Re: [Yaffs] bit error rates --> a vendor speaks

Re: [Yaffs] bit error rates --> a vendor speaks

[Charles Manning]

To mtd-ers... Please forgive the cross-posting, but I think the content is 
sufficiently important  to those who use NAND but are not plugged in to the 
YAFFS list.


On Thursday 16 February 2006 08:53, William Watson wrote:

> I will also note that a NAND vendor who paid us a visit at about that same
> time said that we should expect WORSE soft error behaviour with succeeding
> generations of NAND flash chips.  The geometries would get smaller and
> smaller, the chip dies would get larger and larger, and the amount of time
> for production testing of each chip would not increase, or at least, not
> increase as fast as the total storage of a chip.  Thus, the testing per
> page would only go down in subsequent generations of chips.  These two
> statements seemed to say that we would see both (1) increased rates of ECC
> errors, and (2) an increase in the number of marginal blocks not marked bad
> by the chip vendor.

A vendor on the list contacted me off list, so I asked their permission before 
posting what they said on-list.  I got that permission so long as their name 
was removed.

As William states,  It seems that the reliability has peaked. NAND is expected 
to get worse, and a move to better correction schemes is encouraged.

For the most part, this is not really a YAFFS2 issue since the ECC mechanism 
(on the data) is not part of YAFFS2 per se. THis is now part of the mtd, or 
flash driver for non-Linux (eg. William's case). For YAFFS2, the impact is 
probably limited to:
1) ECC on tags....  Tags are so small that a single-bit correction is probably 
enough. Multibit is probably a good thing to investigate.
2) More OOB being used for multi-bit schemes will probably mean less space 
available for tags.
3) An emerging need for more forgiving block retirement.
4) Thinking about "spreading" to reduce write disturb.

Thus far it has been quite hard to engage NAND vendors but it seems some are 
now willing to talk a bit more. I will try to discuss the issues to better 
understand them.

-- Charles


Without further blaah, the vendor's words, slightly edited:

It's difficult to decide what the best block retirement policy should be.
For a program or erase failure, block retirement should be mandatory
because internally, the chip has already tried to program or erase multiple
times already.  For ECC (soft errors), it's more of an open question.
Should the errors be scrubbed and the data written back (to the same
location or a different location) or should the data be moved and the block
permanently retired?

One thing I can say with certainty, overall NAND flash reliability is
degrading.  At the time the application note was written, 0.16 micron based
NAND flash had a block write/erase endurance of 250k-1M cycles with
reasonable data retention.  However, as lithography shrinks have continued
(0.16um -> 0.13um -> 90nm -> 70nm), the physical cell area has been cut by
roughly half at each die shrink.  The physics of the materials don't
change, therefore, less charge is being stored in the memory cell every
generation.  90nm SLC (single level cell) NAND flash was nominally rated at
100K write/erase cycles per block, and it is very likely that 70nm SLC NAND
flash will be significantly less.  Better ECC is going to be necessary.
The single bit correcting Hamming code that is currently used for SLC NAND
was originally designed for SmartMedia over 10 years ago.  Today, most
memory cards using NAND flash implement Reed Solomon ECC capable of
correcting 4 or more symbol errors (typically 8 or more bits per symbol)
per 512 bytes.  This enables the ability to correct 4 random errors per
sector. However, for speed reasons, the ECC is usually done in hardware.
If a multi-bit ECC was implemented, then it would be possible to implement
a policy like: if there is the maximum number of correctable errors or an
uncorrectable error in a page, then the block can be permanently retired,
otherwise, scrub the data and move to a new location.  But this kind of
policy would be difficult to implement with single bit correcting Hamming
code.   Disturbance errors (read disturb, program disturb) will become more
probable in the future due to increased capacitive coupling between bit
lines and word lines due to decreased separation at each lithography
shrink.

One fact that is somewhat underappreciated is that one cannot have both
high block write/erase endurance and long data retention simultaneously.
One can have better data retention if the block sees fewer write/erase
cycles.  The more the data is spread out across all the blocks, the better
the overall data retention since every block is written and erased as few
times a necessary.  Journaling appears to be one of the best ways to spread
out the writes but some kind of static data wear leveling could improve it
further (however, there might be IP issues).

Keep up the good work,

Sincerely,
[name scrubbed]

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

Charles,

On Thu, 2006-02-16 at 14:32 +1300, Charles Manning wrote:
> 1) ECC on tags....  Tags are so small that a single-bit correction is probably 
> enough. Multibit is probably a good thing to investigate.
> 2) More OOB being used for multi-bit schemes will probably mean less space 
> available for tags.

We really should start to think seriously about oob usage for arbitrary
data storage at all. I know that YAFFS(2) depends on that, but looking
at the required mess in the code to keep this up for all the 9999
variants of ECC/RS whatever mechanisms, bad block marking schemes ...

Some words about Reed Solomon.

Reed Solomon needs hardware support for performance reasons. Efficient
usage of Reed Solomon requires a different Data / RS-code layout:

512 Byte Data
8 Byte RS Code
512 Byte Data
8 Byte RS Code
512 Byte Data
8 Byte RS Code
512 Byte Data
8 Byte RS Code
32 Byte OOB

This layout is supported already (see rtc_from4.c). It requires usage of
flash based bad block tables.

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Vitaly Wool]

Hi folks,

just two small notes from my side.
First, FWIW, YAFFS2 never writes OOB w/o data and that looks more proper 
than JFFS2 style which means cleanmarkers for an empty page.
YAFFS2 just needs means to be agnostic about how OOB bytes are placed 
within a page.

Next, I took a look at rtc_from4.c and I'm not sure how to follow this 
method if the NAND controller just doesn't have means to give the 
caclulated ECC back to user. Thomas, could you please elaborate on that?

Best regards,
  Vitaly

Thomas Gleixner wrote:

>Charles,
>
>On Thu, 2006-02-16 at 14:32 +1300, Charles Manning wrote:
>  
>
>>1) ECC on tags....  Tags are so small that a single-bit correction is probably 
>>enough. Multibit is probably a good thing to investigate.
>>2) More OOB being used for multi-bit schemes will probably mean less space 
>>available for tags.
>>    
>>
>
>We really should start to think seriously about oob usage for arbitrary
>data storage at all. I know that YAFFS(2) depends on that, but looking
>at the required mess in the code to keep this up for all the 9999
>variants of ECC/RS whatever mechanisms, bad block marking schemes ...
>
>Some words about Reed Solomon.
>
>Reed Solomon needs hardware support for performance reasons. Efficient
>usage of Reed Solomon requires a different Data / RS-code layout:
>
>512 Byte Data
>8 Byte RS Code
>512 Byte Data
>8 Byte RS Code
>512 Byte Data
>8 Byte RS Code
>512 Byte Data
>8 Byte RS Code
>32 Byte OOB
>
>This layout is supported already (see rtc_from4.c). It requires usage of
>flash based bad block tables.
>
>	tglx
>
>
>
>______________________________________________________
>Linux MTD discussion mailing list
>http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>
>  
>

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

Vitaly,

On Sat, 2006-02-18 at 19:31 +0300, Vitaly Wool wrote:
> Hi folks,

Can you please stop top posting ?

> just two small notes from my side.
> First, FWIW, YAFFS2 never writes OOB w/o data and that looks more proper 
> than JFFS2 style which means cleanmarkers for an empty page.
> YAFFS2 just needs means to be agnostic about how OOB bytes are placed 
> within a page.

I'm not too happy about this JFFS2 oddity and I would remove it better
today than tomorrow.

The agnostic thing is fine, but it still is not solving the fundamental
flaw of oob usage at all. The only guarantee of NAND is that you can
store data size, but the size of the "free" bytes in OOB (due to ECC,
bad block markers ...) is non constant and depends on hardware design,
chip types ... So any self restriction of a filesystem, block emulation
layer or whatever user of NAND flash to a small number of bytes it wants
to put into OOB can not cover _all_ possible constellations.

The kernel can only provide generic solutions and it makes absolutely no
sense to rely on nifty tricks which require code bloat and can not cover
every corner case.

> Next, I took a look at rtc_from4.c and I'm not sure how to follow this 
> method if the NAND controller just doesn't have means to give the 
> caclulated ECC back to user. Thomas, could you please elaborate on that?

