Re: [PATCH 03/12] KVM: x86: Add a common TSC scaling function

[Paolo Bonzini <pbonzini@redhat.com>]

On 28/09/2015 07:38, Haozhong Zhang wrote:
>  
> -static u64 __scale_tsc(u64 ratio, u64 tsc)
> -{
> -	u64 mult, frac, _tsc;
> -
> -	mult  = ratio >> 32;
> -	frac  = ratio & ((1ULL << 32) - 1);
> -
> -	_tsc  = tsc;
> -	_tsc *= mult;
> -	_tsc += (tsc >> 32) * frac;
> -	_tsc += ((tsc & ((1ULL << 32) - 1)) * frac) >> 32;
> -
> -	return _tsc;
> -}

This is basically

	return mul_u64_u64_shr(ratio, tsc,
			       kvm_tsc_scaling_ratio_frac_bits);

except that Linux has no mul_u64_u64_shr function, only mul_u64_u32_shr.

We should implement that function in include/linux/math64.h instead.
For the x86_64 case (or any other CONFIG_ARCH_SUPPORTS_INT128
architecture) we can just write it directly, as is done already for
mul_u64_u32_shr.

For the 32-bit case, here is an implementation of both the
multiplication and the shift, lifted from QEMU:

static inline void mul64(uint64_t *lo, uint64_t *hi,
                         uint64_t a, uint64_t b)
{
    typedef union {
        uint64_t ll;
        struct {
#ifdef __BIG_ENDIAN
            uint32_t high, low;
#else
            uint32_t low, high;
#endif
        } l;
    } LL;
    LL rl, rm, rn, rh, a0, b0;
    uint64_t c;

    a0.ll = a;
    b0.ll = b;

    rl.ll = (uint64_t)a0.l.low * b0.l.low;
    rm.ll = (uint64_t)a0.l.low * b0.l.high;
    rn.ll = (uint64_t)a0.l.high * b0.l.low;
    rh.ll = (uint64_t)a0.l.high * b0.l.high;

    c = (uint64_t)rl.l.high + rm.l.low + rn.l.low;
    rl.l.high = c;
    c >>= 32;
    c = c + rm.l.high + rn.l.high + rh.l.low;
    rh.l.low = c;
    rh.l.high += (uint32_t)(c >> 32);

    *lo = rl.ll;
    *hi = rh.ll;
}

static inline void rshift128(uint64_t *lo, uint64_t *hi, int n)
{
    uint64_t h;
    if (!n) {
        return;
    }
    h = *hi >> (n & 63);
    if (n >= 64) {
        *hi = 0;
        *lo = h;
    } else {
        *lo = (*lo >> n) | (*hi << (64 - n));
        *hi = h;
    }
}

and you can easily reuse this code in Linux with just uintNN_t types
changed to uNN + some extra cleanups when it's placed in a single functions.

Paolo