svn commit: r298477 - in head/sys: conf riscv/riscv

[Marko Zec]

On Fri, 22 Apr 2016 16:15:59 +0000
Ruslan Bukin <br at freebsd.org> wrote:

> Author: br
> Date: Fri Apr 22 16:15:58 2016
> New Revision: 298477
> URL: https://svnweb.freebsd.org/changeset/base/298477
> 
> Log:
>   Clear the DDR memory. This should be done by bootloaders,
>   but they have no such feature yet.
>   
>   This fixes operation on Rocket Core and lowRISC.
> 
> Modified:
>   head/sys/conf/options.riscv
>   head/sys/riscv/riscv/locore.S
> 
> Modified: head/sys/conf/options.riscv
> ==============================================================================
> --- head/sys/conf/options.riscv	Fri Apr 22 15:12:05
> 2016	(r298476) +++ head/sys/conf/options.riscv	Fri Apr
> 22 16:15:58 2016	(r298477) @@ -2,3 +2,4 @@
>  
>  RISCV				opt_global.h
>  VFP				opt_global.h
> +DDR_CLEAR_SIZE			opt_global.h
> 
> Modified: head/sys/riscv/riscv/locore.S
> ==============================================================================
> --- head/sys/riscv/riscv/locore.S	Fri Apr 22 15:12:05
> 2016	(r298476) +++ head/sys/riscv/riscv/locore.S	Fri
> Apr 22 16:15:58 2016	(r298477) @@ -126,6 +126,17 @@ _start:
>  	csrr	a0, mhartid
>  	bnez	a0, mpentry
>  
> +#if defined(DDR_CLEAR_SIZE)
> +	/* Clear DDR memory */
> +	la	t0, _end
> +	li	t1, DDR_CLEAR_SIZE
> +1:
> +	sd	zero, 0(t0)
> +	addi	t0, t0, 8
> +	bltu	t0, t1, 1b

Usually _end appears to be (double) word aligned, but that happens out
of pure luck, since compilers do not provide firm guarantees about
_end's alignment (at least gcc doesn't, as far as I know).

So depending on which kind of RV this code would run, sometimes it may
end up doing a lot of traps emulating unaligned access in software, or
worse...

Hence it would't hurt to check for _end's alignment first and do a few
byte-by-byte bzero()ing if needed, before proceeding a dword pace.


> +	/* End */
> +#endif
> +
>  	/* Build event queue for current core */
>  	build_ring
>  
>