Re: [PATCH, RFC] x86/boot/compressed/64: Handle 5-level paging boot if kernel is above 4G

[Eric W. Biederman]

"Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> writes:

> On Mon, Oct 09, 2017 at 09:54:53AM -0700, Dave Hansen wrote:
>> On 10/09/2017 09:09 AM, Kirill A. Shutemov wrote:
>> > Apart from trampoline itself we also need place to store top level page
>> > table in lower memory as we don't have a way to load 64-bit value into
>> > CR3 from 32-bit mode. We only really need 8-bytes there as we only use
>> > the very first entry of the page table.
>> 
>> Oh, and this is why you have to move "lvl5_pgtable" out of the kernel image?
>
> Right. I initialize the new location of top level page table directly.

So just a quick note.  I have a fuzzy memory of people loading their
kernels above 4G physical because they did not have any memory below
4G.

That might be a very specialized case if my memory is correct because
cpu startup has to have a trampoline below 1MB.  So I don't know how
that works.  But I do seem to remember someone mentioning it.

Is there really no way to switch to 5 level paging other than to drop to
32bit mode and disable paging?    The x86 architecture does some very
bizarre things so I can believe it but that seems like a lot of work to
get somewhere.

Eric


>
>> > diff --git a/arch/x86/boot/compressed/head_64.S b/arch/x86/boot/compressed/head_64.S
>> > index cefe4958fda9..049a289342bd 100644
>> > --- a/arch/x86/boot/compressed/head_64.S
>> > +++ b/arch/x86/boot/compressed/head_64.S
>> > @@ -288,6 +288,22 @@ ENTRY(startup_64)
>> >  	leaq	boot_stack_end(%rbx), %rsp
>> >  
>> >  #ifdef CONFIG_X86_5LEVEL
>> > +/*
>> > + * We need trampoline in lower memory switch from 4- to 5-level paging for
>> > + * cases when bootloader put kernel above 4G, but didn't enable 5-level paging
>> > + * for us.
>> > + *
>> > + * Here we use MBR memory to store trampoline code.
>> > + *
>> > + * We also have to have top page table in lower memory as we don't have a way
>> > + * to load 64-bit value into CR3 from 32-bit mode. We only need 8-bytes there
>> > + * as we only use the very first entry of the page table.
>> > + *
>> > + * Here we use 0x7000 as top-level page table.
>> > + */
>> > +#define LVL5_TRAMPOLINE	0x7c00
>> > +#define LVL5_PGTABLE	0x7000
>> > +
>> >  	/* Preserve RBX across CPUID */
>> >  	movq	%rbx, %r8
>> >  
>> > @@ -323,29 +339,37 @@ ENTRY(startup_64)
>> >  	 * long mode would trigger #GP. So we need to switch off long mode
>> >  	 * first.
>> >  	 *
>> > -	 * NOTE: This is not going to work if bootloader put us above 4G
>> > -	 * limit.
>> > +	 * We use trampoline in lower memory to handle situation when
>> > +	 * bootloader put the kernel image above 4G.
>> >  	 *
>> >  	 * The first step is go into compatibility mode.
>> >  	 */
>> >  
>> > -	/* Clear additional page table */
>> > -	leaq	lvl5_pgtable(%rbx), %rdi
>> > -	xorq	%rax, %rax
>> > -	movq	$(PAGE_SIZE/8), %rcx
>> > -	rep	stosq
>> > +	/* Copy trampoline code in place */
>> > +	movq	%rsi, %r9
>> > +	leaq	lvl5_trampoline(%rip), %rsi
>> > +	movq	$LVL5_TRAMPOLINE, %rdi
>> > +	movq	$(lvl5_trampoline_end - lvl5_trampoline), %rcx
>> > +	rep	movsb
>> > +	movq	%r9, %rsi
>> 
>> This needs to get more heavily commented, like the use of r9 to stash
>> %rsi.  Why do you do that, btw?  I don't see it getting reused at first
>> glance.
>
> %rsi holds pointer to real_mode_data. It need to be preserved.
>
> I'll add more comments.
>
>> I think it will also be really nice to differentate "lvl5_trampoline"
>> from "LVL5_TRAMPOLINE".  Maybe add "src" and "dst" to them or something.
>
> Makes sense. Thanks.
>
>> >  	/*
>> > -	 * Setup current CR3 as the first and only entry in a new top level
>> > +	 * Setup current CR3 as the first and the only entry in a new top level
>> >  	 * page table.
>> >  	 */
>> >  	movq	%cr3, %rdi
>> >  	leaq	0x7 (%rdi), %rax
>> > -	movq	%rax, lvl5_pgtable(%rbx)
>> > +	movq	%rax, LVL5_PGTABLE
>> > +
>> > +	/*
>> > +	 * Load address of lvl5 into RDI.
>> > +	 * It will be used to return address from trampoline.
>> > +	 */
>> > +	leaq	lvl5(%rip), %rdi
>> 
>> Is there a reason to do a 'lea' here instead of just shoving the address
>> in directly?  Is this a shorter instruction or something?
>
> This code can be loaded anywhere in memory and we need to calculate
> absolute address of the label here.
> AFAIK, "lea <label>(%rip), <register>" is idiomatic way to do this.
>
>> >  	/* Switch to compatibility mode (CS.L = 0 CS.D = 1) via far return */
>> >  	pushq	$__KERNEL32_CS
>> > -	leaq	compatible_mode(%rip), %rax
>> > +	movq	$LVL5_TRAMPOLINE, %rax
>> >  	pushq	%rax
>> >  	lretq
>> >  lvl5:
>> > @@ -488,9 +512,9 @@ relocated:
>> >   */
>> >  	jmp	*%rax
>> >  
>> > -	.code32
>> >  #ifdef CONFIG_X86_5LEVEL
>> > -compatible_mode:
>> > +	.code32
>> > +lvl5_trampoline:
>> >  	/* Setup data and stack segments */
>> >  	movl	$__KERNEL_DS, %eax
>> >  	movl	%eax, %ds
>> > @@ -502,7 +526,7 @@ compatible_mode:
>> >  	movl	%eax, %cr0
>> >  
>> >  	/* Point CR3 to 5-level paging */
>> > -	leal	lvl5_pgtable(%ebx), %eax
>> > +	movl	$LVL5_PGTABLE, %eax
>> >  	movl	%eax, %cr3
>> >  
>> >  	/* Enable PAE and LA57 mode */
>> > @@ -510,14 +534,9 @@ compatible_mode:
>> >  	orl	$(X86_CR4_PAE | X86_CR4_LA57), %eax
>> >  	movl	%eax, %cr4
>> >  
>> > -	/* Calculate address we are running at */
>> > -	call	1f
>> > -1:	popl	%edi
>> > -	subl	$1b, %edi
>> > -
>> >  	/* Prepare stack for far return to Long Mode */
>> >  	pushl	$__KERNEL_CS
>> > -	leal	lvl5(%edi), %eax
>> > +	movl	$(lvl5_enabled - lvl5_trampoline + LVL5_TRAMPOLINE), %eax
>> 
>> This loads the trampoline address of "lvl5_enabled", right?  That'd be
>> handy to spell out explicitly.
>
> Yep, will do.