X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Ftarget%2Ftgt_handler.c;h=d630d85c73b3e574a26ad9cb2ebb95de5b829297;hp=06180874da1d806d6c490406811c5885316d44a6;hb=b0ab95d6133e783acacc6329c025d17fb282775e;hpb=2b294992edce5af7b79d4300ed3aa1ea6a8db850 diff --git a/lustre/target/tgt_handler.c b/lustre/target/tgt_handler.c index 0618087..d630d85 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, 2016, Intel Corporation. */ /* * lustre/target/tgt_handler.c @@ -34,12 +34,17 @@ #define DEBUG_SUBSYSTEM S_CLASS +#include +#ifdef HAVE_UIDGID_HEADER +# include +#endif + #include #include #include -#include #include #include +#include #include "tgt_internal.h" @@ -97,12 +102,7 @@ static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags) if (!IS_ERR(obj)) { if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) { lu_object_put(tsi->tsi_env, obj); - /* for capability renew ENOENT will be handled in - * mdt_renew_capa */ - if (body->mbo_valid & OBD_MD_FLOSSCAPA) - rc = 0; - else - rc = -ENOENT; + rc = -ENOENT; } else { tsi->tsi_corpus = obj; rc = 0; @@ -242,7 +242,6 @@ static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags) { struct ost_body *body; struct req_capsule *pill = tsi->tsi_pill; - struct lustre_capa *capa; struct lu_nodemap *nodemap; int rc; @@ -256,7 +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, @@ -264,15 +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); - - if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { - capa = req_capsule_client_get(pill, &RMF_CAPA1); - if (capa == NULL) { - CERROR("%s: OSSCAPA flag is set without capability\n", - tgt_name(tsi->tsi_tgt)); - RETURN(-EFAULT); - } - } + nodemap_putref(nodemap); tsi->tsi_ost_body = body; tsi->tsi_fid = body->oa.o_oi.oi_fid; @@ -350,10 +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 @@ -384,9 +380,25 @@ static int tgt_handle_request0(struct tgt_session_info *tsi, { int serious = 0; int rc; + __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); ENTRY; + + /* When dealing with sec context requests, no export is associated yet, + * because these requests are sent before *_CONNECT requests. + * A NULL req->rq_export means the normal *_common_slice handlers will + * not be called, because there is no reference to the target. + * So deal with them by hand and jump directly to target_send_reply(). + */ + switch (opc) { + 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); + } + /* * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try * to put same checks into handlers like mdt_close(), mdt_reint(), @@ -400,6 +412,9 @@ static int tgt_handle_request0(struct tgt_session_info *tsi, */ if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE)) RETURN(0); + if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT && + OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET))) + RETURN(0); rc = tgt_request_preprocess(tsi, h, req); /* pack reply if reply format is fixed */ @@ -414,6 +429,10 @@ 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); rc = req_capsule_server_pack(tsi->tsi_pill); } @@ -454,6 +473,7 @@ static int tgt_handle_request0(struct tgt_session_info *tsi, if (likely(rc == 0 && req->rq_export)) target_committed_to_req(req); +out: target_send_reply(req, rc, tsi->tsi_reply_fail_id); RETURN(0); } @@ -468,7 +488,6 @@ static int tgt_filter_recovery_request(struct ptlrpc_request *req, *process = 1; RETURN(0); case MDS_CLOSE: - case MDS_DONE_WRITING: case MDS_SYNC: /* used in unmounting */ case OBD_PING: case MDS_REINT: @@ -483,6 +502,9 @@ static int tgt_filter_recovery_request(struct ptlrpc_request *req, case OST_SETATTR: case OST_SYNC: case OST_WRITE: + case MDS_HSM_PROGRESS: + case MDS_HSM_STATE_SET: + case MDS_HSM_REQUEST: *process = target_queue_recovery_request(req, obd); RETURN(0); @@ -518,11 +540,11 @@ static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id) /* sanity check: if the xid matches, the request must be marked as a * resent or replayed */ - if (req_xid_is_last(req)) { + 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 " - "last_xid, expected REPLAY or RESENT flag " + 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)); req->rq_status = -ENOTCONN; @@ -565,6 +587,11 @@ static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req) ENTRY; tgt = class_exp2tgt(req->rq_export); + if (unlikely(tgt == NULL)) { + DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n", + class_exp2obd(req->rq_export)->obd_name); + RETURN(ERR_PTR(-EINVAL)); + } for (s = tgt->lut_slice; s->tos_hs != NULL; s++) if (s->tos_opc_start <= opc && opc < s->tos_opc_end) @@ -587,6 +614,52 @@ 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); + } + + /* 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); @@ -596,8 +669,9 @@ 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; /* Refill the context, to make sure all thread keys are allocated */ @@ -611,6 +685,7 @@ 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) { @@ -624,6 +699,14 @@ int tgt_request_handle(struct ptlrpc_request *req) } if (unlikely(!class_connected_export(req->rq_export))) { + if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT || + opc == SEC_CTX_FINI) { + /* sec context initialization has to be handled + * by hand in tgt_handle_request0() */ + tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET; + h = NULL; + GOTO(handle_recov, rc = 0); + } CDEBUG(D_HA, "operation %d on unconnected OST from %s\n", opc, libcfs_id2str(req->rq_peer)); req->rq_status = -ENOTCONN; @@ -638,6 +721,32 @@ int tgt_request_handle(struct ptlrpc_request *req) else tsi->tsi_jobid = NULL; + if (tgt == NULL) { + DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n", + class_exp2obd(req->rq_export)->obd_name); + req->rq_status = -EINVAL; + rc = ptlrpc_error(req); + GOTO(out, rc); + } + + /* 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); + } + } + request_fail_id = tgt->lut_request_fail_id; tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id; @@ -648,6 +757,9 @@ int tgt_request_handle(struct ptlrpc_request *req) GOTO(out, rc); } + LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n", + h->th_opc, opc); + if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE)) GOTO(out, rc = 0); @@ -661,10 +773,9 @@ int tgt_request_handle(struct ptlrpc_request *req) GOTO(out, rc); } +handle_recov: rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id); if (likely(rc == 1)) { - LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n", - h->th_opc, opc); rc = tgt_handle_request0(tsi, h, req); if (rc) GOTO(out, rc); @@ -726,129 +837,6 @@ 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 | - OBD_CONNECT_MDS_CAPA | - OBD_CONNECT_OSS_CAPA); -} - -static int tgt_init_sec_level(struct ptlrpc_request *req) -{ - struct lu_target *tgt = class_exp2tgt(req->rq_export); - char *client = libcfs_nid2str(req->rq_peer.nid); - 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); - } - - /* 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) { - if (!tgt->lut_oss_capa) { - CDEBUG(D_SEC, - "client %s -> target %s is set as remote," - " but OSS capabilities are not enabled: %d.\n", - client, tgt_name(tgt), tgt->lut_oss_capa); - RETURN(-EACCES); - } - } else { - 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); - if (!tgt->lut_oss_capa) - reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA; - if (!tgt->lut_mds_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); @@ -877,6 +865,16 @@ int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp spin_lock(&exp->exp_lock); exp->exp_sp_peer = req->rq_sp_from; exp->exp_flvr = flvr; + + /* when on mgs, if no restriction is set, or if client + * is loopback, allow any flavor */ + if ((strcmp(exp->exp_obd->obd_type->typ_name, + LUSTRE_MGS_NAME) == 0) && + (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL || + LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid)) + == LOLND)) + exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY; + if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY && exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) { CERROR("%s: unauthorized rpc flavor %x from %s, " @@ -902,13 +900,18 @@ 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; + if (unlikely(tgt == NULL)) { + CERROR("No target passed"); + return -EINVAL; + } + 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); @@ -934,10 +937,6 @@ 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. @@ -998,13 +997,35 @@ int tgt_obd_log_cancel(struct tgt_session_info *tsi) { return err_serious(-EOPNOTSUPP); } -EXPORT_SYMBOL(tgt_obd_log_cancel); -int tgt_obd_qc_callback(struct tgt_session_info *tsi) +int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf) { - return err_serious(-EOPNOTSUPP); + 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_obd_qc_callback); +EXPORT_SYMBOL(tgt_send_buffer); int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob) { @@ -1020,25 +1041,29 @@ int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob) ENTRY; - desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1, BULK_PUT_SOURCE, - MDS_BULK_PORTAL); + desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1, + PTLRPC_BULK_PUT_SOURCE | + PTLRPC_BULK_BUF_KIOV, + MDS_BULK_PORTAL, + &ptlrpc_bulk_kiov_pin_ops); if (desc == NULL) 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); - ptlrpc_prep_bulk_page_pin(desc, rdpg->rp_pages[i], 0, tmpsize); + tmpsize = min_t(int, tmpcount, PAGE_SIZE); + desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0, + tmpsize); } LASSERT(desc->bd_nob == nob); rc = target_bulk_io(exp, desc, lwi); - ptlrpc_free_bulk_pin(desc); + ptlrpc_free_bulk(desc); RETURN(rc); } EXPORT_SYMBOL(tgt_sendpage); @@ -1092,14 +1117,14 @@ 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); } @@ -1127,12 +1152,10 @@ out: } return rc; } -EXPORT_SYMBOL(tgt_obd_idx_read); struct tgt_handler tgt_obd_handlers[] = { TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping), TGT_OBD_HDL_VAR(0, OBD_LOG_CANCEL, tgt_obd_log_cancel), -TGT_OBD_HDL_VAR(0, OBD_QC_CALLBACK, tgt_obd_qc_callback), TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read) }; EXPORT_SYMBOL(tgt_obd_handlers); @@ -1151,6 +1174,7 @@ 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); } @@ -1159,14 +1183,27 @@ 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; @@ -1174,11 +1211,19 @@ static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, tgt = class_exp2tgt(lock->l_export); + if (unlikely(tgt == NULL)) { + CDEBUG(D_ERROR, "%s: No target for connected export\n", + class_exp2obd(lock->l_export)->obd_name); + RETURN(-EINVAL); + } + 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; @@ -1188,28 +1233,30 @@ 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); } @@ -1240,6 +1287,20 @@ int tgt_enqueue(struct tgt_session_info *tsi) if (rc) RETURN(err_serious(rc)); + switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) { + case LUT_FAIL_MDT: + tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET; + break; + case LUT_FAIL_OST: + tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET; + break; + case LUT_FAIL_MGT: + tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET; + break; + default: + tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY; + break; + } RETURN(req->rq_status); } EXPORT_SYMBOL(tgt_enqueue); @@ -1257,19 +1318,16 @@ int tgt_convert(struct tgt_session_info *tsi) RETURN(req->rq_status); } -EXPORT_SYMBOL(tgt_convert); int tgt_bl_callback(struct tgt_session_info *tsi) { return err_serious(-EOPNOTSUPP); } -EXPORT_SYMBOL(tgt_bl_callback); int tgt_cp_callback(struct tgt_session_info *tsi) { return err_serious(-EOPNOTSUPP); } -EXPORT_SYMBOL(tgt_cp_callback); /* generic LDLM target handler */ struct tgt_handler tgt_dlm_handlers[] = { @@ -1318,7 +1376,6 @@ int tgt_llog_destroy(struct tgt_session_info *tsi) RETURN(rc); } -EXPORT_SYMBOL(tgt_llog_destroy); int tgt_llog_read_header(struct tgt_session_info *tsi) { @@ -1383,15 +1440,27 @@ 