Re: [PATCH v7 net-next 1/3] filter: add Extended BPF interpreter and converter

[Alexei Starovoitov <ast@plumgrid.com>]

On Sun, Mar 9, 2014 at 3:00 PM, Daniel Borkmann <borkmann@iogearbox.net> wrote:
> On 03/09/2014 06:08 PM, Alexei Starovoitov wrote:
>>
>> On Sun, Mar 9, 2014 at 5:29 AM, Daniel Borkmann <borkmann@iogearbox.net>
>> wrote:
>>>
>>> On 03/09/2014 12:15 AM, Alexei Starovoitov wrote:
>>>>
>>>>
>>>> Extended BPF extends old BPF in the following ways:
>>>> - from 2 to 10 registers
>>>>     Original BPF has two registers (A and X) and hidden frame pointer.
>>>>     Extended BPF has ten registers and read-only frame pointer.
>>>> - from 32-bit registers to 64-bit registers
>>>>     semantics of old 32-bit ALU operations are preserved via 32-bit
>>>>     subregisters
>>>> - if (cond) jump_true; else jump_false;
>>>>     old BPF insns are replaced with:
>>>>     if (cond) jump_true; /* else fallthrough */
>>>> - adds signed > and >= insns
>>>> - 16 4-byte stack slots for register spill-fill replaced with
>>>>     up to 512 bytes of multi-use stack space
>>>> - introduces bpf_call insn and register passing convention for zero
>>>>     overhead calls from/to other kernel functions (not part of this
>>>> patch)
>>>> - adds arithmetic right shift insn
>>>> - adds swab32/swab64 insns
>>>> - adds atomic_add insn
>>>> - old tax/txa insns are replaced with 'mov dst,src' insn
>>>>
>>>> Extended BPF is designed to be JITed with one to one mapping, which
>>>> allows GCC/LLVM backends to generate optimized BPF code that performs
>>>> almost as fast as natively compiled code
>>>>
>>>> sk_convert_filter() remaps old style insns into extended:
>>>> 'sock_filter' instructions are remapped on the fly to
>>>> 'sock_filter_ext' extended instructions when
>>>> sysctl net.core.bpf_ext_enable=1
>>>>
>>>> Old filter comes through sk_attach_filter() or
>>>> sk_unattached_filter_create()
>>>>    if (bpf_ext_enable) {
>>>>       convert to new
>>>>       sk_chk_filter() - check old bpf
>>>>       use sk_run_filter_ext() - new interpreter
>>>>    } else {
>>>>       sk_chk_filter() - check old bpf
>>>>       if (bpf_jit_enable)
>>>>           use old jit
>>>>       else
>>>>           use sk_run_filter() - old interpreter
>>>>    }
>>>>
>>>> sk_run_filter_ext() interpreter is noticeably faster
>>>> than sk_run_filter() for two reasons:
>>>>
>>>> 1.fall-through jumps
>>>>     Old BPF jump instructions are forced to go either 'true' or 'false'
>>>>     branch which causes branch-miss penalty.
>>>>     Extended BPF jump instructions have one branch and fall-through,
>>>>     which fit CPU branch predictor logic better.
>>>>     'perf stat' shows drastic difference for branch-misses.
>>>>
>>>> 2.jump-threaded implementation of interpreter vs switch statement
>>>>     Instead of single tablejump at the top of 'switch' statement, GCC
>>>> will
>>>>     generate multiple tablejump instructions, which helps CPU branch
>>>> predictor
>>>>
>>>> Performance of two BPF filters generated by libpcap was measured
>>>> on x86_64, i386 and arm32.
>>>>
>>>> fprog #1 is taken from Documentation/networking/filter.txt:
>>>> tcpdump -i eth0 port 22 -dd
>>>>
>>>> fprog #2 is taken from 'man tcpdump':
>>>> tcpdump -i eth0 'tcp port 22 and (((ip[2:2] - ((ip[0]&0xf)<<2)) -
>>>>      ((tcp[12]&0xf0)>>2)) != 0)' -dd
>>>>
>>>> Other libpcap programs have similar performance differences.
>>>>
