From mboxrd@z Thu Jan 1 00:00:00 1970 From: Theodore Ts'o Subject: Re: [PATCH 3/6] mke2fs: set block_validity as a default mount option Date: Sun, 24 Aug 2014 18:47:21 -0400 Message-ID: <20140824224721.GG6236@thunk.org> References: <20140809042610.2441.6868.stgit@birch.djwong.org> <20140809042630.2441.34661.stgit@birch.djwong.org> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-ext4@vger.kernel.org To: "Darrick J. Wong" Return-path: Received: from imap.thunk.org ([74.207.234.97]:47019 "EHLO imap.thunk.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753483AbaHXWrX (ORCPT ); Sun, 24 Aug 2014 18:47:23 -0400 Content-Disposition: inline In-Reply-To: <20140809042630.2441.34661.stgit@birch.djwong.org> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Fri, Aug 08, 2014 at 09:26:30PM -0700, Darrick J. Wong wrote: > The block_validity mount option spot-checks block allocations against > a bitmap of known group metadata blocks. This helps us to prevent > self-inflicted catastrophic failures such as trying to "share" > critical metadata (think bitmaps) with file data, which usually > results in filesystem destruction. > > In order to test the overhead of the mount option, I re-used the speed > tests in the metadata checksum testing script. In short, the program > creates what looks like 15 copies of a kernel source tree, except that > it uses fallocate to strip out the overhead of writing the file data > so that we can focus on metadata overhead. On a 64G RAM disk, the > overhead was generally about 0.9% and at most 1.6%. On a 160G USB > disk, the overhead was about 0.8% and peaked at 1.2%. I was doing a spot check of the additional memory impact of block_validity mount option, and it's for a 20T file system, assuming the basic flex_bg size of 16 block groups, it's a bit over 400k of kernel memory. That's not a *huge* amount of memory, but it could potentially be noticeable on a bookshelf NAS server. However, I could imagine that for a system with say, two dozen 10T drives (which aren't that far off in the future) in a tray, that's around 4 megabytes of memory, which starts being non-trivial. That being said, I suspect for most users, it's not that big of a deal --- so maybe this is something we should just simply enable by default in the kernel, let those folks who want to disable specify a noblock_validity mount option. The other thing to consider is that for big raid arrays, maybe we should use a larger flex_bg size. The main reason for keeping the size small is to minimize the seek time between the inode table and a block in the flex_bg. But for raid devices, we could probably afford to increase flex_bg size, which would decrease the numer of system zones that the block validity code would need to track. - Ted