GreenFileDescriptorIO.seek should raise IOError, not OSError

[ricardokirkner]

• gevent version: 1.3.7
• Python version: "cPython 2.7.15"
• Operating System: "Arch on amd64"

Description:

GreenFileDescriptorIO should raise an IOError exception instead of an OSError exception during it's seek method in order to remain compatible with the io module (as it's subclassing RawIOBase).

With the current implementation (using os.lseek internally) raising an OSError will break expectations of the file-like interface. For example, when using gevent in conjunction with boto, boto will catch IOError while trying to seek a file and handle it accordingly, but will fail to notice gevent raising an OSError (as it's not what the __builtins__ or io based file-like objects do).

What I've run:

Trying to seek a non-seekable file descriptor raises an OSError

>>> from gevent._fileobjectposix import GreenFileDescriptorIO
>>> gio = GreenFileDescriptorIO(0)
>>> gio.seek(0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python2.7/site-packages/gevent/_fileobjectposix.py", line 143, in seek
    return os.lseek(self._fileno, offset, whence)
OSError: [Errno 29] Illegal seek

but doing the same with the stdlib raises IOError

>>> import io
>>> f = io.open(0, 'rb')
>>> repr(f)
'<_io.BufferedReader name=0>'
>>> f.fileno()
0
>>> f.seek(0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
IOError: [Errno 29] Illegal seek
>>> import os
>>> f2 = os.fdopen(0, 'rb')
>>> repr(f2)
"<open file '<fdopen>', mode 'rb' at 0x7facbdd185d0>"
>>> f2.fileno()
0
>>> f2.seek(0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
IOError: [Errno 29] Illegal seek

For consistency, GreenFileDescriptorIO.seek should also raise IOError and not OSError.

[jamadden]

Thanks for the report! On Python 3, IOError is just a deprecated alias for OSError, so this issue only affects Python 2 (and only on POSIX platforms). Python 2 does document that seek should raise IOError, so we're not in compliance.

I would happily review a PR that makes and tests this change for Python 2. It may be somewhat tricky to get right, as it would need to preserve the errno and other attributes of the underlying exception while changing its type and preserving the original traceback.