xfs
[Top] [All Lists]

Re: [PATCH 1/3] xfs: honor the O_SYNC flag for aysnchronous direct I/O r

To: Jeff Moyer <jmoyer@xxxxxxxxxx>
Subject: Re: [PATCH 1/3] xfs: honor the O_SYNC flag for aysnchronous direct I/O requests
From: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Date: Sat, 28 Jan 2012 09:59:33 -0500
Cc: linux-ext4@xxxxxxxxxxxxxxx, xfs@xxxxxxxxxxx, linux-fsdevel@xxxxxxxxxxxxxxx
In-reply-to: <1327698949-12616-2-git-send-email-jmoyer@xxxxxxxxxx>
References: <1327698949-12616-1-git-send-email-jmoyer@xxxxxxxxxx> <1327698949-12616-2-git-send-email-jmoyer@xxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
This looks pretty good.  Did this past xfstests?  I'd also like to add
tests actually executing this code path just, to be sure.  E.g. variants
of aio-stress actually using O_SYNC.  We can't easily test data really
made it to disk that way, although at least we make sure the code
doesn't break.

On Fri, Jan 27, 2012 at 04:15:47PM -0500, Jeff Moyer wrote:
> Hi,
> 
> If a file is opened with O_SYNC|O_DIRECT, the drive cache does not get
> flushed after the write completion.  Instead, it's flushed *before* the
> I/O is sent to the disk (in __generic_file_aio_write).

XFS doesn't actually use __generic_file_aio_write, so this sentence
isn't correct for XFS.

> +     } else if (xfs_ioend_needs_cache_flush(ioend)) {
> +             struct xfs_inode *ip = XFS_I(ioend->io_inode);
> +             struct xfs_mount *mp = ip->i_mount;
> +             int     err;
> +             int     log_flushed = 0;
> +
> +             /*
> +              * Check to see if we only need to sync data.  If so,
> +              * we can skip the log flush.
> +              */
> +             if (IS_SYNC(ioend->io_inode) ||
> +                 (ioend->io_iocb->ki_filp->f_flags & __O_SYNC)) {

> +                     err = _xfs_log_force(mp, XFS_LOG_SYNC, &log_flushed);

Can you add a TODO comment that this actually is synchronous and thus
will block the I/O completion work queue?

Also you can use _xfs_log_force_lsn here as don't need to flush the
whole log, just up to the last lsn that touched the inode.  Copy, or
better factor the code from xfs_dir_fsync for that.

Last but not least this won't catch timestamp updates.  Given that I'm
about to send a series making timestamp updates transaction I would not
recommend you to bother with that, but if you want to take a look
at how xfs_file_fsync deals with them.  Given that this series touches
the same area I'd also like to take your xfs patch in through the xfs tree
to avoid conflicts.

> @@ -47,6 +47,7 @@ STATIC int xfsbufd(void *);
>  static struct workqueue_struct *xfslogd_workqueue;
>  struct workqueue_struct *xfsdatad_workqueue;
>  struct workqueue_struct *xfsconvertd_workqueue;
> +struct workqueue_struct *xfsflushd_workqueue;
>  
>  #ifdef XFS_BUF_LOCK_TRACKING
>  # define XB_SET_OWNER(bp)    ((bp)->b_last_holder = current->pid)
> @@ -1802,8 +1803,15 @@ xfs_buf_init(void)
>       if (!xfsconvertd_workqueue)
>               goto out_destroy_xfsdatad_workqueue;
>  
> +     xfsflushd_workqueue = alloc_workqueue("xfsflushd",
> +                                           WQ_MEM_RECLAIM, 1);

This should allow a higher concurrently level, it's probably a good
idea to pass 0 and use the default.

<Prev in Thread] Current Thread [Next in Thread>