Re: [PATCH v4] scsi: arm: Convert acornscsi-io.S to C

[Breno Rodrigues Alves]

Hi Jonathan,

Fair question - the change is a conversion, and I do not have a RiscPC
with an Acorn SCSI card to test on real hardware.  Here is what I did
test:

1. Userspace unit test with a mocked MMIO layer

   I built a small harness that compiles acornscsi-io.c unmodified
   against stub <linux/io.h> and <linux/types.h> headers.  The stubs
   map readl()/writel()/writeb() onto a byte-addressable "card" buffer
   and a plain userspace memory buffer.  For each transfer length L in
   {0, 2, 4, 8, 16, 32, 64, 100, 512, 1000}:

     - __acornscsi_in: I filled the "card" with L bytes of a known
       pattern (each 32-bit card word carrying one 16-bit halfword in
       the low half, garbage in the high half) and verified that the
       output buffer matched the pattern byte-for-byte.

     - __acornscsi_out: I filled a userspace buffer with a known
       pattern, called the function, and verified that each 32-bit card
       word held the expected halfword duplicated in both halves.

   All 20 cases pass.  This is the harness that caught the two issues
   the sashiko bot later reported independently - it is not just a
   smoke test.

2. Cross-compile for the target

   Built with arm-linux-gnueabi-gcc on rpc_defconfig with
   CONFIG_SCSI_ACORNSCSI_3=m.  The module links cleanly
   (acornscsi_mod.ko), and nm on acornscsi-io.o shows:

     00000000 T __acornscsi_in
     00000104 T __acornscsi_out

   objdump on the resulting object shows that the alignment step
   compiles to "bic sl, r0, #3", matching the original assembly's
   "bic r0, r0, #3".

3. Apply to a clean tree and rebuild

   Copied the tree, git reset --hard to the parent commit, applied the
   patch with "git am".  It applies cleanly; re-running the same
   compile command produces the same object.  This confirms the patch
   file itself is intact and does not depend on my working tree state.

4. checkpatch.pl --strict --no-tree on the patch itself

   Reports 0 errors, 0 warnings (with FILE_PATH_CHANGES ignored, as
   drivers/scsi/arm/ is already covered by MAINTAINERS).

What I did NOT test:

 - Real hardware.  I do not have an Acorn RiscPC or the SCSI expansion
   card.  The unit test proves equivalence with the original assembly
   semantics, but it cannot prove that the semantics match the actual
   hardware - only that the C version does the same thing the assembly
   did.

 - Odd transfer lengths through acornscsi.c.  All current callers pass
   even lengths, so the trailing-byte path is only exercised by the
   unit test, not by the driver end-to-end.

 - Interrupt handling under load.  This is unchanged by the patch (the
   IRQ path is untouched), so I did not add specific coverage.

If there is a specific test or configuration you would like me to run,
please let me know.

Breno