-ENOPARSE. What does the NAND controller do with the calculated ECC, if
it has no way to let the user read the ECC ?
You mean those controllers which insert the ECC automatically at some
point into the data stream ? A great example why OOB usage is not a good
idea at all. These controllers guarantee data size and nothing else. I
have one on my desk which does not let you use OOB, because it does all
the magic itself.

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Charles Manning]

> > just two small notes from my side.
> > First, FWIW, YAFFS2 never writes OOB w/o data and that looks more proper
> > than JFFS2 style which means cleanmarkers for an empty page.
> > YAFFS2 just needs means to be agnostic about how OOB bytes are placed
> > within a page.
>
> I'm not too happy about this JFFS2 oddity and I would remove it better
> today than tomorrow.

All the NAND fs (yaffs, jffs2 and [in the future] jffs3) use free oob. It is a 
very handy place to put tags/metadata/whatever you want to call it and allows 
for far faster and more efficient flash file systems.

In systems that wrap up NAND as a block device you don't run a flash file 
system.

If a system does have spare oob and can use it for ffs use, then why deny it?

>
> The agnostic thing is fine, but it still is not solving the fundamental
> flaw of oob usage at all. The only guarantee of NAND is that you can
> store data size, but the size of the "free" bytes in OOB (due to ECC,
> bad block markers ...) is non constant and depends on hardware design,
> chip types ... So any self restriction of a filesystem, block emulation
> layer or whatever user of NAND flash to a small number of bytes it wants
> to put into OOB can not cover _all_ possible constellations.
>
> The kernel can only provide generic solutions and it makes absolutely no
> sense to rely on nifty tricks which require code bloat and can not cover
> every corner case.

I think you are looking at this in the wrong way.

We should focus on **clean interfaces** rather than trying to put everything 
into nand_base.c I think nand_base.c is a pretty good generic solution for 
certain classes of nand part, but it should not try to do all things for all 
NAND. Doing so will just slow it down and make it even more complex. I don't 
think anyone wants to make nand_base.c more complex.

Analogy: in serial driver world there are a lot of devices that fit the 16550 
model of operation and it makes sense to treat them similarly and use some 
common code. However, if youwant to support an weird UART like, say, the 
Atmel running in PDC mode then this no longer makes sense. You can however 
still use the same **interfaces**.

Already, many people replace nand_base.c for performance reasons. People with 
really odd-ball hardware write theirs from scratch. I expect that this will 
increase with the proliferaation of hardware assitance for ECC or DMA etc. 
Writing a nand driver is not a huge mission, and is easier with clean 
interfaces. Often people will treat nand_base.c as documentation.

Adding the ability to read fee oob data is trival compared to the rest of the 
effort doing ECC placement etc. Of course it can't be done if the hw denies 
access, but that is another matter.

>
> > Next, I took a look at rtc_from4.c and I'm not sure how to follow this
> > method if the NAND controller just doesn't have means to give the
> > caclulated ECC back to user. Thomas, could you please elaborate on that?
>
> -ENOPARSE. What does the NAND controller do with the calculated ECC, if
> it has no way to let the user read the ECC ?
> You mean those controllers which insert the ECC automatically at some
> point into the data stream ? A great example why OOB usage is not a good
> idea at all. These controllers guarantee data size and nothing else. I
> have one on my desk which does not let you use OOB, because it does all
> the magic itself.

Fine, if the hw makes it into a block device or hides the OOB completely, then 
use it as a block device and it cannot be used with a fs that wants oob 
areas. However, this is no longer NAND even though it might have NAND inside. 
I lose no sleep that it does not work with YAFFS.

I see no point in saying that because *some* systems cannot provide oob we 
should make an effort to deny this on *all* systems. To me this is as silly 
as suggesting we ban USB or ethernet or CAN from Linux because some systems 
don't support it.

Speaking for YAFFS: For the forseeable future, YAFFS will require and use OOB, 
or an adequate simulation. (eg. the way people do YAFFS on NOR).  As NAND 
changes, there will be a need to support new encoding schemes, but that's 
largely seperated off from YAFFS.

YAFFS does not try to be all things for all people. Perhaps YAFFS4 or more 
will not need OOB.  Sure, YAFFS **could** be redesigned to not need OOB but 
that would make it far less effective for its core user base. There are about 
40 Linux filesystems out there, they each exist because they do something 
special.

Having said that though, I guess there is one way to make YAFFS work with no 
OOB and that is to put the tags in the page. eg. 2048-byte page becomes 2032 
bytes of data and 16 bytes of tags. Possible, but would drive up copying 
costs due to the loss of page alignment.

-- CHarles

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Tue, 2006-02-21 at 09:42 +1300, Charles Manning wrote:
> If a system does have spare oob and can use it for ffs use, then why deny it?

Emphasis on "does have spare oob". Thats the whole point. You can not
rely on that. Whats the size you need ? 1, 2, ... 16, 32 bytes? There is
no guarantee.

> Already, many people replace nand_base.c for performance reasons. People with 
> really odd-ball hardware write theirs from scratch. I expect that this will 
> increase with the proliferaation of hardware assitance for ECC or DMA etc. 
> Writing a nand driver is not a huge mission, and is easier with clean 
> interfaces. Often people will treat nand_base.c as documentation.

Great. Whats the benefit for those who put effort into generic solutions
and clean interfaces?

> > You mean those controllers which insert the ECC automatically at some
> > point into the data stream ? A great example why OOB usage is not a good
> > idea at all. These controllers guarantee data size and nothing else. I
> > have one on my desk which does not let you use OOB, because it does all
> > the magic itself.
> 
> Fine, if the hw makes it into a block device or hides the OOB completely, then 

The hardware does not make it a block device. Where did I say that ?

> use it as a block device and it cannot be used with a fs that wants oob 
> areas. However, this is no longer NAND even though it might have NAND inside. 

Err. The controller hides oob. This is still NAND with all its
restrictions. And it does not become a block device magically.

> I lose no sleep that it does not work with YAFFS.

Wake up please. Thats going to be reality for NAND based stuff in the
future. The controllers will expose the raw FLASH but claim the OOB area
for their own purpose - hardware based error correction.

> YAFFS does not try to be all things for all people. Perhaps YAFFS4 or more 
> will not need OOB.  Sure, YAFFS **could** be redesigned to not need OOB but 
> that would make it far less effective for its core user base. There are about 
> 40 Linux filesystems out there, they each exist because they do something 
> special.

Right, but I doubt that the majority of them relies on non guaranteed
hardware features.

> Having said that though, I guess there is one way to make YAFFS work with no 
> OOB and that is to put the tags in the page. eg. 2048-byte page becomes 2032 
> bytes of data and 16 bytes of tags. Possible, but would drive up copying 
> costs due to the loss of page alignment.

Well, whether you like it or not. That topic has to be discussed in the
near future.

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Charles Manning]

On Tuesday 21 February 2006 10:37, Thomas Gleixner wrote:
> On Tue, 2006-02-21 at 09:42 +1300, Charles Manning wrote:
> > If a system does have spare oob and can use it for ffs use, then why deny
> > it?
>
> Emphasis on "does have spare oob". Thats the whole point. You can not
> rely on that. Whats the size you need ? 1, 2, ... 16, 32 bytes? There is
> no guarantee.

Sorry Thomas I don't buy that argument. If a system has a NAND device that 
does have spare OOB available then it does have spare OOB and I can rely on 
that. If a NAND chip is soldered to a board, and the system exposes the OOB  
it is there. YAFFS (or whatever) can then be used on this device.

Just going for the lowest common denominator all the time is like saying "run 
all serial links at 9600 and never use 115200 because 115200 might not be 
supported on all possible serial links", or "you can't run an ftdi USB serial 
port at 230k because most PC serial ports only go up to 115200".

>
> > Already, many people replace nand_base.c for performance reasons. People
> > with really odd-ball hardware write theirs from scratch. I expect that
> > this will increase with the proliferaation of hardware assitance for ECC
> > or DMA etc. Writing a nand driver is not a huge mission, and is easier
> > with clean interfaces. Often people will treat nand_base.c as
> > documentation.
>
> Great. Whats the benefit for those who put effort into generic solutions
> and clean interfaces?

