Re: [PATCH 1/2] boot: increase stack size for kernel boot loader decompressor

["Yinghai Lu" <yhlu.kernel@gmail.com>]

On Wed, Apr 9, 2008 at 8:08 AM, Alexander van Heukelum
<heukelum@fastmail.fm> wrote:
>
>  On Tue, 8 Apr 2008 10:54:15 -0700, "Yinghai Lu" <yhlu.kernel@gmail.com>
>  said:
>
> > On Tue, Apr 8, 2008 at 1:23 AM, Ingo Molnar <mingo@elte.hu> wrote:
>  > >
>  > >  * Alexander van Heukelum <heukelum@mailshack.com> wrote:
>  > >
>  > >  > I did see that the malloc space that the inflate code is using is
>  > >  > taken from _after_ the end of the bss. I don't see how this is
>  > >  > protected from being used/overwritten. Changing the stack size changes
>  > >  > the memory layout a bit... maybe you were so unlucky to create a
>  > >  > vmlinux image that was just barely smaller than some threshold and
>  > >  > increasing the stack size made the decompression/relocation area be
>  > >  > located somewhere else?
>  > >  >
>  > >  > Test patch follows.
>  > >
>  > >  that's a really interesting theory.
>  > >
>  > >  FWIIW, i've been booting allyesconfig bzImages for a long time (with
>  > >  only minimal amount of drivers disabled - mostly old ISA ones that
>  > >  assume the presence of the real hardware), and they boot and work fine
>  > >  on both 32-bit and 64-bit typical whitebox PCs. That means huge bzImages
>  > >  that decompresses into a ~41 MB kernel image. I'd expect that to be a
>  > >  rather severe test of the decompressor.
>  >
>  > i don't that Alexander's patch is needed.
>
>  Hello Yinghai Lu,
>
>  Indeed, I now think it is not needed either. The decompression is
>  done in-place nowadays: the (compressed) image is moved to a high
>  memory address first, then the decompression is done starting at
>  the low end of the buffer. It is guaranteed that the output never
>  overwrites the input, and the decompression code, the stack, and
>  the heap are all at higher addresses than the input buffer. The
>  same goes for the pagetables needed for x86_64.
>
>
>  > also because Alex move heap before _end,
>  > we may need add some extra for buffer offset
>  >
>  >         /* Replace the compressed data size with the uncompressed size */
>  >         subl    input_len(%ebp), %ebx
>  >         movl    output_len(%ebp), %eax
>  >         addl    %eax, %ebx
>  >         /* Add 8 bytes for every 32K input block */
>  >         shrl    $12, %eax
>  >         addl    %eax, %ebx
>  >         /* Add 32K + 18 bytes of extra slack and align on a 4K boundary
>  >         */
>  >         addl    $(32768 + 18 + 4095), %ebx
>  >         andl    $~4095, %ebx =============================> need add
>  > heap size too.
>  > ....
>
>  No, that size is accounted for automatically: the code computes the
>  buffer size needed (including slack) minus the buffer size that is
>  already available (in the embedded gzip-file). The image is moved
>  by this amount (rounded up to a page). So that part is fine.

just wonder if Ingo have very big vmlinux, that +32K + 18 formula still works.

YH