>>>> Raw performance data from BPF micro-benchmark:
>>>> SK_RUN_FILTER on same SKB (cache-hit) or 10k SKBs (cache-miss)
>>>> time in nsec per call, smaller is better
>>>> --x86_64--
>>>>            fprog #1  fprog #1   fprog #2  fprog #2
>>>>            cache-hit cache-miss cache-hit cache-miss
>>>> old BPF     90       101       192       202
>>>> ext BPF     31        71       47         97
>>>> old BPF jit 12        34       17         44
>>>> ext BPF jit TBD
>>>>
>>>> --i386--
>>>>            fprog #1  fprog #1   fprog #2  fprog #2
>>>>            cache-hit cache-miss cache-hit cache-miss
>>>> old BPF    107        136      227       252
>>>> ext BPF     40        119       69       172
>>>>
>>>> --arm32--
>>>>            fprog #1  fprog #1   fprog #2  fprog #2
>>>>            cache-hit cache-miss cache-hit cache-miss
>>>> old BPF    202        300      475       540
>>>> ext BPF    180        270      330       470
>>>> old BPF jit 26        182       37       202
>>>> new BPF jit TBD
>>>>
>>>> Tested with trinify BPF fuzzer
>>>>
>>>> Future work:
>>>>
>>>> 0. add bpf/ebpf testsuite to tools/testing/selftests/net/bpf
>>>>
>>>> 1. add extended BPF JIT for x86_64
>>>>
>>>> 2. add inband old/new demux and extended BPF verifier, so that new
>>>> programs
>>>>      can be loaded through old sk_attach_filter() and
>>>> sk_unattached_filter_create()
>>>>      interfaces
>>>>
>>>> 3. tracing filters systemtap-like with extended BPF
>>>>
>>>> 4. OVS with extended BPF
>>>>
>>>> 5. nftables with extended BPF
>>>>
>>>> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
>>>> Acked-by: Hagen Paul Pfeifer <hagen@jauu.net>
>>>> Reviewed-by: Daniel Borkmann <dborkman@redhat.com>
>>>
>>>
>>>
>>> One more question or possible issue that came through my mind: When
>>> someone attaches a socket filter from user space, and bpf_ext_enable=1
>>> then the old filter will transparently be converted to the new
>>> representation. If then user space (e.g. through checkpoint restore)
>>> will issue a sk_get_filter() and thus we're calling sk_decode_filter()
>>> on sk->sk_filter and, therefore, try to decode what we stored in
>>> insns_ext[] with the assumption we still have the old code. Would that
>>> actually crash (or leak memory, or just return garbage), as we access
>>> decodes[] array with filt->code? Would be great if you could
>>> double-check.
>>
>>
>> ohh. yes. missed that.
>> when bpf_ext_enable=1 I think it's cleaner to return ebpf filter.
>> This way the user space can see how old bpf filter was converted.
>>
>> Of course we can allocate extra memory and keep original bpf code there
>> just to return it via sk_get_filter(), but that seems overkill.
>
>
> Cc'ing Pavel for a8fc92778080 ("sk-filter: Add ability to get socket
> filter program (v2)").
>
> I think the issue can be that when applications could get migrated
> from one machine to another and their kernel won't support ebpf yet,
> then filter could not get loaded this way as it's expected to return
> what the user loaded. The trade-off, however, is that the original
> BPF code needs to be stored as well. :(

I see.
...even on one machine:
bpf_ext=1, attach, get_filter, bpf_ext=0, re-attach...
So we need to save original.
At least we don't need to keep it for 'unattached' filters.
Should memory come from sk_optmem budget or plain kmalloc is enough ?
Latter would have simpler implementation, but former is probably cleaner?

Thanks
Alexei