On Wed, Mar 26, 2025 at 4:41 AM Linus Torvalds
[off-list ref] wrote:

On Tue, 25 Mar 2025 at 05:17, [off-list ref] wrote:

quoted

The rv64ilp32 abi kernel accommodates the lp64 abi userspace and
leverages the lp64 abi Linux interface. Hence, unify the
BITS_PER_LONG = 32 memory layout to match BITS_PER_LONG = 64.

No.

This isn't happening.

You can't do crazy things in the RISC-V code and then expect the rest
of the kernel to just go "ok, we'll do crazy things".

We're not doing crazy __riscv_xlen hackery with random structures
containing 64-bit values that the kernel then only looks at the low 32
bits. That's wrong on *so* many levels.

I'm willing to say "big-endian is dead", but I'm not willing to accept
this kind of crazy hackery.

Not today, not ever.

If you want to run a ilp32 kernel on 64-bit hardware (and support
64-bit ABI just in a 32-bit virtual memory size), I would suggest you

 (a) treat the kernel as natively 32-bit (obviously you can then tell
the compiler to use the rv64 instructions, which I presume you're
already doing - I didn't look)

I used CONFIG_32BIT in v1 and v2, but I've abandoned them because,
based on CONFIG_64BIT, I gain more functionality by inheriting the
lp64-abi kernel. I want the full functionality of the CONFIG_64BIT
Linux kernel, which can be equivalent, used interchangeably, and
seamlessly.

 (b) look at making the compat stuff do the conversion the "wrong way".

And btw, that (b) implies *not* just ignoring the high bits. If
user-space gives 64-bit pointer, you don't just treat it as a 32-bit
one by dropping the high bits. You add some logic to convert it to an
invalid pointer so that user space gets -EFAULT.

Thanks for the advice. I'm looking at how to make the compat stuff.

-- 
Best Regards
 Guo Ren