net: use splice(2) on Linux when reading from UnixConn, rework splice tests

Rework the splice tests and benchmarks. Move the reading and writing of
the spliced connections to child processes so that the I/O is not part
of benchmarks or profiles.

Enable the use of splice(2) when reading from a unix connection and
writing to a TCP connection. The updated benchmarks show a performance
gain when using splice(2) to copy large chunks of data that the original
benchmark did not capture.

  name                          old time/op    new time/op    delta
  Splice/tcp-to-tcp/1024-8        5.01µs ± 2%    5.08µs ± 3%      ~     (p=0.068 n=8+10)
  Splice/tcp-to-tcp/2048-8        4.76µs ± 5%    4.65µs ± 3%    -2.36%  (p=0.015 n=9+8)
  Splice/tcp-to-tcp/4096-8        4.91µs ± 2%    4.98µs ± 5%      ~     (p=0.315 n=9+10)
  Splice/tcp-to-tcp/8192-8        5.50µs ± 4%    5.44µs ± 3%      ~     (p=0.758 n=7+9)
  Splice/tcp-to-tcp/16384-8       7.65µs ± 7%    6.53µs ± 3%   -14.65%  (p=0.000 n=10+9)
  Splice/tcp-to-tcp/32768-8       15.3µs ± 7%     8.5µs ± 5%   -44.21%  (p=0.000 n=10+10)
  Splice/tcp-to-tcp/65536-8       30.0µs ± 6%    15.7µs ± 1%   -47.58%  (p=0.000 n=10+8)
  Splice/tcp-to-tcp/131072-8      59.2µs ± 2%    27.4µs ± 5%   -53.75%  (p=0.000 n=9+9)
  Splice/tcp-to-tcp/262144-8       121µs ± 4%      54µs ±19%   -55.56%  (p=0.000 n=9+10)
  Splice/tcp-to-tcp/524288-8       247µs ± 6%     108µs ±12%   -56.34%  (p=0.000 n=10+10)
  Splice/tcp-to-tcp/1048576-8      490µs ± 4%     199µs ±12%   -59.31%  (p=0.000 n=8+10)
  Splice/unix-to-tcp/1024-8       1.20µs ± 2%    1.35µs ± 7%   +12.47%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/2048-8       1.33µs ±12%    1.57µs ± 4%   +17.85%  (p=0.000 n=9+10)
  Splice/unix-to-tcp/4096-8       2.24µs ± 4%    1.67µs ± 4%   -25.14%  (p=0.000 n=9+10)
  Splice/unix-to-tcp/8192-8       4.59µs ± 8%    2.20µs ±10%   -52.01%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/16384-8      8.46µs ±13%    3.48µs ± 6%   -58.91%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/32768-8      18.5µs ± 9%     6.1µs ± 9%   -66.99%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/65536-8      35.9µs ± 7%    13.5µs ± 6%   -62.40%  (p=0.000 n=10+9)
  Splice/unix-to-tcp/131072-8     79.4µs ± 6%    25.7µs ± 4%   -67.62%  (p=0.000 n=10+9)
  Splice/unix-to-tcp/262144-8      157µs ± 4%      54µs ± 8%   -65.63%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/524288-8      311µs ± 3%     107µs ± 8%   -65.74%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/1048576-8     643µs ± 4%     185µs ±32%   -71.21%  (p=0.000 n=10+10)

