run-fibers returns when initial fiber finishes

* epoll.c (scm_primitive_epoll_wait): Accept timeout in internal time
  units instead of milliseconds.  Avoid some overhead if the timeout is
  zero.
* fibers/epoll.scm (epoll): Adapt to primitive-epoll-wait chance.
  Change get-timeout callback to an "expiry" argument and an
  update-expiry last-minute function.  Tighten the window around the
  "waiting" epoll-state.
* fibers.scm (%run-fibers): Remove current-read-waiter /
  current-write-waiter parameterization, given that fibers individually
  each have this parameterization.
  (run-fibers): Add #:drain? argument.
* fibers/internal.scm (schedule-runnables-for-next-turn): Simplify a
  bit, inlining scheduler-poll-timeout and adapting to epoll change.
  (scheduler-work-pending?): New function.
  (run-scheduler): Simplify finish? logic.
* tests/basic.scm:
* tests/speedup.scm: Update to drain where needed.
* fibers.texi: Update for new run-fibers termination condition.

1 files changed, 3 insertions, 2 deletions

diff --git a/tests/speedup.scm b/tests/speedup.scm
index ea67d6a..6dfe969 100644
--- a/tests/speedup.scm
+++ b/tests/speedup.scm
@@ -40,9 +40,10 @@
     (format #t "speedup for ~s: " 'exp)
     (force-output)
     (let ((thunk (lambda () exp)))
-      (let ((t1 (time (lambda () (run-fibers thunk #:parallelism 1)))))
+      (let ((t1 (time (lambda ()
+                        (run-fibers thunk #:parallelism 1 #:drain? #t)))))
         (format #t "~a s" t1)
-        (let ((t2 (time (lambda () (run-fibers thunk)))))
+        (let ((t2 (time (lambda () (run-fibers thunk #:drain? #t)))))
           (format #t " / ~a s = ~ax (~a cpus)\n" t2 (/ t1 t2)
                   (current-processor-count)))))))