Messages in this thread | | | Date | Thu, 27 Mar 2014 13:30:47 -0700 | Subject | Re: [PATCH net-next v3 8/9] net: filter: rework/optimize internal BPF interpreter's instruction set | From | Alexei Starovoitov <> |
| |
On Thu, Mar 27, 2014 at 12:23 PM, David Miller <davem@davemloft.net> wrote: > From: Daniel Borkmann <dborkman@redhat.com> > Date: Wed, 26 Mar 2014 22:06:09 +0100 > >> - Adds swab insns for 32/64-bit > > I don't like this. > > You don't want a swab instruction, you want "endian X to endian Y". Just > like we have "cpu_to_le32()", "le32_to_cpu()" et al. in the kernel. > > That way the user can be completely oblivious as to the endianness of > the cpu it's running on.
Agreed. If we want to expose ebpf instruction set to the userspace as-is then it completely makes sense.
The reason we picked swab style is two fold: 1. swab32/64 maps as-is to native cpu instructions during jit. 2. assumption was that new_socket_filters/tracing_filters/ovs_filters will be written in C and will use standard macro ntohl() that will be expanded to __builtin_swap or nop by gcc/llvm, so users don't need to worry about underlying endianness while writing filters in C.
> So if you ask for a "to little endian" swab, if the chip is > little-endian then no code needs to be emitted at all, it's a nop. > > There is zero reason for the BPF program emitted by userspace to be > dependant upon the cpu endianness.
Agree. with cpu_to_le and cpu_to_be instructions the programs written in C and written directly in ebpf won't need to worry about endianness. Best of both.
We'll fix and resend.
Thanks Alexei
| |