Re: [patch] get_request starvation fix

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

Rik van Riel wrote:
> 
> On Fri, 8 Feb 2002, Andrew Morton wrote:
> 
> > + *   This all assumes that the rate of taking requests is much, much higher
> > + *   than the rate of releasing them.  Which is very true.
> 
> This is not necessarily true for read requests.
> 
> If each read request is synchronous and the process will
> generate the next read request after the current one
> has finished, then it's quite possible to clog up the
> queue with read requests which are generated at exactly
> the same rate as they're processed.
> 
> Couldn't this still cause starvation, even with your patch?

No, that's fine.

The problem which the comment refers to is: how to provide
per-process request batching without running off and creating
per-process reservation pools or such.

What I'm relying on is that when a sleeper is woken (at low-water),
there are at least (high-water - low-water) requests available before
get_request will again sleep.  And that the woken process will be
able to grab a decent number of those non-blocking requests. I
suspect it's always true, as long as (high-water - low_water) is
"much greater than" the number of CPUs.

The synchronous reader is well-behaved, and should be nicely
FIFO if we're getting low on requests.

-