Re: MIPS eBPF JIT support on pre-32R2

From: Alexander Lobakin
Date: Fri Dec 06 2019 - 02:21:21 EST


Paul Burton wrote 05.12.2019 21:44:
Hi Alexander,

Hi Paul!

On Thu, Dec 05, 2019 at 03:45:27PM +0300, Alexander Lobakin wrote:
Hey all,

I'm writing about lines arch/mips/net/ebpf_jit.c:1806-1807:

if (!prog->jit_requested || MIPS_ISA_REV < 2)
return prog;

Do pre-32R2 architectures (32R1, maybe even R3000-like) actually support
this eBPF JIT code?

No, they don't; the eBPF JIT makes unconditional use of at least the
(d)ins & (d)ext instructions which were added in MIPSr2, so it would
result in reserved instruction exceptions & panics if enabled on
pre-MIPSr2 CPUs.

If they do, then the condition 'MIPS_ISA_REV < 2'
should be removed as it is always true for them and tells CC to remove
JIT completely.

If they don't support instructions from this JIT, then the line
arch/mips/Kconfig:50:

select HAVE_EBPF_JIT if (!CPU_MICROMIPS)

should be changed to something like:

select HAVE_EBPF_JIT if !CPU_MICROMIPS && TARGET_ISA_REV >= 2

(and then the mentioned 'if' condition would become redundant)

Good spot; I agree entirely, this dependency should be reflected in
Kconfig.

At the moment it is possible to build a kernel without both JIT and
interpreter, but with CONFIG_BPF_SYSCALL=y (what should not be allowed
I suppose?) within the following configuration:

- select any pre-32R2 CPU (e.g. CONFIG_CPU_MIPS32_R1);
- enable CONFIG_BPF_JIT (CONFIG_MIPS_EBPF_JIT will be autoselected);
- enable CONFIG_BPF_JIT_ALWAYS_ON (this removes BPF interpreter from
the system).

I may prepare a proper patch by myself if needed (after clarification).

That would be great, thanks!

Great, I'll send it in about ~2-3 hours.

One thing to note is that I hope we'll restore the cBPF JIT with this
patch:

https://lore.kernel.org/linux-mips/20191205182318.2761605-1-paulburton@xxxxxxxxxx/T/#u

The cBPF JIT looks like it should work on older pre-MIPSr2 CPUs, so the
only way this is relevant is that your patch might have a minor
conflict. But I thought I'd mention it anyway :)

Yes, I thought about this too. If pre-32R2 CPUs don't support our eBPF
JIT, we'd better restore cBPF for them, so they could speed-up at least
"classic" instructions. Glad you've decided to do that.

Thanks,
Paul

Regards,
á á á á á á