Generic solutions are good for those that it fits with. It does not mean that 
the generic solution is going to be good for all people all the time.  It is 
impossible to write any body of code that is the best solution to all 
possible uses. It is however possible to write code that is good enough for 
many applications.

However, good clean interfaces are even more valuable because they allow 
people to easily plug in alternative solutions that interact well with the 
rest of the system.

The benefits of good clean interfaces are modularity. If you were to write a 
nand driver from scratch that obeys the interface then you can use it with 
mtdpart, mtdconcat, various fs etc. This is the major reason I don't like to 
have oobinfo hanging through the interface.

>
> > > You mean those controllers which insert the ECC automatically at some
> > > point into the data stream ? A great example why OOB usage is not a
> > > good idea at all. These controllers guarantee data size and nothing
> > > else. I have one on my desk which does not let you use OOB, because it
> > > does all the magic itself.
> >
> > Fine, if the hw makes it into a block device or hides the OOB completely,
> > then
>
> The hardware does not make it a block device. Where did I say that ?
>
> > use it as a block device and it cannot be used with a fs that wants oob
> > areas. However, this is no longer NAND even though it might have NAND
> > inside.
>
> Err. The controller hides oob. This is still NAND with all its
> restrictions. And it does not become a block device magically.

OK I got that wrong. I don't have the specs for the device you're looking at 
so I can't see all the details.

>
> > I lose no sleep that it does not work with YAFFS.
>
> Wake up please. Thats going to be reality for NAND based stuff in the
> future. The controllers will expose the raw FLASH but claim the OOB area
> for their own purpose - hardware based error correction.

Yes that is so for a few parts like OneNAND etc as well as some modules. 
However, there's still a lot of NAND out there that will provide OOB.

So far I think I've only heard one person ask about YAFFS on OneNAND. So far I 
have not heard anyone ask for YAFFS to run on the board sitting on your desk.

Perhaps opening up YAFFS to those people would be valuable. Would you like to 
see YAFFS run on your non-oob board?

I guess it could also make YAFFS easier to use for NOR people. Currently NOR 
folks have to do quite a bit of trickery. Enough people do this for me to 
think this could be valuable.

>
> > Having said that though, I guess there is one way to make YAFFS work with
> > no OOB and that is to put the tags in the page. eg. 2048-byte page
> > becomes 2032 bytes of data and 16 bytes of tags. Possible, but would
> > drive up copying costs due to the loss of page alignment.
>
> Well, whether you like it or not. That topic has to be discussed in the
> near future.

Perhaps this might be the way things go. Thinking about this more, I think it 
would not be so hard and it would be useful for people using OneNAND etc. 
YAFFS already has an internal cache which can take care of the page alignment 
with almost no work at all.

However, if that happens I'd still like to be able to use OOB if it is there 
(just like I want to be able to use 115200 if it is there) because OOB can 
make a more efficient fs with page alignment etc.


-- Charles

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Tue, 2006-02-21 at 11:40 +1300, Charles Manning wrote:
> Just going for the lowest common denominator all the time is like saying "run 
> all serial links at 9600 and never use 115200 because 115200 might not be 
> supported on all possible serial links", or "you can't run an ftdi USB serial 
> port at 230k because most PC serial ports only go up to 115200".

Again, the comparison is still flawed. 

The worst serial device still guarantees a baudrate > 0 and the
effective baudrate has no impact on data storage size.

> The benefits of good clean interfaces are modularity. If you were to write a 
> nand driver from scratch that obeys the interface then you can use it with 
> mtdpart, mtdconcat, various fs etc. This is the major reason I don't like to 
> have oobinfo hanging through the interface.

I'm well aware of interface design, but I see no convincing argument not
to remove oob access at all. At least there is no in kernel user
essentially depending on it. Making JFFS2 oob independend is a no brain
patch.

> OK I got that wrong. I don't have the specs for the device you're looking at 
> so I can't see all the details.

Simply as I said. Standard NAND interface, no oob access. Well, I can
read the raw device (including OOB) for diagnostic purposes in a special
mode.

> > > I lose no sleep that it does not work with YAFFS.
> >
> > Wake up please. Thats going to be reality for NAND based stuff in the
> > future. The controllers will expose the raw FLASH but claim the OOB area
> > for their own purpose - hardware based error correction.
> 
> Yes that is so for a few parts like OneNAND etc as well as some modules. 
> However, there's still a lot of NAND out there that will provide OOB.

That's no reason to keep a dead thing alive. There are still a lot of
cars which need leaded fuel. I dont see any gas station providing it
within a 100km range.

> Perhaps opening up YAFFS to those people would be valuable. Would you like to 
> see YAFFS run on your non-oob board?

At least for a test.

> I guess it could also make YAFFS easier to use for NOR people. Currently NOR 
> folks have to do quite a bit of trickery. Enough people do this for me to 
> think this could be valuable.

:)

> However, if that happens I'd still like to be able to use OOB if it is there 
> (just like I want to be able to use 115200 if it is there) because OOB can 
> make a more efficient fs with page alignment etc.

Go back to top - and as you said before:
> > Already, many people replace nand_base.c for performance reasons.

Nobody is going to prevent this, so they can expose any interface they
want for their private performance gain. 

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Jon Masters]

On 2/20/06, Thomas Gleixner <tglx@linutronix.de> wrote:

> On Tue, 2006-02-21 at 11:40 +1300, Charles Manning wrote:
> > Just going for the lowest common denominator all the time is like saying "run
> > all serial links at 9600 and never use 115200 because 115200 might not be
> > supported on all possible serial links", or "you can't run an ftdi USB serial
> > port at 230k because most PC serial ports only go up to 115200".
>
> Again, the comparison is still flawed.
>
> The worst serial device still guarantees a baudrate > 0 and the
> effective baudrate has no impact on data storage size.

Just let me make sure I'm getting this right:

1). You don't have OOB available to you with your NAND part.
2). You want YAFFS changed to suit your special case.

It's all very well arguing that relying on OOB in all cases is a bad
idea - and indeed, it sounds like a good idea to support packing extra
data into pages on flash - but you seem very keen on pushing the idea
that it's always bad to use OOB. So long as logic is added such that
you can have different behaviour, I fail to see the problem here.

Cheers,

Jon.

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Tue, 2006-02-21 at 00:29 +0000, Jon Masters wrote:
> > The worst serial device still guarantees a baudrate > 0 and the
> > effective baudrate has no impact on data storage size.
> 
> Just let me make sure I'm getting this right:
> 
> 1). You don't have OOB available to you with your NAND part.
> 2). You want YAFFS changed to suit your special case.

No, you get it wrong. Its not my personal problem at all. It's not about
the board on my desk.

It's about some piece of software relying on a non guaranteed hardware
feature.

I'm looking into the variety of hardware which evolves around NAND flash
and I carefully look in which direction this is going.

I see that the near future will require more complexity in the nand code
and I'm looking for a sane solution for that. I'm not saying that its
wrong, when something uses a nice hardware feature, but my point still
stands that it is wrong to rely on a feature which is nowhere
guaranteed. 

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Jörn Engel]

On Tue, 21 February 2006 09:26:13 +0100, Thomas Gleixner wrote:
> 
> It's about some piece of software relying on a non guaranteed hardware
> feature.
> 
> I'm looking into the variety of hardware which evolves around NAND flash
> and I carefully look in which direction this is going.
> 
> I see that the near future will require more complexity in the nand code
> and I'm looking for a sane solution for that. I'm not saying that its
> wrong, when something uses a nice hardware feature, but my point still
> stands that it is wrong to rely on a feature which is nowhere
> guaranteed. 

Ack.  One of the better features of hard disks is that they have a
standard interface.  I can take an ext3 image from a 5 year old disk
from one vendor and write it to a brand new disk from a different
vendor and things just work.  No conversion tool is needed.

For flash, moving images between arbitrary chips is just a pipe dream.
But things will get pushed into that direction.  So for a filesystem
to work on many chips, it should be prepared to work with different
page- and block-sizes and not rely on OOB at all.

Jörn

-- 
You ain't got no problem, Jules. I'm on the motherfucker. Go back in
there, chill them niggers out and wait for the Wolf, who should be
coming directly.
-- Marsellus Wallace

Re: [Yaffs] bit error rates --> YAFFS for devices with no OOB

[Charles Manning]

On Tuesday 21 February 2006 12:18, Thomas Gleixner wrote:

>
> > Perhaps opening up YAFFS to those people would be valuable. Would you
> > like to see YAFFS run on your non-oob board?
>
> At least for a test.

