The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
93c5aecc35