Re: [BUGFIX]{PATCH] flush icache on ia64 take2

[Zoltan Menyhart <Zoltan.Menyhart@bull.net>]

KAMEZAWA Hiroyuki wrote:

> But is it too costly that flushing icache page only if a page is newly
> installed into the system (PG_arch1) && it is mapped as executable ?

Well it was a bit long time ago, I measured on a Tiger box with
CPUs of 1.3 GHz:

Flushing a page of 64 Kbytes, with modified data in D-cache
(it's slower that not having modified data in the D-cache):

13.1 ... 14.7 usec.

You may have quicker machines, but having more CPUs or a NUMA architecture
can slow it down considerably:
- more CPUs have to agree that that's the moment to carry out a flush
- NUMA adds delay

We may have, say 1 Gbyte / sec local i/o activity (using some RAIDs).
Assume a few % of this 1 Gbyte is the program execution, or program swap in.
It gives some hundreds of new exec pages / sec =>
some msec-s can be lost each sec.

I can agree that it should not be a big deal :-)

> I don't want to leak this (stupid) corner case to the file system layer.
> Hmm...can't we do clever flushing (like your idea) in VM layer ?

As the VM layer is designed to be independent of the page read in stuff...

Thanks,

Zoltan