[PATCH] xfs: fix infinite loop by detaching the group/project hints from user dquot

[Dave Chinner]

On Tue, Nov 19, 2013 at 04:42:30PM +0800, Jeff Liu wrote:
> From: Jie Liu <jeff.liu at oracle.com>
> 
> xfs_quota(8) will hang up if trying to turn group quota or project
> quota before the user quota is off, this could be 100% reproduced by:
> # mount -ouquota,gquota /dev/sda7 /xfs
> # mkdir /xfs/test
> # xfs_quota -xc 'off -g' /xfs
> # echo w > /proc/sysrq-trigger
>   SysRq : Show Blocked State
>   task                        PC stack   pid father
>   xfs_quota       D 0000000000000000     0 27574   2551 0x00000000
>   [snip]
>   Call Trace:
>   [<ffffffff81aaa21d>] schedule+0xad/0xc0
>   [<ffffffff81aa327e>] schedule_timeout+0x35e/0x3c0
>   [<ffffffff8114b506>] ? mark_held_locks+0x176/0x1c0
>   [<ffffffff810ad6c0>] ? call_timer_fn+0x2c0/0x2c0
>   [<ffffffffa0c25380>] ? xfs_qm_shrink_count+0x30/0x30 [xfs]
>   [<ffffffff81aa3306>] schedule_timeout_uninterruptible+0x26/0x30
>   [<ffffffffa0c26155>] xfs_qm_dquot_walk+0x235/0x260 [xfs]
>   [<ffffffffa0c059d8>] ? xfs_perag_get+0x1d8/0x2d0 [xfs]
>   [<ffffffffa0c05805>] ? xfs_perag_get+0x5/0x2d0 [xfs]
>   [<ffffffffa0b7707e>] ? xfs_inode_ag_iterator+0xae/0xf0 [xfs]
>   [<ffffffffa0c22280>] ? xfs_trans_free_dqinfo+0x50/0x50 [xfs]
>   [<ffffffffa0b7709f>] ? xfs_inode_ag_iterator+0xcf/0xf0 [xfs]
>   [<ffffffffa0c261e6>] xfs_qm_dqpurge_all+0x66/0xb0 [xfs]
>   [<ffffffffa0c2497a>] xfs_qm_scall_quotaoff+0x20a/0x5f0 [xfs]
>   [<ffffffffa0c2b8f6>] xfs_fs_set_xstate+0x136/0x180 [xfs]
>   [<ffffffff8136cf7a>] do_quotactl+0x53a/0x6b0
>   [<ffffffff812fba4b>] ? iput+0x5b/0x90
>   [<ffffffff8136d257>] SyS_quotactl+0x167/0x1d0
>   [<ffffffff814cf2ee>] ? trace_hardirqs_on_thunk+0x3a/0x3f
>   [<ffffffff81abcd19>] system_call_fastpath+0x16/0x1b
> 
> It's fine if we turn user quota off at first, then turn off other
> kind of quotas if they are enabled since the group/project dquot
> refcount is decreased to zero once the user quota if off.  Otherwse,
> those dquots refcount is non-zero due to the user dquot maybe refer
> to them as hint(s).  Hence, above operation hit an infinite loop at
> xfs_qm_dquot_walk() to purge dquot cache.
> 
> This problem has been around since Linux 3.4, it was introduced by:
> 	b84a3a96751f93071c1863f2962273973c8b8f5e
> 	xfs: remove the per-filesystem list of dquots
> 
> Originally we will release the group dquot pointers because the user
> dquots maybe carrying around as a hint via xfs_qm_detach_gdquots().
> However, with this change, there is no such work to be done before
> purge group/project dquot cache.
> 
> This fix introduce a similar routine to the old xfs_qm_detach_gdquots(),
> it will detach the group/project hints by searching the user dquot radix
> tree and release those hints if they are there.

Ok, we do get stuck in a loop in this case. We need an xfstest to
exercise this - removing group and project quotas while user quotas
are left on and fsstress is active - should be pretty easy to do.

