Repository navigation
Add attributes to os.stat results #35189
Description
| BPO | 462296 |
|---|---|
| Nosy | @mwhudson, @gvanrossum, @loewis, @freddrake |
| Files |
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/freddrake'
closed_at = <Date 2002-03-05.16:46:33.000>
created_at = <Date 2001-09-17.17:57:02.000>
labels = ['library']
title = 'Add attributes to os.stat results'
updated_at = <Date 2002-03-05.16:46:33.000>
user = 'https://bugs.python.org/nickm'bugs.python.org fields:
activity = <Date 2002-03-05.16:46:33.000>
actor = 'loewis'
assignee = 'fdrake'
closed = True
closed_date = None
closer = None
components = ['Library (Lib)']
creation = <Date 2001-09-17.17:57:02.000>
creator = 'nickm'
dependencies = []
files = ['3622', '3623', '3624', '3625', '3626', '3627', '3628', '3629', '3630']
hgrepos = []
issue_num = 462296
keywords = ['patch']
message_count = 35.0
messages = ['37622', '37623', '37624', '37625', '37626', '37627', '37628', '37629', '37630', '37631', '37632', '37633', '37634', '37635', '37636', '37637', '37638', '37639', '37640', '37641', '37642', '37643', '37644', '37645', '37646', '37647', '37648', '37649', '37650', '37651', '37652', '37653', '37654', '37655', '37656']
nosy_count = 5.0
nosy_names = ['mwh', 'gvanrossum', 'loewis', 'fdrake', 'nickm']
pr_nums = []
priority = 'normal'
resolution = 'accepted'
stage = None
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue462296'
versions = []
Activity
nickm commentedon Sep 17, 2001
See bug bpo-111481, and PEP-0042. Both suggest that the
return values for os.{stat,lstat,statvfs,fstatvfs}
ought to be struct-like objects rather than simple tuples.
With this patch, the os module will modify the
aformentioned functions so that their results still
obey the previous tuple protocol, but now have
read-only attributes as well. In other words,
"os.stat('filename')[0]" is now synonymous with
"os.stat('filename').st_mode.
The patch also modifies test_os.py to test the new
behavior.
In order to prevent old code from breaking, these new
return types extend tuple. They also use the new
attribute descriptor interface. (Thanks for
PEP-025[23], Guido!)
Backward compatibility: Code will only break if it
assumes that type(os.stat(...)) == TupleType, or if it
assumes that os.stat(...) has no attributes beyond
those defined in tuple.
nickm commentedon Sep 17, 2001
Logged In: YES
user_id=499
BTW, if this gets in, I have another patch that adds support
for st_blksize, st_blocks, and st_rdev on platforms that
support them. It don't expose these new fields in the
tuple, as that would break all the old code that tries to
unpack all the fields of the tuple. Instead, these fields
are only accessible as attributes.
loewis commentedon Sep 17, 2001
Logged In: YES
user_id=21627
I second the request for supporting additional fields
where available. At the same time, it appears
unimplementable using pure Python.
Consequently, I'd like to see this patch redone in C. The
implementation strategy could probably remain the same,
i.e. inherit from tuple for best compatibility; add the
remaining fields as slots. It may be reasonable to
implement attribute access using a custom getattr
function, though.
I have also my doubts about the naming of the fields. The
st_ prefix originates from the time where struct fields
were living in the global namespace (i.e. across different
structures), so prefixing them for uniqueness was
essential. I'm not sure whether we should inherit this
into Python...
nickm commentedon Sep 17, 2001
Logged In: YES
user_id=499
Martin: I'm not entirely sure what you mean here; while my
patch for extra fields requires a minor chunk of C (to
access the struct fields), the rest still works in pure
python. I'm attaching this second version for reference.
I'm not sure it makes much sense to do this with pure C; it
would certainly take a lot more code, with little benefit I
can descern. But you're more experienced than I; what am I
missing?
I agree that the field naming is suboptimal; I was taking my
lead from the stat and statvfs modules. If people prefer,
we can name the fields whatever we like.
nickm commentedon Sep 17, 2001
Logged In: YES
user_id=499
On further consideration, the approach taken in the second
(example only) patch is indeed too fragile. The C code
should not lengthen the tuple arbitrarily and depend on the
Python code to decode it; instead, it should return a
dictionary of extra fields. I think that this approach uses
a minimum of C, is easily maintainable, and very extensible.
nickm commentedon Sep 18, 2001
Logged In: YES
user_id=499
Here's the revised (example only) patch that takes the
more portable approach I mention below.
gvanrossum commentedon Sep 18, 2001
Logged In: YES
user_id=6380
Haven't had time to review the patch yet, but the idea of
providing a structure with fields that doubles as a tuple is
a good one. It's been tried before and can be done in pure
Python as well.
Regarding the field names: I think the field names should
keep their st_ prefix -- IMO this makes the code more
recognizable and hence readable.
loewis commentedon Sep 18, 2001
Logged In: YES
user_id=21627
The problem with your second and third patch is that it
includes an incompatibility for users of posix.stat (and
friends), since it changes the siye of the tuple. If you
want to continue to return a tuple (as the top-level data
structure), you'll break compatibility for applications
using the C module directly. An example of code that would
be broken is
mode, ino, dev, nlink, uid, gid, size, a, c, m =
posix.stat(filename)
To pass the additional fields, you already need your class
_StatResult available in C.
You may find a way to define it in Python and use it in C,
but that has proven to be very fragile in the past.
nickm commentedon Sep 18, 2001
Logged In: YES
user_id=499
Ah! Now I see. I hadn't realized that anybody used the
posix module directly. (People really do this?)