  name                          old speed      new speed      delta
  Splice/tcp-to-tcp/1024-8       204MB/s ± 2%   202MB/s ± 3%      ~     (p=0.068 n=8+10)
  Splice/tcp-to-tcp/2048-8       430MB/s ± 5%   441MB/s ± 3%    +2.39%  (p=0.014 n=9+8)
  Splice/tcp-to-tcp/4096-8       833MB/s ± 2%   823MB/s ± 5%      ~     (p=0.315 n=9+10)
  Splice/tcp-to-tcp/8192-8      1.49GB/s ± 4%  1.51GB/s ± 3%      ~     (p=0.758 n=7+9)
  Splice/tcp-to-tcp/16384-8     2.14GB/s ± 7%  2.51GB/s ± 3%   +17.03%  (p=0.000 n=10+9)
  Splice/tcp-to-tcp/32768-8     2.15GB/s ± 7%  3.85GB/s ± 5%   +79.11%  (p=0.000 n=10+10)
  Splice/tcp-to-tcp/65536-8     2.19GB/s ± 5%  4.17GB/s ± 1%   +90.65%  (p=0.000 n=10+8)
  Splice/tcp-to-tcp/131072-8    2.22GB/s ± 2%  4.79GB/s ± 4%  +116.26%  (p=0.000 n=9+9)
  Splice/tcp-to-tcp/262144-8    2.17GB/s ± 4%  4.93GB/s ±17%  +127.25%  (p=0.000 n=9+10)
  Splice/tcp-to-tcp/524288-8    2.13GB/s ± 6%  4.89GB/s ±13%  +130.15%  (p=0.000 n=10+10)
  Splice/tcp-to-tcp/1048576-8   2.09GB/s ±10%  5.29GB/s ±11%  +153.36%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/1024-8      850MB/s ± 2%   757MB/s ± 7%   -10.94%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/2048-8     1.54GB/s ±11%  1.31GB/s ± 3%   -15.32%  (p=0.000 n=9+10)
  Splice/unix-to-tcp/4096-8     1.83GB/s ± 4%  2.45GB/s ± 4%   +33.59%  (p=0.000 n=9+10)
  Splice/unix-to-tcp/8192-8     1.79GB/s ± 9%  3.73GB/s ± 9%  +108.05%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/16384-8    1.95GB/s ±13%  4.68GB/s ± 3%  +139.80%  (p=0.000 n=10+9)
  Splice/unix-to-tcp/32768-8    1.78GB/s ± 9%  5.38GB/s ±10%  +202.71%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/65536-8    1.83GB/s ± 8%  4.85GB/s ± 6%  +165.70%  (p=0.000 n=10+9)
  Splice/unix-to-tcp/131072-8   1.65GB/s ± 6%  5.10GB/s ± 4%  +208.77%  (p=0.000 n=10+9)
  Splice/unix-to-tcp/262144-8   1.67GB/s ± 4%  4.87GB/s ± 7%  +191.19%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/524288-8   1.69GB/s ± 3%  4.93GB/s ± 7%  +192.38%  (p=0.000 n=10+10)
  Splice/unix-to-tcp/1048576-8  1.63GB/s ± 3%  5.60GB/s ±44%  +243.26%  (p=0.000 n=10+9)

Change-Id: I1eae4c3459c918558c70fc42283db22ff7e0442c
Reviewed-on: https://go-review.googlesource.com/113997Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

src/net/splice_linux.go

@@ -11,7 +11,7 @@ import (
 
 // splice transfers data from r to c using the splice system call to minimize
 // copies from and to userspace. c must be a TCP connection. Currently, splice
-// is only enabled if r is also a TCP connection.
+// is only enabled if r is a TCP or Unix connection.
 //
 // If splice returns handled == false, it has performed no work.
 func splice(c *netFD, r io.Reader) (written int64, err error, handled bool) {
@@ -23,11 +23,17 @@ func splice(c *netFD, r io.Reader) (written int64, err error, handled bool) {
 			return 0, nil, true
 		}
 	}
-	s, ok := r.(*TCPConn)
-	if !ok {
+
+	var s *netFD
+	if tc, ok := r.(*TCPConn); ok {
+		s = tc.fd
+	} else if uc, ok := r.(*UnixConn); ok {
+		s = uc.fd
+	} else {
 		return 0, nil, false
 	}
-	written, handled, sc, err := poll.Splice(&c.pfd, &s.fd.pfd, remain)
+
+	written, handled, sc, err := poll.Splice(&c.pfd, &s.pfd, remain)
 	if lr != nil {
 		lr.N -= written
 	}