On Wed, Sep 24, 2025, at 17:23, Guntupalli, Manikanta wrote:

quoted

Subject: Re: [PATCH V7 3/4] i3c: master: Add endianness support for i3c_readl_fifo() and i3c_writel_fifo()

quoted

}

With this approach, both little-endian and big-endian cases works as expected.

This version should fix the cases where you have a big-endian kernel with either
I3C_FIFO_BIG_ENDIAN or I3C_FIFO_LITTLE_ENDIAN, as neither combination
does any byte swaps.

However I'm fairly sure it's still broken for little-endian kernels when a driver asks for
a I3C_FIFO_BIG_ENDIAN conversion, same as v7.

We tested using the I3C_FIFO_BIG_ENDIAN flag from the driver on 
little-endian kernels, and it works as expected.

Can you explain how that works? What I see is that your
readsl_be()/writesl_be() functions do a byteswap on
every four bytes, so the bytestream that gets copied
to/from the FIFO gets garbled, in particular the
final (unaligned) bytes of the kernel buffer end up
in the higher bytes of the FIFO register rather than
the first bytes as they do on a big-endian kernel.

Are both the big-endian and little-endian kernels in
your tests on microblaze, using the upstream version
of asm/io.h? Is there a hardware byteswap between the
CPU local bus and the i3c controller? If there is
one, is it set the same way for both kernels?

    Arnd