On Mon, Jul 30, 2018 at 12:30:48PM -0700, Darrick J. Wong wrote:
> On Mon, Jul 30, 2018 at 12:45:06PM -0400, Brian Foster wrote:
> > The current transaction allocation code conditionally initializes
> > the ->t_dfops indirection pointer. Transaction commit/cancel check
> > the validity of the pointer to determine whether to finish/cancel
> > the internal dfops.
> >
> > This disallows the ability to use the internal dfops list as a
> > temporary container (via xfs_trans_alloc_empty()). Refactor
> > transaction allocation to always initialize ->t_dfops and check
> > permanent reservation state on transaction commit/cancel.
> >
> > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> > ---
> > fs/xfs/xfs_trans.c | 12 +++---------
> > 1 file changed, 3 insertions(+), 9 deletions(-)
> >
> > diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
> > index 7bf5c1202719..8d3b7f28b193 100644
> > --- a/fs/xfs/xfs_trans.c
> > +++ b/fs/xfs/xfs_trans.c
> > @@ -281,13 +281,7 @@ xfs_trans_alloc(
> > INIT_LIST_HEAD(&tp->t_items);
> > INIT_LIST_HEAD(&tp->t_busy);
> > tp->t_firstblock = NULLFSBLOCK;
> > - /*
> > - * We only roll transactions with permanent log reservation. Don't init
> > - * ->t_dfops to skip attempts to finish or cancel an empty dfops with a
> > - * non-permanent res.
> > - */
> > - if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES)
> > - xfs_defer_init(tp, &tp->t_dfops_internal);
> > + xfs_defer_init(tp, &tp->t_dfops_internal);
> >
> > error = xfs_trans_reserve(tp, resp, blocks, rtextents);
> > if (error) {
> > @@ -932,7 +926,7 @@ __xfs_trans_commit(
> > trace_xfs_trans_commit(tp, _RET_IP_);
> >
> > /* finish deferred items on final commit */
> > - if (!regrant && tp->t_dfops) {
> > + if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {
>
> The usage model of deferred ops is that one has to create a transaction
> with a permanent reservation, and only then start attaching deferred ops
> to the dfops inside the transaction. It's a programming error if a
> caller tries to finish deferred ops using a non-permanent transaction,
> and prior to this patch t_dfops would be NULL and we'd blow up
> immediately in xfs_defer_add(..., tp->t_dfops, ...);
>
> However, now that we initialize t_dfops unconditionally, won't this
> cause the above programming mistake to leak silently any incorrectly
> queued defer ops?
>
Yeah, I'm not sure the previous behavior was really appropriate in that
regard either. I'll add the xfs_defer_add() assert suggested by
Christoph. We could also add a WARN_ON_ONCE() or something in the commit
path to make sure we never see a !perm transaction with dfops, if you
think that's useful..?
Brian
> --D
>
> > error = xfs_defer_finish_noroll(&tp);
> > if (error) {
> > xfs_defer_cancel(tp);
> > @@ -1029,7 +1023,7 @@ xfs_trans_cancel(
> >
> > trace_xfs_trans_cancel(tp, _RET_IP_);
> >
> > - if (tp->t_dfops)
> > + if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)
> > xfs_defer_cancel(tp);
> >
> > /*
> > --
> > 2.17.1
> >
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > the body of a message to majordomo@xxxxxxxxxxxxxxx
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
