[PATCH v2] arm64: kernel: implement fast refcount checking

[Ard Biesheuvel]

On 26 July 2017 at 05:11, Li Kun <hw.likun at huawei.com> wrote:
> Hi Ard，
>
>
> on 2017/7/26 2:15, Ard Biesheuvel wrote:
>>
>> +#define REFCOUNT_OP(op, asm_op, cond, l, clobber...)                   \
>> +__LL_SC_INLINE int                                                     \
>> +__LL_SC_PREFIX(__refcount_##op(int i, atomic_t *r))                    \
>> +{                                                                      \
>> +       unsigned long tmp;                                              \
>> +       int result;                                                     \
>> +                                                                       \
>> +       asm volatile("// refcount_" #op "\n"                            \
>> +"      prfm            pstl1strm, %2\n"                                \
>> +"1:    ldxr            %w0, %2\n"                                      \
>> +"      " #asm_op "     %w0, %w0, %w[i]\n"                              \
>> +"      st" #l "xr      %w1, %w0, %2\n"                                 \
>> +"      cbnz            %w1, 1b\n"                                      \
>> +       REFCOUNT_CHECK(cond)                                            \
>> +       : "=&r" (result), "=&r" (tmp), "+Q" (r->counter)                \
>> +       : REFCOUNT_INPUTS(r) [i] "Ir" (i)                               \
>> +       clobber);                                                       \
>> +                                                                       \
>> +       return result;                                                  \
>> +}                                                                      \
>> +__LL_SC_EXPORT(__refcount_##op);
>> +
>> +REFCOUNT_OP(add_lt, adds, mi,  , REFCOUNT_CLOBBERS);
>> +REFCOUNT_OP(sub_lt_neg, adds, mi, l, REFCOUNT_CLOBBERS);
>> +REFCOUNT_OP(sub_le_neg, adds, ls, l, REFCOUNT_CLOBBERS);
>> +REFCOUNT_OP(sub_lt, subs, mi, l, REFCOUNT_CLOBBERS);
>> +REFCOUNT_OP(sub_le, subs, ls, l, REFCOUNT_CLOBBERS);
>> +
>
> I'm not quite sure if we use b.lt to judge whether the result of adds is
> less than zero is correct or not.
> The b.lt means N!=V, take an extreme example, if we operate like below, the
> b.lt will also be true.
>
> refcount_set(&ref_c,0x80000000);
> refcount_dec_and_test(&ref_c);
>
> maybe we should use PL/NE/MI/EQ to judge the LT_ZERO or LE_ZERO condition ?
>

The lt/le is confusing here: the actual condition coded used are mi
for negative and ls for negative or zero.

I started out using lt and le, because it matches the x86 code, but I
moved to mi and ls instead. (I don't think it makes sense to deviate
from that just because the flags and predicates work a bit
differently.)

However, I see now that there is one instance of REFCOUNT_CHECK(lt)
remaining (in refcount.h). That should mi as well.

Thanks,
Ard.