gh-86768: check if fd is seekable in os.lseek on Windows

[aisk]

This change will introduce a performance regression:

import os
import timeit
f = open("a.py")
print(timeit.timeit("os.lseek(f.fileno(), 1, os.SEEK_CUR)", number=100000000, globals=globals()))
f.close()

Before the change:

PS C:\Users\xxxxx\Source\cpython> .\python.bat a.py
Running Release|x64 interpreter...
93.8409629999951

After the change:

PS C:\Users\xxxxx\Source\cpython> .\python.bat a.py
Running Release|x64 interpreter...
123.18093929998577

However I think it's acceptable because we added a check in the implementation and os.lseek usually won't been called too many times in the real world.

• Issue: seekable() returns True on pipe objects in Windows #86768

[serhiy-storchaka]

It may be slightly simpler if set result to -1 initially.

[serhiy-storchaka]

LGTM. 👍

[morotti]

hello,

is it the function that is called when doing file.seek() on a file with open("myfile.zip", "r") as f

if yes, I am somewhat concerned about the performance impact of adding a syscall on every call. is it really needed?
the PR doesn't explain why the check is needed? what happened before when it wasn't explicitly checked whether the file is seekable and it was not seekable?

seek is very heavily used by applications and it is performance sensitive.
for example when you install a package with pip install, the python package is merely a zip file to extract.
the zip file can contain thousands of files, the extraction is doing multiple seek+read operations for each file to extract, to locate (multiple) headers and the content.
I wonder if these calls to seek are going to os.lseek?
https://github.com/python/cpython/blob/3.13/Lib/zipfile/__init__.py

[serhiy-storchaka]

For example, you can write a ZIP file to not seekable file. If the output file is seekable, ZipFile seeks back to the local file header to write the compressed and uncompressed file sizes after writing the file data. If the output file is not seekable, ZipFile marks them as "undefined" in the local file header and use other way to write this metainformation. If seekable() falsely returns True and seek() does not have effect for pipe, ZipFile will produce corrupted ZIP file.

It is unfortunate that this change will add 0.3 microseconds for each seek() (this is 0.3 milliseconds for a thousand of files), but we have no other way.

[serhiy-storchaka]

@zooba, could you please take a look? I believe this change should be merged, despite its performance impact.

[zooba]

Why not change seekable instead? That's what the original bug asked for, and it shouldn't occur in tight loops.

My general rule is that I don't like predicting what errors the OS should return. It's weird that it doesn't produce an error here, but I guess it breaks a lot of apps to introduce it (which should also warn us against introducing an error...). But if we have to do extra work for a bool result like seekable, that's fine by me.

[aisk]

Why not change seekable instead? That's what the original bug asked for, and it shouldn't occur in tight loops.

The seekable method for open(os.pipe()[1], 'w') will call lseek on FileIO under the hood, so change os.lseek will also change seekable:

cpython/Modules/_io/fileio.c

Line 655 in cfcfbdd

PyObject *pos = portable_lseek(self, NULL, SEEK_CUR, false);

Changing the seekable method for FileIO to check if it's a pipe on Windows is also a solution, but I think the os module is more like a compatibility layer for Windows to emulate POSIX, so changing os.lseek to let it behave more like Unix or Linux is a better way?

[zooba]

I think the os module is more like a compatibility layer for Windows to emulate POSIX, so changing os.lseek to let it behave more like Unix or Linux is a better way?

Fair, but in that case, I'd prefer we check specifically for a pipe and raise, rather than "anything that isn't disk". It's less disruptive to add new exceptions for specific cases than for negative cases.

[serhiy-storchaka]

Why not change seekable instead? That's what the original bug asked for, and it shouldn't occur in tight loops.

These are the LBYL and EAFP sides of the same problem. Some code uses FileIO.seekable(), other code uses FileIO.seek() and os.lseek(). BTW, we need to update also the FileIO methods, the C implementation does not call os.lseek().

I'd prefer we check specifically for a pipe and raise, rather than "anything that isn't disk".

Does this work for "CON", "NUL" and "\\.\pipe\xxx"?

[zooba]

Does this work for "CON", "NUL" and "\\.\pipe\xxx"?

Maybe? I think NUL claims to be a character device, which may get mapped down to pipe at some point, or might be another case that's worth detecting explicitly.

My point is that we should make a list of things that are definitely not seekable and handle those, rather than making a list of things that definitely are seekable and assume that all others should be an error.

[serhiy-storchaka]

According to the GetFileType() documentation, the only returned values are FILE_TYPE_CHAR, FILE_TYPE_DISK, FILE_TYPE_PIPE and FILE_TYPE_UNKNOWN (which also denotes error). FILE_TYPE_DISK is seekable, FILE_TYPE_PIPE is not seekable.

On Linux, /dev/null, /dev/full, /dev/zero and /dev/urandom are seekable, but seeking is a no-op. /dev/tty and /dev/pts/N are non-seekable. So maybe treat FILE_TYPE_CHAR as seekable?

And what is the correct behavior for FILE_TYPE_UNKNOWN?

[zooba]

So maybe treat FILE_TYPE_CHAR as seekable?

Seems reasonable.

And what is the correct behavior for FILE_TYPE_UNKNOWN?

I think allow it to work, so that cases where seeking work aren't blocked. Cases where seeking is an error can return an error if they want, but if they want it to be a no-op like pipes, then they can do that too.