OK, I'll start investigating an OOB-less YAFFS and give that a priority 
depending on the amount of interest. Hint to everyone: If this sounds 
interesting, respond (on or off list), especially if you'd like to be 
involved in some way.

I think a final solution that exploits OOB if it is available but will work if 
OOB is not available would be best. That way:
1) People could use YAFFS on a wider variety of flash types (inc OneNAND etc) 
without using OOB.
2) If you have OOB, and drivers set up to use it, and wish to use it, then you 
will likely get some performance gains due to page alignment etc.

-- CHarles

Re: [Yaffs] bit error rates --> YAFFS for devices with no OOB

[Jon Masters]

On 2/21/06, Charles Manning <manningc2@actrix.gen.nz> wrote:

> OK, I'll start investigating an OOB-less YAFFS and give that a priority
> depending on the amount of interest. Hint to everyone: If this sounds
> interesting, respond (on or off list), especially if you'd like to be
> involved in some way.

I'd like to play with this, but I need to pick up a board for myself
before I do much more YAFFS work - all the existing stuff was done on
boards in the Monta office.

> I think a final solution that exploits OOB if it is available but will work if
> OOB is not available would be best.

Per my last mail, I think that's the best solution. Might take a bit
of work on MTD too.

Shove stuff my way please Charles and if I'm not frantically writing
chapters then I'll take a look :-)

Jon.

Re: [Yaffs] bit error rates --> YAFFS for devices with no OOB

[Jamie Lokier]

Charles Manning wrote:
> OK, I'll start investigating an OOB-less YAFFS and give that a priority 
> depending on the amount of interest. Hint to everyone: If this sounds 
> interesting, respond (on or off list), especially if you'd like to be 
> involved in some way.

Would that make it more usable as a filesystem for Compact Flash cards?

I've heard of people using JFFS2 on Compact Flash (using blkmtd),
because it's power-fail-safe and is much better for wear than, say,
ext3.

-- Jamie

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Charles Manning wrote:
> Sorry Thomas I don't buy that argument. If a system has a NAND device that 
> does have spare OOB available then it does have spare OOB and I can rely on 
> that. If a NAND chip is soldered to a board, and the system exposes the OOB  
> it is there. YAFFS (or whatever) can then be used on this device.

I understand Thomas's point as as he is fighting for generalization. 
Indeed, this OOB stuff introduces a lot of mess.

Charles' point is - if OOB is there, why not to let users use it? Also 
sounds reasonable.

What I think would be nice to do is to get rid of OOB in MTD stuff, but 
add a possibility to access OOB via some NAND-specific interfaces from 
nand_base. Indeed, if one wants to work with a generalized flash device 
- please use MTD interface. If one still wants to access OOB, use 
lower-layer NAND interfaces. That's all about to have more then one 
layer of Generalization. And I believe this is the right way to go.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Josh Boyer]

On 2/21/06, Artem B. Bityutskiy <dedekind@yandex.ru> wrote:
> Charles Manning wrote:
> > Sorry Thomas I don't buy that argument. If a system has a NAND device that
> > does have spare OOB available then it does have spare OOB and I can rely on
> > that. If a NAND chip is soldered to a board, and the system exposes the OOB
> > it is there. YAFFS (or whatever) can then be used on this device.
>
> I understand Thomas's point as as he is fighting for generalization.
> Indeed, this OOB stuff introduces a lot of mess.
>
> Charles' point is - if OOB is there, why not to let users use it? Also
> sounds reasonable.

But Charles also wants clean interfaces.  I agree that clean
interfaces are definitely a good thing, but trying to come up with a
clean interface for OOB access that won't get bastardized seems to be
unattainable.

> What I think would be nice to do is to get rid of OOB in MTD stuff, but
> add a possibility to access OOB via some NAND-specific interfaces from
> nand_base. Indeed, if one wants to work with a generalized flash device
> - please use MTD interface. If one still wants to access OOB, use
> lower-layer NAND interfaces. That's all about to have more then one
> layer of Generalization. And I believe this is the right way to go.

I think at some time in the not so distant future this whole
conversation will become a moot point.  SLC NAND quality seems to be
degradding as the die sizes go down, and MLC NAND is already of a
degraded quality comparitively.  Better ECC algorithms will be needed
to provide the reliability that people want and I can see that
consuming all of the available OOB area anyway.

josh

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Josh Boyer wrote:
> But Charles also wants clean interfaces.  I agree that clean
> interfaces are definitely a good thing, but trying to come up with a
> clean interface for OOB access that won't get bastardized seems to be
> unattainable.

But this does not mean that we should remove OOB support. Again, the 
right (IMO, of couse) way is to make several levels of generalization.

Just off the top of my head:

1. MTD level: a generic flash model with nothion of eraseblock, a 
mimimal I/O unit, read and write operations, nothing else. This level is 
for properly designed software.

2. NAND flash level: here you have OOB access. Work in terms of NAND 
pages, etc. Here YAFFS could be happy.

3. We could also have a DataFlash layer, where DataFlash guys could make 
use of their blocks and sectores, use features note visible from the 
topmost generic MTD layer.

> I think at some time in the not so distant future this whole
> conversation will become a moot point.  SLC NAND quality seems to be
> degradding as the die sizes go down, and MLC NAND is already of a
> degraded quality comparitively.  Better ECC algorithms will be needed
> to provide the reliability that people want and I can see that
> consuming all of the available OOB area anyway.

Still, the argument with CF cards makes me believe that vendors will 
provide some space for user data in OOB. Why can't they enlarge OOB size?

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Artem B. Bityutskiy wrote:
> But this does not mean that we should remove OOB support. Again, the 
> right (IMO, of couse) way is to make several levels of generalization.
> 
> Just off the top of my head:
> 
> 1. MTD level: a generic flash model with nothion of eraseblock, a 
> mimimal I/O unit, read and write operations, nothing else. This level is 
> for properly designed software.
> 
> 2. NAND flash level: here you have OOB access. Work in terms of NAND 
> pages, etc. Here YAFFS could be happy.
> 
> 3. We could also have a DataFlash layer, where DataFlash guys could make 
> use of their blocks and sectores, use features note visible from the 
> topmost generic MTD layer.
We can put that this way:

The topmost layer is MTD which provides a uniform and generic way to 
access all flash devices: NOR, NAND, AG-AND, DataFlash, OneNAND, ECC-ed 
NORs, and "whataver the hack". Here you don't even notice the divverence 
between them.

Below this MTD layer we could have a NOR layer (generalizes all NOR 
flashes), NAND layer (generalizes all NAND flashes), DataFlash layer 
(generalizes all DataFlash flashes) and so on. YAFFS could then work 
with NAND layer, not with MTD layer, because it is NAND-oriented.

Again, this is a question of proper MTD layering, and just saying that 
we don't support OOB is a lazy way to solve problems, IMO.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Thomas Gleixner wrote:
> Wake up please. Thats going to be reality for NAND based stuff in the
> future. The controllers will expose the raw FLASH but claim the OOB area
> for their own purpose - hardware based error correction.
One of my colleagues said a very interesting argument against this.

Look, consider all those CompactFlash cards. They are NAND flash based. 
They have a kind of block device emulation built-in. And I bet they use 
OOB to store the logical block number corresponding to this physical 
block. The block device over Flash device emulation is so widespread, so 
vendors will never forbid OOB usage.

 From this point of view, OOB is no going to go.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Tue, 2006-02-21 at 14:59 +0300, Artem B. Bityutskiy wrote:
> Thomas Gleixner wrote:
> > Wake up please. Thats going to be reality for NAND based stuff in the
> > future. The controllers will expose the raw FLASH but claim the OOB area
> > for their own purpose - hardware based error correction.
> One of my colleagues said a very interesting argument against this.
> 
> Look, consider all those CompactFlash cards. They are NAND flash based. 
> They have a kind of block device emulation built-in. And I bet they use 
> OOB to store the logical block number corresponding to this physical 
> block. The block device over Flash device emulation is so widespread, so 
> vendors will never forbid OOB usage.

