On Tue, 25 Mar 2025 13:41:30 PDT (-0700), Linus Torvalds 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.

FWIW: this has come up a few times and we've generally said "nobody 
wants this", but that doesn't seem to stick...

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.

OK, maybe that will stick ;)

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)

 (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.

            Linus