Re: [PATCH v3 2/9] mtd: rawnand: denali: refactor syndrome layout handling for raw access

[Masahiro Yamada]

Hi Miquel,


On Fri, Mar 15, 2019 at 5:34 PM Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
>
> Hi Masahiro,
>
> Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote on Thu, 14 Mar
> 2019 17:24:41 +0900:
>
> > Hi Miquel,
> >
> > On Tue, Mar 12, 2019 at 10:13 PM Miquel Raynal
> > <miquel.raynal@xxxxxxxxxxx> wrote:
> > >
> > > Hi Masahiro,
> > >
> > > Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote on Tue, 12 Mar
> > > 2019 20:07:27 +0900:
> > >
> > > > Hi Miquel,
> > > >
> > > >
> > > > On Tue, Mar 12, 2019 at 7:54 PM Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> > > > >
> > > > > Hi Masahiro,
> > > > >
> > > > > Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote on Tue, 12 Mar
> > > > > 2019 19:51:21 +0900:
> > > > >
> > > > > > On Tue, Mar 12, 2019 at 7:28 PM Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > Hi Masahiro,
> > > > > > >
> > > > > > > Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote on Tue, 12 Mar
> > > > > > > 2019 17:44:43 +0900:
> > > > > > >
> > > > > > > > The Denali IP adopts the syndrome page layout (payload and ECC are
> > > > > > > > interleaved). The *_page_raw() and *_oob() callbacks are complicated
> > > > > > > > because they must hide the underlying layout used by the hardware,
> > > > > > > > and always return contiguous in-band and out-of-band data.
> > > > > > > >
> > > > > > > > Currently, similar code is duplicated to reorganize the data layout.
> > > > > > > > For example, denali_read_page_raw() and denali_write_page_raw() look
> > > > > > > > almost the same.
> > > > > > > >
> > > > > > > > The idea for refactoring is to split the code into two parts:
> > > > > > > >   [1] conversion of page layout
> > > > > > > >   [2] what to do at every ECC chunk boundary
> > > > > > > >
> > > > > > > > For [1], I wrote denali_raw_payload_op() and denali_raw_oob_op().
> > > > > > > > They manipulate data for the Denali controller's specific page layout
> > > > > > > > of in-band, out-of-band, respectively.
> > > > > > > >
> > > > > > > > The difference between write and read is just the operation at
> > > > > > > > ECC chunk boundaries. For example, denali_read_oob() calls
> > > > > > > > nand_change_read_column_op(), whereas denali_write_oob() calls
> > > > > > > > nand_change_write_column_op(). So, I implemented [2] as a callback
> > > > > > > > passed into [1].
> > > > > > > >
> > > > > > > > Signed-off-by: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> > > > > > > > ---
> > > > > > > >
> > > > > > >
> > > > > > > [...]
> > > > > > >
> > > > > > > >  static int denali_read_page_raw(struct nand_chip *chip, uint8_t *buf,
> > > > > > > >                               int oob_required, int page)
> > > > > > > >  {
> > > > > > > > +     struct denali_nand_info *denali = to_denali(chip);
> > > > > > > >       struct mtd_info *mtd = nand_to_mtd(chip);
> > > > > > > > -     struct denali_nand_info *denali = mtd_to_denali(mtd);
> > > > > > > > -     int writesize = mtd->writesize;
> > > > > > > > -     int oobsize = mtd->oobsize;
> > > > > > > > -     int ecc_steps = chip->ecc.steps;
> > > > > > > > -     int ecc_size = chip->ecc.size;
> > > > > > > > -     int ecc_bytes = chip->ecc.bytes;
> > > > > > > >       void *tmp_buf = denali->buf;
> > > > > > > > -     int oob_skip = denali->oob_skip_bytes;
> > > > > > > > -     size_t size = writesize + oobsize;
> > > > > > > > -     int ret, i, pos, len;
> > > > > > > > +     size_t size = mtd->writesize + mtd->oobsize;
> > > > > > > > +     int ret;
> > > > > > > > +
> > > > > > > > +     if (!buf)
> > > > > > > > +             return -EINVAL;
> > > > > > > >
> > > > > > > >       ret = denali_data_xfer(chip, tmp_buf, size, page, 1, 0);
> > > > > > > >       if (ret)
> > > > > > > >               return ret;
> > > > > > > >
> > > > > > > > -     /* Arrange the buffer for syndrome payload/ecc layout */
> > > > > > > > -     if (buf) {
> > > > > > > > -             for (i = 0; i < ecc_steps; i++) {
> > > > > > > > -                     pos = i * (ecc_size + ecc_bytes);
> > > > > > > > -                     len = ecc_size;
> > > > > > > > -
> > > > > > > > -                     if (pos >= writesize)
> > > > > > > > -                             pos += oob_skip;
> > > > > > > > -                     else if (pos + len > writesize)
> > > > > > > > -                             len = writesize - pos;
> > > > > > > > -
> > > > > > > > -                     memcpy(buf, tmp_buf + pos, len);
> > > > > > > > -                     buf += len;
> > > > > > > > -                     if (len < ecc_size) {
> > > > > > > > -                             len = ecc_size - len;
> > > > > > > > -                             memcpy(buf, tmp_buf + writesize + oob_skip,
> > > > > > > > -                                    len);
> > > > > > > > -                             buf += len;
> > > > > > > > -                     }
> > > > > > > > -             }
> > > > > > > > -     }
> > > > > > > > +     ret = denali_raw_payload_op(chip, buf, denali_memcpy_in, tmp_buf);
> > > > > > >
> > > > > > > Honestly, I still don't like passing denali_memcpy_in/out as parameter.
> > > > > > >
> > > > > > > Besides that, once you'll have added helpers to avoid abusing the
> > > > > > > ternary operator in 4/9, the rest looks fine by me.
> > > > > > >
> > > > > >
> > > > > >
> > > > > > Do you have any suggestion?
> > > > >
> > > > > Maybe register these two helpers at probe as controller specific hooks,
> > > > > then just pass an in/out boolean to the function?
> > > > >
> > > >
> > > > Sorry, I do not understand.
> > > >
> > > > Are you suggesting to do like follows in probe ?
> > > >
> > > > denali->change_column_read_raw = denali_memcpy_in;
> > > > denali->change_column_write_raw = denali_memcpy_out;
> > > > denali->change_column_read_oob = denali_change_read_column_op;
> > > > denali->change_column_write_oob = denali_change_write_column_op;
> > > >
> > > >
> > > > All the 4 hooks are always needed
> > > > regardless of any probed features.
> > > >
> > > >
> > > > The result is just textual replacement
> > > > denali_* with denali->*.
> > > >
> > > > What's the point of copying fixed function addresses
> > > > to denali structure?
> > > >
> > > >
> > >
> > > What I don't like is the function pointer as a function parameter.
> >
> > This is a usual way to handle callback.
> >
> > > You
> > > can use the functions defined statically if you prefer as long as the
> > > parameter is just a boolean for instance?
> >
> >
> >
> > I still do not understand your concern,
> > but if you ban the use of function pointer,
> > the following is the best I can do
> > since there are 4 hooks depending on the
> > combination of oob/raw, write/read.
> >
> >
> >
> > if (oob) {
> >         if (write)
> >                 return nand_change_write_column_op(chip, offset, buf,
> >                                                    len, false);
> >         else
> >                 return nand_change_read_column_op(chip, offset, buf,
> >                                                   len, false);
> > }
> >
> > if (write)
> >         memcpy(denali->buf + offset, buf, len);
> > else
> >         memcpy(buf, denali->buf + offset, len);
> >
> > return 0;
>
> No, I meant passing a boolean to denali_raw_payload_op() instead of a
> function pointer. Then from denali_raw_payload_op(), intead of doing
>
> ret = cb();
> if (ret)
>         ...
>
> doing:
>
> if (read)
>         ret = denali_memcpy_in()
> else
>         ret = denali_memcpy_out()
>
> if (ret)
>         ...


