Re: [PATCH] x86/boot/KASLR: Skip relocation handling in no kaslr case

[Baoquan He]

On 07/06/17 at 09:28pm, Baoquan He wrote:
> > > Do you mean the handling in boot/compressed/head_64.S? Whatever it does,
> > > it's only for physical address. The virtual address mapping is not
> > > touched. Here virt_addr respresents the offset between
> > > 0xffffffff80000000, 0xffffffffc0000000. And if skip the relocation
> > > handling, we can see that the '_text' will be mapped to
> > > 0xffffffff81000000 forever, no matter where physical address of '_text'
> > > is. So here LOAD_PHYSICAL_ADDR is default value of 'virt_addr'.
> > 
> > So, this isn't always true. The output address is the address assigned
> > by head_*.S for the extraction destination. We can't just change that,
> > since it's based on where to perform the extraction, etc. All the
> > logic in there is designed to make sure we're actually >=
> > LOAD_PHYSICAL_ADDR:
> > 
> >         cmpl    $LOAD_PHYSICAL_ADDR, %ebx
> >         jge     1f
> > #endif
> >         movl    $LOAD_PHYSICAL_ADDR, %ebx
> > 1:
> 
> Yes, here it's trying to do an adjustment. To align the original loading
> address if it's not aligned to CONFIG_PHYSICAL_ALIGN, and to make it be
> LOAD_PHYSICAL_ADDR if it's smaller than LOAD_PHYSICAL_ADDR. Other than
> those, nothing else be done. So let me conclude the possible original
> 'output' value:
> 1) default case
>   It's LOAD_PHYSICAL_ADDR Which is the default loading address.
> 2) kexec/kdump
>   The output will be at the top of the available physical ram.
> 3) Modified bootloader
>   Which could put kernel anywhere by modifying code of bootloader. This
>   is the case that we need do above alignment and adjustment to avoid
>   unpleasant kernel error.
> 
> Except of above 3 cases, physical address of kernel could be changed by
> physical address randomization.
> 4) physical address randomization
> 
> However, for virtual address, there are only two chances to decide:
> 1) arch/x86/kernel/vmlinux.lds.S
>   we just map .text at __START_KERNEL
> #define __PHYSICAL_START        ALIGN(CONFIG_PHYSICAL_START, \                                                                                   
>                                       CONFIG_PHYSICAL_ALIGN)
> 
> #define __START_KERNEL          (__START_KERNEL_map + __PHYSICAL_START)
> 2) handle_relocations :: boot/compressed/misc.c
> In handle_relocations(), we DO pass in the physical 'output', but that
> is used to calculate 'map' which is used to position those memory
> unit where the referenced address need be relocated. 
>   
> So later, in arch/x86/kernel/head_64.S, we will store the delta which is
> between 16M and the actual kernel loading address into phys_base. So as

Here I was wrong about phys_base. It has been changed in the latest code
in Linus's tree. Please check arch/x86/kernel/head64.c: 

void __head __startup_64(unsigned long physaddr)                                                                                                 
{
	...
	load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
	...

	/* Fixup phys_base */
        p = fixup_pointer(&phys_base, physaddr);
        *p += load_delta;
}
Here physaddr should be 'output' which is the place kernel decompressed
at.


> long as we can translate physical address and virtual address of kernel
> by below formula, everything is going very well. 
> 
>     y = x - __START_KERNEL_map + phys_base.
>    
> We really don't need to tie physical address and virtual address of
> kernel together in any cases. They can be randomized or not randomized
> completely separately. As long as we keep above translation formula
> working well, everything need not be worried.