[PATCH v9 2/5] mtd: nand: vf610_nfc: add hardware BCH-ECC support

[stefan@agner.ch (Stefan Agner)]

On 2015-08-01 01:47, Brian Norris wrote:
> On Sat, Aug 01, 2015 at 01:35:52AM +0200, Stefan Agner wrote:
>> On 2015-08-01 01:09, Brian Norris wrote:
> 
>> >> +static int vf610_nfc_read_page(struct mtd_info *mtd, struct nand_chip *chip,
>> >> +				uint8_t *buf, int oob_required, int page)
>> >> +{
>> >> +	int eccsize = chip->ecc.size;
>> >> +	int stat;
>> >> +
>> >> +	vf610_nfc_read_buf(mtd, buf, eccsize);
>> >> +
>> >> +	if (oob_required)
>> >> +		vf610_nfc_read_buf(mtd, chip->oob_poi, mtd->oobsize);
>> >
>> > To fix the bitflips issue above, you'll just want to unconditionally
>> > read the OOB (it's fine to ignore 'oob_required') and...
>> >
>> >> +
>> >> +	stat = vf610_nfc_correct_data(mtd, buf);
>> >
>> > ...pass in chip->oob_poi as a third argument.
>> >
>>
>> Hm, this probably will have an effect on performance, since we usually
>> omit the OOB if not requested.
> 
> You could test :) I don't really like performance claims without tests.
> (I say this because I added the oob_required flag myself, but just for
> functional purposes, not performance. Many drivers got by just fine by
> always copying the OOB data.)

Did the measurement:

As is:
...
[   30.955675] mtd_speedtest: testing eraseblock write speed
[  143.349572] mtd_speedtest: eraseblock write speed is 4641 KiB/s
[  143.355606] mtd_speedtest: testing eraseblock read speed
[  183.816690] mtd_speedtest: eraseblock read speed is 12893 KiB/s
[  185.874702] mtd_speedtest: testing page write speed
[  302.608719] mtd_speedtest: page write speed is 4468 KiB/s
[  302.614229] mtd_speedtest: testing page read speed
[  343.831663] mtd_speedtest: page read speed is 12656 KiB/s
...

Unconditionally read OOB:
...
[   29.076983] mtd_speedtest: testing eraseblock write speed
[  140.829920] mtd_speedtest: eraseblock write speed is 4667 KiB/s
[  140.835960] mtd_speedtest: testing eraseblock read speed
[  181.594498] mtd_speedtest: eraseblock read speed is 12798 KiB/s
[  183.652793] mtd_speedtest: testing page write speed
[  299.772069] mtd_speedtest: page write speed is 4492 KiB/s
[  299.777583] mtd_speedtest: testing page read speed
[  341.283668] mtd_speedtest: page read speed is 12568 KiB/s
...

And with conditional OOB again, reading OOB if required in
vf610_nfc_correct_data.
...
[   29.907147] mtd_speedtest: testing eraseblock write speed
[  141.146171] mtd_speedtest: eraseblock write speed is 4689 KiB/s
[  141.152185] mtd_speedtest: testing eraseblock read speed
[  181.644380] mtd_speedtest: eraseblock read speed is 12883 KiB/s
[  183.703198] mtd_speedtest: testing page write speed
[  299.423179] mtd_speedtest: page write speed is 4507 KiB/s
[  299.428671] mtd_speedtest: testing page read speed
[  340.695925] mtd_speedtest: page read speed is 12640 KiB/s
[  342.747510] mtd_speedtest: testing 2 page write speed
...

The last test is probably pointless since we never read a empty page in
the speedtest. So performance hit is measurable but small (somewhat
below 100KiB/s).

This is with 64 bytes OOB. Since OOB sizes are only getting bigger, I
would rather still consider it... What do you think?

--
Stefan