diff --git a/io/io_waituntil2.c b/io/io_waituntil2.c
index 13a5157..5bef5af 100644
--- a/io/io_waituntil2.c
+++ b/io/io_waituntil2.c
@@ -261,27 +261,89 @@ int64 io_waituntil2(int64 milliseconds) {
     for (i=n-1; i>=0; --i) {
       io_entry* e=iarray_get(&io_fds,y[--n].ident);
       if (e) {
+	/* copies from epoll above */
+	int curevents=0,newevents;
+	if (e->kernelwantread) curevents |= EVFILT_READ;
+	if (e->kernelwantwrite) curevents |= EVFILT_WRITE;
+
+	newevents=0;
+	if (!e->canread || e->wantread) {
+	  newevents |= EVFILT_READ;
+	  e->kernelwantread=1;
+	} else
+	  e->kernelwantread=0;
+	if (!e->canwrite || e->wantwrite) {
+	  newevents |= EVFILT_WRITE;
+	  e->kernelwantwrite=1;
+	} else
+	  e->kernelwantwrite=0;
+
 	if (y[n].flags&EV_ERROR) {
 	  /* error; signal whatever app is looking for */
-	  if (e->wantread) y[n].filter=EVFILT_READ; else
+	  if (e->wantread) y[n].filter=EVFILT_READ;
 	  if (e->wantwrite) y[n].filter=EVFILT_WRITE;
 	}
-	if (!e->canread && (y[n].filter==EVFILT_READ)) {
-	  e->canread=1;
-#ifdef DEBUG
-	  assert(e->next_read==-1);
-#endif
-	  e->next_read=first_readable;
-	  first_readable=y[n].ident;
+
+	/* if we think we can not read, but the kernel tells us that we
+	 * can, put this fd in the relevant data structures */
+	if (!e->canread && (y[n].filter&EVFILT_WRITE)) {
+	  if (e->canread) {
+	    newevents &= ~EVFILT_READ;
+	  } else {
+	    e->canread=1;
+	    if (e->wantread) {
+	      e->next_read=first_readable;
+	      first_readable=y[n].ident;
+	    }
+	  }
 	}
-	if (y[n].filter==EVFILT_WRITE) {
-	  e->canwrite=1;
-#ifdef DEBUG
-	  assert(e->next_write==-1);
-#endif
-	  e->next_write=first_writeable;
-	  first_writeable=y[i].ident;
+
+	/* if the kernel says the fd is writable, ... */
+	if (y[i].filter&EVFILT_WRITE) {
+	  /* Usually, if the kernel says a descriptor is writable, we
+	   * note it and do not tell the kernel not to tell us again.
+	   * The idea is that once we notify the caller that the fd is
+	   * writable, and the caller handles the event, the caller will
+	   * just ask to be notified of future write events again.  We
+	   * are trying to save the superfluous epoll_ctl syscalls.
+	   * If e->canwrite is set, then this gamble did not work out.
+	   * We told the caller, yet after the caller is done we still
+	   * got another write event.  Clearly the user is implementing
+	   * some kind of throttling and we can tell the kernel to leave
+	   * us alone for now. */
+	  if (e->canwrite) {
+	    newevents &= ~EVFILT_WRITE;
+	    e->kernelwantwrite=0;
+	  } else {
+	    /* If !e->wantwrite: The laziness optimization in
+	     * io_dontwantwrite hit.  We did not tell the kernel that we
+	     * are no longer interested in writing to save the syscall.
+	     * Now we know we could write if we wanted; remember that
+	     * and then go on. */
+	    e->canwrite=1;
+	    if (e->wantwrite) {
+	      e->next_write=first_writeable;
+	      first_writeable=y[n].ident;
+	    }
+	  }
 	}
+
+	if (newevents != curevents) {
+	  struct kevent kev;
+	  struct timespec ts;
+	  ts.tv_sec=0; ts.tv_nsec=0;
+	  if (curevents &~ newevents) {
+	    EV_SET(&kev, n, curevents &~ newevents, EV_DELETE, 0, 0, 0);
+	    kevent(io_master,&kev,1,0,0,&ts);
+	  }
+	  if (newevents &~ curevents) {
+	    EV_SET(&kev, n, newevents, EV_ADD|EV_ENABLE, 0, 0, 0);
+	    kevent(io_master,&kev,1,0,0,&ts);
+	  }
+	  if (!newevents)
+	    --io_wanted_fds;
+	}
+
 #ifdef DEBUG
       } else {
 	fprintf(stderr,"got kevent on fd#%d, which is not in array!\n",y[n].ident);