Messages in this thread | | | Date | Fri, 6 Nov 2009 08:42:38 +0900 | From | KAMEZAWA Hiroyuki <> | Subject | Re: [MM] Make mm counters per cpu instead of atomic |
| |
On Thu, 5 Nov 2009 10:10:56 -0500 (EST) Christoph Lameter <cl@linux-foundation.org> wrote:
> On Thu, 5 Nov 2009, KAMEZAWA Hiroyuki wrote: > > > Hmm, I don't fully understand _new_ percpu but... > > In logical (even if not realistic), x86-32 supports up to 512 ? cpus in Kconfig. > > BIGSMP. > > x86-32 only supports 32 processors. Plus per cpu areas are only allocated > for the possible processors. > My number is just from Kconfig.
> > Then, if 65536 process runs, this consumes > > > > 65536(nr_proc) * 8 (size) * 512(cpus) = 256MBytes. > > With 32 possible cpus this results in 16m of per cpu space use. > If swap_usage is added, 24m, 25% of vmalloc area. (But, yes, returning -ENOMEM to fork() is ok to me, 65536 proc are extreme.)
Thanks, -Kame
| |