The transaction tests need both sides to send a sync message after running the test. This ensures that all remote operations are complete before dapltest deregeisters memory and disconnects the endpoints.
Without this logic, we see async errors on iwarp devices because a read or write arrives after the rmr has been destroyed. I believe this is more likely to happen with iWARP than IB because iWARP completions only indicate the local buffer can be reused. It doesn't imply that the message has even arrived at the peer, let alone been placed in memory. I've tested this on cxgb4 only so far. Similar logic is probably needed for the performance tests as well. Signed-off-by: Steve Wise <sw...@opengridcomputing.com> --- test/dapltest/test/dapl_transaction_stats.c | 10 ++ test/dapltest/test/dapl_transaction_test.c | 139 +++++++++++++++++++++++++++ 2 files changed, 149 insertions(+), 0 deletions(-) diff --git a/test/dapltest/test/dapl_transaction_stats.c b/test/dapltest/test/dapl_transaction_stats.c index f9d6377..6422ee3 100644 --- a/test/dapltest/test/dapl_transaction_stats.c +++ b/test/dapltest/test/dapl_transaction_stats.c @@ -59,6 +59,16 @@ DT_transaction_stats_set_ready(DT_Tdep_Print_Head * phead, DT_Mdep_Unlock(&transaction_stats->lock); } +void +DT_transaction_stats_reset_wait_count(DT_Tdep_Print_Head * phead, + Transaction_Stats_t * transaction_stats, + unsigned int num) +{ + DT_Mdep_Lock(&transaction_stats->lock); + transaction_stats->wait_count = num; + DT_Mdep_Unlock(&transaction_stats->lock); +} + bool DT_transaction_stats_wait_for_all(DT_Tdep_Print_Head * phead, Transaction_Stats_t * transaction_stats) diff --git a/test/dapltest/test/dapl_transaction_test.c b/test/dapltest/test/dapl_transaction_test.c index 779ea86..b94a2cc 100644 --- a/test/dapltest/test/dapl_transaction_test.c +++ b/test/dapltest/test/dapl_transaction_test.c @@ -1799,6 +1799,145 @@ DT_Transaction_Run(DT_Tdep_Print_Head * phead, Transaction_Test_t * test_ptr) } /* end loop for each op */ } /* end loop for iteration */ + /* + * Final sync up to ensure all previous remote operations have + * finished. + * + * Without a final sync exchange, we see async errors on iwarp + * devices because a read or write arrives after the rmr has + * been destroyed. I believe this is more likely to happen with + * iWARP than IB because iWARP completions only indicate the + * local buffer can be reused. It doesn't imply that the + * message has even arrived at the peer, let alone been placed in memory. + */ + if (test_ptr->is_server) { + /* + * Server + */ + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x "]: Send Final Sync to Client\n", + test_ptr->base_port)); + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + if (!DT_post_recv_buffer(phead, + test_ptr->ep_context[i].ep_handle, + test_ptr->ep_context[i].bp, + SYNC_RECV_BUFFER_ID, SYNC_BUFF_SIZE)) { + goto bail; + } + if (!DT_post_send_buffer(phead, + test_ptr->ep_context[i]. + ep_handle, + test_ptr->ep_context[i].bp, + SYNC_SEND_BUFFER_ID, + SYNC_BUFF_SIZE)) { + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x + "]: Server final sync send error\n", + test_ptr->base_port)); + goto bail; + } + } + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + DAT_DTO_COMPLETION_EVENT_DATA dto_stat; + + if (!DT_dto_event_wait(phead, + test_ptr->ep_context[i]. + reqt_evd_hdl, &dto_stat)) { + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x + "]: Server final sync send error\n", + test_ptr->base_port)); + + goto bail; + } + } + + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x "]: Wait for Final Sync Message\n", + test_ptr->base_port)); + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + DAT_DTO_COMPLETION_EVENT_DATA dto_stat; + + if (!DT_dto_event_wait(phead, + test_ptr->ep_context[i]. + recv_evd_hdl, &dto_stat)) { + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x + "]: Server final sync recv error\n", + test_ptr->base_port)); + goto bail; + } + } + } else { + + /* + * Client + */ + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x "]: Wait for Final Sync Message\n", + test_ptr->base_port)); + DT_transaction_stats_reset_wait_count(phead, + &test_ptr->pt_ptr-> + Client_Stats, + test_ptr->cmd->eps_per_thread); + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + DAT_DTO_COMPLETION_EVENT_DATA dto_stat; + + if (!DT_post_recv_buffer(phead, + test_ptr->ep_context[i].ep_handle, + test_ptr->ep_context[i].bp, + SYNC_RECV_BUFFER_ID, SYNC_BUFF_SIZE)) { + goto bail; + } + if (!DT_dto_event_wait(phead, + test_ptr->ep_context[i]. + recv_evd_hdl, &dto_stat)) { + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x + "]: Client final sync recv error\n", + test_ptr->base_port)); + goto bail; + } + DT_transaction_stats_set_ready(phead, + &test_ptr->pt_ptr-> + Client_Stats); + } + DT_Tdep_PT_Debug(1, + (phead, "Test[" F64x "]: Send Final Sync Msg\n", + test_ptr->base_port)); + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + if (!DT_post_send_buffer(phead, + test_ptr->ep_context[i]. + ep_handle, + test_ptr->ep_context[i].bp, + SYNC_SEND_BUFFER_ID, + SYNC_BUFF_SIZE)) { + DT_Tdep_PT_Debug(1, + (phead, + "Test[" F64x + "]: Client sync send error\n", + test_ptr->base_port)); + goto bail; + } + } + for (i = 0; i < test_ptr->cmd->eps_per_thread; i++) { + DAT_DTO_COMPLETION_EVENT_DATA dto_stat; + + if (!DT_dto_event_wait(phead, + test_ptr->ep_context[i]. + reqt_evd_hdl, &dto_stat)) { + goto bail; + } + } + } + /* end time and print stats */ test_ptr->stats.end_time = DT_Mdep_GetTime(); if (!test_ptr->pt_ptr->local_is_server) { -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html