Message102720
| Author |
r.david.murray |
| Recipients |
asvetlov, brian.curtin, flox, jafo, jaraco, loewis, r.david.murray, techtonik, tim.golden |
| Date |
2010年04月09日.13:05:31 |
| SpamBayes Score |
1.2478163e-10 |
| Marked as misclassified |
No |
| Message-id |
<1270818334.21.0.114788890224.issue7443@psf.upfronthosting.co.za> |
| In-reply-to |
| Content |
If the problem with the fix is that lots of tests use test_support.unlink, then I don't see why the rename dance can't be implemented in test_support.unlink. (Possibly conditioned on whether or not the tests are running on a windows platform.) Dealing with unlink problems is why that method exists in the first place.
There are probably places in the test suite that *don't* use test_support.unlink, though, so fixing test_support.unlink will not necessarily fix all of the problems. We'll have to fix those other tests (probably by using the new test_support.unlink) as we find them.
An actual patch will need a test that doesn't rely on win32file (ctypes would be OK). It may be necessary to rename to a unique filename, too. (To be clear, I think a unit test that reproduces the problem by doing an open with FILE_SHARE_DELETE is fine, we don't need a test that reproduces the race condition itself. The windows experts will correct me if I'm wrong :)
I'm changing the stage to patch needed because it seems to me that using a technique like rename that doesn't introduce additional delays into the test suite is to be preferred. |
|