If you look at my code closely,
you will notice 4 callbacks passed in
denali_raw_payload_op().

So, if-conditional would end up like follows:


    if (oob) {
            if (write)
                    ret = nand_change_write_column_op(chip, offset, buf,
                                                   len, false);
            else
                    ret = nand_change_read_column_op(chip, offset, buf,
                                                  len, false);
            if (ret)
                    return ret;
     } else {
            if (write)
                  memcpy(denali->buf + offset, buf, len);
            else
                  memcpy(buf, denali->buf + offset, len);
     }


This is extremely ugly.
That's why I passed a function pointer
instead of two boolean parameters 'oob', 'write'.




> But nevermind, if this is bothering you too much let's keep the current
> form, it's fine.
>
> >
> >
> > BTW, when are .read_page_raw / .write_page_raw used?
>
> I'm not sure what is the question here but these hooks are important
> and allow to test the driver. nandbiterrs use them (although we do
> not care about the performance in these hooks).


Currently, I use DMA transfer + memcpy()
in order to get better performance for
.read_page_raw() and .write_page_raw()


nand_change_write_column_op() and
nand_change_read_column_op() are slow
since they are low-level hardware accesses.


If we do not have to care about the performance,
I will only use nand_change_{write,read}_column_op().




> >
> > Currently, I use "whole page access && memcpy" for better performance.
> >
> > If those hooks are rarely used, I use
> > nand_change_write_column_op / nand_change_read_column_op,
> > which will reduce the if-conditional.
>
> Yes you can. We do not care about performance in raw accessors.

OK, I will do this in v4.




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



-- 
Best Regards
Masahiro Yamada