", tgt_name(tgt)); } } - cfs_crypto_hash_update_page(hdesc, - BD_GET_KIOV(desc, i).kiov_page, - BD_GET_KIOV(desc, i).kiov_offset & - ~PAGE_MASK, - BD_GET_KIOV(desc, i).kiov_len); + cfs_crypto_hash_update_page(hdesc, local_nb[i].lnb_page, + local_nb[i].lnb_page_offset & ~PAGE_MASK, + local_nb[i].lnb_len); /* corrupt the data after we compute the checksum, to * simulate an OST->client data error */ if (i == 0 && opc == OST_READ && OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) { - int off = BD_GET_KIOV(desc, i).kiov_offset - & ~PAGE_MASK; - int len = BD_GET_KIOV(desc, i).kiov_len; + int off = local_nb[i].lnb_page_offset & ~PAGE_MASK; + int len = local_nb[i].lnb_len; struct page *np = tgt_page_to_corrupt; - char *ptr = - kmap(BD_GET_KIOV(desc, i).kiov_page) + off; if (np) { - char *ptr2 = kmap(np) + off; + char *ptr = ll_kmap_atomic(local_nb[i].lnb_page, + KM_USER0); + char *ptr2 = page_address(np); - memcpy(ptr2, ptr, len); - memcpy(ptr2, "bad4", min(4, len)); - kunmap(np); - BD_GET_KIOV(desc, i).kiov_page = np; + memcpy(ptr2 + off, ptr + off, len); + memcpy(ptr2 + off, "bad4", min(4, len)); + ll_kunmap_atomic(ptr, KM_USER0); + + /* LU-8376 to preserve original index for + * display in dump_all_bulk_pages() */ + np->index = i; + + cfs_crypto_hash_update_page(hdesc, np, off, + len); + continue; } else { CERROR("%s: can't alloc page for corruption\n", tgt_name(tgt)); @@ -1723,10 +1821,147 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, } } - bufsize = sizeof(cksum); - err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize); + bufsize = sizeof(*cksum); + err = cfs_crypto_hash_final(hdesc, (unsigned char *)cksum, &bufsize); + + return 0; +} + +char dbgcksum_file_name[PATH_MAX]; + +static void dump_all_bulk_pages(struct obdo *oa, int count, + struct niobuf_local *local_nb, + __u32 server_cksum, __u32 client_cksum) +{ + struct file *filp; + int rc, i; + unsigned int len; + char *buf; + + /* will only keep dump of pages on first error for the same range in + * file/fid, not during the resends/retries. */ + snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name), + "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x", + (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ? + libcfs_debug_file_path_arr : + LIBCFS_DEBUG_FILE_PATH_DEFAULT), + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0, + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0, + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0, + local_nb[0].lnb_file_offset, + local_nb[count-1].lnb_file_offset + + local_nb[count-1].lnb_len - 1, client_cksum, server_cksum); + filp = filp_open(dbgcksum_file_name, + O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600); + if (IS_ERR(filp)) { + rc = PTR_ERR(filp); + if (rc == -EEXIST) + CDEBUG(D_INFO, "%s: can't open to dump pages with " + "checksum error: rc = %d\n", dbgcksum_file_name, + rc); + else + CERROR("%s: can't open to dump pages with checksum " + "error: rc = %d\n", dbgcksum_file_name, rc); + return; + } + + for (i = 0; i < count; i++) { + len = local_nb[i].lnb_len; + buf = kmap(local_nb[i].lnb_page); + while (len != 0) { + rc = cfs_kernel_write(filp, buf, len, &filp->f_pos); + if (rc < 0) { + CERROR("%s: wanted to write %u but got %d " + "error\n", dbgcksum_file_name, len, rc); + break; + } + len -= rc; + buf += rc; + CDEBUG(D_INFO, "%s: wrote %d bytes\n", + dbgcksum_file_name, rc); + } + kunmap(local_nb[i].lnb_page); + } + + rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1); + if (rc) + CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc); + filp_close(filp, NULL); + return; +} + +static int check_read_checksum(struct niobuf_local *local_nb, int npages, + struct obd_export *exp, struct obdo *oa, + const struct lnet_process_id *peer, + __u32 client_cksum, __u32 