arm64: vdso: do cntvct workaround in the VDSO

hulk inclusion
category: performance
bugzilla: 16082
CVE: NA

-------------------------------------------------

If a cntvct workaround is enabled, read CNTVCT_EL0 twice
in VDSO to avoid the clock bug.

Test code:

static unsigned long long getcycle(void)
{
    unsigned long long cval;
    asm volatile("isb" : : : "memory");
    asm volatile("mrs %0, cntvct_el0" : "=r" (cval));
}

int main(void)
{
    int i;
    struct  timeval tv;
    struct  timezone tz;
    unsigned long long s = getcycle();

    for (i = 0; i < 100000000; i++) {
            gettimeofday(&tv,&tz);
    }
    printf("cost:%lld\n", getcycle() - s);

    return 0;
}

Before this patchset, it costs 75.78s:
[root@localhost yang]# ./gettime
cost:3789000522 (20ns per cycle)

After this patchset, it costs 3.58s:
[root@localhost yang]# ./gettime
cost:183208254 (20ns per cycle)
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Reviewed-by: NXuefeng Wang <wxf.wang@hisilicon.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>

arch/arm64/kernel/vdso/gettimeofday.c

@@ -29,6 +29,7 @@
 #include <linux/hrtimer.h>
 
 extern struct vdso_data _vdso_data;
+DECLARE_STATIC_KEY_FALSE(vdso_fix);
 
 static notrace int gettimeofday_fallback(struct timeval *_tv,
 					 struct timezone *_tz)
@@ -119,6 +120,17 @@ static notrace u64 get_clock_shifted_nsec(u64 cycle_last, u64 mult)
 	/* Read the virtual counter. */
 	isb();
 	asm volatile("mrs %0, cntvct_el0" : "=r" (res) :: "memory");
+	if (static_branch_unlikely(&vdso_fix)) {
+		u64 new;
+		int retries = 50;
+
+		asm volatile("mrs %0, cntvct_el0" : "=r" (new) :: "memory");
+		while (unlikely((new - res) >> 5) && retries) {
+			asm volatile("mrs %0, cntvct_el0" : "=r" (res) :: "memory");
+			asm volatile("mrs %0, cntvct_el0" : "=r" (new) :: "memory");
+			retries--;
+		}
+	}
 
 	res = res - cycle_last;
 	/* We can only guarantee 56 bits of precision. */