Re: [PATCH 4/4] mtd: nand: mxc_nand: disable subpage reads

[Sascha Hauer]

On Mon, May 06, 2024 at 06:41:08PM +0200, Miquel Raynal wrote:
> Hi Sascha,
> 
> s.hauer@xxxxxxxxxxxxxx wrote on Mon, 22 Apr 2024 12:53:38 +0200:
> 
> > On Fri, Apr 19, 2024 at 11:46:57AM +0200, Miquel Raynal wrote:
> > > Hi Sascha,
> > > 
> > > s.hauer@xxxxxxxxxxxxxx wrote on Thu, 18 Apr 2024 13:43:15 +0200:
> > >   
> > > > On Thu, Apr 18, 2024 at 11:32:44AM +0200, Miquel Raynal wrote:  
> > > > > Hi Sascha,
> > > > > 
> > > > > s.hauer@xxxxxxxxxxxxxx wrote on Thu, 18 Apr 2024 08:48:08 +0200:
> > > > >     
> > > > > > On Wed, Apr 17, 2024 at 09:13:31AM +0200, Sascha Hauer wrote:    
> > > > > > > The NAND core enabled subpage reads when a largepage NAND is used with
> > > > > > > SOFT_ECC. The i.MX NAND controller doesn't support subpage reads, so
> > > > > > > clear the flag again.
> > > > > > > 
> > > > > > > Signed-off-by: Sascha Hauer <s.hauer@xxxxxxxxxxxxxx>
> > > > > > > ---
> > > > > > >  drivers/mtd/nand/raw/mxc_nand.c | 2 ++
> > > > > > >  1 file changed, 2 insertions(+)
> > > > > > > 
> > > > > > > diff --git a/drivers/mtd/nand/raw/mxc_nand.c b/drivers/mtd/nand/raw/mxc_nand.c
> > > > > > > index f44c130dca18d..19b46210bd194 100644
> > > > > > > --- a/drivers/mtd/nand/raw/mxc_nand.c
> > > > > > > +++ b/drivers/mtd/nand/raw/mxc_nand.c
> > > > > > > @@ -1667,6 +1667,8 @@ static int mxcnd_probe(struct platform_device *pdev)
> > > > > > >  	if (err)
> > > > > > >  		goto escan;
> > > > > > >  
> > > > > > > +	this->options &= ~NAND_SUBPAGE_READ;
> > > > > > > +      
> > > > > > 
> > > > > > Nah, it doesn't work like this. It turns out the BBT is read using
> > > > > > subpage reads before we can disable them here.
> > > > > >
> > > > > > This is the code in nand_scan_tail() we stumble upon:
> > > > > > 
> > > > > > 	/* Large page NAND with SOFT_ECC should support subpage reads */
> > > > > > 	switch (ecc->engine_type) {
> > > > > > 	case NAND_ECC_ENGINE_TYPE_SOFT:
> > > > > > 		if (chip->page_shift > 9)
> > > > > > 			chip->options |= NAND_SUBPAGE_READ;
> > > > > > 		break;
> > > > > > 
> > > > > > 	default:
> > > > > > 		break;
> > > > > > 	}
> > > > > > 
> > > > > > So the code assumes subpage reads are ok when SOFT_ECC is in use, which
> > > > > > in my case is not true. I guess some drivers depend on the
> > > > > > NAND_SUBPAGE_READ bit magically be set, so simply removing this code is
> > > > > > likely not an option.  Any ideas what to do?    
> > > > > 
> > > > > Can you elaborate why subpage reads are not an option in your
> > > > > situation? While subpage writes depend on chip capabilities, reads
> > > > > however should always work: it's just the controller selecting the
> > > > > column where to start and then reading less data than it could from the
> > > > > NAND cache. It's a very basic NAND controller feature, and I remember
> > > > > this was working on eg. an i.MX27.    
> > > > 
> > > > On the i.MX27 reading a full 2k page means triggering one read operation
> > > > per 512 bytes in the NAND controller, so it would be possible to read
> > > > subpages by triggering only one read operation instead of four in a row.
> > > > 
> > > > The newer SoCs like i.MX25 always read a full page with a single read
> > > > operation. We could likely read subpages by temporarily configuring the
> > > > controller for a 512b page size NAND.
> > > > 
> > > > I just realized the real problem comes with reading the OOB data. With
> > > > software BCH the NAND layer hardcodes the read_subpage hook to
> > > > nand_read_subpage() which uses nand_change_read_column_op() to read the
> > > > OOB data. This uses NAND_CMD_RNDOUT and I have now idea if/how this can
> > > > be implemented in the i.MX NAND driver. Right now the controller indeed
