Re: [PATCH bpf] selftests/bpf: fix endianness issues in test_sysctl

[Yonghong Song <yhs@fb.com>]

On 8/16/19 4:37 AM, Ilya Leoshkevich wrote:
>> Am 16.08.2019 um 02:05 schrieb Yonghong Song <yhs@fb.com>:
>>
>>> +# define __bpf_constant_be64_to_cpu(x)	___constant_swab64(x)
>>
>> bpf_endian.h is used for both bpf program and native applications.
>> Could you make sure it works for bpf programs? It should be, but want to
>> double check.
> 
> Yes:
> 
> #include <linux/compiler_attributes.h>
> #include "bpf_endian.h"
> u64 answer() { return __bpf_constant_be64_to_cpu(42); }
> 
> compiles to
> 
>          r0 = 3026418949592973312 ll
>          exit
> 
> on x86.
> 
>> The __constant_swab64 looks like a little bit expensive
>> for bpf programs compared to __builtin_bswap64. But
>> __builtin_bswap64 may not be available for all architectures, esp.
>> 32bit system. So macro __bpf__ is required to use it.
> 
> Isn't ___constant_swab64 supposed to be 100% compile-time?
> 
> Also, I think __builtin_bswap64 should be available everywhere for
> userspace. At least the following test does not indicate any problems:
> 
> for cc in "x86_64-linux-gnu-gcc -m32" \
>            "x86_64-linux-gnu-gcc -m64" \
>            "aarch64-linux-gnu-gcc" \
>            "arm-linux-gnueabihf-gcc" \
>            "mips64el-linux-gnuabi64-gcc" \
>            "powerpc64le-linux-gnu-gcc -m32" \
>            "s390x-linux-gnu-gcc -m31" \
>            "s390x-linux-gnu-gcc -m64" \
>            "sparc64-linux-gnu-gcc -m32" \
>            "sparc64-linux-gnu-gcc -m64" \
>            "clang -target bpf -m32" \
>            "clang -target bpf -m64"; do
> 	echo "*** $cc ***"
> 	echo "long long f(long long x) { return __builtin_bswap64(x); }" | \
> 	$cc -x c -S - -O3 -o -;
> done
> 
> Only sparc64 doesn't support it directly, but then it just calls
> libgcc's __bswapdi2. This might not be ok only for kernel native code
> (though even there we have e.g. arch/arm/lib/bswapsdi2.S), but I don't
> think this header is used in such context anyway.

Great to know. Maybe we can define
    __bpf_be64_to_cpu  // using __builtin_bswap64
    __bpf_constant_be64_to_cpu   // use your above definition
    bpf_be64_to_cpu(x)   // check whether x is __builtin_constant_p()
                         // or not, and then call the above two.

bpf_be64_to_cpu() can be used in test_sysctl.c.

> 
>>>
>>>   			BPF_MOV64_REG(BPF_REG_1, BPF_REG_7),
>>> @@ -1344,20 +1379,26 @@ static size_t probe_prog_length(const struct bpf_insn *fp)
>>>   static int fixup_sysctl_value(const char *buf, size_t buf_len,
>>>   			      struct bpf_insn *prog, size_t insn_num)
>>>   {
>>> -	uint32_t value_num = 0;
>>> +	uint64_t value_num = 0;
>>>   	uint8_t c, i;
>>>
>>>   	if (buf_len > sizeof(value_num)) {
>>>   		log_err("Value is too big (%zd) to use in fixup", buf_len);
>>>   		return -1;
>>>   	}
>>> +	if (prog[insn_num].code != (BPF_LD | BPF_DW | BPF_IMM)) {
>>> +		log_err("Can fixup only BPF_LD_IMM64 insns");
>>> +		return -1;
>>> +	}
>>>
>>>   	for (i = 0; i < buf_len; ++i) {
>>>   		c = buf[i];
>>>   		value_num |= (c << i * 8);
>>>   	}
>>> +	value_num = __bpf_le64_to_cpu(value_num);
>>
>> Can we avoid to use __bpf_le64_to_cpu?
>> Look like we already having the value in buf, can we just cast it
>> to get value_num. Note that bpf program and host always have
>> the same endianness. This way, no endianness conversion
>> is needed.
> 
> I think this might be dangerous in case buf is smaller than 8 bytes.

Instead of calculating the value_num as the above, maybe we could
do something like below:

       union {
               uint8_t values[sizeof(__u64)];
               __u64 val;
       } u = {};

       memcpy(u.values, buf, buf_len);

       /* u.val should hold a u64 value which you can use
        * for LD_IMM64 can use.
        */