Re: [patch 1/19] writeback tunables

[Andrew Morton <akpm@zip.com.au>]

Russell King wrote:
> 
> On Mon, Jun 17, 2002 at 12:33:18PM +0200, Martin Dalecki wrote:
> ...
> > > +int dirty_expire_centisecs = 30 * 100;
> > > +
> >
> > Blind guess - didn't the 100 wan't to be HZ?!
> 
> The units are centiseconds (as the name suggests). 5 * 100 centiseconds = 5
> seconds, so the dirty writeback timeout is 5 seconds.  Check the code a
> little further and you'll see HZ gets factored into them on use.
> 

Yup.  Sorry about the "_centisecs" thing.  That's a bit anal, but
I tend to think that it's best to be really explicit about the
units, make it a bit easier to use.  I don't know how many times
I've had to peer in fs/buffer.c to remember what those dang numbers do.

Possibly, "seconds" may be sufficiently high resolution for
these things.  But I wasn't sure - maybe someone wants to
run the kupdate function five times per second?  Dunno.

There are some departures from 2.4 tradition which are worth
mentioning here:

- There is no range checking on the settings.  (But a divide-by
  zero isn't possible, so I think that's OK)

- Unlike the 2.4 bdflush settings, these parameters are not
  updated in a single hit.  So if you modify them by a large
  amount while the system is under heavy writeback load, perhaps
  some whacky things will happen if you create an irrational
  intermediate state.  But that's quite unlikely.

- Unlike 2.4, the settings are scaled by HZ.  So that bdflush
  tuning tool whose name I forget will no longer make kupdate
  run ten times too fast on Alphas.

-