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; } @@ -1399,11 +1468,15 @@ 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; } @@ -1422,7 +1495,7 @@ 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); } @@ -1442,8 +1515,8 @@ 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); } @@ -1505,8 +1578,8 @@ 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; @@ -1515,7 +1588,7 @@ int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, LASSERT(!lustre_handle_is_used(lh)); policy.l_extent.gid = 0; - policy.l_extent.start = start & CFS_PAGE_MASK; + policy.l_extent.start = start & PAGE_MASK; /* * If ->o_blocks is EOF it means "lock till the end of the file". @@ -1524,7 +1597,7 @@ int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, if (end == OBD_OBJECT_EOF || end < start) policy.l_extent.end = OBD_OBJECT_EOF; else - policy.l_extent.end = end | ~CFS_PAGE_MASK; + policy.l_extent.end = end | ~PAGE_MASK; rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode, flags, ldlm_blocking_ast, @@ -1534,7 +1607,7 @@ 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); @@ -1543,7 +1616,7 @@ 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) + struct lustre_handle *lh, enum ldlm_mode mode) { __u64 flags = 0; int nrbufs = obj->ioo_bufcnt; @@ -1554,6 +1627,9 @@ int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, 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); @@ -1566,10 +1642,9 @@ int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, nb[nrbufs - 1].rnb_len - 1, lh, mode, &flags)); } -EXPORT_SYMBOL(tgt_brw_lock); void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob, - struct lustre_handle *lh, int mode) + struct lustre_handle *lh, enum ldlm_mode mode) { ENTRY; @@ -1582,7 +1657,6 @@ void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob, tgt_extent_unlock(lh, mode); EXIT; } -EXPORT_SYMBOL(tgt_brw_unlock); static __u32 tgt_checksum_bulk(struct lu_target *tgt, struct ptlrpc_bulk_desc *desc, int opc, @@ -1594,6 +1668,8 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, 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)) { CERROR("%s: unable to initialize checksum hash %s\n", @@ -1607,10 +1683,11 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, * simulate a client->OST data error */ if (i == 0 && opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) { - int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; - int len = desc->bd_iov[i].kiov_len; + int off = BD_GET_KIOV(desc, i).kiov_offset & + ~PAGE_MASK; + int len = BD_GET_KIOV(desc, i).kiov_len; struct page *np = tgt_page_to_corrupt; - char *ptr = kmap(desc->bd_iov[i].kiov_page) + off; + char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off; if (np) { char *ptr2 = kmap(np) + off; @@ -1618,24 +1695,34 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, memcpy(ptr2, ptr, len); memcpy(ptr2, "bad3", min(4, len)); kunmap(np); - desc->bd_iov[i].kiov_page = np; + + /* LU-8376 to preserve original index for + * display in dump_all_bulk_pages() */ + np->index = BD_GET_KIOV(desc, + i).kiov_page->index; + + BD_GET_KIOV(desc, i).kiov_page = np; } else { CERROR("%s: can't alloc page for corruption\n", tgt_name(tgt)); } } - cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page, - desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK, - desc->bd_iov[i].kiov_len); + 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); /* 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 = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; - int len = desc->bd_iov[i].kiov_len; + int off = BD_GET_KIOV(desc, i).kiov_offset + & ~PAGE_MASK; + int len = BD_GET_KIOV(desc, i).kiov_len; struct page *np = tgt_page_to_corrupt; - char *ptr = kmap(desc->bd_iov[i].kiov_page) + off; + char *ptr = + kmap(BD_GET_KIOV(desc, i).kiov_page) + off; if (np) { char *ptr2 = kmap(np) + off; @@ -1643,7 +1730,13 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, memcpy(ptr2, ptr, len); memcpy(ptr2, "bad4", min(4, len)); kunmap(np); - desc->bd_iov[i].kiov_page = np; + + /* LU-8376 to preserve original index for + * display in dump_all_bulk_pages() */ + np->index = BD_GET_KIOV(desc, + i).kiov_page->index; + + BD_GET_KIOV(desc, i).kiov_page = np; } else { CERROR("%s: can't alloc page for corruption\n", tgt_name(tgt)); @@ -1657,6 +1750,122 @@ static __u32 tgt_checksum_bulk(struct lu_target *tgt, return cksum; } +char dbgcksum_file_name[PATH_MAX]; + +static void dump_all_bulk_pages(struct obdo *oa, int count, + lnet_kiov_t *iov, __u32 server_cksum, + __u32 client_cksum) +{ + struct file *filp; + int rc, i; + unsigned int len; + char *buf; + mm_segment_t oldfs; + + /* 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, + (__u64)iov[0].kiov_page->index << PAGE_SHIFT, + ((__u64)iov[count - 1].kiov_page->index << PAGE_SHIFT) + + iov[count - 1].kiov_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; + } + + oldfs = get_fs(); + set_fs(KERNEL_DS); + for (i = 0; i < count; i++) { + len = iov[i].kiov_len; + buf = kmap(iov[i].kiov_page); + while (len != 0) { + rc = vfs_write(filp, (__force const char __user *)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(iov[i].kiov_page); + } + set_fs(oldfs); + + 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 ptlrpc_bulk_desc *desc, struct obdo *oa, + const lnet_process_id_t *peer, + __u32 client_cksum, __u32 server_cksum, + cksum_type_t server_cksum_type) +{ + char *msg; + cksum_type_t cksum_type; + + /* 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 (desc->bd_export->exp_obd->obd_checksum_dump) + dump_all_bulk_pages(oa, desc->bd_iov_count, + &BD_GET_KIOV(desc, 0), 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"; + + 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", + desc->bd_export->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), + (__u64)BD_GET_KIOV(desc, 0).kiov_page->index << PAGE_SHIFT, + ((__u64)BD_GET_KIOV(desc, + desc->bd_iov_count - 1).kiov_page->index + << PAGE_SHIFT) + + BD_GET_KIOV(desc, desc->bd_iov_count - 1).kiov_len - 1, + client_cksum, cksum_type, server_cksum, server_cksum_type); + return 1; +} + int tgt_brw_read(struct tgt_session_info *tsi) { struct ptlrpc_request *req = tgt_ses_req(tsi); @@ -1725,13 +1934,12 @@ 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); } @@ -1741,12 +1949,15 @@ int tgt_brw_read(struct tgt_session_info *tsi) npages = PTLRPC_MAX_BRW_PAGES; rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, - ioo, remote_nb, &npages, local_nb, NULL, BYPASS_CAPA); + ioo, remote_nb, &npages, local_nb); if (rc != 0) GOTO(out_lock, rc); desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), - BULK_PUT_SOURCE, OST_BULK_PORTAL); + PTLRPC_BULK_PUT_SOURCE | + PTLRPC_BULK_BUF_KIOV, + OST_BULK_PORTAL, + &ptlrpc_bulk_kiov_nopin_ops); if (desc == NULL) GOTO(out_commitrw, rc = -ENOMEM); @@ -1762,9 +1973,10 @@ int tgt_brw_read(struct tgt_session_info *tsi) nob += page_rc; if (page_rc != 0) { /* some data! */ LASSERT(local_nb[i].lnb_page != NULL); - ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].lnb_page, - local_nb[i].lnb_page_offset, - page_rc); + desc->bd_frag_ops->add_kiov_frag + (desc, local_nb[i].lnb_page, + local_nb[i].lnb_page_offset, + page_rc); } if (page_rc != local_nb[i].lnb_len) { /* short read */ @@ -1774,17 +1986,30 @@ int tgt_brw_read(struct tgt_session_info *tsi) 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 = 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); 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(desc, &body->oa, &req->rq_peer, + body->oa.o_cksum, + repbody->oa.o_cksum, cksum_type); } else { repbody->oa.o_valid = 0; } @@ -1793,23 +2018,21 @@ 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))) { + !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2) && + !