Re: [PATCH 5/6] mtd: spi-nor: core; avoid odd length/address reads on 8D-8D-8D mode

[Pratyush Yadav <p.yadav@xxxxxx>]

On 07/05/21 08:14PM, Michael Walle wrote:
> Am 2021-05-07 20:04, schrieb Pratyush Yadav:
> > On 07/05/21 05:51PM, Michael Walle wrote:
> > > Am 2021-05-06 21:18, schrieb Pratyush Yadav:
> > > > On Octal DTR capable flashes like Micron Xcella reads cannot start or
> > > > end at an odd address in Octal DTR mode. Extra bytes need to be read at
> > > > the start or end to make sure both the start address and length remain
> > > > even.
> > > >
> > > > To avoid allocating too much extra memory, thereby putting unnecessary
> > > > memory pressure on the system, the temporary buffer containing the extra
> > > > padding bytes is capped at PAGE_SIZE bytes. The rest of the 2-byte
> > > > aligned part should be read directly in the main buffer.
> > > >
> > > > Signed-off-by: Pratyush Yadav <p.yadav@xxxxxx>
> > > > ---
> > > >
> > > >  drivers/mtd/spi-nor/core.c | 81 +++++++++++++++++++++++++++++++++++++-
> > > >  1 file changed, 80 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/mtd/spi-nor/core.c b/drivers/mtd/spi-nor/core.c
> > > > index 5cc206b8bbf3..3d66cc34af4d 100644
> > > > --- a/drivers/mtd/spi-nor/core.c
> > > > +++ b/drivers/mtd/spi-nor/core.c
> > > > @@ -1904,6 +1904,82 @@ static const struct flash_info
> > > > *spi_nor_read_id(struct spi_nor *nor)
> > > >  	return ERR_PTR(-ENODEV);
> > > >  }
> > > >
> > > > +/*
> > > > + * On Octal DTR capable flashes like Micron Xcella reads cannot start
> > > > or
> > > > + * end at an odd address in Octal DTR mode. Extra bytes need to be read
> > > > + * at the start or end to make sure both the start address and length
> > > > + * remain even.
> > > > + */
> > > > +static int spi_nor_octal_dtr_read(struct spi_nor *nor, loff_t from,
> > > > size_t len,
> > > > +				  u_char *buf)
> > > > +{
> > > > +	u_char *tmp_buf;
> > > > +	size_t tmp_len;
> > > > +	loff_t start, end;
> > > > +	int ret, bytes_read;
> > > > +
> > > > +	if (IS_ALIGNED(from, 2) && IS_ALIGNED(len, 2))
> > > > +		return spi_nor_read_data(nor, from, len, buf);
> > > > +	else if (IS_ALIGNED(from, 2) && len > PAGE_SIZE)
> > > > +		return spi_nor_read_data(nor, from, round_down(len, PAGE_SIZE),
> > > > +					 buf);
> > > > +
> > > > +	tmp_buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
> > > > +	if (!tmp_buf)
> > > > +		return -ENOMEM;
> > > > +
> > > > +	start = round_down(from, 2);
> > > > +	end = round_up(from + len, 2);
> > > > +
> > > > +	/*
> > > > +	 * Avoid allocating too much memory. The requested read length might
> > > > be
> > > > +	 * quite large. Allocating a buffer just as large (slightly bigger, in
> > > > +	 * fact) would put unnecessary memory pressure on the system.
> > > > +	 *
> > > > +	 * For example if the read is from 3 to 1M, then this will read from 2
> > > > +	 * to 4098. The reads from 4098 to 1M will then not need a temporary
> > > > +	 * buffer so they can proceed as normal.
> > > > +	 */
> > > > +	tmp_len = min_t(size_t, end - start, PAGE_SIZE);
> > > > +
> > > > +	ret = spi_nor_read_data(nor, start, tmp_len, tmp_buf);
> > > > +	if (ret == 0) {
> > > > +		ret = -EIO;
> > > > +		goto out;
> > > > +	}
> > > > +	if (ret < 0)
> > > > +		goto out;
> > > > +
> > > > +	/*
> > > > +	 * More bytes are read than actually requested, but that number can't
