[LTP] [PATCH v2] BPF: Regression test for 64bit arithmetic

[Richard Palethorpe <rpalethorpe@suse.de>]

Hello,

Cyril Hrubis <chrubis@suse.cz> writes:

> Hi!
>> > Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com>
>> > +static int load_prog(int fd)
>> > +{
>> > +	struct bpf_insn *prog;
>> > +	struct bpf_insn insn[] = {
>> > +		BPF_MOV64_IMM(BPF_REG_6, 1),            /* r6 = 1 */
>> > +
>> > +		BPF_LD_MAP_FD(BPF_REG_1, fd),	        /* r1 = &fd */
>> > +		BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),   /* r2 = fp */
>> > +		BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -8),  /* r2 = r2 - 8 */
>> > +		BPF_ST_MEM(BPF_DW, BPF_REG_2, 0, 0),    /* *r2 = 0 */
>> > +		BPF_EMIT_CALL(BPF_FUNC_map_lookup_elem),
>> > +		BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 17), /* if(!r0) goto exit */
>>
>> Patch looks good to me.
>>
>> But I keep thinking if there's way to make it more obvious where
>> offset (e.g. 17) came from.
>>
>> Idea 1: use multiple lines per instruction to denote length
>>   BPF_LD_IMM64(BPF_REG_4,
>>                A64INT),
>>
>> Idea 2: prefix commented instructions with offset
>>         /* 1: r3 = r0 */
>>         /* 2: r4 = 2^61 */
>
> I guess I like the Idea 2 better.
>
> Another option would be having eBPF assembler included in the LTP build
> system. I guess that it may be useful later on and there seems to be one
> written in python:
>
> https://github.com/solarflarecom/ebpf_asm
>
> But for the short term I would go with adding the offset to the
> comments.

Another idea I had was to use place holder values in the instruction
array which can be substituted.

Infact we could use an invalid instruction code to indicate a goto
instruction and another code for a tag. Then replace the tag with a NOP
and replace the goto with a valid jump statement.

I decided not to try any of that because it seemed like overkill at the
time. However for a more complex program I can see this getting very
confusing.

I wonder if an assembler will make things better in some tests and worse
in others. Sometimes the assembler has pseudo instructions where it is
not clear what the resulting machine instructions will be.

--
Thank you,
Richard.