Re: [RFC] mlock/stack guard interaction fixup

[Ian Campbell <ijc@hellion.org.uk>]

[-- Attachment #1: Type: text/plain, Size: 2571 bytes --]

On Sun, 2010-08-22 at 08:33 +0100, Ian Campbell wrote:
> On Sun, 2010-08-22 at 07:57 +0100, Ian Campbell wrote:
> > On Sat, 2010-08-21 at 08:48 -0700, Linus Torvalds wrote:
> > > On Sat, Aug 21, 2010 at 4:56 AM, Ian Campbell <ijc@hellion.org.uk> wrote:
> > > >
> > > > I don't know that they are particularly good tests for this change but I
> > > > also ran allmodconfig kernel build and ltp on 2.6.35.3+fixes without
> > > > issue. Are there any good mlock heavy workloads?
> > > 
> > > mlock itself isn't very interesting, I think more interesting is
> > > testing that the doubly linked list handles all the cases correctly.
> > > Something that splits mappings, unmaps partial ones etc etc. Running
> > > something like Electric Fence is probably a good idea.
> > 
> > EF_DISABLE_BANNER=1 EF_ALLOW_MALLOC_0=1 LD_PRELOAD=libefence.so.0.0 make
> > 
> > craps out pretty quickly with:
> > 
> >           CC      init/main.o
> >         
> >         ElectricFence Exiting: mprotect() failed: Cannot allocate memory
> >         make[1]: *** [init/main.o] Error 255
> >         make: *** [init] Error 2
> > 
> > but it does that with 2.6.35.3, 2.6.35.2, 2.6.35.1 and 2.6.35 too so it
> > doesn't seem to be breakage relating to any of the stack guard stuff
> 
> I increased the vm.max_map_count sysctl and now things are rolling
> along. Will let you know how I get on.

So its slow and memory intensive as hell due to efence so my test box is
struggling[0] but it has compiled 270+ .o files successfully so I think
it's OK from that perspective. I think it'll be quite a while before I
can say its passed an allmodconfig under efence though.

In the meantime I notice you've committed the patches. Can we get them
queued up for stable backports at some point? I appreciate you might
want them to bake for a bit longer in 2.6.36-rc first.

Greg, we are talking about:
0e8e50e20c837eeec8323bba7dcd25fe5479194c mm: make stack guard page logic use vm_prev pointer
7798330ac8114c731cfab83e634c6ecedaa233d7 mm: make the mlock() stack guard page checks stricter
297c5eee372478fc32fec5fe8eed711eedb13f3d mm: make the vma list be doubly linked

Cheers,
Ian.

[0] It's a 4G x 4 core box which normally does -j16 builds with no swap.
Now a -j4 will send it deep into the 16G of swap I added this morning,
and even a -j3 is a bit tight. I knew efence was resource intensive but
this still surprised me.

-- 
Ian Campbell

Who does not love wine, women, and song,
Remains a fool his whole life long.
		-- Johann Heinrich Voss

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 198 bytes --]