Re: [PATCH] mtd: nand: squash struct nand_buffers into struct nand_chip

[Boris Brezillon]

On Tue, 5 Dec 2017 19:02:26 +0900
Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote:

> Hi Boris,
> 
> 2017-12-04 18:10 GMT+09:00 Boris Brezillon
> <boris.brezillon@xxxxxxxxxxxxxxxxxx>:
> 
> >>       }
> >>
> >>       if (!(chip->options & NAND_OWN_BUFFERS)) {
> >> -             nbuf = kzalloc(sizeof(*nbuf), GFP_KERNEL);
> >> -             if (!nbuf)
> >> +             chip->ecccalc = kmalloc(mtd->oobsize, GFP_KERNEL);
> >> +             if (!chip->ecccalc)
> >>                       return -ENOMEM;
> >>
> >> -             nbuf->ecccalc = kmalloc(mtd->oobsize, GFP_KERNEL);
> >> -             if (!nbuf->ecccalc) {
> >> +             chip->ecccode = kmalloc(mtd->oobsize, GFP_KERNEL);
> >> +             if (!chip->ecccode) {
> >>                       ret = -ENOMEM;
> >>                       goto err_free_nbuf;
> >>               }  
> >
> > Hm, again not directly related to this patch, but I wonder if we
> > couldn't allocate those buffers only when they are really needed.
> > For example, most NAND controllers do the ECC calculation/correct
> > in HW and simply don't need those buffers.  
> 
> 
> The only idea I came up with is to add a new flag,
> but I am not sure if you are happy with it
> because we are removing NAND_OWN_BUFFERS.

All drivers using ->calc/code_buf are providing a ->correct() and/or
->calculate() method, so I thought we could make the allocation
dependent on the presence of one of these hooks [1].

The only exception is the denali driver, but I think we can patch it
to not use the ->code_buf buffer [2].

[1]http://code.bulix.org/2ks7yp-236649
[2]http://code.bulix.org/sxqx7o-236650