I did nowhere say, that oob usage will be forbidden.

>  From this point of view, OOB is no going to go.

The CF controller does its own closed proprietary magic and looking at
the robustness of those cards I dont want to know what it does.

The OOB usage of a closed device is in no way relevant for a discussion
about a robust, sane and quite generic solution for handling NAND flash
devices inside of Linux.

I don't care what those chips do unless they run Linux inside.

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

BTW, talking about generalization, that mtd->eraseregions stuff is as 
bad as OOB. IMO, it is better to treat different regions as different 
partitions or as different MTD devices.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Josh Boyer]

On 2/25/06, Artem B. Bityutskiy <dedekind@yandex.ru> wrote:
> BTW, talking about generalization, that mtd->eraseregions stuff is as
> bad as OOB. IMO, it is better to treat different regions as different
> partitions or as different MTD devices.

Why is that?

Often times the eraseregions have no practical use for the software
involved.  Take the Intel P30 for example.  It has a few eraseblocks
at either the top or bottom of the chip that are different eraseblock
size.  Some software may want to use this feature, but by having the
MTD layer abstract them away it makes a lot of things simpler.

If you have different eraseregions show up as different MTD devices
(or partitions which are essentially the same thing in this
discussion), then you have to manually concatenate them back into a
single device.  It takes a) more RAM overhead and b) more layers of
complexity to do that.

josh

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Josh Boyer wrote:
> Often times the eraseregions have no practical use for the software
> involved.  Take the Intel P30 for example.  It has a few eraseblocks
> at either the top or bottom of the chip that are different eraseblock
> size.  
This explains why putting this all together in one mtd-> structure is 
bad  (they have different eraseblocks size).

> If you have different eraseregions show up as different MTD devices
> (or partitions which are essentially the same thing in this
> discussion),
Err, actually partitions are better, because they are handeled by the 
same flash driver.

> then you have to manually concatenate them back into a
> single device.
What for do you need to concatenate them back? Different erase regions 
are used for completely different purposes, right? So if you work with 
the small "boot" region, you don't normally need the rest of the flash 
and vice versa.

My point is that there should be a generalized flash model like this:

1. there are only 3 operations: read, write, erase;
2. erase is done in terms of eraseblocks, eraseblocks are all equivalent 
in size;
4. there is a minimal input/output unit exists;

That's all. Erase regions and OOB is out of this simple flash model, do 
you see what I mean? Add your erase regions to this model and you'll end 
up with a mess. Organize these regions as different instances of the 
above generalized model, and you have a very nice picture.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Josh Boyer]

On 2/27/06, Artem B. Bityutskiy <dedekind@yandex.ru> wrote:
> Josh Boyer wrote:
> > Often times the eraseregions have no practical use for the software
> > involved.  Take the Intel P30 for example.  It has a few eraseblocks
> > at either the top or bottom of the chip that are different eraseblock
> > size.
> This explains why putting this all together in one mtd-> structure is
> bad  (they have different eraseblocks size).

The chip driver abstracts that away.  The users of the device
typically don't _know_ there are different eraseregions.

>
> > If you have different eraseregions show up as different MTD devices
> > (or partitions which are essentially the same thing in this
> > discussion),
> Err, actually partitions are better, because they are handeled by the
> same flash driver.

I meant that for all intents and purposes, partitions show up as
completely different MTD devices.  They aren't traditional partitions
as on a hard drive.  It's unrelated to this topic, so ignore that.

>
> > then you have to manually concatenate them back into a
> > single device.
> What for do you need to concatenate them back? Different erase regions
> are used for completely different purposes, right? So if you work with
> the small "boot" region, you don't normally need the rest of the flash
> and vice versa.

Different erase regions _can_ be used for completely different
purposes.  Often, they aren't used at all.  Which means that if the
show up as a separate device, you have to concatenate them back
together.

> My point is that there should be a generalized flash model like this:
>
> 1. there are only 3 operations: read, write, erase;
> 2. erase is done in terms of eraseblocks, eraseblocks are all equivalent
> in size;
> 4. there is a minimal input/output unit exists;
>
> That's all. Erase regions and OOB is out of this simple flash model, do
> you see what I mean? Add your erase regions to this model and you'll end
> up with a mess. Organize these regions as different instances of the
> above generalized model, and you have a very nice picture.

A very nice picture that complicates things for end users and wastes
DRAM when there is no intention to use the eraseregions differently.

I can buy the OOB argument.  I think we'll have to agree to disagree
on the eraseregion stuff.

josh

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Josh, I believe you're only *partially* right

Please, switch to "architecture mode", leave this "low-level programmer" 
mode. Forget about few bytes of DRAM for now, please! :-)

The bellow are different points which should help me to change your mind.

1. The region is an attribute of *flash chip*. The region is *not* an 
attribute of MTD device. MTD device may describe the whole flash chip, 
it may describe a part of flash chip (partitions), and it may describe 
many flash chips (concatenation).

So, I state: the notion of region has nothing to do with the notion of 
MTD device.

If you have a specific application, which works with a *flash chip* not 
an *MTD device*, please, introduce another software object, sort of 
"struct flash_chip". Let your applications to work with these "struct 
flash_chip" objects.

Putting references to regions to the MTD device structure (struct 
mtd_info) is *bad*.

2. Well, you're in ecstasy of the current approach (kidding :-) ). Then 
explain me, why if, let's say JFFS3, opens an MTD device, it has access 
to some weird "regions"? What's on earth is going on? This is definitely 
error-prone. This is messy. This is bad design. JFFS3 does not want to 
see regions.

3. Regions... How do you look at them?

We could look at them like this:
a). They are different entities, used for completely different purposes, 
a rare program needs to use all of them simultaneously, they may have 
different eraseblock size and other properties, they are just different. 
Yes, it happened to be that they are on one single chip, but they are 
still different.

Or we could look at them narrower:
b). A region is a kind of "appendix" to the flash chip, where the 
bootcode or whatever may be put. There is not more then one (or really 
few) regions per flash chip. There is only one big and main "data" 
region, and one or few small auxiliary regions.

I reckon that the reality is more like b). And in this respect I see 
your point. Indeed, what for should we spend more DRAM to describe this 
unworthy "appendix" (aka region) by a distinct software object? Let's 
push this all together!

I can agree with you *partially*, providing that chips like a). are 
*not* going to appear. Partially means that I only agree that there *may 
be* no reason to create a distinct MTD device for this poor appendix.

So I state: Ok, in b)-like chips, do not create a distinct MTD device. 
Let's invent something different. But this must not be in struct 
mtd_info anyway.

Do you see my points?

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [Yaffs] bit error rates --> a vendor speaks

[Josh Boyer]

On 2/27/06, Artem B. Bityutskiy <dedekind@yandex.ru> wrote:
> Josh, I believe you're only *partially* right

I'm surprised I'm right at all :).  Normally I just make an idiot of myself.

> 1. The region is an attribute of *flash chip*. The region is *not* an
> attribute of MTD device. MTD device may describe the whole flash chip,
> it may describe a part of flash chip (partitions), and it may describe
> many flash chips (concatenation).
>
> So, I state: the notion of region has nothing to do with the notion of
> MTD device.

Ok.  I agree with that.  Regions are indeed a chip thing.

> If you have a specific application, which works with a *flash chip* not
> an *MTD device*, please, introduce another software object, sort of
> "struct flash_chip". Let your applications to work with these "struct
> flash_chip" objects.
>
> Putting references to regions to the MTD device structure (struct
> mtd_info) is *bad*.

Yes.  And that is where we are getting our signals crossed.  Or mabye
it was just me.  I was thinking that you wanted to take the region
info and make it manditory for a new MTD to be created.  If it's still
valid for the _chip_ driver to abstract away the region stuff, I'm all
for that.

> 2. Well, you're in ecstasy of the current approach (kidding :-) ). Then
> explain me, why if, let's say JFFS3, opens an MTD device, it has access
> to some weird "regions"? What's on earth is going on? This is definitely
> error-prone. This is messy. This is bad design. JFFS3 does not want to
> see regions.

Right.  This is indeed not really wanted.  Are there even users of the
region info in the mtd_info structure?