server_cksum, + enum cksum_types server_cksum_type) +{ + char *msg; + enum cksum_types cksum_type; + loff_t start, end; + + /* unlikely to happen and only if resend does not occur due to cksum + * control failure on Client */ + if (unlikely(server_cksum == client_cksum)) { + CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n", + client_cksum); + return 0; + } + + if (exp->exp_obd->obd_checksum_dump) + dump_all_bulk_pages(oa, npages, local_nb, server_cksum, + client_cksum); + + cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ? + oa->o_flags : 0); + + if (cksum_type != server_cksum_type) + msg = "the server may have not used the checksum type specified" + " in the original request - likely a protocol problem"; + else + msg = "should have changed on the client or in transit"; + + start = local_nb[0].lnb_file_offset; + end = local_nb[npages-1].lnb_file_offset + + local_nb[npages-1].lnb_len - 1; + + LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode " + DFID " object "DOSTID" extent [%llu-%llu], client returned csum" + " %x (type %x), server csum %x (type %x)\n", + exp->exp_obd->obd_name, + msg, libcfs_nid2str(peer->nid), + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL, + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0, + oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0, + POSTID(&oa->o_oi), + start, end, client_cksum, cksum_type, server_cksum, + server_cksum_type); + + return 1; +} + +static int tgt_pages2shortio(struct niobuf_local *local, int npages, + unsigned char *buf, int size) +{ + int i, off, len, copied = size; + char *ptr; + + for (i = 0; i < npages; i++) { + off = local[i].lnb_page_offset & ~PAGE_MASK; + len = local[i].lnb_len; + + CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n", + i, off, len, size); + if (len > size) + return -EINVAL; - return cksum; + ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0); + memcpy(buf + off, ptr, len); + ll_kunmap_atomic(ptr, KM_USER0); + buf += len; + size -= len; + } + return copied - size; } int tgt_brw_read(struct tgt_session_info *tsi) @@ -1740,12 +1975,14 @@ int tgt_brw_read(struct tgt_session_info *tsi) struct ost_body *body, *repbody; struct l_wait_info lwi; struct lustre_handle lockh = { 0 }; - int npages, nob = 0, rc, i, no_reply = 0; + int npages, nob = 0, rc, i, no_reply = 0, + npages_read; struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data; ENTRY; - if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL && + ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) { CERROR("%s: deny read request from %s to portal %u\n", tgt_name(tsi->tsi_tgt), obd_export_nid2str(req->rq_export), @@ -1788,8 +2025,8 @@ int tgt_brw_read(struct tgt_session_info *tsi) local_nb = tbc->local; - rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo, - remote_nb, &lockh, LCK_PR); + rc = tgt_brw_lock(exp, &tsi->tsi_resid, ioo, remote_nb, &lockh, + LCK_PR); if (rc != 0) RETURN(rc); @@ -1797,13 +2034,12 @@ int tgt_brw_read(struct tgt_session_info *tsi) * If getting the lock took more time than * client was willing to wait, drop it. b=11330 */ - if (cfs_time_current_sec() > req->rq_deadline || + if (ktime_get_real_seconds() > req->rq_deadline || OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { no_reply = 1; - CERROR("Dropping timed-out read from %s because locking" - "object "DOSTID" took %ld seconds (limit was %ld).