eBPF Talk: 修复了 bpftool 中存在了 7 年的 BUG

在 eBPF Talk: 自制查看 bpf prog 反汇编的工具里提到的 bpftool BUG Wrong callq address displayed，花了两三天时间将它给修了。

TL;DR [PATCH bpf v3] bpf, bpftool: Fix incorrect disasm pc 已合入 bpf 仓库。

BUG 复现

使用 bpftool 查看 bpf prog 的反汇编：

1
2
3
4
5
6
7
8
9


# bpftool prog dump jited name kprobe_skb_1
...
; event->cpu_id = bpf_get_smp_processor_id();
 c70:   call   0xffffffffd2702ab4
...
# echo "0xffffffffd2702ab4 is wrong for bpf_get_smp_processor_id"
# grep ffffffffd2702ab4 /proc/kallsyms
# grep bpf_get_smp_processor_id /proc/kallsyms
ffffffff92b0b490 T bpf_get_smp_processor_id

翻看了一下 commit 历史，确认所有版本的 bpftool 都存在这个问题，因为从一开始就是错的。

BUG 修复之尝试篇

翻看一下 bpftool prog dump jited 的源代码，找到反汇编的核心代码逻辑：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66


// ${KERNEL}/tools/bpf/bpftool/jit_disasm.c

#ifdef HAVE_LLVM_SUPPORT

static int
disassemble_insn(disasm_ctx_t *ctx, unsigned char *image, ssize_t len, int pc)
{
    char buf[256];
    int count;

    count = LLVMDisasmInstruction(*ctx, image + pc, len - pc, pc,
                                  buf, sizeof(buf));
    if (json_output)
        printf_json(buf);
    else
        printf("%s", buf);

    return count;
}
#endif /* HAVE_LLVM_SUPPORT */

#ifdef HAVE_LIBBFD_SUPPORT

static int
disassemble_insn(disasm_ctx_t *ctx, __maybe_unused unsigned char *image,
                 __maybe_unused ssize_t len, int pc)
{
    return ctx->disassemble(pc, ctx->info);
}

#endif /* HAVE_LIBBPFD_SUPPORT */

int disasm_print_insn(unsigned char *image, ssize_t len, int opcodes,
                      const char *arch, const char *disassembler_options,
                      const struct btf *btf,
                      const struct bpf_prog_linfo *prog_linfo,
                      __u64 func_ksym, unsigned int func_idx,
                      bool linum)
{
    const struct bpf_line_info *linfo = NULL;
    unsigned int nr_skip = 0;
    int count, i, pc = 0;
    disasm_ctx_t ctx;

    if (!len)
        return -1;

    if (init_context(&ctx, arch, disassembler_options, image, len))
        return -1;

    // ...
    do {
        // ...

        count = disassemble_insn(&ctx, image, len, pc);

        // ...

        pc += count;
    } while (count > 0 && pc < len);
    // ...

    destroy_context(&ctx);

    return 0;
}

找到 LLVMDisasmInstruction() 的文档：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13


/**
 * Disassemble a single instruction using the disassembler context specified in
 * the parameter DC.  The bytes of the instruction are specified in the
 * parameter Bytes, and contains at least BytesSize number of bytes.  The
 * instruction is at the address specified by the PC parameter.  If a valid
 * instruction can be disassembled, its string is returned indirectly in
 * OutString whose size is specified in the parameter OutStringSize.  This
 * function returns the number of bytes in the instruction or zero if there was
 * no valid instruction.
 */
size_t LLVMDisasmInstruction(LLVMDisasmContextRef DC, uint8_t *Bytes,
                             uint64_t BytesSize, uint64_t PC,
                             char *OutString, size_t OutStringSize);

这儿写明了 PC 参数需要是 insn 的地址。

所以，将 disassemble_insn(&ctx, image, len, pc); 里的 pc 参数改成 func_ksym + pc 试试：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26


