Re: [RFC PATCH 25/27] mtd: nand: Add helpers to manage ECC engines and configurations

[Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx>]

On Mon, 25 Feb 2019 17:34:22 +0100
Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> wrote:

> On Mon, 25 Feb 2019 17:01:18 +0100
> Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> 
> > Hi Boris,
> > 
> > Boris Brezillon <bbrezillon@xxxxxxxxxx> wrote on Fri, 22 Feb 2019
> > 15:44:31 +0100:
> >   
> > > On Thu, 21 Feb 2019 13:58:04 +0100
> > > Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> > >     
> > > > Add the logic in the NAND core to find the right ECC engine depending
> > > > on the NAND chip requirements and the user desires. Right now, the
> > > > choice may be made between (more will come):
> > > > * software Hamming
> > > > * software BCH
> > > > * on-die (SPI-NAND devices only)
> > > > 
> > > > Once the ECC engine has been found, the ECC engine must be
> > > > configured.
> > > > 
> > > > Signed-off-by: Miquel Raynal <miquel.raynal@xxxxxxxxxxx>
> > > > ---
> > > >  drivers/mtd/nand/core.c  | 107 +++++++++++++++++++++++++++++++++++++++
> > > >  include/linux/mtd/nand.h |   4 ++
> > > >  2 files changed, 111 insertions(+)
> > > > 
> > > > diff --git a/drivers/mtd/nand/core.c b/drivers/mtd/nand/core.c
> > > > index 872d46b5fc0f..9feb118c9f68 100644
> > > > --- a/drivers/mtd/nand/core.c
> > > > +++ b/drivers/mtd/nand/core.c
> > > > @@ -207,6 +207,113 @@ int nanddev_mtd_max_bad_blocks(struct mtd_info *mtd, loff_t offs, size_t len)
> > > >  }
> > > >  EXPORT_SYMBOL_GPL(nanddev_mtd_max_bad_blocks);
> > > >  
> > > > +/**
> > > > + * nanddev_find_ecc_engine() - Find a suitable ECC engine
> > > > + * @nand: NAND device
> > > > + */
> > > > +static int nanddev_find_ecc_engine(struct nand_device *nand)      
> > > 
> > > Can we pass the conf in argument instead of reading it from
> > > nand->ecc.user_conf?
> > >     
> > > > +{
> > > > +	bool is_spinand = mtd_type_is_spinand(&nand->mtd);      
> > > 
> > > And here is the reason for the SPINAND type.
> > >     
> > > > +
> > > > +	/* Read the user desires in terms of ECC engine/configuration */
> > > > +	nand_ecc_read_user_conf(nand);
> > > > +
> > > > +	/* No ECC engine requestedn, let's return without error */
> > > > +	if (nand->ecc.user_conf.mode == NAND_ECC_NONE)
> > > > +		return 0;
> > > > +
> > > > +	/* Raw NAND default mode is hardware */
> > > > +	if (!is_spinand && nand->ecc.user_conf.mode < 0)
> > > > +		nand->ecc.user_conf.mode = NAND_ECC_HW;      
> > > 
> > > We should let the raw NAND layer take this decision (actually, it's
> > > even a raw NAND controller driver decision). Please complain if
> > > user_conf.mode is invalid.
> > > This way you won't need the SPINAND type you added in one of your
> > > previous patch.
> > >     
> > > > +
> > > > +	/* SPI-NAND default mode is on-die */
> > > > +	if (is_spinand && nand->ecc.user_conf.mode < 0)
> > > > +		nand->ecc.user_conf.mode = NAND_ECC_ON_DIE;
> > > > +
> > > > +	switch (nand->ecc.user_conf.mode) {
> > > > +	case NAND_ECC_SOFT:
> > > > +		nand->ecc.engine = nand_ecc_sw_get_engine(nand);
> > > > +		break;
> > > > +	case NAND_ECC_ON_DIE:
> > > > +		if (is_spinand)
> > > > +			nand->ecc.engine = spinand_ondie_ecc_get_engine();      
> > > 
> > > So, maybe it's worth having the ondie ECC engine instance directly
> > > embedded in nand_device instead of spinand, or maybe just a pointer, so
> > > that you don't reserve extra space when the NAND device does not have
> > > on-die ECC.
> > >     
> > > > +		else
> > > > +			pr_err("On-die ECC engines for non SPI devices not supported yet\n");
> > > > +		break;
> > > > +	case NAND_ECC_HW:
> > > > +		pr_err("Hardware ECC engines not supported yet\n");
> > > > +		break;
> > > > +	default:
> > > > +		pr_err("Missing ECC engine property\n");
> > > > +	}
> > > > +
> > > > +	if (!nand->ecc.engine)
> > > > +		return  -EINVAL;
> > > > +
> > > > +	return 0;
> > > > +}      
> > >     
> > 
> > I think this is the root patch were our ideas diverge. For me, each
> > 'ECC engine' has a _get() helper and the NAND core decides which one to
> > call to retrieve the right engine. Can you please explain what was your
> > idea if this one does not fit?  
> 
> Each class of ECC engine has its own way of getting a pointer to an ECC
> engine instance:
> 
> - ondie: the engine is directly attached to the device and can be
>   retrieved by accessing a nand_device field (nanddev->ondie_ecc?).
> - sw ECC: you can create a new instance for each device and possibly
>   pass the OOB layout you want to use (assuming you don't want to use
>   the default one)
> - HW-controller-side engine: refers to the controller device (parent
>   node) or the device pointed by the ecc-engine DT prop (we'll probably
>   need a way to pass this info when for non-DT platforms). For this one
>   we'll add a nanddev_get_hw_ecc_engine() which will search for all
>   registered HW ECC engines and try to match with some search keys (in
>   case of DT, it's the ->of_node pointer address)
> 
> Clearly, for SW-based ECC, you'll need more than just the nanddev
> object, as layouts can differ depending on the controller driver. So,
> what I'm suggesting is to have a
> 
> 	nand_create_sw_ecc_engine(algo, layout) 
> 
> funtion that returns this ECC engine instance which the driver will
> then attach to the NAND device.

I'm reconsidering what I said here. I guess having the layout set by
the driver before nand_ecc_sw_get_engine() is called could do the
trick. We just need to make sure the layout provided has enough space
to store ECC bytes at context creation time.

> 
> > 
> > Also, the parsing of the DT (in nand_ecc_read_user_conf()) gives me the
> > user ECC mode and algo, so I cannot let the raw NAND core (or a raw
> > NAND controller driver) or the SPI NAND core decide which mode is the
> > default if not provided by the user.  
> 
> Except this prop is optional in most cases, and the default value is
> not always the same, which is why I think this ECC engine retrieval step
> should be left to each sub-layer (and sometimes to the controller driver
> behind it). Maybe you can provide helpers to help with that, but I
> don't think taking this decision here, based on the bus type, is a good
> idea. And I also don't like the idea of adding a new SPINAND type.

Hm, after thinking a bit more about it, maybe we could have something
in-between: let the controller driver or sub-layer specify what the
default values are (provider/mode and possibly algorithm if that makes
sense) so that this generic function can use these defaults when
nand->ecc.user_conf.{mode,algo} == UNKNOWN.

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/