commit d40a4e46b6e4b0ab2f544fe844f6b3523203b550 Author: David Goulet <dgoulet@torproject.org> Date: Thu Feb 1 15:48:28 2018 -0500 test: KIST Scheduler unit tests to test the pending list state This tests many cases of the KIST scheduler with the pending list state by calling entry point in the scheduler while channels are scheduled or not. Also, it adds a test for the bug #24700. Signed-off-by: David Goulet <dgoulet@torproject.org> --- src/test/test_scheduler.c | 267 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 267 insertions(+) diff --git a/src/test/test_scheduler.c b/src/test/test_scheduler.c index 26b412f89..eac66dd27 100644 --- a/src/test/test_scheduler.c +++ b/src/test/test_scheduler.c @@ -1059,6 +1059,271 @@ test_scheduler_ns_changed(void *arg) return; } +/* + * Mocked functions for the kist_pending_list test. + */ + +static int mock_flush_some_cells_num = 1; +static int mock_more_to_flush = 0; +static int mock_update_socket_info_limit = 0; + +static ssize_t +channel_flush_some_cells_mock_var(channel_t *chan, ssize_t num_cells) +{ + (void) chan; + (void) num_cells; + return mock_flush_some_cells_num; +} + +/* Because when we flush cells, it is possible that the connection outbuf gets + * fully drained, the wants to write scheduler event is fired back while we + * are in the scheduler loop so this mock function does it for us. + * Furthermore, the socket limit is set to 0 so once this is triggered, it + * informs the scheduler that it can't write on the socket anymore. */ +static void +channel_write_to_kernel_mock_trigger_24700(channel_t *chan) +{ + static int chan_id_seen[2] = {0}; + if (++chan_id_seen[chan->global_identifier - 1] > 1) { + tt_assert(0); + } + + scheduler_channel_wants_writes(chan); + + done: + return; +} + +static int +channel_more_to_flush_mock_var(channel_t *chan) +{ + (void) chan; + return mock_more_to_flush; +} + +static void +update_socket_info_impl_mock_var(socket_table_ent_t *ent) +{ + ent->cwnd = ent->unacked = ent->mss = ent->notsent = 0; + ent->limit = mock_update_socket_info_limit; +} + +static void +test_scheduler_kist_pending_list(void *arg) +{ + (void) arg; + +#ifndef HAVE_KIST_SUPPORT + return; +#endif + + /* This is for testing the channel flow with the pending list that is + * depending on the channel state, what will be the expected behavior of the + * scheduler with that list. + * + * For instance, we want to catch double channel add or removing a channel + * that doesn't exists, or putting a channel in the list in a wrong state. + * Essentially, this will articifically test cases of the KIST main loop and + * entry point in the channel subsystem. + * + * In part, this is to also catch things like #24700 and provide a test bed + * for more testing in the future like so. */ + + /* Mocking a series of scheduler function to control the flow of the + * scheduler loop to test every use cases and assess the pending list. */ + MOCK(get_options, mock_get_options); + MOCK(channel_flush_some_cells, channel_flush_some_cells_mock_var); + MOCK(channel_more_to_flush, channel_more_to_flush_mock_var); + MOCK(update_socket_info_impl, update_socket_info_impl_mock_var); + MOCK(channel_write_to_kernel, channel_write_to_kernel_mock); + MOCK(channel_should_write_to_kernel, channel_should_write_to_kernel_mock); + + /* Setup options so we're sure about what sched we are running */ + mocked_options.KISTSchedRunInterval = 10; + set_scheduler_options(SCHEDULER_KIST); + + /* Init scheduler. */ + scheduler_init(); + + /* Initialize a channel. We'll need a second channel for the #24700 bug + * test. */ + channel_t *chan1 = new_fake_channel(); + channel_t *chan2 = new_fake_channel(); + tt_assert(chan1); + tt_assert(chan2); + chan1->magic = chan2->magic = TLS_CHAN_MAGIC; + channel_register(chan1); + channel_register(chan2); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_IDLE); + tt_int_op(chan1->sched_heap_idx, OP_EQ, -1); + tt_int_op(chan2->scheduler_state, OP_EQ, SCHED_CHAN_IDLE); + tt_int_op(chan2->sched_heap_idx, OP_EQ, -1); + + /* Once a channel becomes OPEN, it always have at least one cell in it so + * the scheduler is notified that the channel wants to write so this is the + * first step. Might not make sense to you but it is the way it is. */ + scheduler_channel_wants_writes(chan1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_FOR_CELLS); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + + /* Signal the scheduler that it has waiting cells which means the channel + * will get scheduled. */ + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + /* Subsequent call should not add it more times. It is possible we add many + * cells in rapid succession before the channel is scheduled. */ + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + + /* We'll flush one cell and make it that the socket can write but no more to + * flush else we end up in an infinite loop. We expect the channel to be put + * in waiting for cells state and the pending list empty. */ + mock_update_socket_info_limit = INT_MAX; + mock_more_to_flush = 0; + the_scheduler->run(); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_FOR_CELLS); + + /* Lets make believe that a cell is now in the channel but this time the + * channel can't write so obviously it has more to flush. We expect the + * channel to be back in the pending list. */ + scheduler_channel_has_waiting_cells(chan1); + mock_update_socket_info_limit = 0; + mock_more_to_flush = 1; + the_scheduler->run(); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + + /* Channel is in the pending list now, during that time, we'll trigger a + * wants to write event because maybe the channel buffers were emptied in + * the meantime. This is possible