[U-Boot] [PATCH RFC] NAND: Improve read performance from Large Page NAND devices

[Scott Wood]

Nick Thompson wrote:
> On 01/12/09 10:13, Nick Thompson wrote:
>> On 01/12/09 00:55, Scott Wood wrote:
> 
>>> This change will break drivers that support large page and use the 
>>> default read_page functions, but do not implement cmd_ctrl (they replace 
>>> cmdfunc instead).  This includes fsl_elbc_nand, mxc_nand, and 
>>> mpc5121_nfc.  While I'd like to move them to implementing their own 
>>> read_page-type functions instead of cmdfunc, is there any way to make it 
>>> a smoother transition?
>> Yes, as it stands they would need modifying simultaneously and I have no
>> way to test such a change myself. The only required change in cmdfunc is
>> not to wait after a read0 request. You maybe in a better position to decide
>> if this has wider repercussions, but I will take a look at the above
>> drivers as well. [This is the main reason I made this an RFC].
> 
> How about, if nand_wait_cache_load was replaceable (by a no-op in your case)

Or it could be off by default, and enabled only on those platforms where 
it works and is beneficial.

> and the pre-fetch optimisation could be disabled by setting rstate to 
> (INIT | NO_REQ) on every page read function call #ifdef
> CONFIG_NAND_NO_PREFETCH_READS?
> 
> This leaves a problem with NAND_CMD_RNDOUT which is used by oob_first
> page reads but not supported by fsl_elbc_cmdfunc. I expect you don't use
> oob_first though..?

Right.  It's also used on swecc, but we don't use that either.

> I believe this would allow you to restore the original sequences and keep you
> going until you can define your own page read functions.
> 
> [BTW these changes applied quite cleanly to my Linux tree and give similar
> performance gains there as well. If we can reach agreement here, I will
> make patches for Linux as well. Unfortunately, my tree is 2.6.18 with
> backported NAND support from 2.6.32rc1, but I can deal with that.]

If possible, it would be nice to run these patches by the Linux list 
now, so we get their feedback earlier rather than later.

-Scott