[PATCH] bpf: use the expected instruction for NOPs

[PATCH] bpf: use the expected instruction for NOPs

[David Faust]

The BPF ISA does not have a no-op instruction. There are different ways
to achieve no-op, but in practice the Linux kernel verifier expects a
particular form, namely "ja 0". Update the define md_single_noop_insn
for BPF to reflect this.

For more detail, see the corresponding GCC back end patch:
https://gcc.gnu.org/pipermail/gcc-patches/2020-September/553893.html

gas/
	* config/tc-bpf.h (md_single_noop_insn): Use 'ja 0' for no-op.

---
 gas/config/tc-bpf.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/gas/config/tc-bpf.h b/gas/config/tc-bpf.h
index cb02d6c1338..e7e505cfaa5 100644
--- a/gas/config/tc-bpf.h
+++ b/gas/config/tc-bpf.h
@@ -48,4 +48,6 @@
 /* Values passed to md_apply_fix don't include the symbol value.  */
 #define MD_APPLY_SYM_VALUE(FIX) 0
 
-#define md_single_noop_insn "mov %r1,%r1"
+/* The Linux kernel verifier expects NOPs to be encoded in this way;
+   a jump to offset 0 means jump to the next instruction.  */
+#define md_single_noop_insn "ja 0"
-- 
2.26.2

Re: [PATCH] bpf: use the expected instruction for NOPs

[Nick Clifton]

Hi David,

> The BPF ISA does not have a no-op instruction.

Just an aside - wouldn't it be a good idea if the ISA did define an explicit
encoding for no-ops ?

> There are different ways
> to achieve no-op, but in practice the Linux kernel verifier expects a
> particular form, namely "ja 0". Update the define md_single_noop_insn
> for BPF to reflect this.

Thanks for pointing this out.  I took the original encoding from the
gas/testsuite/gas/all/org-1.s file, in the hopes that it would be correct
one.
 
> gas/
> 	* config/tc-bpf.h (md_single_noop_insn): Use 'ja 0' for no-op.

Approved and applied.

Cheers
  Nick

Re: [PATCH] bpf: use the expected instruction for NOPs

[David Faust]

Hi Nick!

>> The BPF ISA does not have a no-op instruction.
> 
> Just an aside - wouldn't it be a good idea if the ISA did define an explicit
> encoding for no-ops ?

In my opinion, yes that would be good. But, eBPF is effectively defined
by its implementation in the kernel, so we have to support whatever the
kernel decides for the ISA.

> 
>> There are different ways
>> to achieve no-op, but in practice the Linux kernel verifier expects a
>> particular form, namely "ja 0". Update the define md_single_noop_insn
>> for BPF to reflect this.
> 
> Thanks for pointing this out.  I took the original encoding from the
> gas/testsuite/gas/all/org-1.s file, in the hopes that it would be correct
> one.
>  

I wasn't actually aware of this until yesterday. It was your patch
adding the .nop directive that prompted digging into precisely what the
kernel expects for no-op; thank you.

(The kernel eBPF verifier has an optimization built into it to remove
no-ops, and it defines no-op as 'ja 0'.)

>> gas/
>> 	* config/tc-bpf.h (md_single_noop_insn): Use 'ja 0' for no-op.
> 
> Approved and applied.
> 
> Cheers
>   Nick
> 

Thanks!

David