Re: simple_alloc space tramples initrd

On Mon, 2016-01-11 at 15:07 -0800, dwalker@fifo99.com wrote:

On Tue, Jan 12, 2016 at 09:17:53AM +1100, Michael Ellerman wrote:

quoted

On Mon, 2016-01-11 at 08:49 -0800, dwalker@fifo99.com wrote:

quoted

On Mon, Jan 11, 2016 at 02:09:34PM +1100, Michael Ellerman wrote:

quoted

On Fri, 2016-01-08 at 09:45 -0800, dwalker@fifo99.com wrote:

quoted

A powerpc machine I'm working on has this problem where the
simple_alloc_init() area is trampling the initrd. The two are placed fairly
close together.

Which machine / platform?

It's not upstream yet. I'm still putting the patches together, that's when this
issue came up. I can send an RFC if you want to look at the patches.

OK. Thanks but I don't need more patches to look at :)

I was just trying to narrow down which code you were talking about.

It's coming eventually anyways ;) ..

Hah, yeah I know :)

quoted

quoted

quoted

I don't really know that code very well. But ideally either the boot loader
gives you space, or the platform boot code is smart enough to detect that there
is insufficient room and puts the heap somewhere else.

It seems like the kernel should be able to handle it. I believe the bootloader passes
the initrd location , but I don't think it's evaluated till later in the boot up. For
simple_alloc_init() it seems all platforms just assume the space is empty without checking.

Yeah that's what I see too, which seems like it's liable to break, but
obviously hasn't for anyone else yet.

The bootloader must pass the initrd location, otherwise the kernel can't use
it, so it seems like the kernel should be able to notice when they are too
close. But it may be complicated by the sequencing of the code.

I found a similar one,

arch/powerpc/boot/ps3.c:platform_init()

I realized that in platform_init() your discovering the initrd location, so you do have
access to the values. In ps3 you can see how if the initrd is placed in the 16megs after
the kernel image then the simple_alloc code could corrupt it.

I think it would be appropriate to check the initrd location in that function
(since it's available) and make a choice to put the simple_alloc area after
the initrd if the areas overlap. Does that make sense ?

Hmm, maybe. I think the ps3 code knows at link time where the initrd is, so
that's kind of cheating.

But if on your platform you can find out the initrd location early enough, then
yes ideally you take that into account when initialising the allocator.

cheers