NAND pages cache?

NAND pages cache?

[Artem Bityuckiy]

Hello,

It is known that MTD always reads/writes in fractions of the whole NAND 
pages. An this is right I believe.

But while working with JFFS2, I have mentioned that very often it reads 
the same page several times consequently. This happens for example, when 
JFFS2 performs Garbage Collection, and there are several small nodes in 
the NAND page. Thus, when JFFS2 reads these nodes, it re-reads the same 
NAND page several times. This is bad.

So, the obvious idea how to avoid this is to introduce something like 
NAND pages cache. The cache will keep several NAND pages which were last 
accessed.

Obviously, this will require at least one more page buffer copy 
operation and, for example, will only decrease the sequential read speed.

So, It is possible to introduce calls like:

nand_read_buffered
nand_read_unbuffered

I don not know is it good to do this only in JFFS2 or on the MTD NAND layer?

Comments?

Thanks.

-- 
Best regards, Artem B. Bityuckiy
Oktet Labs (St. Petersburg), Software Engineer.
+78124286709 (office) +79112449030 (mobile)
E-mail: dedekind@oktetlabs.ru, web: http://www.oktetlabs.ru

Re: NAND pages cache?

[David Woodhouse]

On Mon, 2004-11-15 at 11:49 +0300, Artem Bityuckiy wrote:
> So, the obvious idea how to avoid this is to introduce something like 
> NAND pages cache. The cache will keep several NAND pages which were last 
> accessed.
> 
> I don not know is it good to do this only in JFFS2 or on the MTD NAND layer?

I thought we already did this in the MTD NAND layer, since we had to
read whole pages there to do ECC anyway. Thomas?

-- 
dwmw2

Re: NAND pages cache?

[Thomas Gleixner]

On Mon, 2004-11-15 at 08:49 +0000, David Woodhouse wrote:
> On Mon, 2004-11-15 at 11:49 +0300, Artem Bityuckiy wrote:
> > So, the obvious idea how to avoid this is to introduce something like 
> > NAND pages cache. The cache will keep several NAND pages which were last 
> > accessed.
> > 
> > I don not know is it good to do this only in JFFS2 or on the MTD NAND layer?
> 
> I thought we already did this in the MTD NAND layer, since we had to
> read whole pages there to do ECC anyway. Thomas?

We do caching only if a page is read partially. If we read a full page
we read directly into the fs buffer.

The partial reads happen when we read nodes. The full page reads happen
when we read data. That's sufficient and captures most of the double
reads on the same page.

tglx

Re: NAND pages cache?

[Artem Bityuckiy]

Thomas Gleixner wrote:
> On Mon, 2004-11-15 at 08:49 +0000, David Woodhouse wrote:
> 
>>On Mon, 2004-11-15 at 11:49 +0300, Artem Bityuckiy wrote:
>>
>>>So, the obvious idea how to avoid this is to introduce something like 
>>>NAND pages cache. The cache will keep several NAND pages which were last 
>>>accessed.
>>>
>>>I don not know is it good to do this only in JFFS2 or on the MTD NAND layer?
>>
>>I thought we already did this in the MTD NAND layer, since we had to
>>read whole pages there to do ECC anyway. Thomas?
> 
> 
> We do caching only if a page is read partially. If we read a full page
> we read directly into the fs buffer.
> 
> The partial reads happen when we read nodes. The full page reads happen
> when we read data. That's sufficient and captures most of the double
> reads on the same page.
> 
> tglx
Hmm, It seems I do not know something about MTD. I thought it *always* 
reads whole
pages in orger to always check ECC... I'll see this more close.

Ok, thanks for comment.

-- 
Best regards, Artem B. Bityuckiy
Oktet Labs (St. Petersburg), Software Engineer.
+78124286709 (office) +79112449030 (mobile)
E-mail: dedekind@oktetlabs.ru, web: http://www.oktetlabs.ru

Re: NAND pages cache?

[Thomas Gleixner]

On Mon, 2004-11-15 at 12:43 +0300, Artem Bityuckiy wrote:
> Thomas Gleixner wrote:
> > On Mon, 2004-11-15 at 08:49 +0000, David Woodhouse wrote:
> > 
> >>On Mon, 2004-11-15 at 11:49 +0300, Artem Bityuckiy wrote:
> >>
> >>>So, the obvious idea how to avoid this is to introduce something like 
> >>>NAND pages cache. The cache will keep several NAND pages which were last 
> >>>accessed.
> >>>
> >>>I don not know is it good to do this only in JFFS2 or on the MTD NAND layer?
> >>
> >>I thought we already did this in the MTD NAND layer, since we had to
> >>read whole pages there to do ECC anyway. Thomas?
> > 
> > 
> > We do caching only if a page is read partially. If we read a full page
> > we read directly into the fs buffer.
> > 
> > The partial reads happen when we read nodes. The full page reads happen
> > when we read data. That's sufficient and captures most of the double
> > reads on the same page.
> > 
> > tglx
> Hmm, It seems I do not know something about MTD. I thought it *always* 
> reads whole
> pages in orger to always check ECC... I'll see this more close.

The nand driver always reads whole pages due to ECC. But if the fs
driver requests only a part of the page then the page is read into the
cache buffer of the nand driver, ECC is checked and the part which was
requested by the fs driver is copied into the fs buffer. If the next
read requests another part of this page then we read out of the cache
buffer instead of reading the FLASH again.

tglx

Re: NAND pages cache?

[Artem Bityuckiy]

Thomas Gleixner wrote:
> The nand driver always reads whole pages due to ECC. But if the fs
> driver requests only a part of the page then the page is read into the
> cache buffer of the nand driver, ECC is checked and the part which was
> requested by the fs driver is copied into the fs buffer. If the next
> read requests another part of this page then we read out of the cache
> buffer instead of reading the FLASH again.
> 
> tglx
O, thank you for the hint!

-- 
Best regards, Artem B. Bityuckiy
Oktet Labs (St. Petersburg), Software Engineer.
+78124286709 (office) +79112449030 (mobile)
E-mail: dedekind@oktetlabs.ru, web: http://www.oktetlabs.ru