I'll try to write up a patch in C tonight or tomorrow
morning. A couple of questions on which I could use advice:
(1) Where is the proper place to put this kind of
tuple-with-fields hybrid? Modules? Objects? In a new file
or an existing one?
(2) Should I try to make it general enough for non-stat use?
12 remaining items
nickm commentedon Oct 1, 2001
Logged In: YES
user_id=499
I've sent my email address to 'guido at python.org'. For
reference, it's 'nickm at alum.mit.edu'.
nickm commentedon Oct 2, 2001
Logged In: YES
user_id=499
The fifth all-C (!) version, with changes as suggested by
Guido's comments via email.
Big changes: This version no longer subclasses tuple.
Instead, it creates a general-purpose mechanism for making
struct/sequence hybrids in C.
It now includes a patch for timemodule.c as well.
Shortcomings:
(1) As before, macmodule and riscosmodule aren't tested.
(2) These new classes don't participate in GC and aren't
subclassable. (Famous last words: "I don't think this will
matter." :) )
(3) This isn't a brand-new metaclass; it's just a quick bit
of C. As such, you can't use this mechanism to create new
struct/tuple hybrids from Python. (I claim this isn't a
drawback, since it's way easier to reimplement this in
python than it is to make it accessible from python.)
So, how's *this* one?
mwhudson commentedon Oct 3, 2001
Logged In: YES
user_id=6656
If this goes in, I'd like to see it used for termios.tc
{get,set}attr too.
I could probably implement this (but not *right* now...).
gvanrossum commentedon Oct 3, 2001
Logged In: YES
user_id=6380
Patience, please. I'm behind reviewing this, probably won't
have time today either.
gvanrossum commentedon Oct 18, 2001
Logged In: YES
user_id=6380
I'm looking at this now.
gvanrossum commentedon Oct 18, 2001
Logged In: YES
user_id=6380
Thanks, Nick! Good job.
Checked in, just in time for 2.2b1. I'm passing this
tracker entry on to Fred for documentation. (Fred, feel
free to pester Nick for docs. Nick, feel free to upload
approximate patches to Doc/libos.tex and Doc/libtime.tex.
:-)
nickm commentedon Oct 18, 2001
Logged In: YES
user_id=499
Here's a documentation patch for libos.tex. I don't know
the TeX macros well enough to write an analogous one for
libtime.tex; fortunately, it should be fairly easy to
extrapolate from the included patch.
freddrake commentedon Nov 29, 2001
Logged In: YES
user_id=3066
This has been checked in, edited, and checked in again.
mwhudson commentedon Mar 5, 2002
Logged In: YES
user_id=6656
I know this patch is closed, but it seems a vaguely sane
place to ask the question: why do we vary the number of
field of os.stat_result across platforms? Wouldn't it be
better to let it always have the same values & fill in one's
that don't exists locally with -1 or something?
It's hard to pickle os.stat_results portably the way things
are at the moment...
loewis commentedon Mar 5, 2002
Logged In: YES
user_id=21627
Adding all fields is both difficult and undesirable. It is
difficult because you may not know in advance what fields
will be added in future versions, and it is undesirable
because applications may think that there is a value even
though the is none.
What problem does that cause for pickling, and why would a
complete list of all attributes solve this problem?
mwhudson commentedon Mar 5, 2002
Logged In: YES
user_id=6656
I'm not worried about cross version problems.
The problem with pickling is that stat_results (as of today)
get pickled as "os.stat_result" and a tuple of arguments.
The number of arguments os.stat_result takes varies by
platform (it seems to be 10 on this NT box, but it's 13 on
the starship, f'ex). So if a stat_result gets pickled on
the starship and shoved down a socket to an NT machine, it
can't be unpickled. I don't know if this sort of thing ever
happens, but I could see it being surprising & annoying if I
ran into it.
If os.stat_result took 13 arguments everywhere, this problem
obviously wouldn't arise.
loewis commentedon Mar 5, 2002
Logged In: YES
user_id=21627
To support pickling, I think structseq objects should
implement a __reduce__ method, returning the type and a
dictionary. The type's tp_new should accept dictionaries,
and reconstruct the instance from the dictionary.
Alternatively, copy_reg could grow support for stat_result,
which seems desirable anyway, since os.stat returns a
'nt.stat_result' instance on Windows.
Furthermore, fixing the number of arguments does not help at
all in pickling; __reduce__ will return an argument tuple
which includes the original object; in turn, pickle will
recurse until the stack overflows.
mwhudson commentedon Mar 5, 2002
Logged In: YES
user_id=6656
Martin, I may not have been 100% clear in my last note, but
please run
cvs up Objects/structseq.c
structseq objects *do* now implement a __reduce__ method,
but it returns a tuple. Using a dictionary would be more
complicated, and not solve the issue completely: what
happens when you go from a platform with less fields to one
with more? What value does the not-prepared-for field have?
Hmm, the point about nt.stat_result is a good one.
Getting support into copy_reg.py leads to interesting
bootstrapping problems when using uninstalled builds,
unfortunately (site.py imports distutils imports re imports
copy_reg; try to import, say, time, and you can't, because
the whole reason to import distutils was to set up the path
to find dynamically linked libraries...).
loewis commentedon Mar 5, 2002
Logged In: YES
user_id=21627
I'd not put the copyreg support into copy_reg, but into
os.py. Pickling would save a reference to
os._load_stat_result (or some such). When pickle tries to
restore the value, it would first restore
os.load_stat_result. For that, it would import os, which
would register the copy_reg support.
As for constructing structseq objects from dictionaries: it
would be a ValueError if fields within [:n_sequence_fields]
are not filled out; leaving out other fields is fine.