Re: [PATCH kvm-unit-tests 1/8] arm/arm64: Reorganize cstart assembler

[Andrew Jones <drjones@redhat.com>]

On Wed, Apr 14, 2021 at 04:15:10PM +0100, Alexandru Elisei wrote:
> Hi Drew,
> 
> On 4/14/21 9:59 AM, Andrew Jones wrote:
> > On Tue, Apr 13, 2021 at 05:34:24PM +0100, Alexandru Elisei wrote:
> >> Hi Drew,
> >>
> >> On 4/7/21 7:59 PM, Andrew Jones wrote:
> >>> Move secondary_entry helper functions out of .init and into .text,
> >>> since secondary_entry isn't run at "init" time.
> >> The tests aren't loaded using the loader, so as far as I can tell the reason for
> >> having an .init section is to make sure the code from the start label is put at
> >> offset 0 in the test binary. As long as the start label is kept at the beginning
> >> of the .init section, and the loader script places the section first, I don't see
> >> any issues with this change.
> >>
> >> The only hypothetical problem that I can think of is that the code from .init
> >> calls code from .text, and if the text section grows very large we might end up
> >> with a PC offset larger than what can be encoded in the BL instruction. That's
> >> unlikely to happen (the offset is 16MB for arm and 64MB for arm64), and the .init
> >> code already calls other functions (like setup) which are in .text, so we would
> >> have this problem regardless of this change. And the compiler will emit an error
> >> if that happens.
> >>
> >>> Signed-off-by: Andrew Jones <drjones@redhat.com>
> >>> ---
> >>>  arm/cstart.S   | 62 +++++++++++++++++++++++++++-----------------------
> >>>  arm/cstart64.S | 22 +++++++++++-------
> >>>  2 files changed, 48 insertions(+), 36 deletions(-)
> >>>
> >>> diff --git a/arm/cstart.S b/arm/cstart.S
> >>> index d88a98362940..653ab1e8a141 100644
> >>> --- a/arm/cstart.S
> >>> +++ b/arm/cstart.S
> >>> @@ -96,32 +96,7 @@ start:
> >>>  	bl	exit
> >>>  	b	halt
> >>>  
> >>> -
> >>> -.macro set_mode_stack mode, stack
> >>> -	add	\stack, #S_FRAME_SIZE
> >>> -	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
> >>> -	isb
> >>> -	mov	sp, \stack
> >>> -.endm
> >>> -
> >>> -exceptions_init:
> >>> -	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
> >>> -	bic	r2, #CR_V		@ SCTLR.V := 0
> >>> -	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
> >>> -	ldr	r2, =vector_table
> >>> -	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
> >>> -
> >>> -	mrs	r2, cpsr
> >>> -
> >>> -	/* first frame reserved for svc mode */
> >>> -	set_mode_stack	UND_MODE, r0
> >>> -	set_mode_stack	ABT_MODE, r0
> >>> -	set_mode_stack	IRQ_MODE, r0
> >>> -	set_mode_stack	FIQ_MODE, r0
> >>> -
> >>> -	msr	cpsr_cxsf, r2		@ back to svc mode
> >>> -	isb
> >>> -	mov	pc, lr
> >>> +.text
> >> Hm... now we've moved enable_vfp from .init to .text, and enable_vfp *is* called
> >> from .init code, which doesn't fully match up with the commit message. Is the
> >> actual reason for this change that the linker script for EFI will discard the
> >> .init section? Maybe it's worth mentioning that in the commit message, because it
> >> will explain this change better.
> > Right, the .init section may not exist when linking with other linker
> > scripts. I'll make the commit message more clear.
> >
> >> Or is it to align arm with arm64, where only
> >> start is in the .init section?
> >>
> >>>  
> >>>  enable_vfp:
> >>>  	/* Enable full access to CP10 and CP11: */
> >>> @@ -133,8 +108,6 @@ enable_vfp:
> >>>  	vmsr	fpexc, r0
> >>>  	mov	pc, lr
> >>>  
> >>> -.text
> >>> -
> >>>  .global get_mmu_off
> >>>  get_mmu_off:
> >>>  	ldr	r0, =auxinfo
> >>> @@ -235,6 +208,39 @@ asm_mmu_disable:
> >>>  
> >>>  	mov     pc, lr
> >>>  
> >>> +/*
> >>> + * Vectors
> >>> + */
> >>> +
> >>> +.macro set_mode_stack mode, stack
> >>> +	add	\stack, #S_FRAME_SIZE
> >>> +	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
> >>> +	isb
> >>> +	mov	sp, \stack
> >>> +.endm
> >>> +
> >>> +exceptions_init:
> >>> +	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
> >>> +	bic	r2, #CR_V		@ SCTLR.V := 0
> >>> +	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
> >>> +	ldr	r2, =vector_table
> >>> +	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
> >>> +
> >>> +	mrs	r2, cpsr
> >>> +
> >>> +	/*
> >>> +	 * Input r0 is the stack top, which is the exception stacks base
> >>> +	 * The first frame is reserved for svc mode
> >>> +	 */
> >>> +	set_mode_stack	UND_MODE, r0
> >>> +	set_mode_stack	ABT_MODE, r0
> >>> +	set_mode_stack	IRQ_MODE, r0
> >>> +	set_mode_stack	FIQ_MODE, r0
> >>> +
> >>> +	msr	cpsr_cxsf, r2		@ back to svc mode
> >>> +	isb
> >>> +	mov	pc, lr
> >>> +
> >>>  /*
> >>>   * Vector stubs
> >>>   * Simplified version of the Linux kernel implementation
> >>> diff --git a/arm/cstart64.S b/arm/cstart64.S
> >>> index 0a85338bcdae..d39cf4dfb99c 100644
> >>> --- a/arm/cstart64.S
> >>> +++ b/arm/cstart64.S
> >>> @@ -89,10 +89,12 @@ start:
> >>>  	msr	cpacr_el1, x4
> >>>  
> >>>  	/* set up exception handling */
> >>> +	mov	x4, x0				// x0 is the addr of the dtb
> >> I suppose changing exceptions_init to use x0 as a scratch register instead of x4
> >> makes some sense if you look at it from the perspective of it being called from
> >> secondary_entry, where all the functions use x0 as a scratch register. But it's
> >> still called from start, where using x4 as a scratch register is preferred because
> >> of the kernel boot protocol (x0-x3 are reserved).
> >>
> >> Is there an actual bug that this is supposed to fix (I looked for it and couldn't
> >> figure it out) or is it just a cosmetic change?
> > Now that exceptions_init isn't a private function of start (actually it
> > hasn't been in a long time, considering secondary_entry calls it) I would
> > like it to better conform to calling conventions. I guess I should have
> > used x19 here instead of x4 to be 100% correct. Or, would you rather I
> > just continue using x4 in exceptions_init in order to avoid the
> > save/restore?
> 
> To be honest, for this patch, I think it would be best to leave exceptions_init
> unchanged:
> 
> - We switch to using x0 like the rest of the code from secondary_entry, but
> because of that we need to save and restore the DTB address from x0 in start, so I
> don't think we've gained anything.
> - It makes the diff larger.
> - It runs the risk of introducing regressions (like all changes).
> 
> Maybe this can be left for a separate patch that changes code called from C to
> follow aapcs64.
>

OK, I'll switch it back to x4 and add a comment.

Thanks,
drew