xfs
[Top] [All Lists]

Re: [PATCH 3/5] xfs: reduce ioend latency

To: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Subject: Re: [PATCH 3/5] xfs: reduce ioend latency
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 22 Aug 2011 16:43:45 +1000
Cc: xfs@xxxxxxxxxxx
In-reply-to: <20110814222518.001177781@xxxxxxxxxxxxxxxxxxxxxx>
References: <20110814222412.359079843@xxxxxxxxxxxxxxxxxxxxxx> <20110814222518.001177781@xxxxxxxxxxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Sun, Aug 14, 2011 at 06:24:15PM -0400, Christoph Hellwig wrote:
> There is no reason to queue up ioends for processing in user context
> unless we actually need it.  Just complete ioends that do not convert
> unwritten extents or need a size update from the end_io context.
> 
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> 
> Index: xfs/fs/xfs/xfs_aops.c
> ===================================================================
> --- xfs.orig/fs/xfs/xfs_aops.c        2011-08-13 10:57:57.559366326 -0700
> +++ xfs/fs/xfs/xfs_aops.c     2011-08-13 10:57:57.979364052 -0700
> @@ -150,6 +150,15 @@ xfs_ioend_new_eof(
>  }
>  
>  /*
> + * Fast and loose check if this write could update the on-disk inode size.
> + */
> +static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
> +{
> +     return ioend->io_offset + ioend->io_size >
> +             XFS_I(ioend->io_inode)->i_d.di_size;
> +}
> +
> +/*
>   * Update on-disk file size now that data has been written to disk.  The
>   * current in-memory file size is i_size.  If a write is beyond eof 
> i_new_size
>   * will be the intended file size until i_size is updated.  If this write 
> does
> @@ -186,6 +195,9 @@ xfs_setfilesize(
>  
>  /*
>   * Schedule IO completion handling on the final put of an ioend.
> + *
> + * If there is no work to do we might as well call it a day and free the
> + * ioend right now.
>   */
>  STATIC void
>  xfs_finish_ioend(
> @@ -194,8 +206,10 @@ xfs_finish_ioend(
>       if (atomic_dec_and_test(&ioend->io_remaining)) {
>               if (ioend->io_type == IO_UNWRITTEN)
>                       queue_work(xfsconvertd_workqueue, &ioend->io_work);
> -             else
> +             else if (xfs_ioend_is_append(ioend))
>                       queue_work(xfsdatad_workqueue, &ioend->io_work);
> +             else
> +                     xfs_destroy_ioend(ioend);
>       }
>  }

That's similar to a check I added in a previous patch series to
avoid taking the ILOCK in IO completion if it wasn't necessary. THis
just checks earlier to avoid the workqueue switch, so it definitely
better than what I did.

Reviewed-by: Dave Chinner <dchinner@xxxxxxxxxx>

-- 
Dave Chinner
david@xxxxxxxxxxxxx

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