On Tue, Jun 17, 2014 at 4:21 AM, Daniel Borkmann <firstname.lastname@example.org> wrote:
> On 06/17/2014 01:09 PM, Markos Chandras wrote:
>> Thanks for these instructions. I will try them myself once I find some
>> time since I don't think bpf_jit for MIPS has ever been tested with all
>> the opcodes.
> Sounds great! If you find some tests are missing, please feel free to
> submit them as well via netdev.
thank you for taking care of it so quickly :)
from the BPF perspective the fix looks good:
Acked-by: Alexei Starovoitov <email@example.com>
please do run the testsuite.
Doing quick code review of mips jit, it looks like:
- your version of pkt_type_offset() will work for little endian only.
(we've recently fixed it in net/core/filter.c)
- vlan tag handling is incorrect, since it's missing shifts.
classic BPF standard for vlan_tag_present has to return 1 or 0
and not just emit_and(r_A, r_s0, VLAN_TAG_PRESENT, ctx);
- pr_warn("%s: Unhandled opcode: 0x%02x\n", __FILE__,
is way too heavy, since when jit is on, unprivileged user can spam log.
- /* sa is 5-bits long */
BUG_ON(sa >= BIT(5));
is wrong too. Malicious user can cause kernel crash…
Also shift A>>=33 was always allowed by classic BPF checker, so
JITs have to silently do C-equivalent version of such shift.
- /* Determine if immediate is within the 16-bit signed range */
static inline bool is_range16(s32 imm)
if (imm >= SBIT(15) || imm < -SBIT(15))
the function name and comment are doing the opposite of
actual code, which makes harder to follow.
- the rest looks pretty good!
Also you'll get a lot more mileage out of mips jit if you use eBPF
instruction set as a base for JITing. You wouldn't need to worry
about vlan, pkt_type and other classic extensions. You'll get all
extensions for free, plus seccomp, tracing, etc.