(gdb) run p d j i 378 linum
Starting program: /root/bpftool/src/bpftool p d j i 378 linum
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
int kprobe_skb_1(struct pt_regs * ctx):
bpf_prog_1e1ae32f79e6c86e_kprobe_skb_1:
; PWRU_ADD_KPROBE(1) [file:bpf/kprobe_pwru.c line_num:530 line_col:0]
   0:bptool disas, func ksym=ffffffffc03fffc8, pc=ffffffffc03fffc8(0)

Program received signal SIGSEGV, Segmentation fault.
llvm::support::endian::read<unsigned char, 1ul> (memory=0x555517997578, endian=llvm::endianness::little) at /root/llvm-project/llvm/include/llvm/Support/Endian.h:61
61        memcpy(&ret,
(gdb) bt
#0  llvm::support::endian::read<unsigned char, 1ul> (memory=0x555517997578, endian=llvm::endianness::little) at /root/llvm-project/llvm/include/llvm/Support/Endian.h:61
#1  0x000055555575456b in consume<unsigned char> (insn=0x7fffffffdb70, ptr=@0x7fffffffdacb: 0 '\000') at /root/llvm-project/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp:203
#2  0x000055555574628e in readPrefixes (insn=0x7fffffffdb70) at /root/llvm-project/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp:231
#3  0x000055555574ad72 in (anonymous namespace)::X86GenericDisassembler::getInstruction (this=0x5555575a1190, Instr=..., Size=@0x7fffffffdca8: 93824994447723, Bytes=..., Address=18446744072640004040, CStream=...)
    at /root/llvm-project/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp:1879
#4  0x0000555555ff5d95 in LLVMDisasmInstruction (DCR=0x5555575a1280, Bytes=0x555517997578 <error: Cannot access memory at address 0x555517997578>, BytesSize=1069547794, PC=18446744072640004040, OutString=0x7fffffffdf60 "0",
    OutStringSize=256) at /root/llvm-project/llvm/lib/MC/MCDisassembler/Disassembler.cpp:267
#5  0x00005555556f56c6 in disassemble_insn (image=image@entry=0x5555575975b0 "\017\037D", len=len@entry=218, pc=pc@entry=-1069547576, ctx=<optimized out>) at jit_disasm.c:117
#6  0x00005555556f5927 in disasm_print_insn (image=image@entry=0x5555575975b0 "\017\037D", len=218, opcodes=opcodes@entry=0, arch=arch@entry=0x0, disassembler_options=<optimized out>, btf=btf@entry=0x555557593070,
    prog_linfo=0x55555759c560, func_ksym=18446744072640004040, func_idx=0, linum=true) at jit_disasm.c:364
#7  0x0000555555701f48 in prog_dump (linum=<optimized out>, visual=<optimized out>, opcodes=<optimized out>, filepath=<optimized out>, mode=DUMP_JITED, info=0x7fffffffe1e0) at prog.c:825
#8  do_dump (argc=<optimized out>, argv=<optimized out>) at prog.c:986
#9  0x00005555556e6c9a in main (argc=<optimized out>, argv=<optimized out>) at main.c:539

能看到所有 debug info，是编译 LLVM 和 bpftool 的时候都带上 -g 选项：

1
2
3
4
5
6
7
8


# echo "Build and install llvm, -DCMAKE_BUILD_TYPE=Debug is the key option"
mkdir llvm_build
cmake -S llvm-project/llvm -B llvm_build -DCMAKE_BUILD_TYPE=Debug -DLLVM_ENABLE_PROJECTS="clang;lldb" -DLLVM_TARGETS_TO_BUILD="X86"
make -j64 -C llvm_build llvm-config llvm-libraries clang lldb
sudo make install -j64 -C llvm_build

# echo "Build bpftool with -g"
EXTRA_CFLAGS=-g make -j64 -C src

gdb 里已提示 Bytes=0x555517997578 <error: Cannot access memory at address 0x555517997578>，即提供的 insn buffer 有误。