\n", + CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n", libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid), - cfs_time_current_sec() - req->rq_arrival_time.tv_sec, + ktime_get_real_seconds() - req->rq_arrival_time.tv_sec, req->rq_deadline - req->rq_arrival_time.tv_sec); GOTO(out_lock, rc = -ETIMEDOUT); } @@ -1817,50 +2053,76 @@ int tgt_brw_read(struct tgt_session_info *tsi) if (rc != 0) GOTO(out_lock, rc); - desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), - PTLRPC_BULK_PUT_SOURCE | - PTLRPC_BULK_BUF_KIOV, - OST_BULK_PORTAL, - &ptlrpc_bulk_kiov_nopin_ops); - if (desc == NULL) - GOTO(out_commitrw, rc = -ENOMEM); + if (body->oa.o_flags & OBD_FL_SHORT_IO) { + desc = NULL; + } else { + desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), + PTLRPC_BULK_PUT_SOURCE | + PTLRPC_BULK_BUF_KIOV, + OST_BULK_PORTAL, + &ptlrpc_bulk_kiov_nopin_ops); + if (desc == NULL) + GOTO(out_commitrw, rc = -ENOMEM); + } nob = 0; + npages_read = npages; for (i = 0; i < npages; i++) { int page_rc = local_nb[i].lnb_rc; if (page_rc < 0) { rc = page_rc; + npages_read = i; break; } nob += page_rc; - if (page_rc != 0) { /* some data! */ + if (page_rc != 0 && desc != NULL) { /* some data! */ LASSERT(local_nb[i].lnb_page != NULL); desc->bd_frag_ops->add_kiov_frag (desc, local_nb[i].lnb_page, - local_nb[i].lnb_page_offset, + local_nb[i].lnb_page_offset & ~PAGE_MASK, page_rc); } if (page_rc != local_nb[i].lnb_len) { /* short read */ + local_nb[i].lnb_len = page_rc; + npages_read = i + (page_rc != 0 ? 1 : 0); /* All subsequent pages should be 0 */ while (++i < npages) LASSERT(local_nb[i].lnb_rc == 0); break; } } + if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) && + nob != cfs_fail_val) + rc = -E2BIG; if (body->oa.o_valid & OBD_MD_FLCKSUM) { - cksum_type_t cksum_type = + enum cksum_types cksum_type = cksum_type_unpack(body->oa.o_valid & OBD_MD_FLFLAGS ? body->oa.o_flags : 0); + repbody->oa.o_flags = cksum_type_pack(cksum_type); repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS; - repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc, - OST_READ, cksum_type); + rc = tgt_checksum_niobuf(tsi->tsi_tgt, local_nb, + npages_read, OST_READ, cksum_type, + &repbody->oa.o_cksum); + if (rc < 0) + GOTO(out_commitrw, rc); + CDEBUG(D_PAGE, "checksum at read origin: %x\n", repbody->oa.o_cksum); + + /* if a resend it could be for a cksum error, so check Server + * cksum with returned Client cksum (this should even cover + * zero-cksum case) */ + if ((body->oa.o_valid & OBD_MD_FLFLAGS) && + (body->oa.o_flags & OBD_FL_RECOV_RESEND)) + check_read_checksum(local_nb, npages_read, exp, + &body->oa, &req->rq_peer, + body->oa.o_cksum, + repbody->oa.o_cksum, cksum_type); } else { repbody->oa.o_valid = 0; } @@ -1868,18 +2130,37 @@ int tgt_brw_read(struct tgt_session_info *tsi) /* Check if client was evicted while we were doing i/o before touching * network */ - if (likely(rc == 0 && - !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) { - rc = target_bulk_io(exp, desc, &lwi); + if (rc == 0) { + if (body->oa.o_flags & OBD_FL_SHORT_IO) { + unsigned char *short_io_buf; + int short_io_size; + + short_io_buf = req_capsule_server_get(&req->rq_pill, + &RMF_SHORT_IO); + short_io_size = req_capsule_get_size(&req->rq_pill, + &RMF_SHORT_IO, + RCL_SERVER); + rc = tgt_pages2shortio(local_nb, npages_read, + short_io_buf, short_io_size); + if (rc >= 0) + req_capsule_shrink(&req->rq_pill, + &RMF_SHORT_IO, rc, + RCL_SERVER); + rc = rc > 0 ? 0 : rc; + } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) { + rc = target_bulk_io(exp, desc, &lwi); + } no_reply = rc != 0; + } else { + if (body->oa.o_flags & OBD_FL_SHORT_IO) + req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0, + RCL_SERVER); } out_commitrw: /* Must commit after prep above in all cases */ rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo, remote_nb, npages, local_nb, rc); - if (rc == 0) - tgt_drop_id(exp, &repbody->oa); out_lock: tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR); @@ -1901,8 +2182,10 @@ out_lock: obd_export_nid2str(exp), rc); } /* send a bulk after reply to simulate a network delay or reordering - * by a router */ - if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) { + * by