[PATCH 2/2] ARM: LPAE: reduce damage caused by idmap to virtual memory layout

[will.deacon@arm.com (Will Deacon)]

On Mon, Jul 28, 2014 at 07:57:00PM +0100, Konstantin Khlebnikov wrote:
> On Mon, Jul 28, 2014 at 10:41 PM, Will Deacon <will.deacon@arm.com> wrote:
> > On Mon, Jul 28, 2014 at 07:25:14PM +0100, Konstantin Khlebnikov wrote:
> >> On Mon, Jul 28, 2014 at 10:14 PM, Will Deacon <will.deacon@arm.com> wrote:
> >> > On Tue, Jul 22, 2014 at 04:36:35PM +0100, Konstantin Khlebnikov wrote:
> >> >> idmap layout combines both phisical and virtual addresses.
> >> >> Everything works fine if ram physically lays below PAGE_OFFSET.
> >> >> Otherwise idmap starts punching huge holes in virtual memory layout.
> >> >> It maps ram by 2MiB sections, but when it allocates new pmd page it
> >> >> cuts 1GiB at once.
> >> >>
> >> >> This patch makes a copy of all affected pmds from init_mm.
> >> >> Only few (usually one) 2MiB sections will be lost.
> >> >> This is not eliminates problem but makes it 512 times less likely.
> >> >
> >> > I'm struggling to understand your commit message, but making a problem `512
> >> > times less likely' does sound like a bit of a hack to me. Can't we fix this
> >> > properly instead?
> >>
> >> Yep, my comment sucks.
> >>
> >> Usually idmap looks like this:
> >>
> >> |0x00000000 -- <chunk of physical memory in identical mapping > --- |
> >> TASK_SIZE -- <kernel space vm layoyt> --- 0xFFFFFFFF |
> >>
> >> But when that physical memory chunk starts from 0xE8000000 or even
> >> 0xF2000000 evenything becomes very complicated.
> >
> > Why? As long as we don't clobber the kernel text (which would require
> > PHYS_OFFSET to be at a really weird alignment and very close to
> > PAGE_OFFSET), then you should be alright. Sure, you'll lose things like your
> > stack and the vmalloc area etc, but you're running in the idmap, so don't
> > use those things.
> 
> Yep, we have piece of hardware with really weird aligned PHYS_OFFSET,
> mostly all ram is above 4gb and we was lucky enough to get into the trouble.
> 
> It seems keystone has all memory above 4gb but small piece is mapped below
> especially for booting. I suppose it's below PAGE_OFFSET so Cyril
> hadn't seen that problem.

Sure, I remember when the keystone support was merged. There's a low alias
of memory which isn't I/O coherent, so when we enable the MMU we rewrite the
page tables to use the high alias (which is also broken, as I pointed out on
the list today).

> Also I seen comment somewhere in the code which tells that idrmap pgd is
> always below 4gb which isn't quite true. Moreover, I had some experiments with
> mapping ram to random places in qemu and seen that kernel cannot boot if
> PHYS_OFFSET isn't alligned to 128mb which is strange.
> So, it seems there is plenty bugs anound.
> 
> >
> > soft_restart is an example of code that deals with these issues. Which code
> > is causing you problems?
> 
> That was booting of secondary cpus, all of them.

Ok. I think we need more specifics to progress with this patch. It's not
clear to me what's going wrong or why your platform is causing the issues
you're seeing.

Will