More comments below.

> Signed-off-by: Jie Liu <jeff.liu at oracle.com>
> ---
>  fs/xfs/xfs_qm.c |   79 +++++++++++++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 77 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/xfs/xfs_qm.c b/fs/xfs/xfs_qm.c
> index 14a4996..410adf4 100644
> --- a/fs/xfs/xfs_qm.c
> +++ b/fs/xfs/xfs_qm.c
> @@ -60,6 +60,77 @@ STATIC void	xfs_qm_dqfree_one(struct xfs_dquot *dqp);
>   */
>  #define XFS_DQ_LOOKUP_BATCH	32
>  
> +/*
> + * Release the group or project dquot pointers the user dquots may be
> + * carrying around as a hint.
> + */
> +STATIC void
> +xfs_qm_dqdetach_hint(
> +	struct xfs_mount	*mp,
> +	int			type)
> +{
> +	struct xfs_quotainfo	*qi = mp->m_quotainfo;
> +	struct radix_tree_root	*tree = xfs_dquot_tree(qi, XFS_DQ_USER);
> +	uint32_t		next_index;
> +	int			skipped;
> +	int			nr_found;
> +
> +	ASSERT(type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
> +
> +restart:
> +	next_index = 0;
> +	skipped = 0;
> +	nr_found = 0;
> +
> +	while (1) {
> +		struct xfs_dquot *batch[XFS_DQ_LOOKUP_BATCH];
> +		int		i;
> +
> +		mutex_lock(&qi->qi_tree_lock);
> +		nr_found = radix_tree_gang_lookup(tree, (void **)batch,
> +					next_index, XFS_DQ_LOOKUP_BATCH);
> +		if (!nr_found) {
> +			mutex_unlock(&qi->qi_tree_lock);
> +			break;
> +		}

Why reimplement xfs_qm_dquot_walk?

>From a quick scan of the code, this could be done with
xfs_qm_dquot_walk(XFS_DQ_USER) and an execute function that looks
like the hint handling at the start of xfs_qm_dqpurge()....

>  STATIC int
>  xfs_qm_dquot_walk(
>  	struct xfs_mount	*mp,
> @@ -224,10 +295,14 @@ xfs_qm_dqpurge_all(
>  {
>  	if (flags & XFS_QMOPT_UQUOTA)
>  		xfs_qm_dquot_walk(mp, XFS_DQ_USER, xfs_qm_dqpurge, NULL);
> -	if (flags & XFS_QMOPT_GQUOTA)
> +	if (flags & XFS_QMOPT_GQUOTA) {
> +		xfs_qm_dqdetach_hint(mp, XFS_DQ_GROUP);
>  		xfs_qm_dquot_walk(mp, XFS_DQ_GROUP, xfs_qm_dqpurge, NULL);

and called here like:

	if (flags & XFS_QMOPT_GQUOTA) {
		xfs_qm_dquot_walk(mp, XFS_DQ_USER, xfs_qm_detach_group_hint, NULL);
		xfs_qm_dquot_walk(mp, XFS_DQ_GROUP, xfs_qm_dqpurge, NULL);
	}

In effect, this means we always need to walk the user quota tree,
so we could do all the hint freeing and purging in a single pass by
passing the flags as an argument to a special execution function for
the user dquot tree walk that handles releasing hints and purging
user dquots accordingly. i.e:

	xfs_qm_dquot_walk(mp, XFS_DQ_USER, xfs_qm_dqpurge_hints, &flags);
	if (flags & XFS_QMOPT_GQUOTA)
		xfs_qm_dquot_walk(mp, XFS_DQ_GROUP, xfs_qm_dqpurge, NULL);
	if (flags & XFS_QMOPT_PQUOTA)
		xfs_qm_dquot_walk(mp, XFS_DQ_PROJ, xfs_qm_dqpurge, NULL);

But either way, we don't need to duplicate the radix tree walk code
to release the hints...

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com