Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
net: filter: rework/optimize internal BPF interpreter's instruction set
This patch replaces/reworks the kernel-internal BPF interpreter with
an optimized BPF instruction set format that is modelled closer to
mimic native instruction sets and is designed to be JITed with one to
one mapping. Thus, the new interpreter is noticeably faster than the
current implementation of sk_run_filter(); mainly for two reasons:
1. Fall-through jumps:
BPF jump instructions are forced to go either 'true' or 'false'
branch which causes branch-miss penalty. The new BPF jump
instructions have only one branch and fall-through otherwise,
which fits the CPU branch predictor logic better. `perf stat`
shows drastic difference for branch-misses between the old and
new code.
2. Jump-threaded implementation of interpreter vs switch
statement:
Instead of single table-jump at the top of 'switch' statement,
gcc will now generate multiple table-jump instructions, which
helps CPU branch predictor logic.
Note that the verification of filters is still being done through
sk_chk_filter() in classical BPF format, so filters from user- or
kernel space are verified in the same way as we do now, and same
restrictions/constraints hold as well.
We reuse current BPF JIT compilers in a way that this upgrade would
even be fine as is, but nevertheless allows for a successive upgrade
of BPF JIT compilers to the new format.
The internal instruction set migration is being done after the
probing for JIT compilation, so in case JIT compilers are able to
create a native opcode image, we're going to use that, and in all
other cases we're doing a follow-up migration of the BPF program's
instruction set, so that it can be transparently run in the new
interpreter.
In short, the *internal* format extends BPF in the following way (more
details can be taken from the appended documentation):
- Number of registers increase from 2 to 10
- Register width increases from 32-bit to 64-bit
- Conditional jt/jf targets replaced with jt/fall-through
- Adds signed > and >= insns
- 16 4-byte stack slots for register spill-fill replaced
with up to 512 bytes of multi-use stack space
- Introduction of bpf_call insn and register passing convention
for zero overhead calls from/to other kernel functions
- Adds arithmetic right shift and endianness conversion insns
- Adds atomic_add insn
- Old tax/txa insns are replaced with 'mov dst,src' insn
Performance of two BPF filters generated by libpcap resp. bpf_asm
was measured on x86_64, i386 and arm32 (other libpcap programs
have similar performance differences):
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
Raw performance data from BPF micro-benchmark: SK_RUN_FILTER on the
same SKB (cache-hit) or 10k SKBs (cache-miss); time in ns 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
new BPF 31 71 47 97
old BPF jit 12 34 17 44
new 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
new 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
new BPF 180 270 330 470
old BPF jit 26 182 37 202
new BPF jit TBD
Thus, without changing any userland BPF filters, applications on
top of AF_PACKET (or other families) such as libpcap/tcpdump, cls_bpf
classifier, netfilter's xt_bpf, team driver's load-balancing mode,
and many more will have better interpreter filtering performance.
While we are replacing the internal BPF interpreter, we also need
to convert seccomp BPF in the same step to make use of the new
internal structure since it makes use of lower-level API details
without being further decoupled through higher-level calls like
sk_unattached_filter_{create,destroy}(), for example.
Just as for normal socket filtering, also seccomp BPF experiences
a time-to-verdict speedup:
05-sim-long_jumps.c of libseccomp was used as micro-benchmark:
seccomp_rule_add_exact(ctx,...
seccomp_rule_add_exact(ctx,...
rc = seccomp_load(ctx);
for (i = 0; i < 10000000; i++)
syscall(199, 100);
'short filter' has 2 rules
'large filter' has 200 rules
'short filter' performance is slightly better on x86_64/i386/arm32
'large filter' is much faster on x86_64 and i386 and shows no
difference on arm32
--x86_64-- short filter
old BPF: 2.7 sec
39.12% bench libc-2.15.so [.] syscall
8.10% bench [kernel.kallsyms] [k] sk_run_filter
6.31% bench [kernel.kallsyms] [k] system_call
5.59% bench [kernel.kallsyms] [k] trace_hardirqs_on_caller
4.37% bench [kernel.kallsyms] [k] trace_hardirqs_off_caller
3.70% bench [kernel.kallsyms] [k] __secure_computing
3.67% bench [kernel.kallsyms] [k] lock_is_held
3.03% bench [kernel.kallsyms] [k] seccomp_bpf_load
new BPF: 2.58 sec
42.05% bench libc-2.15.so [.] syscall
6.91% bench [kernel.kallsyms] [k] system_call
6.25% bench [kernel.kallsyms] [k] trace_hardirqs_on_caller
6.07% bench [kernel.kallsyms] [k] __secure_computing
5.08% bench [kernel.kallsyms] [k] sk_run_filter_int_seccomp
--arm32-- short filter
old BPF: 4.0 sec
39.92% bench [kernel.kallsyms] [k] vector_swi
16.60% bench [kernel.kallsyms] [k] sk_run_filter
14.66% bench libc-2.17.so [.] syscall
5.42% bench [kernel.kallsyms] [k] seccomp_bpf_load
5.10% bench [kernel.kallsyms] [k] __secure_computing
new BPF: 3.7 sec
35.93% bench [kernel.kallsyms] [k] vector_swi
21.89% bench libc-2.17.so [.] syscall
13.45% bench [kernel.kallsyms] [k] sk_run_filter_int_seccomp
6.25% bench [kernel.kallsyms] [k] __secure_computing
3.96% bench [kernel.kallsyms] [k] syscall_trace_exit
--x86_64-- large filter
old BPF: 8.6 seconds
73.38% bench [kernel.kallsyms] [k] sk_run_filter
10.70% bench libc-2.15.so [.] syscall
5.09% bench [kernel.kallsyms] [k] seccomp_bpf_load
1.97% bench [kernel.kallsyms] [k] system_call
new BPF: 5.7 seconds
66.20% bench [kernel.kallsyms] [k] sk_run_filter_int_seccomp
16.75% bench libc-2.15.so [.] syscall
3.31% bench [kernel.kallsyms] [k] system_call
2.88% bench [kernel.kallsyms] [k] __secure_computing
--i386-- large filter
old BPF: 5.4 sec
new BPF: 3.8 sec
--arm32-- large filter
old BPF: 13.5 sec
73.88% bench [kernel.kallsyms] [k] sk_run_filter
10.29% bench [kernel.kallsyms] [k] vector_swi
6.46% bench libc-2.17.so [.] syscall
2.94% bench [kernel.kallsyms] [k] seccomp_bpf_load
1.19% bench [kernel.kallsyms] [k] __secure_computing
0.87% bench [kernel.kallsyms] [k] sys_getuid
new BPF: 13.5 sec
76.08% bench [kernel.kallsyms] [k] sk_run_filter_int_seccomp
10.98% bench [kernel.kallsyms] [k] vector_swi
5.87% bench libc-2.17.so [.] syscall
1.77% bench [kernel.kallsyms] [k] __secure_computing
0.93% bench [kernel.kallsyms] [k] sys_getuid
BPF filters generated by seccomp are very branchy, so the new
internal BPF performance is better than the old one. Performance
gains will be even higher when BPF JIT is committed for the
new structure, which is planned in future work (as successive
JIT migrations).
BPF has also been stress-tested with trinity's BPF fuzzer.
Joint work with Daniel Borkmann.
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Cc: Kees Cook <keescook@chromium.org>
Cc: Paul Moore <pmoore@redhat.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: H. Peter Anvin <hpa@linux.intel.com>
Cc: linux-kernel@vger.kernel.org
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: David S. Miller <davem@davemloft.net>- Loading branch information
