xfs
[Top] [All Lists]

Re: [PATCH] xfs: enforce imax_pct when we have per-cpu SB accounting

To: "Darrick J. Wong" <darrick.wong@xxxxxxxxxx>
Subject: Re: [PATCH] xfs: enforce imax_pct when we have per-cpu SB accounting
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Wed, 6 May 2015 08:14:34 +1000
Cc: xfs <xfs@xxxxxxxxxxx>
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20150505220208.GC11601@xxxxxxxxxxxxxxxx>
References: <20150505220208.GC11601@xxxxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Tue, May 05, 2015 at 03:02:08PM -0700, Darrick J. Wong wrote:
> When per-CPU superblock counters are enabled (i.e. SMP=y), we don't
> update the in-core superblock's sb_icount value prior to checking if
> we're already at maximum inodes.  Therefore, it's possible for a tight
> file creation loop (e.g. "for i in {0..100000}; do touch /mnt/$i;
> done") to allocate far more inodes than the administrator specified at
> mkfs time.  Fix this by synchronizing the counters every time we think
> we need to allocate another chunk of inodes.
> 
> Signed-off-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
> ---
>  fs/xfs/libxfs/xfs_ialloc.c |    4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/fs/xfs/libxfs/xfs_ialloc.c b/fs/xfs/libxfs/xfs_ialloc.c
> index 116ef1d..a8aaab6 100644
> --- a/fs/xfs/libxfs/xfs_ialloc.c
> +++ b/fs/xfs/libxfs/xfs_ialloc.c
> @@ -375,6 +375,8 @@ xfs_ialloc_ag_alloc(
>        * at one time.
>        */
>       newlen = args.mp->m_ialloc_inos;
> +     if (args.mp->m_maxicount)
> +             xfs_icsb_sync_counters(args.mp, 0);
>       if (args.mp->m_maxicount &&
>           args.mp->m_sb.sb_icount + newlen > args.mp->m_maxicount)
>               return -ENOSPC;
> @@ -1339,6 +1341,8 @@ xfs_dialloc(
>        * okalloc so we scan all available agi structures for a free
>        * inode.
>        */
> +     if (mp->m_maxicount)
> +             xfs_icsb_sync_counters(mp, 0);
>       if (mp->m_maxicount &&
>           mp->m_sb.sb_icount + mp->m_ialloc_inos > mp->m_maxicount) {
>               noroom = 1;

As i mentioned on IRC, sb_icount is now using the generic percpu
counter infrastrcture, so shouldn't have this problem - overshoot
should be, at maximum, (Ncpus * 32).

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

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