Some observations and hunches....
> So I'm curious to see what performance others are getting (can you also
> include your proc and bus speeds?).

The WinCE stuff I have played with uses the following hw:
* SA1110 @ 206MHz
* SDRAM: 100MHz bus
* 128MB NAND (2x64MB Samsung parts). Write cycle ~90ns, read ~115ns.

Write performance all the way through YAFFS is around 1MB/s.
Under Linux I'd expect approx the same performance with a custom NAND access 
layer (ie. not going through mtd). This is not optimised that much (eg. loops 
to write data).

OK, now the observations:
* ECC: ECC speed is largely affected by raw CPU/RAM speed and the algorithm 
used. nand_ecc.c uses table look-ups and is pretty fast. Slower CPUs will 
obviously be slower. devices using DRAM instead of SDRAM will probably also 
get punished on RAM bandwdth.
* We're pushing a lot of data. The copy time is critical. NAND can handle 
50ns cycles, see how close you can get.
* Unwinding loops will give speed. ie. 
instead of
  for(i = 0; i < 512; i++) *nand = *buf++;
try
 *nand = *buf++;
 *nand = *buf++;
 *nand = *buf++;
 *nand = *buf++;
 *nand = *buf++;
 *nand = *buf++;
  ... 512 times.
This isn't what they taught us in computer science but it should go approx 4x 
the speed since there are no looping overheads. [The WinCE code could benefit 
from this too].

* The mtd nand.c buffers the writes internally (JFFS2 friendly). Bypassing 
the buffering would be faster for YAFFS.
* Schedules etc within the mtd layer will yield and probably slow things 
down. 
* The erasure model used in mtd could probably be a lot simpler/faster for 
NAND. Much of the complexity, IMHO, is there to try support both NOR (very 
slow erase) and NAND through one interface.
* Don't do any ECC or verification in the mtd layer since YAFFS does this 
already.

As I don't actually have any nand hw running Linux (and my time is better 
spent in yaffs_guts) I would appreciate someone else having a play.  Keeping 
the current mtd char device interface is useful for formatting, dumping etc. 
Yous might find my nandemul code to be a good kickoff point since this has 
been stripped back pretty hard already.

-- Charles


---------------------------------------------------------------------------------------
This mailing list is hosted by Toby Churchill open software (www.toby-churchill.org).
If mailing list membership is no longer wanted you can remove yourself from the list by 
sending an email to yaffs-request@toby-churchill.org with the text "unsubscribe" 
(without the quotes) as the subject.