[PATCH] mtd: nand: default bitflip-reporting threshold to 75% of correction strength

[Boris Brezillon]

Hi Brian,

On Wed, 21 Jan 2015 00:22:57 -0800
Brian Norris <computersforpeace at gmail.com> wrote:

> I'm sorry, I just noticed your reply as I was applying the patch. I can
> back it out if we find a real objection.

No, I'm fine with this patch, I was just trying to draw attention on
several problems I noticed while working on MLC NANDs.
We'll need to sort this out at some point, but I won't prevent sane
SLC/MLC NANDs from having this ECC threshold changed for something that
is not well supported yet ;-).

> 
> On Sat, Jan 17, 2015 at 08:01:37PM +0100, Boris Brezillon wrote:
> > On Mon, 12 Jan 2015 12:51:29 -0800
> > Brian Norris <computersforpeace at gmail.com> wrote:
> > > The MTD API reports -EUCLEAN only if the maximum number of bitflips
> > > found in any ECC block exceeds a certain threshold. This is done to
> > > avoid excessive -EUCLEAN reports to MTD users, which may induce
> > > additional scrubbing of data, even when the ECC algorithm in use is
> > > perfectly capable of handling the bitflips.
> > > 
> > > This threshold can be controlled by user-space (via sysfs), to allow
> > > users to determine what they are willing to tolerate in their
> > > application. But it still helps to have sane defaults.
> > > 
> > > In recent discussion [1], it was pointed out that our default threshold
> > > is equal to the correction strength. That means that we won't actually
> > > report any -EUCLEAN (i.e., "bitflips were corrected") errors until there
> > > are almost too many to handle. It was determined that 3/4 of the
> > > correction strength is probably a better default.
> > > 
> > > [1] http://lists.infradead.org/pipermail/linux-mtd/2015-January/057259.html
> > > 
> > > Signed-off-by: Brian Norris <computersforpeace at gmail.com>
> > > ---
> > >  drivers/mtd/nand/nand_base.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/mtd/nand/nand_base.c b/drivers/mtd/nand/nand_base.c
> > > index 816b5c1fd416..3f24b587304f 100644
> > > --- a/drivers/mtd/nand/nand_base.c
> > > +++ b/drivers/mtd/nand/nand_base.c
> > > @@ -4171,7 +4171,7 @@ int nand_scan_tail(struct mtd_info *mtd)
> > >  	 * properly set.
> > >  	 */
> > >  	if (!mtd->bitflip_threshold)
> > > -		mtd->bitflip_threshold = mtd->ecc_strength;
> > > +		mtd->bitflip_threshold = DIV_ROUND_UP(mtd->ecc_strength * 3, 4);
> > 
> > Just sharing my experience with MLC NANDs requiring read-retry: the
> > number of reported bitflips often raise ecc_strength value (at least
> > with the current read-retry approach).
> 
> I did not have fun when testing a few MLC NAND which required read
> retry. I ended up recommending that most of our customers move off of it
> onto another solution where possible. But yes, I can understand your
> issue.

Neither did I, but that's the world we live in, and I guess board
manufacturers will keep using crapy MLC chips so this is definitely
something we have to work on.

> 
> > This patch will definitely make UBI move NAND blocks over and over
> > again considering the threshold has been raised and the block is not
> > reliable anymore.
> 
> Personally, we found that we just needed to lower the
> MTD_UBI_WL_THRESHOLD and let UBI move away from blocks with high bitflip
> counts. I suppose we essentially treat the block as bad.

The problem I have is that a lot of them reach this ECC threshold limit
(at least with the current implementation).

> 
> > While I like the idea of limiting the threshold to something smaller
> > than what's recommended on the datasheet (or reported by ONFI) I wonder
> > if it won't make things worst in some cases.
> 
> I wouldn't exactly say that there is any threshold recommended on the
> datasheet or in ONFI. They simply specify a required correction
> strength,

Yes, that's what I was talking about.

> with no word about any intermediate handling -- what do they
> expect software to do when bitflips exceed the ECC strength? We
> immediately lose data. So we need to preemptively move such data.

Yep, I agree.

> 
> So I don't think it's a good idea at all to say that
> threshold == ecc_strength. That renders your ECC nearly useless w.r.t.
> the original design of UBI. UBI intends to use -EUCLEAN as a signal that
> high # of bitflips. I would suggest that 23 bitflips on a 24-bit
> correction algorithm is a "high # of bitflips." So as I see it, this
> patch is just restoring that UBI assumption.

Yes.

> 
> > Regarding the read-retry code, it currently stops retrying reading the
> > page once the page has been successfully retrieved (or in other terms
> > all bitflips have been fixed). But it might stop to soon, because by
> > changing the bit level threshold (in other term retrying one more time)
> > it might successfully read the page with less bitflips than the
> > previous attempt (these are just supposition, I haven't tested it yet).
> > If we can achieve that we could retry until we reach something below
> > the bitflips threshold value, and if we fail to find any, just consider
> > the lower number of bitflips found during those read-retry operations.
> 
> I believe I suggested scenarios like this to some flash vendors when
> speaking to reps in person, but they didn't seem to consider that
> likely. I think they were implying that there would be only one read
> retry mode that gives a reasonable result. I'm not sure if they were
> really the experts on that particular topic, though, or if they were
> just giving me an answer to make me happy.

Okay, good to know. I'll try to do some more testing to verify that.

> 
> Honestly, there's a lot of work that goes into MLC NAND for use by
> serious applications. For instance, SSD controller vendors, and even
> eMMC makers, use plenty of low-level knowledge that we simply do not
> have. From whatever I've gleaned about Toshiba MLC NAND (I don't think
> they even try to sell much raw NAND outside of eMMC and SSD solutions),
> they actually expose fine-grained voltage threshold controls through
> their (non-standard, obviously) command interface, and their clients
> likely use this to chart a targeted plan on how to adjust thresholds
> over the lifetime of a block, rather than just using a dumb sequential
> search like we do.

Yep, but they do sell these unreliable MLC chips (at least Micron and
Hynix do) to board manufacturers, and a lot of people would really like
to use a mainline kernel on these boards.

> 
> So anyway, that drifted a little away from the topic at hand. Are you
> suggesting that we should not change this default in nand_base, because
> of potential issues with highly-unreliable NAND that need read-retry? I
> can back out the patch while we finish this discussion, although I'm not
> very convinced so far.

As I said I don't expect you to back it out, just discussing the MLC
chip problems in an almost unrelated thread ;-).

Best Regards,

Boris


-- 
Boris Brezillon, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com