Re: [PATCH] Fix 'git clean' failure on NFS.

From: Erik Faye-Lund <kusmabite@googlemail.com>
Date: Jun 18, 2010, 13:26 UTC
Message-ID: <AANLkTinEGpB567-v8TXFv5ghXfjY9ZtkOMYcGOhzePj3@mail.gmail.com>
In-Reply-To: <4C1B6E13.1030309@gmail.com>

2010/6/18 Török Edwin <edwintorok@gmail.com>:

Show 19 quoted lines

> On 06/18/2010 03:29 PM, Erik Faye-Lund wrote:
>> 2010/6/18 Török Edwin <edwintorok@gmail.com>:
>>> readdir() used to do a single pass, which is not always enough.
>>> If you unlink some files/rmdir some subdirs then there might be some
>>> files you haven't seen yet *before* the readdir cursor (files get rearranged
>>> in the directory, etc.).
>>>
>>> The fix is to do an additional readdir() pass if we unlinked/rmdired something.
>>> This is easily accomplished by using rewinddir.
>>>
>>
>> Won't this break on Windows? mingw_readdir() doesn't seem to have a
>> maching mingw_rewinddir() in compat/mingw.c/h....
>>
>
> rewinddir() is used in ClamAV too, and it builds natively on win32.
> I think we have some compat function that map rewinddir to the Win32
> API, somebody will probably have to do that for git too.
>

There's a rewinddir function in msvcrt.dll. However, we're not using msvcrt.dll's opendir() implementation for performance reasons, so this patch would require a rewinddir implementation as a preparation path, no? If such a patch could be extracted from ClamAV (and ClamAV is license compatible with git), would you mind including such a patch for the next round?

> Another possibility is to do closedir/opendir again on mingw.

Yes. Yet another possibility is to enable the NO_MINGW_REPLACE_READDIR-switch, but this has negative performance effects...

-- 
Erik "kusma" Faye-Lund