a router - Note that !desc implies short io, so there is no bulk + * to reorder. */ + if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) && + desc) { wait_queue_head_t waitq; struct l_wait_info lwi1; @@ -1919,6 +2202,32 @@ out_lock: } EXPORT_SYMBOL(tgt_brw_read); +static int tgt_shortio2pages(struct niobuf_local *local, int npages, + unsigned char *buf, int size) +{ + int i, off, len; + char *ptr; + + for (i = 0; i < npages; i++) { + off = local[i].lnb_page_offset & ~PAGE_MASK; + len = local[i].lnb_len; + + if (len == 0) + continue; + + CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n", + i, off, len, size); + ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0); + if (ptr == NULL) + return -EINVAL; + memcpy(ptr + off, buf, len < size ? len : size); + ll_kunmap_atomic(ptr, KM_USER0); + buf += len; + size -= len; + } + return 0; +} + static void tgt_warn_on_cksum(struct ptlrpc_request *req, struct ptlrpc_bulk_desc *desc, struct niobuf_local *local_nb, int npages, @@ -1933,19 +2242,23 @@ static void tgt_warn_on_cksum(struct ptlrpc_request *req, body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); LASSERT(body != NULL); - if (req->rq_peer.nid != desc->bd_sender) { + if (desc && req->rq_peer.nid != desc->bd_sender) { via = " via "; router = libcfs_nid2str(desc->bd_sender); } + if (exp->exp_obd->obd_checksum_dump) + dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum, + client_cksum); + if (mmap) { CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n", client_cksum, server_cksum); return; } - LCONSOLE_ERROR_MSG(0x168, "BAD WRITE CHECKSUM: %s from %s%s%s inode " - DFID" object "DOSTID" extent ["LPU64"-"LPU64 + LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode " + DFID" object "DOSTID" extent [%llu-%llu" "]: client csum %x, server csum %x\n", exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer), via, router, @@ -1976,13 +2289,15 @@ int tgt_brw_write(struct tgt_session_info *tsi) __u32 *rcs; int objcount, niocount, npages; int rc, i, j; - cksum_type_t cksum_type = OBD_CKSUM_CRC32; + enum cksum_types cksum_type = OBD_CKSUM_CRC32; bool no_reply = false, mmap; struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data; + bool wait_sync = false; ENTRY; - if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL && + ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) { CERROR("%s: deny write request from %s to portal %u\n", tgt_name(tsi->tsi_tgt), obd_export_nid2str(req->rq_export), @@ -2032,7 +2347,7 @@ int tgt_brw_write(struct tgt_session_info *tsi) RETURN(err_serious(-EPROTO)); if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) && - (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)) + ptlrpc_connection_is_local(exp->exp_connection)) memory_pressure_set(); req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER, @@ -2046,8 +2361,8 @@ int tgt_brw_write(struct tgt_session_info *tsi) local_nb = tbc->local; - rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo, - remote_nb, &lockh, LCK_PW); + rc = tgt_brw_lock(exp, &tsi->tsi_resid, ioo, remote_nb, &lockh, + LCK_PW); if (rc != 0) GOTO(out, rc); @@ -2055,14 +2370,13 @@ int tgt_brw_write(struct tgt_session_info *tsi) * If getting the lock took more time than * client was willing to wait, drop it. b=11330 */ - if (cfs_time_current_sec() > req->rq_deadline || + if (ktime_get_real_seconds() > req->rq_deadline || OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { no_reply = true; - CERROR("%s: Dropping timed-out write from %s because locking " - "object "DOSTID" took %ld seconds (limit was %ld).