[1/2] bus: mbus: fix window size calculation for 4GB windows

At least the Armada XP SoC supports 4GB on a single DRAM window. Because
the size register values contain the actual size - 1, the MSB is set in
that case. For example, the SDRAM window's control register's value is
0xffffffe1 for 4GB (bits 31 to 24 contain the size).

The MBUS driver reads back each window's size from registers and
calculates the actual size as (control_reg | ~DDR_SIZE_MASK) + 1, which
overflows for 32 bit values, resulting in other miscalculations further
on (a bad RAM window for the CESA crypto engine calculated by
mvebu_mbus_setup_cpu_target_nooverlap() in my case).

This patch changes the type in 'struct mbus_dram_window' from u32 to
u64, which allows us to keep using the same register calculation code in
most MBUS-using drivers (which calculate ->size - 1 again).

Signed-off-by: Jan Luebbe <jlu@pengutronix.de>
---
 drivers/bus/mvebu-mbus.c | 2 +-
 include/linux/mbus.h     | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

Hi Jan,
 
 On lun., août 28 2017, Jan Luebbe <jlu@pengutronix.de> wrote:

> At least the Armada XP SoC supports 4GB on a single DRAM window. Because
> the size register values contain the actual size - 1, the MSB is set in
> that case. For example, the SDRAM window's control register's value is
> 0xffffffe1 for 4GB (bits 31 to 24 contain the size).
>
> The MBUS driver reads back each window's size from registers and
> calculates the actual size as (control_reg | ~DDR_SIZE_MASK) + 1, which
> overflows for 32 bit values, resulting in other miscalculations further
> on (a bad RAM window for the CESA crypto engine calculated by
> mvebu_mbus_setup_cpu_target_nooverlap() in my case).
>
> This patch changes the type in 'struct mbus_dram_window' from u32 to
> u64, which allows us to keep using the same register calculation code in
> most MBUS-using drivers (which calculate ->size - 1 again).
>

Your patch looks good, but as it is a fix we should also apply it on
stable, could you provide the commit to fix?

Thanks,

Gregory

> Signed-off-by: Jan Luebbe <jlu@pengutronix.de>
> ---
>  drivers/bus/mvebu-mbus.c | 2 +-
>  include/linux/mbus.h     | 4 ++--
>  2 files changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/bus/mvebu-mbus.c b/drivers/bus/mvebu-mbus.c
> index c7f396903184..70db4d5638a6 100644
> --- a/drivers/bus/mvebu-mbus.c
> +++ b/drivers/bus/mvebu-mbus.c
> @@ -720,7 +720,7 @@ mvebu_mbus_default_setup_cpu_target(struct mvebu_mbus_state *mbus)
>  			if (mbus->hw_io_coherency)
>  				w->mbus_attr |= ATTR_HW_COHERENCY;
>  			w->base = base & DDR_BASE_CS_LOW_MASK;
> -			w->size = (size | ~DDR_SIZE_MASK) + 1;
> +			w->size = (u64)(size | ~DDR_SIZE_MASK) + 1;
>  		}
>  	}
>  	mvebu_mbus_dram_info.num_cs = cs;
> diff --git a/include/linux/mbus.h b/include/linux/mbus.h
> index 0d3f14fd2621..4773145246ed 100644
> --- a/include/linux/mbus.h
> +++ b/include/linux/mbus.h
> @@ -31,8 +31,8 @@ struct mbus_dram_target_info
>  	struct mbus_dram_window {
>  		u8	cs_index;
>  		u8	mbus_attr;
> -		u32	base;
> -		u32	size;
> +		u64	base;
> +		u64	size;
>  	} cs[4];
>  };
>  
> -- 
> 2.11.0
>

On Tue, Sep 19, 2017 at 04:43:18PM +0200, Gregory CLEMENT wrote:
> Hi Jan,
>  
>  On lun., août 28 2017, Jan Luebbe <jlu@pengutronix.de> wrote:
> 
> > At least the Armada XP SoC supports 4GB on a single DRAM window. Because
> > the size register values contain the actual size - 1, the MSB is set in
> > that case. For example, the SDRAM window's control register's value is
> > 0xffffffe1 for 4GB (bits 31 to 24 contain the size).
> >
> > The MBUS driver reads back each window's size from registers and
> > calculates the actual size as (control_reg | ~DDR_SIZE_MASK) + 1, which
> > overflows for 32 bit values, resulting in other miscalculations further
> > on (a bad RAM window for the CESA crypto engine calculated by
> > mvebu_mbus_setup_cpu_target_nooverlap() in my case).
> >
> > This patch changes the type in 'struct mbus_dram_window' from u32 to
> > u64, which allows us to keep using the same register calculation code in
> > most MBUS-using drivers (which calculate ->size - 1 again).
> >
> 
> Your patch looks good, but as it is a fix we should also apply it on
> stable, could you provide the commit to fix?

It was there just from the start: the .c file was introduced in
v3.10-rc1~64^2~1^2~8^2~2 and already did that 32 bit calculus.

Best regards
Uwe