接着，梳理一下整个函数调用流程，发现不能直接将 pc 调整为 func_ksym + pc；需要通过以下方式才能修复。

BUG 修复之 LLVM 篇

bpftool 支持使用 LLVM 和 libbfd 作为反汇编的后端；不过，都存在这问题。

bpftool: Add LLVM as default library for disassembling JIT-ed programs since 6.2

其中出问题的地方在于：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12


static int
disassemble_insn(disasm_ctx_t *ctx, unsigned char *image, ssize_t len, int pc)
{
    char buf[256];
    int count;

    count = LLVMDisasmInstruction(*ctx, image + pc, len - pc, pc,
                                  buf, sizeof(buf));
    // ...

    return count;
}

提供给 LLVMDisasmInstruction() 的 pc 参数是有错误的。

在 disasm_print_insn() 里，pc 变量指的是当前 bpf prog 里 insn 的索引。

修复方式如下：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10


-disassemble_insn(disasm_ctx_t *ctx, unsigned char *image, ssize_t len, int pc)
+disassemble_insn(disasm_ctx_t *ctx, unsigned char *image, ssize_t len, int pc,
+                 __u64 func_ksym)
 {
     char buf[256];
     int count;

-   count = LLVMDisasmInstruction(*ctx, image + pc, len - pc, pc,
+   count = LLVMDisasmInstruction(*ctx, image + pc, len - pc, func_ksym + pc,
                                  buf, sizeof(buf));

func_ksym + pc 才是 insn 对应的真实地址。

注：func_ksym 是当前 bpf prog 的 image 的起始地址。

BUG 修复之 libbfd 篇

使用 libbfd 作为反汇编的后端，也存在这个问题：

tools: bpf: add bpftool since 4.15

使用 libbfd 反汇编时没处理 relative address 的问题；比如 call, je 等指令使用 relative address 的时候，反汇编出来的地址是错误的。

注：这个 7 年前的 commit，使用 libbfd 作为反汇编的后端，但却忽略了 relative address 的处理。

直接翻看 libbfd 的源代码：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34


// https://github.com/bminor/binutils-gdb/blob/master/opcodes/disassemble.c

disassembler_ftype
disassembler (enum bfd_architecture a,
              bool big ATTRIBUTE_UNUSED,
              unsigned long mach ATTRIBUTE_UNUSED,
              bfd *abfd ATTRIBUTE_UNUSED)
{
    disassembler_ftype disassemble;

    // ...

#ifdef ARCH_ia64
        case bfd_arch_ia64:
            disassemble = print_insn_ia64;
            break;
#endif

    // ...
}

// https://github.com/bminor/binutils-gdb/blob/master/opcodes/ia64-dis.c

int
print_insn_ia64 (bfd_vma memaddr, struct disassemble_info *info)
{
    // ...

        case IA64_OPND_CLASS_REL:
            (*info->print_address_func) (memaddr + value, info);
            break;

    // ...
}

啊哈，当看到 info->print_address_func 用来处理 relative address 时，bpftool 的问题就有解了：提供自己的 print_address_func，然后给传过来的 memaddr 加上 func_ksym。

不过，先看看 print_address_func 初始化：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13


// https://github.com/bminor/binutils-gdb/blob/master/opcodes/dis-init.c

void
init_disassemble_info (struct disassemble_info *info, void *stream,
                       fprintf_ftype fprintf_func,
                       fprintf_styled_ftype fprintf_styled_func)
{
    memset (info, 0, sizeof (*info));

    // ...
    info->print_address_func = generic_print_address;
    // ...
}

所以，解决办法如下：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12


struct disasm_info {
    struct disassemble_info info;
    __u64 func_ksym;
};

static void disasm_print_addr(bfd_vma addr, struct disassemble_info *info)
{
    struct disasm_info *dinfo = container_of(info, struct disasm_info, info);

    addr += dinfo->func_ksym;
    generic_print_address(addr, info);
}

总结

Gray 大佬一句感叹：这些自古以来的 bug 无人在意会感到孤单北半球。

eBPF Talk: 修复了 bpftool 中存在了 7 年的 BUG

文章目录

BUG 复现

BUG 修复之尝试篇

BUG 修复之 LLVM 篇

BUG 修复之 libbfd 篇

总结

知识星球

星球里的专栏：

《XDP 进阶手册》