\n", + CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n", tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid), - cfs_time_current_sec() - req->rq_arrival_time.tv_sec, + ktime_get_real_seconds() - req->rq_arrival_time.tv_sec, req->rq_deadline - req->rq_arrival_time.tv_sec); GOTO(out_lock, rc = -ETIMEDOUT); } @@ -2085,26 +2399,45 @@ int tgt_brw_write(struct tgt_session_info *tsi) objcount, ioo, remote_nb, &npages, local_nb); if (rc < 0) GOTO(out_lock, rc); + if (body->oa.o_flags & OBD_FL_SHORT_IO) { + int short_io_size; + unsigned char *short_io_buf; + + short_io_size = req_capsule_get_size(&req->rq_pill, + &RMF_SHORT_IO, + RCL_CLIENT); + short_io_buf = req_capsule_client_get(&req->rq_pill, + &RMF_SHORT_IO); + CDEBUG(D_INFO, "Client use short io for data transfer," + " size = %d\n", short_io_size); + + /* Copy short io buf to pages */ + rc = tgt_shortio2pages(local_nb, npages, short_io_buf, + short_io_size); + desc = NULL; + } else { + desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), + PTLRPC_BULK_GET_SINK | + PTLRPC_BULK_BUF_KIOV, + OST_BULK_PORTAL, + &ptlrpc_bulk_kiov_nopin_ops); + if (desc == NULL) + GOTO(skip_transfer, rc = -ENOMEM); + + /* NB Having prepped, we must commit... */ + for (i = 0; i < npages; i++) + desc->bd_frag_ops->add_kiov_frag(desc, + local_nb[i].lnb_page, + local_nb[i].lnb_page_offset & ~PAGE_MASK, + local_nb[i].lnb_len); + + rc = sptlrpc_svc_prep_bulk(req, desc); + if (rc != 0) + GOTO(skip_transfer, rc); - desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), - PTLRPC_BULK_GET_SINK | PTLRPC_BULK_BUF_KIOV, - OST_BULK_PORTAL, - &ptlrpc_bulk_kiov_nopin_ops); - if (desc == NULL) - GOTO(skip_transfer, rc = -ENOMEM); - - /* NB Having prepped, we must commit... */ - for (i = 0; i < npages; i++) - desc->bd_frag_ops->add_kiov_frag(desc, - local_nb[i].lnb_page, - local_nb[i].lnb_page_offset, - local_nb[i].lnb_len); - - rc = sptlrpc_svc_prep_bulk(req, desc); - if (rc != 0) - GOTO(skip_transfer, rc); + rc = target_bulk_io(exp, desc, &lwi); + } - rc = target_bulk_io(exp, desc, &lwi); no_reply = rc != 0; skip_transfer: @@ -2117,8 +2450,12 @@ skip_transfer: repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS; repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL; repbody->oa.o_flags |= cksum_type_pack(cksum_type); - repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc, - OST_WRITE, cksum_type); + rc = tgt_checksum_niobuf(tsi->tsi_tgt, local_nb, + npages, OST_WRITE, cksum_type, + &repbody->oa.o_cksum); + if (rc < 0) + GOTO(out_commitrw, rc); + cksum_counter++; if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) { @@ -2137,6 +2474,7 @@ skip_transfer: } } +out_commitrw: /* Must commit after prep above in all cases */ rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo, remote_nb, npages, local_nb, rc); @@ -2146,6 +2484,12 @@ skip_transfer: * has timed out the request already */ no_reply = true; + for (i = 0; i < niocount; i++) { + if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) { + wait_sync = true; + break; + } + } /* * Disable sending mtime back to the client. If the client locked the * whole object, then it has already updated the mtime on its side, @@ -2173,23 +2517,22 @@ skip_transfer: } LASSERT(j == npages); ptlrpc_lprocfs_brw(req, nob); - - tgt_drop_id(exp, &repbody->oa); } out_lock: tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW); if (desc) ptlrpc_free_bulk(desc); out: - if (no_reply) { + if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) { req->rq_no_reply = 1; /* reply out callback would free */ ptlrpc_req_drop_rs(req); - LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), " - "client will retry: rc %d\n", - exp->exp_obd->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), rc); + if (!exp->exp_obd->obd_no_transno) + LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s)," + " client will retry: rc = %d\n", + exp->exp_obd->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), rc); } memory_pressure_clr(); RETURN(rc);