io.FileIO calls flush() after file closed

[brianquinlan]

BPO	5700
Nosy	@terryjreedy, @brianquinlan, @pitrou, @vadmium, @serhiy-storchaka
Files	remove_flush.diff: Patch to remove the call to flush by _io._IOBase cooperation.diff: Use cooperative inheritance fileio_flush_closed.patch fileio_flush_closed_2.patch fileio_flush_closed-2.7.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/serhiy-storchaka'
closed_at = <Date 2015-02-20.22:37:46.140>
created_at = <Date 2009-04-05.15:35:55.732>
labels = ['type-bug', 'library', 'expert-IO']
title = 'io.FileIO calls flush() after file closed'
updated_at = <Date 2015-02-22.22:33:59.994>
user = 'https://github.com/brianquinlan'

bugs.python.org fields:

activity = <Date 2015-02-22.22:33:59.994>
actor = 'python-dev'
assignee = 'serhiy.storchaka'
closed = True
closed_date = <Date 2015-02-20.22:37:46.140>
closer = 'serhiy.storchaka'
components = ['Library (Lib)', 'IO']
creation = <Date 2009-04-05.15:35:55.732>
creator = 'bquinlan'
dependencies = []
files = ['13620', '13677', '36889', '38147', '38148']
hgrepos = []
issue_num = 5700
keywords = ['patch']
message_count = 22.0
messages = ['85521', '85616', '85890', '85894', '85895', '85900', '85901', '85902', '85913', '85914', '85915', '85917', '139205', '228234', '229143', '234880', '236039', '236040', '236074', '236337', '236339', '236417']
nosy_count = 6.0
nosy_names = ['terry.reedy', 'bquinlan', 'pitrou', 'python-dev', 'martin.panter', 'serhiy.storchaka']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue5700'
versions = ['Python 2.7', 'Python 3.4', 'Python 3.5']

[brianquinlan]

>>> import io
>>> class MyIO(io.FileIO):
...     def flush(self):
...             print('closed:', self.closed)
...
>>> f = MyIO('test.out', 'wb')
>>> f.close()
closed: True

IMHO, calling flush() after the file has already been closed is
incorrect behaviour. Search for "Possible py3k io wierdness" on
python-dev for discussion.

[brianquinlan]

Discussion about this bug is ongoing in python-dev.

[brianquinlan]

cooperation.diff:

• change the close method to call .flush() and then ._close()
• only IOBase implements close() (though a subclass can override close
  without causing problems - so long as it calls super().close())
• .flush() invokes super().flush()
• ._close() invokes super()._close()
• FileIO is implemented in Python in _pyio.py so that it can have the
  same base class as the other Python-implemented files classes
• tests verify that .flush() is not called after the file is closed
• tests verify that ._close()/.flush() calls move correctly are
  propagated correctly

[pitrou]

• FileIO is implemented in Python in _pyio.py so that it can have the
  same base class as the other Python-implemented files classes

Is it really necessary (e.g. to pass the tests)?

[brianquinlan]

> - FileIO is implemented in Python in _pyio.py so that it can have the
> same base class as the other Python-implemented files classes

Is it really necessary (e.g. to pass the tests)?

It is necessary to make MI work. With out it the inheritance graph looks
like this (using _pyio):

io.IOBase _pyio.IOBase
| |
io.FileIO MyMixin
| |
\ /
\ /
\ /
MyClass

If you call MyClass.flush() with this hierarchy then it will propagate
to io.IOBase. Since io.IOBase doesn't call super().flush() in its flush
implementation (because it has no super class), flush propagation would
stop and MyMixin.flush wouldn't be called.

There are other ways you could solve this, of course, like getting
io.IOBase to call super().flush and ignore any errors that it gets.

But it seems like this is the right way to fix this problem anyway - as
a user, I would expect isinstance(FileIO(...), IOBase) but that is not
currently the case with _pyio.

[pitrou]

It is necessary to make MI work. With out it the inheritance graph looks
like this (using _pyio):

io.IOBase _pyio.IOBase
| |
io.FileIO MyMixin
| |
\ /
\ /
\ /
MyClass

MyMixin doesn't need to inherit from IOBase, since precisely it is a
mixin. It's just a piece of functionality that you drop into an already
existing inheritance tree.

But it seems like this is the right way to fix this problem anyway - as
a user, I would expect isinstance(FileIO(...), IOBase) but that is not
currently the case with _pyio.

That's because the reference ABCs are defined in the io module, not in
_pyio:

True
>>> issubclass(_pyio.FileIO, io.IOBase)
True
>>> issubclass(io.TextIOWrapper, io.IOBase)
True
>>> issubclass(_pyio.TextIOWrapper, io.IOBase)
True

_pyio.IOBase and friends are just concrete implementations of those
ABCs:

True
>>> issubclass(_pyio.TextIOBase, io.TextIOBase)
True

I know it looks a bit non-intuitive at first, but the point is that the
ABCs are unified, even there are two different implementations. I think
it's better than having two different sets of ABCs depending on whether
you import the pure-Python version or the C version.

[brianquinlan]

Antoine Pitrou wrote:

Antoine Pitrou <pitrou@free.fr> added the comment:

> It is necessary to make MI work. With out it the inheritance graph looks
> like this (using _pyio):
>
>
> io.IOBase _pyio.IOBase
> | |
> io.FileIO MyMixin
> | |
> \ /
> \ /
> \ /
> MyClass

MyMixin doesn't need to inherit from IOBase, since precisely it is a
mixin. It's just a piece of functionality that you drop into an already
existing inheritance tree.

