# About close() in commit_lock_file() 4 messages from 2013-08-05 to 2013-08-06. Participants: Duy Nguyen, Junio C Hamano, Johannes Sixt. Thread: https://gitlist.dev/t/34611 ## Duy Nguyen, 2013-08-05 14:23 Subject: About close() in commit_lock_file() Message-ID: URL: https://gitlist.dev/e/CACsJy8By1cpPZ5QyVd6VhKSkd-y_E6pTYdDimK9P0wXia-uMqg%40mail.gmail.com ``` close() is added in commit_lock_file(), before rename(), by 4723ee9 (Close files opened by lock_file() before unlinking. - 2007-11-13), which is needed by Windows. But doesn't that create a gap between close() and rename() on other platforms where another process can replace .lock file with something else before rename() is executed? Should we enclose close() in #ifdef __MINGW32__ (and maybe __CYGWIN__)? -- Duy ``` ## Junio C Hamano, 2013-08-05 16:56 Subject: Re: About close() in commit_lock_file() Message-ID: <7vli4gkrf0.fsf@alter.siamese.dyndns.org> URL: https://gitlist.dev/e/7vli4gkrf0.fsf%40alter.siamese.dyndns.org In-Reply-To: ``` Duy Nguyen writes: > close() is added in commit_lock_file(), before rename(), by 4723ee9 > (Close files opened by lock_file() before unlinking. - 2007-11-13), > which is needed by Windows. But doesn't that create a gap between > close() and rename() on other platforms where another process can > replace .lock file with something else before rename() is executed? Interesting. > Should we enclose close() in #ifdef __MINGW32__ (and maybe > __CYGWIN__)? Or just have "close and retry" code after seeing rename() fails with some known errno, without singling out a particular platform? ``` ## Johannes Sixt, 2013-08-06 06:41 Subject: Re: About close() in commit_lock_file() Message-ID: <52009A90.7050701@viscovery.net> URL: https://gitlist.dev/e/52009A90.7050701%40viscovery.net In-Reply-To: ``` Am 8/5/2013 16:23, schrieb Duy Nguyen: > close() is added in commit_lock_file(), before rename(), by 4723ee9 > (Close files opened by lock_file() before unlinking. - 2007-11-13), > which is needed by Windows. But doesn't that create a gap between > close() and rename() on other platforms where another process can > replace .lock file with something else before rename() is executed? First, files manipulated by commit_lock_file() are to be opened only using lock_file() by definition. Opening such a file in with open() or fopen() or renaming it via rename() without using the lockfile.[ch] API is possible regardless of whether commit_lock_file() has closed the file or not. Such manipulation is already undefined behavior (from Git's point of view), and there is nothing we can do about "misbehaving" processes. Second, lock_file() uses O_CREAT|O_EXCL, which fails when the file exists, regardless of whether it is open or not. Conclusion: There is no problem. -- Hannes ``` ## Duy Nguyen, 2013-08-06 08:36 Subject: Re: About close() in commit_lock_file() Message-ID: URL: https://gitlist.dev/e/CACsJy8DquJj1qVawKfrbZY%2B2WLC54Gskizs1sxZ_j1yutBkbdA%40mail.gmail.com In-Reply-To: <52009A90.7050701@viscovery.net> ``` On Tue, Aug 6, 2013 at 1:41 PM, Johannes Sixt wrote: > Am 8/5/2013 16:23, schrieb Duy Nguyen: >> close() is added in commit_lock_file(), before rename(), by 4723ee9 >> (Close files opened by lock_file() before unlinking. - 2007-11-13), >> which is needed by Windows. But doesn't that create a gap between >> close() and rename() on other platforms where another process can >> replace .lock file with something else before rename() is executed? > > First, files manipulated by commit_lock_file() are to be opened only using > lock_file() by definition. Opening such a file in with open() or fopen() > or renaming it via rename() without using the lockfile.[ch] API is > possible regardless of whether commit_lock_file() has closed the file or > not. Such manipulation is already undefined behavior (from Git's point of > view), and there is nothing we can do about "misbehaving" processes. > > Second, lock_file() uses O_CREAT|O_EXCL, which fails when the file exists, > regardless of whether it is open or not. The second process could unlink(.lock file) first, then open it, even using lock_file(), but I think that falls into your "first" and is an invalid case from git point of view, so.. > Conclusion: There is no problem. Agreed. False alarm. -- Duy ```