On 5/19/20 3:52 PM, Bijan Mottahedeh wrote: > This patch set and a corresponding kernel patch set are fixes and > optimizations resulting from running unit test 500f9fbadef8-test. > > - Patch 1 is a fix to the test hanging when it runs on a non-mq queue. > > The patch preserves the value of wait_nr if SETUP_IOPOLL is set > since otherwise __sys_io_uring_enter() could never be called > __io_uring_peek_cqe() could never find new completions. > > With this patch applied, two problems were hit in the kernel as described > in the kernel patch set, which caused 500f9fbadef8-test to fail and > to hang. With all three patches, 500f9fbadef8-test either passes > successfully or skips the test gracefully with the following message: > > Polling not supported in current dir, test skipped > > - Patch 2 is an optimization for io_uring_enter() system calls. > > If we want to wait for completions (wait_nr > 0), account for the > completion we might fetch with __io_uring_peek_cqe(). For example, > with wait_nr=1 and submit=0, there is no need to call io_uring_enter() > if the peek call finds a completion. > > Below are the perf results for 500f9fbadef8-test without/with the fix: > > perf stat -e syscalls:sys_enter_io_uring_enter 500f9fbadef8-test > > 12,289 syscalls:sys_enter_io_uring_enter > 8,193 syscalls:sys_enter_io_uring_enter > > - Patch 3 is a cleanup with no functional changes. > > Since we always have > > io_uring_wait_cqe_nr() > -> __io_uring_get_cqe() > -> __io_uring_peek_cqe() > > remove the direct call from io_uring_wait_cqe_nr() to __io_uring_peek_cqe(). > > After the removal, __io_uring_peek_cqe() is called only from > __io_uring_get_cqe() so move the two routines together(). Without the > move, compilation fails with a 'defined but not used' error. LGTM, thanks! I've applied them. -- Jens Axboe
