> Am 09.07.2019 um 19:48 schrieb Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx>:
>
> On Tue, Jul 9, 2019 at 8:19 AM Ilya Leoshkevich <iii@xxxxxxxxxxxxx> wrote:
>>
>> Right now, on certain architectures, these macros are usable only with
>> kernel headers. This patch makes it possible to use them with userspace
>> headers and, as a consequence, not only in BPF samples, but also in BPF
>> selftests.
>>
>> On s390, provide the forward declaration of struct pt_regs and cast it
>> to user_pt_regs in PT_REGS_* macros. This is necessary, because instead
>> of the full struct pt_regs, s390 exposes only its first member
>> user_pt_regs to userspace, and bpf_helpers.h is used with both userspace
>> (in selftests) and kernel (in samples) headers. It was added in commit
>> 466698e654e8 ("s390/bpf: correct broken uapi for
>> BPF_PROG_TYPE_PERF_EVENT program type").
>>
>> Ditto on arm64.
>>
>> On x86, provide userspace versions of PT_REGS_* macros. Unlike s390 and
>> arm64, x86 provides struct pt_regs to both userspace and kernel, however,
>> with different member names.
>>
>> Signed-off-by: Ilya Leoshkevich <iii@xxxxxxxxxxxxx>
>> ---
>
> Just curious, what did you use as a reference for which register
> corresponds to which PARM, RET, etc for different archs? I've tried to
> look it up the other day, and it wasn't as straightforward to find as
> I hoped for, so maybe I'm missing something obvious.
For this particular change I did not have to look it up, because it all
was already in the code, I just needed to adapt it to userspace headers.
Normally I would google for „abi supplement“ to find this information.
A lazy way would be to simply ask the (cross-)compiler:
cat <<HERE | aarch64-linux-gnu-gcc -x c -O3 -S - -o -
int f(int a, int b, int c, int d, int e, int f, int g, int h, int i, int j);
int g() { return -f(1, 2, 3, 4, 5, 6, 7, 8, 9, 10); }
HERE
I’ve just double checked the supported arches, and noticed that:
#define PT_REGS_PARM5(x) ((x)->uregs[4])
for bpf_target_arm (arm-linux-gnueabihf) looks wrong:
the 5th parameter should be passed on stack. This observation matches
[1].
#define PT_REGS_RC(x) ((x)->regs[1])
for bpf_target_mips (mips64el-linux-gnuabi64) also looks wrong:
the return value should be in register 2. This observation matches [2].
Since I’m not an expert on those architectures, my conclusions could be
incorrect (e.g. becase a different ABI is normally used in practice).
Adrian and David, could you please correct me if I’m wrong?
[1] http://infocenter.arm.com/help/topic/com.arm.doc.ihi0042f/IHI0042F_aapcs.pdf
[2] ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/psABI_mips3.0.pdf
>> tools/testing/selftests/bpf/bpf_helpers.h | 61 +++++++++++++++--------
>> 1 file changed, 41 insertions(+), 20 deletions(-)
>>
>> diff --git a/tools/testing/selftests/bpf/bpf_helpers.h b/tools/testing/selftests/bpf/bpf_helpers.h
>> index 73071a94769a..212ec564e5c3 100644
>> --- a/tools/testing/selftests/bpf/bpf_helpers.h
>> +++ b/tools/testing/selftests/bpf/bpf_helpers.h
>> @@ -358,6 +358,7 @@ static int (*bpf_skb_adjust_room)(void *ctx, __s32 len_diff, __u32 mode,
>>
>> #if defined(bpf_target_x86)
>>
>> +#ifdef __KERNEL__
>> #define PT_REGS_PARM1(x) ((x)->di)
>> #define PT_REGS_PARM2(x) ((x)->si)
>> #define PT_REGS_PARM3(x) ((x)->dx)
>> @@ -368,19 +369,35 @@ static int (*bpf_skb_adjust_room)(void *ctx, __s32 len_diff, __u32 mode,
>> #define PT_REGS_RC(x) ((x)->ax)
>> #define PT_REGS_SP(x) ((x)->sp)
>> #define PT_REGS_IP(x) ((x)->ip)
>> +#else
>> +#define PT_REGS_PARM1(x) ((x)->rdi)
>> +#define PT_REGS_PARM2(x) ((x)->rsi)
>> +#define PT_REGS_PARM3(x) ((x)->rdx)
>> +#define PT_REGS_PARM4(x) ((x)->rcx)
>> +#define PT_REGS_PARM5(x) ((x)->r8)
>> +#define PT_REGS_RET(x) ((x)->rsp)
>> +#define PT_REGS_FP(x) ((x)->rbp)
>> +#define PT_REGS_RC(x) ((x)->rax)
>> +#define PT_REGS_SP(x) ((x)->rsp)
>> +#define PT_REGS_IP(x) ((x)->rip)
>
> Will this also work for 32-bit x86?
Thanks, this is a good catch: this builds, but makes 64-bit accesses, as if it used the 64-bit
variant of pt_regs. I will fix this.
