[U-Boot] [PATCH] ARM: socfpga: Configure PL310 latencies

Marek Vasut marex at denx.de
Fri Mar 1 09:40:32 UTC 2019


On 3/1/19 12:59 AM, Dinh Nguyen wrote:
> Hi Marek,
> 
> On 2/19/19 4:01 AM, Simon Goldschmidt wrote:
>> On Tue, Feb 19, 2019 at 1:44 AM Marek Vasut <marex at denx.de> wrote:
>>>
>>> Configure the PL310 tag and data latency registers, which slightly
>>> improves performance and aligns the behavior with Linux.
>>>
>>> Signed-off-by: Marek Vasut <marex at denx.de>
>>> Cc: Dalon Westergreen <dwesterg at gmail.com>
>>> Cc: Dinh Nguyen <dinguyen at kernel.org>
>>> ---
>>>  arch/arm/mach-socfpga/misc.c | 3 +++
>>>  1 file changed, 3 insertions(+)
>>>
>>> diff --git a/arch/arm/mach-socfpga/misc.c b/arch/arm/mach-socfpga/misc.c
>>> index 78fbe28724..1ea4e32c11 100644
>>> --- a/arch/arm/mach-socfpga/misc.c
>>> +++ b/arch/arm/mach-socfpga/misc.c
>>> @@ -62,6 +62,9 @@ void v7_outer_cache_enable(void)
>>>         /* Disable the L2 cache */
>>>         clrbits_le32(&pl310->pl310_ctrl, L2X0_CTRL_EN);
>>>
>>> +       writel(0x111, &pl310->pl310_tag_latency_ctrl);
>>> +       writel(0x121, &pl310->pl310_data_latency_ctrl);
>>
>> Would it make sense to add defines as named constants for this?
>> OTOH, in Linux, the values in the devicetree aren't really described,
>> either, so:
> 
> I was thinking the same, so I'm working on a patch to get these values
> from the device tree.
> 
> So while I was doing that, I realized that this patch is wrong.
> 
> The patch should be like this:
> 
> 	writel(0x0, &pl310->pl310_tag_latency_ctrl);
> 	writel(0x010, &pl310->pl310_data_latency_ctrl);
> 
> The reason is for the latency values:
> 
> 000 = 1 cycle of latency, there is no additional latency.
> 001 = 2 cycles of latency.
> 010 = 3 cycles of latency.
> 011 = 4 cycles of latency.
> 100 = 5 cycles of latency.
> 101 = 6 cycles of latency.
> 110 = 7 cycles of latency.
> 111 = 8 cycles of latency.
> 
> So from the values in the device tree, they should be n-1.
> 
> It looks like you've already sent the patch to Tom. I'll send a follow
> up patch to fix that when it lands.

Drat, thanks.

Better yet, pull the latency config into a function, so it can be used
by other platforms. The prototype should take 7 parameters, address and
latency in cycles, so that it shields the users from this n-1 stuff.

-- 
Best regards,
Marek Vasut


More information about the U-Boot mailing list