[U-Boot] [PATCH] arm: mx6: Reduce SPL malloc pool size

[Tim Harvey]

On Wed, Nov 25, 2015 at 5:32 AM, Marek Vasut <marex at denx.de> wrote:
> On Wednesday, November 25, 2015 at 02:15:13 PM, Przemyslaw Marczak wrote:
>> On 11/25/2015 01:16 PM, Marek Vasut wrote:
>> > On Wednesday, November 25, 2015 at 01:00:41 PM, Przemyslaw Marczak wrote:
>> >> Hello Marek,
>> >>
>> >> On 11/25/2015 11:56 AM, Marek Vasut wrote:
>> >>> On Wednesday, November 25, 2015 at 11:40:36 AM, Przemyslaw Marczak wrote:
>> >>>> Hello Tim, Marek
>> >>>>
>> >>>> On 11/20/2015 10:40 PM, Tim Harvey wrote:
>> >>>>> On Fri, Nov 20, 2015 at 12:43 PM, Marek Vasut <marex at denx.de> wrote:
>> >>>>>> Using 50 MiB malloc pool in SPL is nonsense. Since the caches are
>> >>>>>> not enabled in SPL, it takes 2 seconds to init the pool and has no
>> >>>>>> obvious benefit. Reduce the size to 1 MiB.
>> >>>>>>
>> >>>>>> Signed-off-by: Marek Vasut <marex at denx.de>
>> >>>>>> Cc: Stefano Babic <sbabic at denx.de>
>> >>>>>> Cc: Tim Harvey <tharvey at gateworks.com>
>> >>>>>> ---
>> >>>>>>
>> >>>>>>     include/configs/imx6_spl.h | 6 +++---
>> >>>>>>     1 file changed, 3 insertions(+), 3 deletions(-)
>> >>>>>>
>> >>>>>> diff --git a/include/configs/imx6_spl.h b/include/configs/imx6_spl.h
>> >>>>>> index 1744f2c..43ce7fe 100644
>> >>>>>> --- a/include/configs/imx6_spl.h
>> >>>>>> +++ b/include/configs/imx6_spl.h
>> >>>>>> @@ -63,15 +63,15 @@
>> >>>>>>
>> >>>>>>     #if defined(CONFIG_MX6SX) || defined(CONFIG_MX6UL) ||
>> >>>>>>     defined(CONFIG_MX6SL) #define CONFIG_SPL_BSS_START_ADDR
>> >>>>>>     0x88200000
>> >>>>>>
>> >>>>>> -#define CONFIG_SPL_BSS_MAX_SIZE        0x100000        /* 1 MB */
>> >>>>>> +#define CONFIG_SPL_BSS_MAX_SIZE        0x100000                /* 1
>> >>>>>> MB */
>> >>>>>>
>> >>>>>>     #define CONFIG_SYS_SPL_MALLOC_START    0x88300000
>> >>>>>>
>> >>>>>> -#define CONFIG_SYS_SPL_MALLOC_SIZE     0x3200000       /* 50 MB */
>> >>>>>> +#define CONFIG_SYS_SPL_MALLOC_SIZE     0x100000                /* 1
>> >>>>>> MB */
>> >>>>>>
>> >>>>>>     #define CONFIG_SYS_TEXT_BASE           0x87800000
>> >>>>>>     #else
>> >>>>>>     #define CONFIG_SPL_BSS_START_ADDR      0x18200000
>> >>>>>>     #define CONFIG_SPL_BSS_MAX_SIZE                0x100000
>> >>>>>>     /* 1 MB */ #define CONFIG_SYS_SPL_MALLOC_START    0x18300000
>> >>>>>>
>> >>>>>> -#define CONFIG_SYS_SPL_MALLOC_SIZE     0x3200000       /* 50 MB */
>> >>>>>> +#define CONFIG_SYS_SPL_MALLOC_SIZE     0x100000        /* 1 MB */
>> >>>>>>
>> >>>>>>     #define CONFIG_SYS_TEXT_BASE           0x17800000
>> >>>>>>     #endif
>> >>>>>>     #endif
>> >>>>>>
>> >>>>>> --
>> >>>>>> 2.1.4
>> >>>>>
>> >>>>> Acked-by: Tim Harvey <tharvey at gateworks.com>
>> >>>>>
>> >>>>> thanks for dropping 2 secs off our time to boot!
>> >>>>>
>> >>>>> Tim
>> >>>>
>> >>>> The boot time for SPL and U-Boot can be reduced more if
>> >>>> CONFIG_SYS_MALLOC_CLEAR_ON_INIT is unset (default is set).
>> >>>
>> >>> This might confuse DM, so I don't want this. DM checks if it's
>> >>> structures were already inited and if we don't clear the malloc area,
>> >>> they would be.
>> >>
>> >> DM is safe - it uses calloc() for it's structures and also driver's
>> >> static structures are zeroed. Only some private data which is not
>> >> allocated by DM can be risky to use.
>> >>
>> >> To be precise - we don't use malloc() to get pointer for zeroed memory,
>> >> this is the job for the calloc().
>> >>
>> >> So any code with bad use of malloc() should be fixed ASAP.
>> >>
>> >> If you assume that clear the malloc pool at boot is safe to use, then
>> >> I'm asking you, what about the sequence malloc/free/malloc for
>> >> potentially the same area? You can't be sure that the second returned
>> >> pointer after calling free() - points to the initially zeroed pool.
>> >>
>> >> We don't clean the malloc pool for configs that we maintain and we
>> >> didn't noticed any issues related to this. And the boot time is reduced
>> >> significantly.
>> >
>> > Do you use DM in SPL ?
>>
>> No, we don't use the SPL - for the boards which we maintain.
>
> OK
>
>> > I just checked and I see it's probably only the GD which needs to be
>> > cleared out, so we should indeed be safe.  In which case, patch is
>> > welcome to enable CONFIG_SYS_MALLOC_CLEAR_ON_INIT .
>>
>> Yes I can see also for the Exynos SPL, that GD is manually cleared.
>>
>> But disabling the default CONFIG_SYS_MALLOC_CLEAR_ON_INIT, will help you
>> in reducing the U-Boot init time.
>> Just try and if you can easy measure the real difference for
>> enable/disable this config - share the results - it can give us an
>> interesting conclusions.
>
> Sure, thanks for the hint, but I still don't see any reason for having 50 MiB
> malloc area in SPL ;-)

Perhaps the use case for a >1MB malloc area is falcon mode. While a
1MB area certainly works for loading and bootstrapping to u-boot.img I
have used anywhere from 1MB to 8MB images for falcon mode.

Tim