> 3. Regions... How do you look at them?
>
> We could look at them like this:
> a). They are different entities, used for completely different purposes,
> a rare program needs to use all of them simultaneously, they may have
> different eraseblock size and other properties, they are just different.
> Yes, it happened to be that they are on one single chip, but they are
> still different.
>
> Or we could look at them narrower:
> b). A region is a kind of "appendix" to the flash chip, where the
> bootcode or whatever may be put. There is not more then one (or really
> few) regions per flash chip. There is only one big and main "data"
> region, and one or few small auxiliary regions.
>
> I reckon that the reality is more like b). And in this respect I see
> your point. Indeed, what for should we spend more DRAM to describe this
> unworthy "appendix" (aka region) by a distinct software object? Let's
> push this all together!
>
> I can agree with you *partially*, providing that chips like a). are
> *not* going to appear. Partially means that I only agree that there *may
> be* no reason to create a distinct MTD device for this poor appendix.
>
> So I state: Ok, in b)-like chips, do not create a distinct MTD device.
> Let's invent something different. But this must not be in struct
> mtd_info anyway.

Or not invent anything at all unless it's really, really, really needed :).

>
> Do you see my points?

Yep.  I think we agree for the most part now.

josh

Re: [Yaffs] bit error rates --> a vendor speaks

[Russ Dill]

> Some words about Reed Solomon.
>
> Reed Solomon needs hardware support for performance reasons. Efficient
> usage of Reed Solomon requires a different Data / RS-code layout:

At what level is hardware support required? I'm involved in the design
of a new system with a 466Mhz 80200. Should fpga considerations be
mode for ecc correction? What sort of logic would be best to put in
the fpga?

Re: [Yaffs] bit error rates --> a vendor speaks

[Charles Manning]

On Sunday 19 February 2006 07:11, Russ Dill wrote:
> > Some words about Reed Solomon.
> >
> > Reed Solomon needs hardware support for performance reasons. Efficient
> > usage of Reed Solomon requires a different Data / RS-code layout:
>
> At what level is hardware support required? I'm involved in the design
> of a new system with a 466Mhz 80200. Should fpga considerations be
> mode for ecc correction? What sort of logic would be best to put in
> the fpga?

If you have NAND going past or through the FPGA then thinking about RS or ECC 
is a GoodIdea.

ECC is pretty expensive (yaffs_ecc.c is a bit faster than nand_ecc.c, but 
still requires quite a bit of computation). From what I've heard, RS is a lot 
more expensive.

I have not yet looked at RS yet, but I "pencil designed" a fast and simple ECC 
scheme that needs approx 22 flipflops + some other gates a while ago. This 
mechanism is used pretty much as follows:

1) Clear flipflops.
2) Transfer data
3) Read calculated ECC out of flipflops.

So the actual correction is still done in software, but the expensive part - 
the calculation - is done in hw. This gets you most of the way there with 
limited hw costs.

I hunch that you could do something similar for RS.

-- CHarles

Re: [Yaffs] bit error rates --> a vendor speaks

[Jon Masters]

On 2/19/06, Charles Manning <manningc2@actrix.gen.nz> wrote:

> If you have NAND going past or through the FPGA then thinking about RS or ECC
> is a GoodIdea.

On a related note, does anyone know of a good Xilinx reference board
that comes with a good amount of NAND already soldered on?

Having just quit working with Montavista stuff, I'm looking to pick up
a board for myself at home - something like an ML403 but with a good
amount of NAND would be nice :-)

Jon.

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Sat, 2006-02-18 at 10:11 -0800, Russ Dill wrote:
> > Some words about Reed Solomon.
> >
> > Reed Solomon needs hardware support for performance reasons. Efficient
> > usage of Reed Solomon requires a different Data / RS-code layout:
> 
> At what level is hardware support required? I'm involved in the design
> of a new system with a 466Mhz 80200. Should fpga considerations be
> mode for ecc correction? What sort of logic would be best to put in
> the fpga?

Depends. The 1bit correction/ 2bit detection Hamming ECC algorithm found
in the kernel is not too bad, but Reed Solomon is a quite conmputation
expensive algorithm. Look into the encoder / decoder code in
lib/reed_solomon.

In general you have to iterate over the data buffer and compute on each
step. The performance penalty depends on the complexitiy of the
algortihm. If you have enough space in your FPGA then its definitely a
good idea to put some ECC calculation mechanism into it. There are
implementations for both ECC and Reed Solomon available.

	tglx

Re: [Yaffs] bit error rates --> a vendor speaks

[Russ Dill]

> Depends. The 1bit correction/ 2bit detection Hamming ECC algorithm found
> in the kernel is not too bad, but Reed Solomon is a quite conmputation
> expensive algorithm. Look into the encoder / decoder code in
> lib/reed_solomon.


Do NAND errors tend to clump in sequences of bits (ie, something reed
solomon is paticularly good at)?

> In general you have to iterate over the data buffer and compute on each
> step. The performance penalty depends on the complexitiy of the
> algortihm. If you have enough space in your FPGA then its definitely a
> good idea to put some ECC calculation mechanism into it. There are
> implementations for both ECC and Reed Solomon available.

Care to recommend any?

Re: [Yaffs] bit error rates --> a vendor speaks

[Thomas Gleixner]

On Wed, 2006-02-22 at 17:46 -0700, Russ Dill wrote:
> > Depends. The 1bit correction/ 2bit detection Hamming ECC algorithm found
> > in the kernel is not too bad, but Reed Solomon is a quite conmputation
> > expensive algorithm. Look into the encoder / decoder code in
> > lib/reed_solomon.
> 
> 
> Do NAND errors tend to clump in sequences of bits (ie, something reed
> solomon is paticularly good at)?

In tests with AG-AND, which is known to be less reliable, I saw the bits
flip randomly in several consecutive bytes.

> > In general you have to iterate over the data buffer and compute on each
> > step. The performance penalty depends on the complexitiy of the
> > algortihm. If you have enough space in your FPGA then its definitely a
> > good idea to put some ECC calculation mechanism into it. There are
> > implementations for both ECC and Reed Solomon available.
> 
> Care to recommend any?

I dont remember where the Hamming ECC was from, but a Reed Solomon
encoder is available on opencores.org. And there are several generators
out there which produce RS encoder VHDL code from the given parameters.
Someone used this one with success
http://home.arcor.de/christianschuler/software/genenc_v1_2_tar.gz
You need to build the interfaces around though.

Parameters used back then:

-b 10	data width
-n 518	total length of code word
-k 512  number of information words
-m 10	symbol width
-p 10000001001	polynomial

Result below.

Some hints:

Expand the 8 bit data bus to 10 bits by setting the bit 8/9 hard to 1.
Before feeding the resulting 10 bit word into the generator invert them.
Also invert the resulting RS code. The reason is:
code for all data 0x00 is 0 0 0 0 0 0. Empty flash (0xff) expanded with
bit 8/9 feeds 0x3ff to the inverter, which results in 0x000 for the
generator input. Now you invert the resulting code and get 6 words of
0x3ff. That way you need no special checking for all empty flash as the
resulting code will be correct.

	tglx

-------------------------------------------------------------------------------
-- Parallel Reed-Solomon Encoder (ENTITY)
-- automatically generated by program 'genenc_solaris'
-- ./genenc_solaris -e 2 -b 10 -n 518 -k 512 -m 10 -p 10000001001 >
rs_ecc_nand1.vhdl
-- 
-- Date Thu Jun 23 08:07:57 2005
-- 
-------------------------------------------------------------------------------

LIBRARY ieee;
USE ieee.std_logic_1164.ALL;

ENTITY rs_enc IS
    PORT (
        clk     : IN    STD_LOGIC;
        enable  : IN    STD_LOGIC;
        reset   : IN    STD_LOGIC;
        out_enb : OUT   STD_LOGIC;
        d_in    : IN    STD_LOGIC_VECTOR(9 DOWNTO 0);
        d_out   : OUT   STD_LOGIC_VECTOR(9 DOWNTO 0)
    );
END rs_enc;
-------------------------------------------------------------------------------
-- Parallel Reed-Solomon Encoder (ARCHITECTURE)
-- automatically generated by program 'genenc_solaris'
-- 
-- Date Thu Jun 23 08:07:57 2005
-- 
-------------------------------------------------------------------------------

