From: Junio C Hamano Date: Fri, 07 Nov 2014 16:51:47 GMT Subject: Re: [RFC] git checkout $tree -- $path always rewrites files Message-ID: In-Reply-To: Duy Nguyen writes: > On Fri, Nov 7, 2014 at 3:38 PM, Jeff King wrote: >> On Fri, Nov 07, 2014 at 03:13:24AM -0500, Jeff King wrote: >> >>> I noticed that "git checkout $tree -- $path" will _always_ unlink and >>> write a new copy of each matching path, even if they are up-to-date with >>> the index and the content in $tree is the same. >> >> By the way, one other thing I wondered while looking at this code: when >> we checkout a working tree file, we unlink the old one and write the new >> one in-place. Is there a particular reason we do this versus writing to >> a temporary file and renaming it into place? That would give >> simultaneous readers a more atomic view. >> >> I suspect the answer is something like: you cannot always do a rename, >> because you might have a typechange, directory becoming a file, or vice >> versa; so anyone relying on an atomic view during a checkout operation >> is already Doing It Wrong. Handling a content-change of an existing >> path would complicate the code, so we do not bother. > > Not a confirmation, but it looks like Linus did it just to make sure > he had new permissions right, in e447947 (Be much more liberal about > the file mode bits. - 2005-04-16). I think you are referring to the "... to get the new one with the right permissions" comment in that patch, but I do not think that affects the choice between (1) unlink and write anew to the final and (2) create a new temporary and rename. Either way, you do not update the existing file in-place and try to checkout the permission bits with chmod(2).