because once the connection outbuf is + * flushed down the low watermark, the scheduler is notified. + * + * We expect the channel to NOT be added in the pending list again and stay + * in PENDING state. */ + scheduler_channel_wants_writes(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + + /* Make it that the channel can write now but has nothing else to flush. We + * expect that it is removed from the pending list and waiting for cells. */ + mock_update_socket_info_limit = INT_MAX; + mock_more_to_flush = 0; + the_scheduler->run(); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_FOR_CELLS); + + /* While waiting for cells, lets say we were able to write more things on + * the connection outbuf (unlikely that this can happen but let say it + * does). We expect the channel to stay in waiting for cells. */ + scheduler_channel_wants_writes(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_FOR_CELLS); + + /* We'll not put it in the pending list and make the flush cell fail with 0 + * cell flushed. We expect that it is put back in waiting for cells. */ + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + mock_flush_some_cells_num = 0; + the_scheduler->run(); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_FOR_CELLS); + + /* Set the channel to a state where it doesn't want to write more. We expect + * that the channel becomes idle. */ + scheduler_channel_doesnt_want_writes(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_IDLE); + + /* Some cells arrive on the channel now. We expect it to go back in waiting + * to write. You might wonder why it is not put in the pending list? Because + * once the channel becomes OPEN again (the doesn't want to write event only + * occurs if the channel goes in MAINT mode), if there are cells in the + * channel, the wants to write event is triggered thus putting the channel + * in pending mode. + * + * Else, if no cells, it stays IDLE and then once a cell comes in, it should + * go in waiting to write which is a BUG itself because the channel can't be + * scheduled until a second cell comes in. Hopefully, #24554 will fix that + * for KIST. */ + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_TO_WRITE); + + /* Second cell comes in, unfortunately, it won't get scheduled until a wants + * to write event occurs like described above. */ + scheduler_channel_has_waiting_cells(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 0); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_WAITING_TO_WRITE); + + /* Unblock everything putting the channel in the pending list. */ + scheduler_channel_wants_writes(chan1); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 1); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + + /* Testing bug #24700 which is the situation where we have at least two + * different channels in the pending list. The first one gets flushed and + * bytes are written on the wire which triggers a wants to write event + * because the outbuf is below the low watermark. The bug was that this + * exact channel was added back in the pending list because its state wasn't + * PENDING. + * + * The following does some ninja-tsu to try to make it happen. We need two + * different channels so we create a second one and add it to the pending + * list. Then, we have a custom function when we write to kernel that does + * two important things: + * + * 1) Calls scheduler_channel_wants_writes(chan) on the channel. + * 2) Keeps track of how many times it sees the channel going through. If + * that limit goes > 1, it means we've added the channel twice in the + * pending list. + * + * In the end, we expect both channels to be in the pending list after this + * scheduler run. */ + + /* Put the second channel in the pending list. */ + scheduler_channel_wants_writes(chan2); + scheduler_channel_has_waiting_cells(chan2); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 2); + tt_int_op(chan2->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + + /* This makes it that the first pass on socket_can_write() will be true but + * then when a single cell is flushed (514 + 29 bytes), the second call to + * socket_can_write() will be false. If it wasn't sending back false on the + * second run, we end up in an infinite loop of the scheduler. */ + mock_update_socket_info_limit = 600; + /* We want to hit "Case 3:" of the scheduler so channel_more_to_flush() is + * true but socket_can_write() has to be false on the second check on the + * channel. */ + mock_more_to_flush = 1; + mock_flush_some_cells_num = 1; + MOCK(channel_write_to_kernel, channel_write_to_kernel_mock_trigger_24700); + the_scheduler->run(); + tt_int_op(smartlist_len(get_channels_pending()), OP_EQ, 2); + tt_int_op(chan1->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + tt_int_op(chan2->scheduler_state, OP_EQ, SCHED_CHAN_PENDING); + + done: + chan1->state = chan2->state = CHANNEL_STATE_CLOSED; + chan1->registered = chan2->registered = 0; + channel_free(chan1); + channel_free(chan2); + scheduler_free_all(); + + UNMOCK(get_options); + UNMOCK(channel_flush_some_cells); + UNMOCK(channel_more_to_flush); + UNMOCK(update_socket_info_impl); + UNMOCK(channel_write_to_kernel); + UNMOCK(channel_should_write_to_kernel); +} + struct testcase_t scheduler_tests[] = { { "compare_channels", test_scheduler_compare_channels, TT_FORK, NULL, NULL }, @@ -1068,6 +1333,8 @@ struct testcase_t scheduler_tests[] = { { "loop_kist", test_scheduler_loop_kist, TT_FORK, NULL, NULL }, { "ns_changed", test_scheduler_ns_changed, TT_FORK, NULL, NULL}, { "should_use_kist", test_scheduler_can_use_kist, TT_FORK, NULL, NULL }, + { "kist_pending_list", test_scheduler_kist_pending_list, TT_FORK, + NULL, NULL }, END_OF_TESTCASES };