ARCHITECTURE rtl OF rs_enc IS

    CONSTANT nn:     INTEGER := 518;
        -- Number of code symbols

    CONSTANT kk:     INTEGER := 512;
        -- Number of information symbols

    CONSTANT bw:     INTEGER := 10;
        -- Bussize in bits

    CONSTANT mm:     INTEGER := 10;
        -- symbol size in bits

    CONSTANT last_in:  INTEGER := 512;
        -- Last input clock cycle

    CONSTANT last_out: INTEGER := 518;
        -- Last output clock cycle


-------------------------------------------------------------------------------

-- Galois field multiplier functions for Generator Polynomial
-- RS(518,512) encoder:

-- G(0):
    FUNCTION gf_mul_21 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(3) XOR bb(9);
            cc_next(1) := bb(0) XOR bb(4);
            cc_next(2) := bb(1) XOR bb(5);
            cc_next(3) := bb(2) XOR bb(3) XOR bb(6) XOR bb(9);
            cc_next(4) := bb(3) XOR bb(4) XOR bb(7);
            cc_next(5) := bb(4) XOR bb(5) XOR bb(8);
            cc_next(6) := bb(5) XOR bb(6) XOR bb(9);
            cc_next(7) := bb(0) XOR bb(6) XOR bb(7);
            cc_next(8) := bb(1) XOR bb(7) XOR bb(8);
            cc_next(9) := bb(2) XOR bb(8) XOR bb(9);
        RETURN cc_next;
    END gf_mul_21;

-- G(1):
    FUNCTION gf_mul_981 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(3) XOR bb(6) XOR bb(7) XOR bb(8) XOR bb(9);
            cc_next(1) := bb(0) XOR bb(4) XOR bb(7) XOR bb(8) XOR bb(9);
            cc_next(2) := bb(0) XOR bb(1) XOR bb(5) XOR bb(8) XOR bb(9);
            cc_next(3) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(7)
XOR 
                bb(8);
            cc_next(4) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(4)
XOR 
                bb(8) XOR bb(9);
            cc_next(5) := bb(1) XOR bb(2) XOR bb(3) XOR bb(4) XOR bb(5)
XOR 
                bb(9);
            cc_next(6) := bb(2) XOR bb(3) XOR bb(4) XOR bb(5) XOR bb(6);
            cc_next(7) := bb(0) XOR bb(3) XOR bb(4) XOR bb(5) XOR bb(6)
XOR 
                bb(7);
            cc_next(8) := bb(1) XOR bb(4) XOR bb(5) XOR bb(6) XOR bb(7)
XOR 
                bb(8);
            cc_next(9) := bb(2) XOR bb(5) XOR bb(6) XOR bb(7) XOR bb(8)
XOR 
                bb(9);
        RETURN cc_next;
    END gf_mul_981;

-- G(2):
    FUNCTION gf_mul_312 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(2) XOR bb(6) XOR bb(9);
            cc_next(1) := bb(3) XOR bb(7);
            cc_next(2) := bb(4) XOR bb(8);
            cc_next(3) := bb(2) XOR bb(5) XOR bb(6);
            cc_next(4) := bb(0) XOR bb(3) XOR bb(6) XOR bb(7);
            cc_next(5) := bb(1) XOR bb(4) XOR bb(7) XOR bb(8);
            cc_next(6) := bb(2) XOR bb(5) XOR bb(8) XOR bb(9);
            cc_next(7) := bb(3) XOR bb(6) XOR bb(9);
            cc_next(8) := bb(0) XOR bb(4) XOR bb(7);
            cc_next(9) := bb(1) XOR bb(5) XOR bb(8);
        RETURN cc_next;
    END gf_mul_312;

-- G(3):
    FUNCTION gf_mul_606 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(0) XOR bb(1) XOR bb(2) XOR bb(4) XOR bb(7);
            cc_next(1) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(5)
XOR 
                bb(8);
            cc_next(2) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(4)
XOR 
                bb(6) XOR bb(9);
            cc_next(3) := bb(0) XOR bb(3) XOR bb(5);
            cc_next(4) := bb(1) XOR bb(4) XOR bb(6);
            cc_next(5) := bb(2) XOR bb(5) XOR bb(7);
            cc_next(6) := bb(0) XOR bb(3) XOR bb(6) XOR bb(8);
            cc_next(7) := bb(1) XOR bb(4) XOR bb(7) XOR bb(9);
            cc_next(8) := bb(0) XOR bb(2) XOR bb(5) XOR bb(8);
            cc_next(9) := bb(0) XOR bb(1) XOR bb(3) XOR bb(6) XOR bb(9);
        RETURN cc_next;
    END gf_mul_606;

-- G(4):
    FUNCTION gf_mul_305 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(0) XOR bb(3) XOR bb(9);
            cc_next(1) := bb(0) XOR bb(1) XOR bb(4);
            cc_next(2) := bb(1) XOR bb(2) XOR bb(5);
            cc_next(3) := bb(2) XOR bb(6) XOR bb(9);
            cc_next(4) := bb(3) XOR bb(7);
            cc_next(5) := bb(4) XOR bb(8);
            cc_next(6) := bb(5) XOR bb(9);
            cc_next(7) := bb(0) XOR bb(6);
            cc_next(8) := bb(1) XOR bb(7);
            cc_next(9) := bb(2) XOR bb(8);
        RETURN cc_next;
    END gf_mul_305;

-- G(5):
    FUNCTION gf_mul_967 ( 
        bb: STD_LOGIC_VECTOR(9 DOWNTO 0))
        RETURN STD_LOGIC_VECTOR IS
        VARIABLE cc_next: STD_LOGIC_VECTOR(9 DOWNTO 0);
            BEGIN
            cc_next(0) := bb(4) XOR bb(5) XOR bb(6) XOR bb(7) XOR bb(8)
XOR 
                bb(9);
            cc_next(1) := bb(0) XOR bb(5) XOR bb(6) XOR bb(7) XOR bb(8)
XOR 
                bb(9);
            cc_next(2) := bb(0) XOR bb(1) XOR bb(6) XOR bb(7) XOR bb(8)
XOR 
                bb(9);
            cc_next(3) := bb(0) XOR bb(1) XOR bb(2) XOR bb(4) XOR bb(5)
XOR 
                bb(6);
            cc_next(4) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(5)
XOR 
                bb(6) XOR bb(7);
            cc_next(5) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(4)
XOR 
                bb(6) XOR bb(7) XOR bb(8);
            cc_next(6) := bb(0) XOR bb(1) XOR bb(2) XOR bb(3) XOR bb(4)
XOR 
                bb(5) XOR bb(7) XOR bb(8) XOR bb(9);
            cc_next(7) := bb(1) XOR bb(2) XOR bb(3) XOR bb(4) XOR bb(5)
XOR 
                bb(6) XOR bb(8) XOR bb(9);
            cc_next(8) := bb(2) XOR bb(3) XOR bb(4) XOR bb(5) XOR bb(6)
XOR 
                bb(7) XOR bb(9);
            cc_next(9) := bb(3) XOR bb(4) XOR bb(5) XOR bb(6) XOR bb(7)
XOR 
                bb(8);
        RETURN cc_next;
    END gf_mul_967;

-- G(6): input equals output !

