From: Jeff Moyer Subject: Re: [PATCH 4/9] xfs: honor the O_SYNC flag for aysnchronous direct I/O requests Date: Tue, 20 Nov 2012 14:42:48 -0500 Message-ID: References: <20121120074116.24645.36369.stgit@blackbox.djwong.org> <20121120075114.25270.40680.stgit@blackbox.djwong.org> <20121120112038.GC2591@dastard> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: "Darrick J. Wong" , axboe@kernel.dk, tytso@mit.edu, bpm@sgi.com, viro@zeniv.linux.org.uk, jack@suse.cz, linux-fsdevel@vger.kernel.org, hch@infradead.org, linux-ext4@vger.kernel.org, linux-kernel@vger.kernel.org, xfs@oss.sgi.com, djwong+kernel@djwong.org To: Dave Chinner Return-path: Received: from mx1.redhat.com ([209.132.183.28]:12543 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751122Ab2KTTnG (ORCPT ); Tue, 20 Nov 2012 14:43:06 -0500 In-Reply-To: <20121120112038.GC2591@dastard> (Dave Chinner's message of "Tue, 20 Nov 2012 22:20:38 +1100") Sender: linux-ext4-owner@vger.kernel.org List-ID: Dave Chinner writes: > And requeuing work from one workqueue to the next is something that > we can avoid. We know at IO submission time (i.e. > xfs_vm_direct_io)) whether an fsync completion is going to be needed > during Io completion. The ioend->io_needs_fsync flag can be set > then, and the first pass through xfs_finish_ioend() can queue it to > the correct workqueue. i.e. it only needs to be queued if it's not > already an unwritten or append ioend and it needs an fsync. > > As it is, all the data completion workqueues run the same completion > function so all you need to do is handle the fsync case at the end > of the existing processing - it's not an else case. i.e the end of > xfs_end_io() becomes: > > if (ioend->io_needs_fsync) { > error = xfs_ioend_fsync(ioend); > if (error) > ioend->io_error = -error; > goto done; > } > done: > xfs_destroy_ioend(ioend); Works for me, that makes things simpler. > As it is, this code is going to change before these changes go in - > there's a nasty regression in the DIO code that I found this > afternoon that requires reworking this IO completion logic to > avoid. The patch will appear on the list soon.... I'm not on the xfs list, so if you haven't already sent it, mind Cc-ing me? >> --- a/fs/xfs/xfs_mount.h >> +++ b/fs/xfs/xfs_mount.h >> @@ -209,6 +209,7 @@ typedef struct xfs_mount { >> struct workqueue_struct *m_data_workqueue; >> struct workqueue_struct *m_unwritten_workqueue; >> struct workqueue_struct *m_cil_workqueue; >> + struct workqueue_struct *m_aio_blkdev_flush_wq; > > struct workqueue_struct *m_aio_fsync_wq; For the record, m_aio_blkdev_flush_wq is the name you chose previously. ;-) Thanks for the review! Cheers, Jeff