[PATCH] Increase kernel text size for x86_64

[Simon Horman]

On Thu, Aug 13, 2009 at 10:12:37AM +0200, Hannes Reinecke wrote:
> Simon Horman wrote:
> > On Wed, Aug 12, 2009 at 09:41:23AM +0200, Hannes Reinecke wrote:
> >> Hi all,
> >>
> >> I cannot load a x86_64 kernel with kexec on 2.6.31; the error message is:
> >>
> >> Can't find kernel text map area from kcore
> >> Cannot load /boot/vmlinuz
> >>
> >> Digging through the source I found a mismatch between the assumed
> >> kernel text size; kexec has:
> >>
> >> #define KERNEL_TEXT_SIZE  (40UL*1024*1024)
> >>
> >> but on the kernel side we have:
> >>
> >> include/asm/page_64_types.h:
> >> #define KERNEL_IMAGE_SIZE   (512 * 1024 * 1024)
> >>
> >> And, indeed, changing the definition in kexec-tools to the kernel one
> >> fixed the problem.
> >>
> >> Not sure if this has been reported before, if so please ignore
> >> the noise.
> > 
> > Hi,
> > 
> > I'm wondering what effect this would have when using
> > kexec-tools with older kernels, which presumably have
> > a smaller value for KERNEL_IMAGE_SIZE.
> > 
> >> diff --git a/kexec/arch/x86_64/crashdump-x86_64.h b/kexec/arch/x86_64/crashdump-x86_64.h
> >> index 9f4dee9..0e83527 100644
> >> --- a/kexec/arch/x86_64/crashdump-x86_64.h
> >> +++ b/kexec/arch/x86_64/crashdump-x86_64.h
> >> @@ -11,7 +11,7 @@ int load_crashdump_segments(struct kexec_info *info, char *mod_cmdline,
> >>  #define MAXMEM           0x3fffffffffffUL
> >>  
> >>  /* Kernel text size */
> >> -#define KERNEL_TEXT_SIZE  (40UL*1024*1024)
> >> +#define KERNEL_TEXT_SIZE  (512UL*1024*1024)
> >>  
> >>  #define CRASH_MAX_MEMMAP_NR	(KEXEC_MAX_SEGMENTS + 1)
> >>  #define CRASH_MAX_MEMORY_RANGES	(MAX_MEMORY_RANGES + 2)
> > 
> Well, the actual point of failure is in kexec/arch/x86_64/crashdump-x86_64.c:125
> 
> 			/* Look for kernel text mapping header. */
> 			if ((saddr >= __START_KERNEL_map) &&
> 			    (eaddr <= __START_KERNEL_map + KERNEL_TEXT_SIZE)) {
> 				saddr = (saddr) & (~(KERN_VADDR_ALIGN - 1));
> 				info->kern_vaddr_start = saddr;
> 				size = eaddr - saddr;
> 
> which basically should look for a ELF header into which the kernel fits. As we're
> effectively just moving the upper limit I fail to see any harm to older kernels.
> 
> Hmm. Haven't checked for other occurrences of KERNEL_TEXT_SIZE, though.
> 
> Would be nice we could be getting these values off somewhere; rely on
> some kernel defines is going to break now and then.

Agreed, but I think that your patch is worth having in the mean time.
I will apply it.