On Wed, Nov 13, 2019 at 8:53 AM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
> + /* Pack the residue into a 32bit word */
> + for (i = 0; i < host->pio_write_residue_sz; i++) {
> + val |= host->pio_write_residue[i];
> + val <<= 8;
> + }
> + /* Top up with new data */
> + for (i = 0; i < fill; i++) {
> + val |= *ptr;
> + val <<= 8;
> + ptr++;
> + remain--;
> + }
I'm worried that I might have gotten this wrong.
iowrite32_rep() reads the data little-endian (native endianness)
from memory does it not?
Bytes [0 1 2 3] end up in the FIFO like [3 2 1 0].
So it will pack the first byte into the lowest 8 bits, second byte into
bits 8-15 etc.
So I should rewrite all the loops like this:
for (i = 0; i < host->pio_write_residue_sz; i++) {
val |= (host->pio_write_residue[i] << 24);
val >>= 8;
}
So I shift the value down from the high bits instead of the
other way around.
This also gives a pretty plausible hint att what might be wrong with the
DMA in non-divisible by 4.
As suggested by Stephan in another context, I will try to set up my
own test rig for this.
Yours,
Linus Walleij
