[U-Boot] [PATCH 1/9] dm: allow 4GB of DRAM on 32bit systems

[Alexander Graf]

> Am 02.09.2018 um 18:04 schrieb Vagrant Cascadian <vagrant at debian.org>:
> 
>> On 2018-09-02, Alexander Graf <agraf at suse.de> wrote:
>>> On 02.08.18 23:31, Dr. Philipp Tomsich wrote:
>>> 
>>>> On 2 Aug 2018, at 22:36, Simon Glass <sjg at chromium.org> wrote:
>>>> 
>>>> On 26 July 2018 at 07:59, Philipp Tomsich
>>>> <philipp.tomsich at theobroma-systems.com <mailto:philipp.tomsich at theobroma-systems.com>> wrote:
>>>>> Even on 32bit systems a full 4GB of DRAM may be installed and reported
>>>>> by the DRAM controller.  Whether these 4GB are larger available
>>>>> depends on the size/configuration of address decoding windows and
>>>>> architectural features (e.g. consider a hypothetical architecture that
>>>>> uses dedicated instructions to access the 'memory-mapped' peripheral
>>>>> IO ranges).  In U-Boot, the available DRAM, as reported by the
>>>>> device-model is independent of the accessible DRAM (i.e. adjusted top
>>>>> and effective size).
>>>>> 
>>>>> This was first reported against the RK3288, which may have 4GB DRAM
>>>>> attached, but will have a small (e.g. 128MB) window not accessible due
>>>>> to address decoding limitations.
>>>>> 
>>>>> The solution is to increase the types used for storing the ram_size to
>>>>> have at least 33 bits even on 32bit systems... i.e. we need to use a
>>>>> u64 for these always (this was previously only the case for
>>>>> PHYS_64BIT, which will have unwanted side-effects for 32bit systems
>>>>> and would require more intrusive changes).
>>>>> 
>>>>> This commit changes the size-field in 'struct ram' and the ram_size in
>>>>> 'gd' to be a u64.
>>>>> 
>>>>> Reported-by: Marty E. Plummer <hanetzer at startmail.com>
>>>>> Signed-off-by: Philipp Tomsich <philipp.tomsich at theobroma-systems.com>
>>>>> ---
>>>>> 
>>>>> include/asm-generic/global_data.h | 2 +-
>>>>> include/ram.h                     | 9 ++++++++-
>>>>> 2 files changed, 9 insertions(+), 2 deletions(-)
>>>>> 
>>>> 
>>>> Reviewed-by: Simon Glass <sjg at chromium.org <mailto:sjg at chromium.org>>
>>>> 
>>>>> diff --git a/include/asm-generic/global_data.h b/include/asm-generic/global_data.h
>>>>> index 0fd4900..f3d687c 100644
>>>>> --- a/include/asm-generic/global_data.h
>>>>> +++ b/include/asm-generic/global_data.h
>>>>> @@ -55,7 +55,7 @@ typedef struct global_data {
>>>>>       unsigned long ram_base;         /* Base address of RAM used by U-Boot */
>>>>>       unsigned long ram_top;          /* Top address of RAM used by U-Boot */
>>>>>       unsigned long relocaddr;        /* Start address of U-Boot in RAM */
>>>>> -       phys_size_t ram_size;           /* RAM size */
>>>>> +       u64 ram_size;                   /* RAM size */
>>>>>       unsigned long mon_len;          /* monitor len */
>>>>>       unsigned long irq_sp;           /* irq stack pointer */
>>>>>       unsigned long start_addr_sp;    /* start_addr_stackpointer */
>>>>> diff --git a/include/ram.h b/include/ram.h
>>>>> index 67e22d7..1fe024f 100644
>>>>> --- a/include/ram.h
>>>>> +++ b/include/ram.h
>>>>> @@ -9,7 +9,14 @@
>>>>> 
>>>>> struct ram_info {
>>>>>       phys_addr_t base;
>>>>> -       size_t size;
>>>>> +       /*
>>>>> +        * We use a 64bit type for the size to correctly handle 32bit
>>>>> +        * systems with 4GB of DRAM (which would overflow a 32bit type
>>>>> +        * and read back as 0).  For 64bit systems, we assume (for now)
>>>> 
>>>> forever :-)
>>>> 
>>>>> +        * that they will always have less than 2^65 byte of DRAM
>>>>> +        * installed. -- prt
>>>> 
>>>> Is the prt your signature? I suggest dropping it.
>>> 
>>> In fact it is. But I’ll need to rewrite the entire comment anyway for the next
>>> version of this series as there’s even more functions and places that the
>>> memory size is stored in...
>>> 
>>>> 
>>>>> +        */
>>>>> +       u64 size;
>> 
>> With LPAE available in all modern ARM cores, shouldn't phys_addr_t just
>> be u64? And then we'd probably want to use that throughout the code, right?
> 
> Quite a few currently supported boards do not support LPAE, e.g. imx6.

What I'm trying to say is that we probably want to make phys_addr_t be u64 when CONFIG_LPAE is set.

Alex

> 
> 
> live well,
>  vagrant