Given a .zip file with more than one inner file, when reading those inner files using zipfile.Path the zip module closes the .zip file prematurely, causing an error.

Given the following code (example zipfile is attached, but any should work).

with zipfile.ZipFile('foo.zip') as file:
    for name in ['file-1.txt', 'file-2.txt']:
        p = zipfile.Path(file, name)
        with p.open() as inner:
            print(list(inner)) # ValueError: seek of closed file

But the following code does not fail:

with zipfile.ZipFile('foo.zip') as file:
    for name in ['file-1.txt', 'file-2.txt']:
        with file.open(name) as inner:
            print(list(inner)) # Works!

Python 3.8.2 macOS 10.15.4.

Bisection show this to be caused due to e5bd736 . See also bpo-41640 for a similar report.

I've attempted to replicate the issue in the zipp test suite with this test:

diff --git a/test_zipp.py b/test_zipp.py
index a6fbf39..dc7c8aa 100644
--- a/test_zipp.py
+++ b/test_zipp.py
@@ -259,3 +259,10 @@ class TestPath(unittest.TestCase):
     def test_implied_dirs_performance(self):
         data = ['/'.join(string.ascii_lowercase + str(n)) for n in range(10000)]
         zipp.CompleteDirs._implied_dirs(data)
+
+    def test_read_does_not_close(self):
+        for alpharep in self.zipfile_alpharep():
+            for rep in range(2):
+                p_ = zipp.Path(alpharep, 'a.txt')
+                with p_.open() as inner:
+                    print(list(inner))

But the test passes.

I am able to replicate the failure using the ondisk fixture:

diff --git a/test_zipp.py b/test_zipp.py
index a6fbf39..539d0a9 100644
--- a/test_zipp.py
+++ b/test_zipp.py
@@ -259,3 +259,11 @@ class TestPath(unittest.TestCase):
     def test_implied_dirs_performance(self):
         data = ['/'.join(string.ascii_lowercase + str(n)) for n in range(10000)]
         zipp.CompleteDirs._implied_dirs(data)
+
+    def test_read_does_not_close(self):
+        for alpharep in self.zipfile_ondisk():
+            with zipfile.ZipFile(alpharep) as file:
+                for rep in range(2):
+                    p_ = zipp.Path(file, 'a.txt')
+                    with p_.open() as inner:
+                        print(list(inner))

I suspect the issue lies in how the CompleteDirs.make replaces one instance with another in order to provide a complete list of implied directories and to memoize the names lookup.

Because constructing a zipfile.Path object around a zipfile.ZipFile copies the underlying state, closing one will have the affect of closing the other.

I believe this issue is the same root issue as bpo-41350.

I see a few options here:

• Implement CompleteDirs/FastLookup as adapters instead of subclasses, allowing the original ZipFile object to represent the state in a single place. This approach would likely be slower due to the indirection on all operations through the wrapper.
• Instead of constructing a new object and copying the state, CompleteDirs.make could mutate the existing ZipFile class, replacing source.__class__ with the new class. This approach is messy and the caller would still need to be aware that this change could be applied to the zipfile object.
• Consider this use-case unsupported and document that any ZipFile object passed into Path is no longer viable and should not be referenced for another purpose.
• Eliminate the class-based performance optimizations and replace them with some functional/imperative form. This approach may provide less separation of concerns.
• Disable the close-on-delete behavior for the subclasses when state is copied from another.

Implementing that last option:

diff --git a/zipp.py b/zipp.py
index 697d4a9..f244cba 100644
--- a/zipp.py
+++ b/zipp.py
@@ -111,6 +111,7 @@ class CompleteDirs(zipfile.ZipFile):
 
         res = cls.__new__(cls)
         vars(res).update(vars(source))
+        res.close = lambda: None
         return res
 
 

Does appear to address the issue. I'm not super happy about the implementation, though.

In jaraco/zipp, I've implemented three of the options above:

• https://github.com/jaraco/zipp/tree/bugfix/bpo-40564-option-1
• https://github.com/jaraco/zipp/tree/bugfix/bpo-40564-option-2
• https://github.com/jaraco/zipp/tree/bugfix/bpo-40564-option-5

I've also created this alternative to option 2:

• https://github.com/jaraco/zipp/tree/bugfix/bpo-40564-mixin

This alternative approach uses a mix-in rather than subclasses, creating a new class on-demand. I was hoping this approach would enable just augmenting the instance rather than affecting source.__class__, but when I got to the implementation, I found that source.__class__ had to be mutated regardless.

This approach does have an advantage over option 2 in that it would support other ZipFile subclasses for source. It has the disadvantage in that it creates a new subclass for every instance created.

I've thought about it a lot and while I'm not particularly happy with any of the approaches, I'm leaning toward option 2.

I've released zipp 3.2.0 that includes a fix for this issue (option 2). Please test it out.

Because this change affects the user's expectation about the effect of what's passed in, I'm hesitant to backport it to 3.8 and 3.9. It's too late to go into 3.9.0, so the earliest it can appear is in 3.9.1.

Please advise - does the undesirable behavior warrant a bugfix backport, or is it sufficient to direct users impacted by this issue prior to Python 3.10 to use the zipp backport?

New changeset ebbe803 by Jason R. Coombs in branch 'master':
bpo-40564: Avoid copying state from extant ZipFile. (GH-22371)
ebbe803

Fix merged into master. Please use zipp 3.2.0 on Python 3.9 and earlier for the improved behavior or report back here if a backport is needed (and why).