[PATCH] xfs: force background CIL push under sustained load

[Alex Elder]

On Fri, 2010-09-24 at 11:18 +1000, Dave Chinner wrote:
> On Thu, Sep 23, 2010 at 01:53:55PM -0500, Alex Elder wrote:
> > On Thu, 2010-09-23 at 12:27 +1000, Dave Chinner wrote:

. . .

> > I don't really care much about this, but I'll take this
> > opportunity for a small rant.
> > 
> > The difference in calculation cost/speed between "x >> 3" and
> > "x / 8" is vanishingly small.
> 
> That is architecture dependent, but in most cases these days the
> compiler will optimise a divide-by-power-of−2-constant into a shift
> operation anyway. I'm pretty sure that optimisation is done on even
> on x86 as a shift is a single cycle operation while an integer
> divide still takes several cycles and consumes more power.

I know, it'll take about a nanosecond longer than necessary,
and an extra 25 picowatt hours (IF the compiler doesn't optimize
it away).  This being done to determine whether to start a series
of I/O operations...

> > I think it is meaningful to use
> > a shift in places where a power-of-two is mandated, but in places
> > like this it suggests there is a constraint that simply doesn't
> > exist.  So for example, you could have chosen (log->logsize / 10)
> > as the "try pushing" value, and (log->logsize / 4 - 1) as the
> > "must push" value.
> 
> It's more the fact that XFS uses power-of-2 logic (i.e shifts)
> everywhere. I just tend to be consistent with what is already there.
> In this case, the AIL push thresholds are calculated using shifts:
> 
> 	free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
> 
> and so when you compare that to the XLOG_CIL_SPACE_LIMIT()
> definitions, it is immediately clear that the CIL limits are smaller
> than the AIL push threshold...

I get it.  And as I said before, I truly don't care much about
this--it's an observation more than even an annoyance, and it's
already taken a lot more time and mental energy (not to mention
electric power) than it ever really deserved.

					-Alex