[U-Boot] am335x: GPMC: reading speed with prefetch mode

[Yegor Yefremov]

On Fri, Mar 20, 2015 at 3:54 PM, Daniel Mack <zonque at gmail.com> wrote:
> Hi,
>
> On 03/20/2015 12:24 PM, Yegor Yefremov wrote:
>> I've also put printf() into omap_nand_read_prefetch8() just to make
>> sure it is called - it was called.
>
> Does it fall back to polled mode because the engine is busy maybe? See
> the comment in the code that deals with the return value of
> __read_prefetch_aligned(). Not that I'd be aware of such an issue, but
> maybe you need to enable to enable some module or clock?

I've seen this fallback comment and added printf to ret < 0 case and
in __read_prefetch_aligned(). I had no cases, where fallback was
invoked.

I've also a suspicion, that some clocks settings are missing. Will
have to compare register values with Linux ones.

>> Further ideas?
>
> It's been a while, but when I worked on cutting down the boot times of a
> custom AM335x board months ago, these were the steps that I took:
>
>
> * I tweaked the NAND timing values in the DTS, and then looked at the
> GPMC timing values Linux calculated from that and copied it over to
> U-Boot. This gave a NAND read speed-up of ~20-30% in comparison to the
> defaults U-Boot ships with.
>
> * Implementing the GPMC prefetch operations boosted the speed by of raw
> NAND reads by approximately factor 2.
>
> * The NAND bad block table scan was reduced to the first 64 blocks,
> because this is all I cared for from U-Boot anyway, and Linux will do
> the same anyway at a later point. This brought down the scan process
> from ~2s to some milliseconds.
>
> * In the SPL code, the CPU was put in GHz mode as early as possible.
> That improved load times again, and also reduced the time spent in the
> kernel decompressor.

Do you mean am33xx_spl_board_init()? In theory it reads ID and sets
desired MPU/Core voltage. But I'll look at this once again.

>> I Linux I had ti,nand-xfer-type = "polled";. After replacing it with
>> ti,nand-xfer-type = "prefetch-polled"; I now get
>>
>> # dd if=/dev/mtdblock5 of=/dev/null bs=2M count=8
>> 8+0 records in
>> 8+0 records out
>> 16777216 bytes (17 MB) copied, 2.58744 s, 6.5 MB/s
>>
>> instead of:
>>
>> # dd if=/dev/mtdblock5 of=/dev/null bs=2M count=8
>> 8+0 records in
>> 8+0 records out
>> 16777216 bytes (17 MB) copied, 6.05157 s, 2.8 MB/s
>>
>> Do I see it right, that DMA support is not implemented in am33xx.dtsi?
>
> Right, I just double-checked with the DTS we were using. It had no
> specific "ti,nand-xfer-type" setting, so it should default to
> "prefetch-polled". So it seems DMA is not in place for this board and
> the gpmc-nand driver. Sorry for the confusion.
>
>
> Daniel