• Ian Lance Taylor's avatar
    runtime: call amd64 VDSO entry points on large stack · a158382b
    Ian Lance Taylor authored
    If the Linux kernel was built with CONFIG_OPTIMIZE_INLINING=n and was
    built with hardening options turned on, GCC will insert a stack probe
    in the VDSO function that requires a full page of stack space.
    The stack probe can corrupt memory if another thread is using it.
    Avoid sporadic crashes by calling the VDSO on the g0 or gsignal stack.
    
    While we're at it, align the stack as C code expects. We've been
    getting away with a misaligned stack, but it's possible that the VDSO
    code will change in the future to break that assumption.
    
    Benchmarks show a 11% hit on time.Now, but it's only 6ns.
    
    name                      old time/op  new time/op  delta
    AfterFunc-12              1.66ms ± 0%  1.66ms ± 1%     ~     (p=0.905 n=9+10)
    After-12                  1.90ms ± 6%  1.86ms ± 0%   -2.05%  (p=0.012 n=10+8)
    Stop-12                    113µs ± 3%   115µs ± 2%   +1.60%  (p=0.017 n=9+10)
    SimultaneousAfterFunc-12   145µs ± 1%   144µs ± 0%   -0.68%  (p=0.002 n=10+8)
    StartStop-12              39.5µs ± 3%  40.4µs ± 5%   +2.19%  (p=0.023 n=10+10)
    Reset-12                  10.2µs ± 0%  10.4µs ± 0%   +2.45%  (p=0.000 n=10+9)
    Sleep-12                   190µs ± 1%   190µs ± 1%     ~     (p=0.971 n=10+10)
    Ticker-12                 4.68ms ± 2%  4.64ms ± 2%   -0.83%  (p=0.043 n=9+10)
    Now-12                    48.4ns ±11%  54.0ns ±11%  +11.42%  (p=0.017 n=10+10)
    NowUnixNano-12            48.5ns ±13%  56.9ns ± 8%  +17.30%  (p=0.000 n=10+10)
    Format-12                  489ns ±11%   504ns ± 6%     ~     (p=0.289 n=10+10)
    FormatNow-12               436ns ±23%   480ns ±13%  +10.25%  (p=0.026 n=9+10)
    MarshalJSON-12             656ns ±14%   587ns ±24%     ~     (p=0.063 n=10+10)
    MarshalText-12             647ns ± 7%   638ns ± 9%     ~     (p=0.516 n=10+10)
    Parse-12                   348ns ± 8%   328ns ± 9%   -5.66%  (p=0.030 n=10+10)
    ParseDuration-12           136ns ± 9%   140ns ±11%     ~     (p=0.425 n=10+10)
    Hour-12                   14.8ns ± 6%  15.6ns ±11%     ~     (p=0.085 n=10+10)
    Second-12                 14.0ns ± 6%  14.3ns ±12%     ~     (p=0.443 n=10+10)
    Year-12                   32.4ns ±11%  33.4ns ± 6%     ~     (p=0.492 n=10+10)
    Day-12                    41.5ns ± 9%  42.3ns ±12%     ~     (p=0.239 n=10+10)
    
    Fixes #20427
    
    Change-Id: Ia395cbb863215f4499b8e7ef95f4b99f51090911
    Reviewed-on: https://go-review.googlesource.com/76990Reviewed-by: 's avatarAustin Clements <austin@google.com>
    a158382b
sys_linux_amd64.s 15.4 KB