Re: PCI Express support for 2.4 kernel

[Jeff Garzik <jgarzik@pobox.com>]

Vladimir Kondratiev wrote:
> Linus, FIXMAP helps, it is lighter then ioremap, but it still requires 
> page table walk. In addition, since several operations should be done 
> atomically, lock/unlock required as well. Can it be faster method, 
> without page table walk for each transaction? To what extent should one 
> concern about performance here?
> 
> As alternative between 1 page and 256M, I see also lazy allocation on 
> per-bus basis: when bus is first accessed, ioremap 1Mb for it. On real 
> system, it is no more then 3-4 buses. This way, we will end with several 
> 1MB mappings. Finer granularity do not looks feasible, since bus 
> scanning procedure tries to access all devices.


Well, two things to consider:
* probing is not a performance-intensive operation
* even with a system loaded down with many PCI Express devices, I doubt 
you will need more than 1-2MB mapped total, during runtime

So, one alternative could be to keep a cache of ioremap() regions -- 
smaller than 1MB in my opinion -- and update that as needed.

Anybody doing large numbers of PCI config register reads/writes in a hot 
path should be shot (or the h/w designers, depending on situation), and 
PCI Express doesn't change that ;-)

	Jeff