[PATCH] arm64: fix CONFIG_ZONE_DMA on systems with no 32-bit addressable DRAM

[PATCH] arm64: fix CONFIG_ZONE_DMA on systems with no 32-bit addressable DRAM

[Mark Salter]

Commit 2d5a5612bc (arm64: Limit the CMA buffer to 32-bit if ZONE_DMA)
forces the CMA buffer to be 32-bit addressable if CONFIG_ZONE_DMA is
defined. This breaks CMA on platforms with no 32-bit addressable DRAM.
This patch checks to make sure there is 32-bit addressable DRAM before
setting the 32-bit limit. If there is none, no limit is placed on the
CMA buffer. This allows a single kernel (with CONFIG_ZONE_DMA defined)
to support platforms requiring the 32-bit limit and platforms with no
32-bit limit.

Signed-off-by: Mark Salter <msalter@redhat.com>
---
 arch/arm64/mm/init.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
index c5415e2..2925576 100644
--- a/arch/arm64/mm/init.c
+++ b/arch/arm64/mm/init.c
@@ -149,8 +149,17 @@ void __init arm64_memblock_init(void)
 	early_init_fdt_scan_reserved_mem();
 
 	/* 4GB maximum for 32-bit only capable devices */
-	if (IS_ENABLED(CONFIG_ZONE_DMA))
+	if (IS_ENABLED(CONFIG_ZONE_DMA)) {
 		dma_phys_limit = dma_to_phys(NULL, DMA_BIT_MASK(32)) + 1;
+		/*
+		 * If platform doesn't have DRAM within the dma_phys_limit,
+		 * remove the limit altogether. This allows one kernel (with
+		 * CONFIG_ZONE_DMA defined) to support platforms with 32-bit
+		 * only devices and platforms with no 32-bit DRAM.
+		 */
+		if (dma_phys_limit <= memblock_start_of_DRAM())
+			dma_phys_limit = 0;
+	}
 	dma_contiguous_reserve(dma_phys_limit);
 
 	memblock_allow_resize();
-- 
1.8.3.1

[PATCH] arm64: fix CONFIG_ZONE_DMA on systems with no 32-bit addressable DRAM

[David Rientjes]

On Mon, 23 Jun 2014, Mark Salter wrote:

> Commit 2d5a5612bc (arm64: Limit the CMA buffer to 32-bit if ZONE_DMA)
> forces the CMA buffer to be 32-bit addressable if CONFIG_ZONE_DMA is
> defined. This breaks CMA on platforms with no 32-bit addressable DRAM.
> This patch checks to make sure there is 32-bit addressable DRAM before
> setting the 32-bit limit. If there is none, no limit is placed on the
> CMA buffer. This allows a single kernel (with CONFIG_ZONE_DMA defined)
> to support platforms requiring the 32-bit limit and platforms with no
> 32-bit limit.
> 
> Signed-off-by: Mark Salter <msalter@redhat.com>
> ---
>  arch/arm64/mm/init.c | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> index c5415e2..2925576 100644
> --- a/arch/arm64/mm/init.c
> +++ b/arch/arm64/mm/init.c
> @@ -149,8 +149,17 @@ void __init arm64_memblock_init(void)
>  	early_init_fdt_scan_reserved_mem();
>  
>  	/* 4GB maximum for 32-bit only capable devices */
> -	if (IS_ENABLED(CONFIG_ZONE_DMA))
> +	if (IS_ENABLED(CONFIG_ZONE_DMA)) {
>  		dma_phys_limit = dma_to_phys(NULL, DMA_BIT_MASK(32)) + 1;
> +		/*
> +		 * If platform doesn't have DRAM within the dma_phys_limit,
> +		 * remove the limit altogether. This allows one kernel (with
> +		 * CONFIG_ZONE_DMA defined) to support platforms with 32-bit
> +		 * only devices and platforms with no 32-bit DRAM.
> +		 */
> +		if (dma_phys_limit <= memblock_start_of_DRAM())
> +			dma_phys_limit = 0;
> +	}
>  	dma_contiguous_reserve(dma_phys_limit);
>  
>  	memblock_allow_resize();

It is a shame that we have very similar logic, all dependent on 
memblock_start_of_DRAM(), in both zone_sizes_init() and 
arm64_memblock_init() and the latter having to be called before the 
former.  I assume there's no way to make this cleaner in the bootstrap 
code so the dma_to_phys(NULL, DMA_BIT_MASK(32)) + 1 limit can be handled 
properly in a single location?

[PATCH] arm64: fix CONFIG_ZONE_DMA on systems with no 32-bit addressable DRAM

[Catalin Marinas]

On Wed, Jun 25, 2014 at 01:26:02AM +0100, David Rientjes wrote:
> On Mon, 23 Jun 2014, Mark Salter wrote:
> 
> > Commit 2d5a5612bc (arm64: Limit the CMA buffer to 32-bit if ZONE_DMA)
> > forces the CMA buffer to be 32-bit addressable if CONFIG_ZONE_DMA is
> > defined. This breaks CMA on platforms with no 32-bit addressable DRAM.
> > This patch checks to make sure there is 32-bit addressable DRAM before
> > setting the 32-bit limit. If there is none, no limit is placed on the
> > CMA buffer. This allows a single kernel (with CONFIG_ZONE_DMA defined)
> > to support platforms requiring the 32-bit limit and platforms with no
> > 32-bit limit.
> > 
> > Signed-off-by: Mark Salter <msalter@redhat.com>
> > ---
> >  arch/arm64/mm/init.c | 11 ++++++++++-
> >  1 file changed, 10 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > index c5415e2..2925576 100644
> > --- a/arch/arm64/mm/init.c
> > +++ b/arch/arm64/mm/init.c
> > @@ -149,8 +149,17 @@ void __init arm64_memblock_init(void)
> >  	early_init_fdt_scan_reserved_mem();
> >  
> >  	/* 4GB maximum for 32-bit only capable devices */
> > -	if (IS_ENABLED(CONFIG_ZONE_DMA))
> > +	if (IS_ENABLED(CONFIG_ZONE_DMA)) {
> >  		dma_phys_limit = dma_to_phys(NULL, DMA_BIT_MASK(32)) + 1;
> > +		/*
> > +		 * If platform doesn't have DRAM within the dma_phys_limit,
> > +		 * remove the limit altogether. This allows one kernel (with
> > +		 * CONFIG_ZONE_DMA defined) to support platforms with 32-bit
> > +		 * only devices and platforms with no 32-bit DRAM.
> > +		 */
> > +		if (dma_phys_limit <= memblock_start_of_DRAM())
> > +			dma_phys_limit = 0;
> > +	}
> >  	dma_contiguous_reserve(dma_phys_limit);
> >  
> >  	memblock_allow_resize();
> 
> It is a shame that we have very similar logic, all dependent on 
> memblock_start_of_DRAM(), in both zone_sizes_init() and 
> arm64_memblock_init() and the latter having to be called before the 
> former.  I assume there's no way to make this cleaner in the bootstrap 
> code so the dma_to_phys(NULL, DMA_BIT_MASK(32)) + 1 limit can be handled 
> properly in a single location?

Probably and I suggested something along the lines of only calculating
it once in reply to the "[PATCH] arm64: make CONFIG_ZONE_DMA user
settable" message. I'll follow up on the other thread (Mark considers
there's a penalty issue with all memory in the DMA zone).

-- 
Catalin