On Friday 28 April 2006 09:40, Jesse Off wrote:
> Recently our founder had lunch with a STMicro engineer that knew quite a
> bit about NAND and specifically the read-disturb failure modes. He told us
> that after 100,000 or so reads of a sector, the stored data is actually
> very likely to experience corruption. Subsequent reads of the sector will
> return the same corrupted data, but the flash sector can be erased and
> reprogrammed (with the same data or otherwise) for another 100,000 or so
> reads. Towards the end of the useable life of the flash, a written sector
> may only be good for 1000 (or less) reads before it looses value. He also
> reaffirmed that the quality of NAND chips have peaked and we should expect
> the NAND "quirks" to get worse in the future.
>
> When asked for market predictions on alternative NAND chips with embedded
> controller logic in the chip (such as OneNAND, etc..) STMicro seems to
> believe those chips will be confined to niche markets and the current
> architecture of the 2k sector NAND chip will be where all the volume will
> be. I'm not sure if YAFFS is currently pursuing support of OneNAND, but
> from the sound of it it may not be a big deal either way.
>
> Is it still the current behavior of YAFFS2 to retire a sector permanently
> that experiences a read-disturb? We still seem to have a return rate of
> our boards about 1 every couple months with YAFFS having marked a large
> proportion of the flash sectors bad. This is much worse than the failure
> rates we have seen for CompactFlash cards-- which are usign the same NAND
> chips inside as YAFFS. In every case we've done failure analysis, the
> flash is almost 100% good after wiping the flash clean and unmarking the
> sectors bad.
Thanx Jess
As the read-disturb becomes more prevalent with the newer parts, the current
retirement policy is becoming too aggressive. I have put an investigation of
this on my todo list.
It is not a 100% styraightforward decision as to what to do here: too relaxed
a retirement poolicy and we could end up losing data.
-- CHarles
(text/plain)