Re: [PATCH bpf-next 1/2] bpf: Stop JIT passes only when no insn moves

[Mykyta Yatsenko <mykyta.yatsenko5@gmail.com>]

On 10/21/26 11:15 PM, Alexei Starovoitov wrote:
> On Wed, Oct 21, 2026 at 07:21 AM Mykyta Yatsenko <mykyta.yatsenko5@gmail.com> wrote:
>> A backward jump is sized from its target's current-pass address and its
>> own previous-pass address, so it can grow while other jumps shrink. Two
>> passes may then give the same image length with jumps still changing,
>> and the image pass fails with "bpf_jit: fatal error".
> 
> can we fix the sizing of the backward jump instead ?
> 
> It's the third time it bites us.
> commit 1612a981b766 ("bpf, x64: fix JIT emission for dead code")
> special cased ja -1 and
> commit c8831bdbfbab ("bpf, x64: Fix a jit convergence issue")
> limited imm8 to 123 because of it.
> With this patch jumps still grow, and the one that grows in a padding
> pass still ends in "unexpected jmp_cond padding: -4 bytes".
> 
> do_jit() knows how far the current pass moved insn i. With that taken
> into account a backward jump is sized by the insns it jumps over, the
> same way as a forward one. Insn sizes start at 64 and only go down, so
> the image shrinks with every pass like the comment in
> bpf_int_jit_compile() says, and proglen == oldproglen means that
> nothing moved.
> Then no need for addrs_changed, for the ja -1 special case and
> for is_imm8_jmp_offset().
> 
> Something like this (compile tested only):

I'll send v2. Thanks for the suggestion.

> pw-bot: cr