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_BULK))) { rc = target_bulk_io(exp, desc, &lwi); no_reply = rc != 0; } 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, - NULL, rc); - if (rc == 0) - tgt_drop_id(exp, &repbody->oa); + rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo, + remote_nb, npages, local_nb, rc); out_lock: tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR); if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) - ptlrpc_free_bulk_nopin(desc); + ptlrpc_free_bulk(desc); LASSERT(rc <= 0); if (rc == 0) { @@ -1837,7 +2060,7 @@ out_lock: lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL); l_wait_event(waitq, 0, &lwi1); target_bulk_io(exp, desc, &lwi); - ptlrpc_free_bulk_nopin(desc); + ptlrpc_free_bulk(desc); } RETURN(rc); @@ -1852,27 +2075,30 @@ static void tgt_warn_on_cksum(struct ptlrpc_request *req, { struct obd_export *exp = req->rq_export; struct ost_body *body; - char *router; - char *via; + char *router = ""; + char *via = ""; body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); LASSERT(body != NULL); - if (req->rq_peer.nid == desc->bd_sender) { - via = router = ""; - } else { + if (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, desc->bd_iov_count, + &BD_GET_KIOV(desc, 0), 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, @@ -1906,6 +2132,7 @@ int tgt_brw_write(struct tgt_session_info *tsi) cksum_type_t 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; @@ -1959,7 +2186,7 @@ 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, @@ -1982,14 +2209,13 @@ 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); } @@ -2009,21 +2235,23 @@ int tgt_brw_write(struct tgt_session_info *tsi) npages = PTLRPC_MAX_BRW_PAGES; rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa, - objcount, ioo, remote_nb, &npages, local_nb, NULL, - BYPASS_CAPA); + objcount, ioo, remote_nb, &npages, local_nb); if (rc < 0) GOTO(out_lock, rc); desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), - BULK_GET_SINK, OST_BULK_PORTAL); + 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++) - ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].lnb_page, - local_nb[i].lnb_page_offset, - local_nb[i].lnb_len); + 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) @@ -2064,14 +2292,19 @@ skip_transfer: /* 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, NULL, - rc); + objcount, ioo, remote_nb, npages, local_nb, rc); if (rc == -ENOTCONN) /* quota acquire process has been given up because * either the client has been evicted or the client * 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, @@ -2099,25 +2332,65 @@ 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_nopin(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); } EXPORT_SYMBOL(tgt_brw_write); + +/* Check if request can be reconstructed from saved reply data + * A copy of the reply data is returned in @trd if the pointer is not NULL + */ +bool req_can_reconstruct(struct ptlrpc_request *req, + struct tg_reply_data *trd) +{ + struct tg_export_data *ted = &req->rq_export->exp_target_data; + struct lsd_client_data *lcd = ted->ted_lcd; + bool found; + + if (tgt_is_multimodrpcs_client(req->rq_export)) + return tgt_lookup_reply(req, trd); + + mutex_lock(&ted->ted_lcd_lock); + found = req->rq_xid == lcd->lcd_last_xid || + req->rq_xid == lcd->lcd_last_close_xid; + + if (found && trd != NULL) { + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) { + trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid; + trd->trd_reply.lrd_transno = + lcd->lcd_last_close_transno; + trd->trd_reply.lrd_result = lcd->lcd_last_close_result; + } else { + trd->trd_reply.lrd_xid = lcd->lcd_last_xid; + trd->trd_reply.lrd_transno = lcd->lcd_last_transno; + trd->trd_reply.lrd_result = lcd->lcd_last_result; + trd->trd_reply.lrd_data = lcd->lcd_last_data; + trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0]; + trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1]; + trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2]; + trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3]; + } + } + mutex_unlock(&ted->ted_lcd_lock); + + return found; +} +EXPORT_SYMBOL(req_can_reconstruct); +