> > In preceeding versions of 2.0 I had the same problem with ppp and
> > cdrom, rare and not reproducible. Another person menions the problem
>
> PPP doesnt use DMA memory , so PPP isnt going to be subject to that. The PPP
> problems would have to have been something totally different.
PPP probably does some (too much ??) GFP_ATOMIC allocations that may
fail, and probably is not able to recover from those failures.
I noticed that modules do GFP_ATOMIC allocations at initialisation (load).
Is it a constraint under Linux or could they use some cooler allocation
priorities? (or is it just untrue?)
Why, when I switch some benchmark from a 1K fs to a 4K fs, the process
doing IOs eats about 100% CPU for about 1 IO every 2 seconds, during about
20 seconds?
OK. That is not a normal case and fortunately Linux behaviour is OK in
normal situations.
In my opinion, that shows that memory stress may often happen under
Linux.
> Alan
Gerard.