> > > > be
> > > > +	 * reported to the calling function or it will confuse its
> > > > calculations.
> > > > +	 * Calculate how many of the _requested_ bytes were read.
> > > > +	 */
> > > > +	bytes_read = ret;
> > > > +
> > > > +	if (from != start)
> > > > +		ret -= from - start;
> > > > +
> > > > +	/*
> > > > +	 * Only account for extra bytes at the end if they were actually read.
> > > > +	 * For example, if the total length was truncated because of temporary
> > > > +	 * buffer size limit then the adjustment for the extra bytes at the
> > > > end
> > > > +	 * is not needed.
> > > > +	 */
> > > > +	if (start + bytes_read == end)
> > > > +		ret -= end - (from + len);
> > > > +
> > > > +	if (ret < 0) {
> > > > +		ret = -EIO;
> > > > +		goto out;
> > > > +	}
> > > > +
> > > > +	memcpy(buf, tmp_buf + (from - start), ret);
> > > > +out:
> > > > +	kfree(tmp_buf);
> > > > +	return ret;
> > > > +}
> > > > +
> > > >  static int spi_nor_read(struct mtd_info *mtd, loff_t from, size_t len,
> > > >  			size_t *retlen, u_char *buf)
> > > >  {
> > > > @@ -1921,7 +1997,10 @@ static int spi_nor_read(struct mtd_info *mtd,
> > > > loff_t from, size_t len,
> > > >
> > > >  		addr = spi_nor_convert_addr(nor, addr);
> > > >
> > > > -		ret = spi_nor_read_data(nor, addr, len, buf);
> > > > +		if (nor->read_proto == SNOR_PROTO_8_8_8_DTR)
> > > > +			ret = spi_nor_octal_dtr_read(nor, addr, len, buf);
> > > > +		else
> > > > +			ret = spi_nor_read_data(nor, addr, len, buf);
> > > >  		if (ret == 0) {
> > > >  			/* We shouldn't see 0-length reads */
> > > >  			ret = -EIO;
> > > 
> > > Reviewed-by: Michael Walle <michael@xxxxxxxx>
> > 
> > Thanks.
> > 
> > > 
> > > I wonder how much performance is lost if this would just split
> > > one transfer into up to three ones: 2 byte, size - 2, 2 bytes.
> > 
> > This case is not really possible since it would try to read PAGE_SIZE
> > whenever it can. But there is a situation possible where one transfer is
> > split into three. It would look something like: 4096 bytes, size - 4096
> > bytes, 2 bytes.
> 
> Ah no, I wasn't talking about your implementation, but just having a naive
> one where you don't move around up to PAGE_SIZE of data but just read
> 2 bytes in the beginning (if unaligned) and 2 bytes at the end (if
> unaligned)
> and reading the part in between just as usual because its then aligend.
> 
> > I am trying to find a balance between minimizing number of reads while
> > keeping the size of the temporary buffer to a reasonable limit. This is
> > the best I could come up with. It optimizes for smaller transfers so
> > while the absolute amount of overhead remains roughly the same, the
> > ratio of it relative to read size is smaller.
> 
> Yes, with this you will have that memcpy() and one transfer for transfers
> up to PAGE_SIZE; the "naive" one above would have up to three depending on
> the aligment.

Right. Smaller transfers lose much more performance to the overhead than 
the larger ones do. So I think the optimization is worth the extra code 
complexity.

> 
> > You can optimize for read performance if you are willing to waste memory
> > by simple allocating a size + 2 bytes long buffer. Then the read can
> > proceed in one transaction. But IMO memory is much more important
> > compared to read throughput.
> 
> -michael

-- 
Regards,
Pratyush Yadav
Texas Instruments Inc.