In my implementation, it does. Otherwise the linearization of the class
hierarchy becomes:

MyClass -> io.FileIO -> io.IOBase -> MyMixin (-> _pyio.IOBase)

But io.IOBase doesn't make super calls so MyMixin method won't be called.

I think that this linearization is probably more useful:

MyClass -> io.FileIO -> MyMixin -> IOBase

And you get that when io.FileIO and MyMixin share the same IOBase as a
base class.

I don't mind changing io.IOBase to make super calls in ._close() and
.flush() and suppressing the attributes errors that will be generated if
a mixin is not present - but it feels a bit untidy to me.

[snipped]

I know it looks a bit non-intuitive at first, but the point is that the
ABCs are unified, even there are two different implementations. I think
it's better than having two different sets of ABCs depending on whether
you import the pure-Python version or the C version.

I'm not trying to change the ABC unification at all - and if I did then
there is a bug in my code. I just think that the immediate parent class
of _pyio.FileIO should be _pyio.IOBase (just like _io.FileIO has
_io.IOBase as an immediate parent). That will make the Python and C
class hierarchies completely consistent within themselves.

Cheers,
Brian

[pitrou]

I think that this linearization is probably more useful:

MyClass -> io.FileIO -> MyMixin -> IOBase

But why not simply:

MyClass -> MyMixin -> io.FileIO -> IOBase

?
Is there something I'm missing that prevents you from doing this?

I'm not trying to change the ABC unification at all - and if I did then
there is a bug in my code. I just think that the immediate parent class
of _pyio.FileIO should be _pyio.IOBase (just like _io.FileIO has
_io.IOBase as an immediate parent). That will make the Python and C
class hierarchies completely consistent within themselves.

I understand, but that's at the price of an otherwise useless
indirection layer, which will also make _pyio even slower that it
already is :-)
(I admit, however, that _pyio shouldn't be used in normal circumstances,
so this is not a showstopper argument)

[brianquinlan]

Antoine Pitrou wrote:

Antoine Pitrou <pitrou@free.fr> added the comment:

> I think that this linearization is probably more useful:
>
> MyClass -> io.FileIO -> MyMixin -> IOBase

But why not simply:

MyClass -> MyMixin -> io.FileIO -> IOBase

?
Is there something I'm missing that prevents you from doing this?

No, doing this is trivial. But shouldn't it be up to the implementor of
MyClass to decide whether MyMixin or io.FileIO methods are evaluated first?

> I'm not trying to change the ABC unification at all - and if I did then
> there is a bug in my code. I just think that the immediate parent class
> of _pyio.FileIO should be _pyio.IOBase (just like _io.FileIO has
> _io.IOBase as an immediate parent). That will make the Python and C
> class hierarchies completely consistent within themselves.

I understand, but that's at the price of an otherwise useless
indirection layer, which will also make _pyio even slower that it
already is :-)

Maybe I misunderstand the purpose of _pyio. The Python 3.1 says that its
purpose is for experimentation. For experimentation, having a Python
implementation where you can add methods and change behavior (though
perhaps not in as deep as way is if this class were completely written
in Python) is useful. It is also useful for the behavior of the Python
implementation to match that of the C implementation as closely as
possible - this patch makes the inheritance graph for _pyio.FileIO more
consistent.

I, for example, want to add a sync() method to FileIO that will call
fsync() on the file's file descriptor. With this change, I have a place
to plug in that change in Python and I can write the C implementation
when I have the Python implementation right.

Cheers,
Brian

[pitrou]

No, doing this is trivial. But shouldn't it be up to the implementor of
MyClass to decide whether MyMixin or io.FileIO methods are evaluated first?

Is there a concrete use case, though?

By the way, what if _pyio.FileIO inherited from io.FileIO instead of
delegating all methods?

[brianquinlan]

Antoine Pitrou wrote:

Antoine Pitrou <pitrou@free.fr> added the comment:

> No, doing this is trivial. But shouldn't it be up to the implementor of
> MyClass to decide whether MyMixin or io.FileIO methods are evaluated first?

Is there a concrete use case, though?

I don't have a good one in mind - though

By the way, what if _pyio.FileIO inherited from io.FileIO instead of
delegating all methods?

I don't think that would work - but reasoning about MRO hurts my brain ;-)

You'd have to declare the class like this:

class FileIO(io.FileIO, IOBase):
   pass

to get the io.File methods to resolve first. But that means that the
method invocation chain would be broken by io.IOBase, which doesn't make
super calls.

Cheers,
Brian

[brianquinlan]

Oops, I didn't finish my thought:

> No, doing this is trivial. But shouldn't it be up to the implementor of
> MyClass to decide whether MyMixin or io.FileIO methods are evaluated
first?

Is there a concrete use case, though?

I don't have a good one in mind - though writing the unit test
implementation required that the mixin class methods be called last.

[terryjreedy]

The original snippet works the same on 3.2.0. Was their any conclusion as to whether or not a change should be made?

[serhiy-storchaka]

There is complete implementation of FileIO in pure Python in bpo-21859. It is free from this bug.

>>> import _pyio as io
>>> class MyIO(io.FileIO):
...     def flush(self):
...             print('closed:', self.closed)
... 
>>> f = MyIO('test.out', 'wb')
>>> f.close()
closed: False