copy_file_range_linux.go

     1  // Copyright 2020 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package poll
     6  
     7  import (
     8  	"internal/syscall/unix"
     9  	"sync/atomic"
    10  	"syscall"
    11  )
    12  
    13  var copyFileRangeSupported int32 = -1 // accessed atomically
    14  
    15  const maxCopyFileRangeRound = 1 << 30
    16  
    17  func kernelVersion() (major int, minor int) {
    18  	var uname syscall.Utsname
    19  	if err := syscall.Uname(&uname); err != nil {
    20  		return
    21  	}
    22  
    23  	rl := uname.Release
    24  	var values [2]int
    25  	vi := 0
    26  	value := 0
    27  	for _, c := range rl {
    28  		if '0' <= c && c <= '9' {
    29  			value = (value * 10) + int(c-'0')
    30  		} else {
    31  			// Note that we're assuming N.N.N here.  If we see anything else we are likely to
    32  			// mis-parse it.
    33  			values[vi] = value
    34  			vi++
    35  			if vi >= len(values) {
    36  				break
    37  			}
    38  			value = 0
    39  		}
    40  	}
    41  	switch vi {
    42  	case 0:
    43  		return 0, 0
    44  	case 1:
    45  		return values[0], 0
    46  	case 2:
    47  		return values[0], values[1]
    48  	}
    49  	return
    50  }
    51  
    52  // CopyFileRange copies at most remain bytes of data from src to dst, using
    53  // the copy_file_range system call. dst and src must refer to regular files.
    54  func CopyFileRange(dst, src *FD, remain int64) (written int64, handled bool, err error) {
    55  	if supported := atomic.LoadInt32(&copyFileRangeSupported); supported == 0 {
    56  		return 0, false, nil
    57  	} else if supported == -1 {
    58  		major, minor := kernelVersion()
    59  		if major > 5 || (major == 5 && minor >= 3) {
    60  			atomic.StoreInt32(&copyFileRangeSupported, 1)
    61  		} else {
    62  			// copy_file_range(2) is broken in various ways on kernels older than 5.3,
    63  			// see issue #42400 and
    64  			// https://man7.org/linux/man-pages/man2/copy_file_range.2.html#VERSIONS
    65  			atomic.StoreInt32(&copyFileRangeSupported, 0)
    66  			return 0, false, nil
    67  		}
    68  	}
    69  	for remain > 0 {
    70  		max := remain
    71  		if max > maxCopyFileRangeRound {
    72  			max = maxCopyFileRangeRound
    73  		}
    74  		n, err := copyFileRange(dst, src, int(max))
    75  		switch err {
    76  		case syscall.ENOSYS:
    77  			// copy_file_range(2) was introduced in Linux 4.5.
    78  			// Go supports Linux >= 2.6.33, so the system call
    79  			// may not be present.
    80  			//
    81  			// If we see ENOSYS, we have certainly not transferred
    82  			// any data, so we can tell the caller that we
    83  			// couldn't handle the transfer and let them fall
    84  			// back to more generic code.
    85  			//
    86  			// Seeing ENOSYS also means that we will not try to
    87  			// use copy_file_range(2) again.
    88  			atomic.StoreInt32(&copyFileRangeSupported, 0)
    89  			return 0, false, nil
    90  		case syscall.EXDEV, syscall.EINVAL, syscall.EIO, syscall.EOPNOTSUPP, syscall.EPERM:
    91  			// Prior to Linux 5.3, it was not possible to
    92  			// copy_file_range across file systems. Similarly to
    93  			// the ENOSYS case above, if we see EXDEV, we have
    94  			// not transferred any data, and we can let the caller
    95  			// fall back to generic code.
    96  			//
    97  			// As for EINVAL, that is what we see if, for example,
    98  			// dst or src refer to a pipe rather than a regular
    99  			// file. This is another case where no data has been
   100  			// transferred, so we consider it unhandled.
   101  			//
   102  			// If src and dst are on CIFS, we can see EIO.
   103  			// See issue #42334.
   104  			//
   105  			// If the file is on NFS, we can see EOPNOTSUPP.
   106  			// See issue #40731.
   107  			//
   108  			// If the process is running inside a Docker container,
   109  			// we might see EPERM instead of ENOSYS. See issue
   110  			// #40893. Since EPERM might also be a legitimate error,
   111  			// don't mark copy_file_range(2) as unsupported.
   112  			return 0, false, nil
   113  		case nil:
   114  			if n == 0 {
   115  				// If we did not read any bytes at all,
   116  				// then this file may be in a file system
   117  				// where copy_file_range silently fails.
   118  				// https://lore.kernel.org/linux-fsdevel/20210126233840.GG4626@dread.disaster.area/T/#m05753578c7f7882f6e9ffe01f981bc223edef2b0
   119  				if written == 0 {
   120  					return 0, false, nil
   121  				}
   122  				// Otherwise src is at EOF, which means
   123  				// we are done.
   124  				return written, true, nil
   125  			}
   126  			remain -= n
   127  			written += n
   128  		default:
   129  			return written, true, err
   130  		}
   131  	}
   132  	return written, true, nil
   133  }
   134  
   135  // copyFileRange performs one round of copy_file_range(2).
   136  func copyFileRange(dst, src *FD, max int) (written int64, err error) {
   137  	// The signature of copy_file_range(2) is:
   138  	//
   139  	// ssize_t copy_file_range(int fd_in, loff_t *off_in,
   140  	//                         int fd_out, loff_t *off_out,
   141  	//                         size_t len, unsigned int flags);
   142  	//
   143  	// Note that in the call to unix.CopyFileRange below, we use nil
   144  	// values for off_in and off_out. For the system call, this means
   145  	// "use and update the file offsets". That is why we must acquire
   146  	// locks for both file descriptors (and why this whole machinery is
   147  	// in the internal/poll package to begin with).
   148  	if err := dst.writeLock(); err != nil {
   149  		return 0, err
   150  	}
   151  	defer dst.writeUnlock()
   152  	if err := src.readLock(); err != nil {
   153  		return 0, err
   154  	}
   155  	defer src.readUnlock()
   156  	var n int
   157  	for {
   158  		n, err = unix.CopyFileRange(src.Sysfd, nil, dst.Sysfd, nil, max, 0)
   159  		if err != syscall.EINTR {
   160  			break
   161  		}
   162  	}
   163  	return int64(n), err
   164  }
   165
View as plain text