From: Linus Torvalds Date: Mon, 14 Sep 2009 16:27:59 -0700 (PDT) Subject: [Ocfs2-devel] [GIT PULL] ocfs2 changes for 2.6.32 In-Reply-To: <20090914221434.GA4507@mail.oracle.com> References: <20090911200458.GA15416@mail.oracle.com> <20090914221434.GA4507@mail.oracle.com> Message-ID: List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Joel Becker Cc: Mark Fasheh , Andrew Morton , linux-kernel@vger.kernel.org, ocfs2-devel@oss.oracle.com On Mon, 14 Sep 2009, Joel Becker wrote: > > It's a link(2) analogue. symlink(2) has the loosest coupling, > and reflink(2) the highest. We're not going to add freflink[at](). > It's a snap, not a copy. It can be used to implement a copy, and > copyfile() in libc can be written with reflinkat(2), but it isn't just a > copy. From all but a performance standpoint, it's a copy. It has absolutely _zero_ "link" semantics. When you do a symlink or a hardlink, you see it in the resulting semantics: changing one changes the other. This 'reflink' has no such semantics that I can tell. It has purely copy semantics, never mind that it's optimized. And the thing to note is that it doesn't even have to be optimized as a "link". Think about network filesystems: maybe they want to implement this thing as a server-side "copy" operation (with atomicity guarantees). In other words, I can well imagine that for some filesystems, there really is no refcounting or linking implied, and that's why I think naming should be about semantics, not some random implementation issue that just happens to be true for some particular class of filesystems. So tell me - are there actually any non-copying semantics as far as the _user_ is concerned? Is there some reason why a NFS server might not implement this as a server-side copy? Is there something fundamentally in this all that is about reference counting as far as a user is concerned? I also still didn't get any answer to the "freflink()" question. You just said that we wouldn't do it, with no explanation. Why? We've discussed 'flink()' in the past, I just want to know that when we do a new system call there is some _reason_ why it's not going to explode into many different variants later... Linus