> > > > reads some data and then the SRAM buffer really contains part of the
> > > > desired OOB data, but also part of the user data.  
> > > 
> > > NAND_CMD_RNDOUT is impossible to avoid,  
> > 
> > Apparently it has been possible until now. NAND_CMD_RNDOUT has never
> > been used with this driver and it also doesn't work like expected.
> > 
> > One problem is that the read_page_raw() and write_page_raw() are not
> > implemented like supposed by the NAND layer. The i.MX NAND controller
> > uses a syndrome type ECC layout, meaning that the user data and OOB data
> > is interleaved, so the raw r/w functions should normally pass/expect the
> > page data in interleaved format. Unfortunately the raw functions are not
> > implemented like that, instead they detangle the data themselves. This
> > also means that setting the cursor using NAND_CMD_RNDOUT will not put
> > the cursor at a meaningful place, as the raw functions are not really
> > exect/return the raw page data.
> > 
> > This could be fixed, but the raw operations are also exposed to
> > userspace, so fixing these would mean that we might break some userspace
> > applications.
> 
> As answered to patch 3/4 I believe you need other raw page helpers for
> the software ECC path, just because the existing functions are tight to
> the on-host ECC logic and do what they are expected to do (I believe).
> 
> Creating another set of raw page helpers should be straightforward to
> do if that's really needed (mainly for performance purposes, but we're
> not yet there). Using the core helpers should work, the only thing is
> supporting properly the NAND_CMD_RNDOUT path, which should be possible
> at a rather low cost, it really is a very very basic command, I know no
> controller without this feature, even old ones.
> 
> > The other point is that with using software BCH ecc the NAND layer
> > requests me to read 7 bytes at offset 0x824. This can't be really
> > implemented in the i.MX NAND driver. It only allows us to read a full
> > 512 byte subpage, so whenever the NAND layer requests me to read a few
> > bytes the controller will always transfer 512 bytes from which I then
> > ignore most of it (and possibly trigger another 512 bytes transfer when
> > reading the ECC for the next subpage).
> 
> If you manage to get the NAND_CMD_RNDOUT op working I believe you'll be
> tempted to use memcpy32_fromio() with just a slightly rounded up length.

The overhead due to word rounding in memcpy32_fromio() is negligible, my
concern is that the controller will read 512 bytes from the NAND chip SRAM
to the controllers internal SRAM each time this command is used.

> 
> > I think these issues can all be handled somehow, but this comes at a
> > rather high price, both in effort and the risk of breaking userspace.
> > It would be far easier to tell the NAND layer not to do subpage reads.
> 
> I understand it may feel like that but that is likely not the right
> approach. I just think about another possibility: using monolithic
> reads if the controller is too constrained this way you might end up
> avoiding the RNDOUT command (might require a bit of tweaking in the
> core, I no longer remember exactly).

Not sure if I understand you correctly. I think using monolithic reads
is exactly what I am trying to archieve with disallowing subpage reads.

Subpage reads are rather unnecessary anyway. They are a performance
improvement when scanning the NAND for bad blocks and while reading
the UBI EC and VID headers. The former can be avoided with a
BBT and the latter with UBI fastmap (at least for the boot time critical
path when attaching a UBI).

> 
> Good luck, I really appreciate this effort.

Thanks. I'll do my best to get this done.

Sascha

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |