do_io_getevents() is the only user of set/restore_user_sigmask which can return success or error and deliver a signal which was temporary unblocked by set_user_sigmask().
Change it to keep the modified sigmask (pass true to restore_unless) only if the syscall returns ERESTARTNOHAND/EINTR. This matches all other syscalls which modify current->blocked before wait-for-event. Note that it doesn't even need the signal_pending() check, read_event() wait_event_interruptible_hrtimeout() returns ERESTARTSYS if interrupted. But it seems that read_events() needs some unrelated cleanups which should be done first. Signed-off-by: Oleg Nesterov <[email protected]> --- fs/aio.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/fs/aio.c b/fs/aio.c index 944eef7..5bdbef0 100644 --- a/fs/aio.c +++ b/fs/aio.c @@ -2034,7 +2034,6 @@ static long do_io_getevents(aio_context_t ctx_id, { ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX; struct kioctx *ioctx = lookup_ioctx(ctx_id); - bool interrupted; long ret = -EINVAL; if (likely(ioctx)) { @@ -2043,10 +2042,9 @@ static long do_io_getevents(aio_context_t ctx_id, percpu_ref_put(&ioctx->users); } - interrupted = signal_pending(current); - restore_saved_sigmask_unless(interrupted); - if (interrupted && !ret) + if (!ret && signal_pending(current)) ret = -ERESTARTNOHAND; + restore_saved_sigmask_unless(ret == -ERESTARTNOHAND); return ret; } -- 2.5.0
