From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from ipmail06.adl6.internode.on.net ([150.101.137.145]:5478 "EHLO ipmail06.adl6.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S935248AbeEXD7L (ORCPT ); Wed, 23 May 2018 23:59:11 -0400 Date: Thu, 24 May 2018 13:59:08 +1000 From: Dave Chinner To: Jeff Layton Cc: "Darrick J. Wong" , Matthew Wilcox , "Theodore Y. Ts'o" , linux-fsdevel@vger.kernel.org, fstests@vger.kernel.org, xfs Subject: Re: [PATCH] fs: clear writeback errors in inode_init_always Message-ID: <20180524035908.GD23861@dastard> References: <20180518225037.GA26206@thunk.org> <630faadb74f608aa5a42649b81657e8b62d46bc3.camel@kernel.org> <20180519152700.GB4507@magnolia> <20180519153619.GA24698@bombadil.infradead.org> <20180521175457.GI23858@magnolia> <20180522220926.GW10363@dastard> <2917996756c844877e220b9f9e8740182d3aa2f9.camel@kernel.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <2917996756c844877e220b9f9e8740182d3aa2f9.camel@kernel.org> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Wed, May 23, 2018 at 06:56:00AM -0400, Jeff Layton wrote: > On Wed, 2018-05-23 at 08:09 +1000, Dave Chinner wrote: > > On Tue, May 22, 2018 at 06:30:40AM -0400, Jeff Layton wrote: > > > On Mon, 2018-05-21 at 10:54 -0700, Darrick J. Wong wrote: > > > > On Sat, May 19, 2018 at 08:36:19AM -0700, Matthew Wilcox wrote: > > > > > On Sat, May 19, 2018 at 08:27:00AM -0700, Darrick J. Wong wrote: > > > > > > From: Darrick J. Wong > > > > > > > > > > > > In inode_init_always(), we clear the inode mapping flags, which clears > > > > > > any retained error (AS_EIO, AS_ENOSC) bits. Unfortunately, we do not > > > > > > > > > > typo of ENOSPC in case you do a new version > > > > > > > > Fixed, thanks. > > > > > > > > > > also clear wb_err, which means that old mapping errors can leak through > > > > > > to new inodes. > > > > > > > > > > > > This is crucial for the XFS inode allocation path because we recycle old > > > > > > in-core inodes and we do not want error state from an old file to leak > > > > > > into the new file. This bug was discovered by running generic/036 and > > > > > > generic/047 in a loop and noticing that the EIOs generated by the > > > > > > collision of direct and buffered writes in generic/036 would survive the > > > > > > remount between 036 and 047, and get reported to the fsyncs (on > > > > > > different files on a reformatted fs!) in generic/047. > > > > > > > > > > > > Since we're changing the semantics of inode_init_always, we must also > > > > > > change xfs_reinit_inode to retain the writeback error state when we go > > > > > > to recover an inode that has been torn down in the vfs but not yet > > > > > > disposed of by XFS. > > > > > > > > > > Don't you also need to preserve inode->i_mapping->flags across the > > > > > reinitialisation for the users which aren't yet using ->wb_err? > > > > > > > > At first I thought (mistakenly) that xfs had been completely converted, > > > > but digging further we still use the old filemap_check_errors. It seems > > > > reasonable to me that if we're going to resurrect an incore inode then > > > > we should try to hang on to AS_EIO/AS_ENOSPC for as long as we can. > > > > > > > > > > > > > > Yes, I think the patch you sent makes sense. > > > > > > Most of the XFS callers are using filemap_write_and_wait{_range}, and > > > most of those seem to be called from ioctls (->setattr op being the > > > notable exception). > > > > > > We could (in principle) pass a pointer to file->f_wb_err to most of > > > those callers, and use that to check for errors instead of looking for > > > AS_EIO/AS_ENOSPC. We wouldn't want to advance the error cursor for those > > > (as that should only be done in the context of an fsync), but it might > > > be more reliable than using the flags here. > > > > Why wouldn't we advance the error pointer? The data writeback error > > caused an operation to fail and the error has been reported to > > userspace, so how is that different to fsync() failing and reporting > > the error to userspace? > > > > This seems like a slippery slope of inconsistency to me, where data > > writeback errors are only reported once on fsync(), but can be > > reported multiple times through different filesystem operations... > > > > If I get back an error on ioctl, why should I think that that has > anything to do with writeback? For example, suppose I do: > > write = 0 > ioctl = -EIO > fsync = 0 > > The ioctl returns an error but the fsync returns 0. It would be > reasonable to assume that my writes had made it to disk. > > I think that we really do want to ensure that fsync always returns an > error if writeback failed since the last fsync. That was sort of the > point of the errseq_t work in the first place. This needs documentation explaining all these intricacies. Because I'm certain that we'll get it wrong again if the expected behaviour isn't all spelled out very clearly. Cheers, Dave. -- Dave Chinner david@fromorbit.com