SEEK_DATA should fail at EOF (was: coreutils-9.6-1 (TEST): cp: infinite SEEK_SET/DATA/HOLE loop if file is compressed)

[Christian Franke]

Christian Franke via Cygwin wrote:
> Testcase:
>
> $ uname -r
> 3.5.7-1.x86_64
>
> $ cygcheck -f /bin/cp.exe
> coreutils-9.6-1
>
> $ for i in 1 2 3; do cat /bin/cygwin1.dll > file$i; done
>
> $ compact /C file2 # NTFS compression
> ... (1.7 : 1) ...
>
> $ compact /C /EXE:LZX file3 # Compact OS LZX compression
> ... (2.8 : 1) ...
>
> $ stat -c '%b %s %n' file?
> 2928 2995253 file1
> 1720 2995253 file2
> 1044 2995253 file3
>
> $ cp file1 copy1 # OK
>
> $ cp file2 copy2 # Hangs
> ...[^C]
>
> $ cp file3 copy3 # Hangs
> ...[^C]
>
> $ md5sum file? copy?
> 2954646a9a0fe4579c3fc1f44dd4bb6a *file1
> 2954646a9a0fe4579c3fc1f44dd4bb6a *file2
> 2954646a9a0fe4579c3fc1f44dd4bb6a *file3
> 2954646a9a0fe4579c3fc1f44dd4bb6a *copy1
> 2954646a9a0fe4579c3fc1f44dd4bb6a *copy2
> 2954646a9a0fe4579c3fc1f44dd4bb6a *copy3
>
> $ (sleep 2; pskill strace) & strace cp file3 copy3
> ...
>    47 2004141 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 0) # 
> SEEK_SET
>    46 2004187 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253 # EOF
>    47 2004234 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 3) # 
> SEEK_DATA
>    46 2004280 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253
>    47 2004327 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 4) # 
> SEEK_HOLE
>    46 2004373 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253
>    46 2004419 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 0)
>    51 2004470 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253
>    47 2004517 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 3)
>    47 2004564 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253
>    47 2004611 [main] cp 5546 lseek: 2995253 = lseek(3, 2995253, 4)
>    46 2004657 [main] cp 5546 fhandler_base::lseek: setting file 
> pointer to 2995253
> Process strace killed.
>
>
> file1/2 are detected as a possible sparse files but the optimized copy 
> algorithm does not properly handle the non-sparse case. 

Should be "file2/3" of course.

> Upstream bug?
>

Possibly not. A closer look shows that the main loop in 
copy.c:lseek_copy() expects that SEEK_DATA fails with ENXIO at EOF.

https://github.com/coreutils/coreutils/blob/v9.6/src/copy.c#L543

  lseek_copy(..., off_t ext_start, ...)
  {
    ...
    while (0 <= ext_start) {
      {
       ...
       ext_start = lseek (src_fd, dest_pos, SEEK_DATA);
       if (ext_start < 0 && errno != ENXIO)
         goto cannot_lseek;
      }
    ...
}

This works on Linux (checked on Debian 12) but Cygwin returns the offset 
if it is equal to the file size.

Recent POSIX says:
"[ENXIO] The whence argument is SEEK_HOLE or SEEK_DATA, and offset is 
greater than or equal to the file size"
https://pubs.opengroup.org/onlinepubs/9799919799/functions/lseek.html

But (at least older) Linux man pages suggest that Cygwin behavior may be 
correct also:
"In the simplest implementation, a filesystem can support the operations 
by making ... SEEK_DATA always return offset."
"ENXIO - whence is SEEK_DATA or SEEK_HOLE, and offset is beyond the end 
of the file"
https://man7.org/linux/man-pages/man2/lseek.2.html

Hmm... does "beyond" mean '>=' or '>' ?

-- 
Regards,
Christian