cc @serhiy-storchaka

@serhiy-storchaka Both test_open_undecodable_uri and test_open_undecodable_path are failing in CI at present (ref).

FWIW - in both cases, the test path being used is b'@test_42_tmp\xe7w\xf0' - the path is defined, and it is successfully opened, so the get_undecodable_path() check is passing where previously the test was explicitly skipped on Emscripten.

When I run this locally, I get a warning in the console:

Warning -- files was modified by test.test_sqlite3.test_dbapi
Warning --   Before: []
Warning --   After:  ['@test_42_tmp�w�'] 
test test.test_sqlite3.test_dbapi failed

which suggests to me that the bad unicode isn't round-tripping consistently through the filesystem.

I agree that this should be a "skip both or skip neither" situation; and I'd definitely prefer to understand better why it needs to be skipped explicitly before restoring the skip.

I suppose that paths are Unicode strings on Emscripten. Python and SQLite can use different ways to decode bytes path to Unicode, so os.path.exists() does not see the file created by SQLite, and unlink() cannot remove it.

This is similar to Windows where we need to keep a separate skip.

Well '@test_42_tmp\xe7w\xf0' does round trip through JavaScript, MEMFS, and NODEFS correctly. So I agree, this test should be fixable. I will investigate more.

Both test_open_undecodable_uri and test_open_undecodable_path are failing in CI at present

Interesting. Locally only test_open_with_undecodable_path is failing for me.

I changed the test to this and added syscall tracing:

    def get_undecodable_path(self):
        path = TESTFN_UNDECODABLE
        print("open", path)
        f = open(path, 'wb')
        print("close", path)
        f.close()
        print("unlink", path)
        unlink(path)
        return path

    @unittest.skipIf(sys.platform == "win32", "skipped on Windows")
    def test_open_with_undecodable_path(self):
        path = self.get_undecodable_path()
        self.addCleanup(unlink, path)
        print("sqlite.connect", path)
        c = sqlite.connect(path)
        with contextlib.closing(c) as cx:
            print("exists", path)
            exists = os.path.exists(path)
            print(" .. ", exists)
            self.assertTrue(exists)

The relevant part of the log looks like this:

open b'@test_42_tmp\xe7w\xf0'
___syscall_openat @test_42_tmp緰
close b'@test_42_tmp\xe7w\xf0'
_fd_close 3 /home/.../test_python_worker_601748æ/@test_42_tmp緰
unlink b'@test_42_tmp\xe7w\xf0'
___syscall_unlinkat @test_42_tmp緰
sqlite.connect b'@test_42_tmp\xe7w\xf0'
___syscall_openat /home/.../test_python_worker_601748æ/@test_42_tmp�w�
___syscall_stat64 /home/.../test_python_worker_601748æ/@test_42_tmp�w�
exists b'@test_42_tmp\xe7w\xf0'
___syscall_stat64 @test_42_tmp緰
 ..  False
___syscall_stat64 /home/.../test_python_worker_601748æ/@test_42_tmp�w�
_fd_close 3 /home/.../test_python_worker_601748æ/@test_42_tmp�w�
<unrelated teardown syscalls>
FAIL
___syscall_unlinkat @test_42_tmp緰

Okay I got it: the problem is that UTF8ArrayToString uses a different code path on strings with more than 16 bytes and the behaviors don't exactly match:

    // When using conditional TextDecoder, skip it for short strings as the overhead of the native call is not worth it.
    if (endPtr - idx > 16 && heapOrArray.buffer && UTF8Decoder) {
      return UTF8Decoder.decode({{{ getUnsharedTextDecoderView('heapOrArray', 'idx', 'endPtr') }}});
    }

https://github.com/emscripten-core/emscripten/blob/main/src/lib/libstrings.js#L57-L59

So when the file system decodes the string, it checks the length. If it's short, it uses the JS decoder, if it's long it uses the native decoder. sqlite fully resolves the path before opening the file, and the absolute path is longer than 16 bytes. Whereas the not fully resolved path is 15 bytes long and gets the correct slow path. Adding two extra bytes to TESTFN_UNDECODABLE makes the test pass.

Upstream report:
emscripten-core/emscripten#24690