-------------------------------------------------------------------------------

    TYPE bb_t IS ARRAY (0 TO nn-kk-1) OF STD_LOGIC_VECTOR(mm-1 DOWNTO
0);

    TYPE prod_t IS ARRAY (0 TO nn-kk-1) OF STD_LOGIC_VECTOR(mm-1 DOWNTO
0);

-------------------------------------------------------------------------------

    SIGNAL bb: bb_t;


    SIGNAL rs_calc :  STD_LOGIC;

    SIGNAL rs_ins  :  STD_LOGIC;


    BEGIN

    main:
    PROCESS (clk,bb,d_in,reset)

    VARIABLE feedback: STD_LOGIC_VECTOR(mm-1 DOWNTO 0);
    VARIABLE product: bb_t;

    BEGIN

        FOR i IN 0 TO mm - 1 LOOP
            feedback:= bb(0) XOR d_in;
        END LOOP;

        product(0) := gf_mul_967(feedback);
        product(1) := gf_mul_305(feedback);
        product(2) := gf_mul_606(feedback);
        product(3) := gf_mul_312(feedback);
        product(4) := gf_mul_981(feedback);
        product(5) := gf_mul_21(feedback);

        IF reset = '1' THEN

            FOR i IN 0 TO nn-kk-1 LOOP
                bb(i)    <= (others => '0');
            END LOOP;  

            out_enb <= '0';

        ELSIF clk = '1' AND clk'EVENT THEN

            IF enable = '1' THEN

                    out_enb <= '1';

                IF rs_ins = '0' AND rs_calc = '1' THEN

                        -- calculate:

                    FOR i IN 0 TO 4 LOOP
                        bb(i) <= bb(i+1) XOR product(i);
                    END LOOP;

                    bb(5) <= product(5); 
                    d_out <= d_in;

                ELSIF rs_ins = '1' AND rs_calc = '0' THEN

                        -- insert and shift:

                    d_out <= bb(0);

                    FOR i IN 0 TO nn-kk-2  LOOP
                        bb(i) <= bb(i+1);
                    END LOOP;

                    bb(nn-kk-1) <= (others =>'0');

                ELSIF rs_ins = '0' AND rs_calc = '0' THEN

                        -- bypass:

                    d_out <= d_in;

                END IF;

            ELSE
                out_enb <= '0';
            END IF;
        END IF;
    END PROCESS;  --main


    control:
    PROCESS (clk,reset)
    VARIABLE b_cnt : INTEGER RANGE 0 TO 518;
    BEGIN
        IF reset = '1' THEN
            rs_calc <= '1';
            rs_ins  <= '0';
            b_cnt   :=  0;
        ELSE
            IF clk = '1' AND clk'EVENT THEN
                IF enable = '1' THEN
                    IF b_cnt = 0 THEN
                        rs_calc <= '1';
                        rs_ins  <= '0';
                        b_cnt   := b_cnt + 1;
                    ELSIF b_cnt =  last_in - 1 THEN
                        rs_calc <= '0';
                        rs_ins  <= '1';
                        b_cnt   := b_cnt + 1;
                    ELSIF b_cnt = last_out - 1 THEN
                        rs_calc <= '1';
                        rs_ins  <= '0';
                        b_cnt   :=  0;
                    ELSE
                        b_cnt := b_cnt + 1;
                    END IF;
                END IF;
            END IF;
        END IF;
    END PROCESS;  -- ctrl

END rtl;  -- rs_enc

Re: [Yaffs] bit error rates --> a vendor speaks

[Vitaly Wool]

Hello folks,

my two cents...

The way out I'd like to suggest is to [re]design the NAND filesystems 
(jffs{2|3}, yaffs[2]) in such a way that no sensible information is 
stored in OOB area. That will mean no compatibility/applicability issues.
But how to deal with speed decrease within this approach? I think it's 
reasonable to store speedup information (indexes, hash values etc.) in 
the OOB area. Well, this speedup info will need to be designed in an 
adaptable way as the amount of spare OOB bytes can be whatever.
Also, the tools for writing the filesystem image onto the flash better 
be capable of generating this speedup info as they write the image, 
otherwise first mount will take a lot of time (probably as it happens 
once, it's also acceptable but less convenient).

And... I do not like the idea to remove OOB-handling stuff from the MTD 
layer. I also do not like the idea to make a separate interface for NAND 
layer, I'm afraid this will lead to confusion and hard to track errors.

Hope that makes sense,

Best regards,
   Vitaly

Charles Manning wrote:

>To mtd-ers... Please forgive the cross-posting, but I think the content is 
>sufficiently important  to those who use NAND but are not plugged in to the 
>YAFFS list.
>
>
>On Thursday 16 February 2006 08:53, William Watson wrote:
>
>  
>
>>I will also note that a NAND vendor who paid us a visit at about that same
>>time said that we should expect WORSE soft error behaviour with succeeding
>>generations of NAND flash chips.  The geometries would get smaller and
>>smaller, the chip dies would get larger and larger, and the amount of time
>>for production testing of each chip would not increase, or at least, not
>>increase as fast as the total storage of a chip.  Thus, the testing per
>>page would only go down in subsequent generations of chips.  These two
>>statements seemed to say that we would see both (1) increased rates of ECC
>>errors, and (2) an increase in the number of marginal blocks not marked bad
>>by the chip vendor.
>>    
>>
>
>A vendor on the list contacted me off list, so I asked their permission before 
>posting what they said on-list.  I got that permission so long as their name 
>was removed.
>
>As William states,  It seems that the reliability has peaked. NAND is expected 
>to get worse, and a move to better correction schemes is encouraged.
>
>For the most part, this is not really a YAFFS2 issue since the ECC mechanism 
>(on the data) is not part of YAFFS2 per se. THis is now part of the mtd, or 
>flash driver for non-Linux (eg. William's case). For YAFFS2, the impact is 
>probably limited to:
>1) ECC on tags....  Tags are so small that a single-bit correction is probably 
>enough. Multibit is probably a good thing to investigate.
>2) More OOB being used for multi-bit schemes will probably mean less space 
>available for tags.
>3) An emerging need for more forgiving block retirement.
>4) Thinking about "spreading" to reduce write disturb.
>
>Thus far it has been quite hard to engage NAND vendors but it seems some are 
>now willing to talk a bit more. I will try to discuss the issues to better 
>understand them.
>
>-- Charles
>
>
>Without further blaah, the vendor's words, slightly edited:
>
>It's difficult to decide what the best block retirement policy should be.
>For a program or erase failure, block retirement should be mandatory
>because internally, the chip has already tried to program or erase multiple
>times already.  For ECC (soft errors), it's more of an open question.
>Should the errors be scrubbed and the data written back (to the same
>location or a different location) or should the data be moved and the block
>permanently retired?
>
>One thing I can say with certainty, overall NAND flash reliability is
>degrading.  At the time the application note was written, 0.16 micron based
>NAND flash had a block write/erase endurance of 250k-1M cycles with
>reasonable data retention.  However, as lithography shrinks have continued
>(0.16um -> 0.13um -> 90nm -> 70nm), the physical cell area has been cut by
>roughly half at each die shrink.  The physics of the materials don't
>change, therefore, less charge is being stored in the memory cell every
>generation.  90nm SLC (single level cell) NAND flash was nominally rated at
>100K write/erase cycles per block, and it is very likely that 70nm SLC NAND
>flash will be significantly less.  Better ECC is going to be necessary.
>The single bit correcting Hamming code that is currently used for SLC NAND
>was originally designed for SmartMedia over 10 years ago.  Today, most
>memory cards using NAND flash implement Reed Solomon ECC capable of
>correcting 4 or more symbol errors (typically 8 or more bits per symbol)
>per 512 bytes.  This enables the ability to correct 4 random errors per
>sector. However, for speed reasons, the ECC is usually done in hardware.
>If a multi-bit ECC was implemented, then it would be possible to implement
>a policy like: if there is the maximum number of correctable errors or an
>uncorrectable error in a page, then the block can be permanently retired,
>otherwise, scrub the data and move to a new location.  But this kind of
>policy would be difficult to implement with single bit correcting Hamming
>code.   Disturbance errors (read disturb, program disturb) will become more
>probable in the future due to increased capacitive coupling between bit
>lines and word lines due to decreased separation at each lithography
>shrink.
>
>One fact that is somewhat underappreciated is that one cannot have both
>high block write/erase endurance and long data retention simultaneously.
>One can have better data retention if the block sees fewer write/erase
>cycles.  The more the data is spread out across all the blocks, the better
>the overall data retention since every block is written and erased as few
>times a necessary.  Journaling appears to be one of the best ways to spread
>out the writes but some kind of static data wear leveling could improve it
>further (however, there might be IP issues).
>
>Keep up the good work,
>
>Sincerely,
>[name scrubbed]
>
>______________________________________________________
>Linux MTD discussion mailing list
>http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>
>  
>

Re: [Yaffs] bit error rates --> a vendor speaks

[Artem B. Bityutskiy]

Vitaly Wool wrote:
> And... I do not like the idea to remove OOB-handling stuff from the MTD 
> layer. I also do not like the idea to make a separate interface for NAND 
> layer, I'm afraid this will lead to confusion and hard to track errors.
Having a generalize MTD layer on top of NAND, NOR, etc layers cannot 
lead to any confusion. What can confuse you? This is just a nice layering.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.