From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mx3-rdu2.redhat.com ([66.187.233.73]:58824 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1727378AbeIGPp6 (ORCPT ); Fri, 7 Sep 2018 11:45:58 -0400 Date: Fri, 7 Sep 2018 07:05:29 -0400 From: Brian Foster Subject: Re: [PATCH 1/4] mkfs: stop zeroing old superblocks excessively Message-ID: <20180907110529.GA33388@bfoster> References: <20180905081932.27478-1-david@fromorbit.com> <20180905081932.27478-2-david@fromorbit.com> <20180906133108.GA3311@bfoster> <20180907000432.GH27618@dastard> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20180907000432.GH27618@dastard> Sender: linux-xfs-owner@vger.kernel.org List-ID: List-Id: xfs To: Dave Chinner Cc: linux-xfs@vger.kernel.org On Fri, Sep 07, 2018 at 10:04:32AM +1000, Dave Chinner wrote: > On Thu, Sep 06, 2018 at 09:31:08AM -0400, Brian Foster wrote: > > On Wed, Sep 05, 2018 at 06:19:29PM +1000, Dave Chinner wrote: > .... > > > @@ -1220,15 +1221,68 @@ zero_old_xfs_structures( > > > > > > /* > > > * block size and basic geometry seems alright, zero the secondaries. > > > + * > > > + * Don't be insane when it comes to overwriting really large filesystems > > > + * as it could take millions of IOs to zero every secondary > > > + * superblock. If we are remaking a huge filesystem, then do the > > > + * zeroing, but if we are replacing it with a small one (typically done > > > + * in test environments, limit the zeroing to: > > > + * > > > + * - around the range of the new filesystem > > > + * - the middle of the old filesystem > > > + * - the end of the old filesystem > > > + * > > > + * Killing the middle and end of the old filesystem will prevent repair > > > + * from finding it with it's fast secondary sb scan algorithm. The slow > > > + * scan algorithm will then confirm the small filesystem geometry by > > > + * brute force scans. > > > */ > > > memset(buf, 0, new_sb->sb_sectsize); > > > + > > > + /* this carefully avoids integer overflows */ > > > + end = sb.sb_dblocks; > > > + if (sb.sb_agcount > 10000 && > > > + new_sb->sb_dblocks < end / 10) > > > + end = new_sb->sb_dblocks * 10; > > > > ... but what's with the 10k agcount cutoff? Just a number out of a hat > > to demonstrate the improvement..? > > yeah, I pulled it from a hat, but mainly so it only triggers the new > "partial zeroing" code on really large devices that had a large > filesystem and now we're making a much smaller filesystem on it. > > There's no real reason for it except for the fact I haven't done the > verification needed to make this the default behaviour (hence the > RFCRAP status :). > Ok... > > Given that you've done the work to rough in an AIO buffered write > > mechanism for mkfs, have you considered whether we can find a way to > > apply that mechanism here? > > It would have be another AIO context because this loop doesn't > use xfs_bufs. > Yup. I wasn't suggesting to necessarily use xfs_bufs, but I suppose that could be an option if we could find a clever way to do so. It's probably not worth getting into the weeds of an implementation until the core I/O infrastructure is more fleshed out. > > I'm guessing that the result of using AIO > > wouldn't be quite as impressive of not doing I/O at all, but as you've > > already noted, this algorithmic optimization is more targeted at test > > environments than production ones. The AIO approach sounds like it could > > be more broadly beneficial, even if not quite as fast in those > > particular test cases. > > I just don't think it's worth the hassle as, like you said, it's > easier just to avoid the IO altogether. > Well, I said it probably wouldn't be as fast as not doing the I/O at all. I'm not necessarily convinced it's easier (or better). ;P This patch is kind of a case in point. It's more code than the current dead-simple pwrite() loop, it picks an arbitrary cutoff which suggests it needs more work (and/or complexity) to be more generically acceptable, and it ultimately requires some kind of evaluation against (or cooperation with) the repair secondary scan to ensure it's done properly. If this function could do exactly what it does now, but much faster (based on the mkfs results you've achieved), then that could be a more simple approach. All I'm really saying is that if we're going through the work to create a fast/efficient I/O mechanism for mkfs, it's worth looking into applying that mechanism here because it's the same fundamental problem (applied to the more narrow case of zeroing out that old, huge fs rather than creating it in the first place). If the code complexity is comparable (an open/valid question), then at least we stand to benefit regardless of whether the old -> new fs is an 8x delta rather than 10x for example, without having to come up with (and maintain) new/rigid heuristics. If the benefits aren't as great or there's just too much of an impedence mismatch between this algorithm and the underlying I/O mechanism to justify the complexity, then we can always fall back to doing something like this. I'd just prefer to see some of that analysis before comitting to this approach. Brian > > > > > off = 0; > > > - for (i = 1; i < sb.sb_agcount; i++) { > > > + for (i = 1; i < sb.sb_agcount && off < end; i++) { > > > + off += sb.sb_agblocks; > > > + if (pwrite(xi->dfd, buf, new_sb->sb_sectsize, > > > + off << sb.sb_blocklog) == -1) > > > + break; > > > + } > > > + > > > + if (end == sb.sb_dblocks) > > > + return; > > > + > > > + /* > > > + * Trash the middle 1000 AGs of the old fs, which we know has at least > > > + * 10000 AGs at this point. Cast to make sure we are doing 64bit > > > + * multiplies, otherwise off gets truncated to 32 bit. I hate C. > > > + */ > > > + i = (sb.sb_agcount / 2) - 500; > > > + off = (xfs_off_t)sb.sb_agblocks * i; > > > + off = (xfs_off_t)sb.sb_agblocks * ((sb.sb_agcount / 2) - 500); > > > > Looks like a couple lines of dead code there. > > Yup, didn't clean it up properly, did I? > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com