Re: [PATCH V4] Switch from select() to poll() in xenconsoled's IO loop

[Wei Liu <Wei.Liu2@citrix.com>]

On Mon, 2013-01-07 at 15:16 +0000, Ian Campbell wrote:
> On Mon, 2013-01-07 at 15:01 +0000, Wei Liu wrote:
> > On Mon, 2013-01-07 at 14:41 +0000, Mats Petersson wrote:
> > 
> > > >                          return;
> > > > @@ -982,11 +1024,7 @@ void handle_io(void)
> > > >                  /* Re-calculate any event counter allowances & unblock
> > > >                     domains with new allowance */
> > > >                  for (d = dom_head; d; d = d->next) {
> > > > -                       /* Add 5ms of fuzz since select() often returns
> > > > -                          a couple of ms sooner than requested. Without
> > > > -                          the fuzz we typically do an extra spin in select()
> > > > -                          with a 1/2 ms timeout every other iteration */
> > > > -                       if ((now+5) > d->next_period) {
> > > > +                       if (now > d->next_period) {
> > > Is poll more accurate than select? I would have thought that they were
> > > based on the same timing, and thus equally "fuzzy"?
> > 
> > Is there any actual proof that the fuzz is needed? Specs of both
> > select() and poll() don't seem to mention this behaviour at all.
> 
> I agree that it seems pretty dubious but it might be interesting to see
> what strace shows, specifically if it shows this extra spin every other
> iteration.
> 
> The fuzz was introduced in 16257:955ee4fa1345 "Rate-limit activity
> caused by each domU." but the commit log doesn't make any reference to
> the reason for it.
> 
> It may just have been a kernel bug at around the time that patch was
> authored, but google doesn't seem to show any evidence of such a bug
> ever being widespread (i.e. I don't find any references to it).
> 

Then I will leave the fuzz here, and do another patch to remove it.

I will also add reference to CS 16257 in the comment.


Wei.