On Thu, Sep 22, 2016 at 10:41:50AM -0700, Hugh Dickins wrote:
> init_tlb_ubc() looked unnecessary to me: tlb_ubc is statically initialized
> with zeroes in the init_task, and copied from parent to child while it is
> quiescent in arch_dup_task_struct(); so I went to delete it.
>
> But inserted temporary debug WARN_ONs in place of init_tlb_ubc() to check
> that it was always empty at that point, and found them firing: because
> memcg reclaim can recurse into global reclaim (when allocating biosets
> for swapout in my case), and arrive back at the init_tlb_ubc() in
> shrink_node_memcg().
>
> Resetting tlb_ubc.flush_required at that point is wrong: if the upper
> level needs a deferred TLB flush, but the lower level turns out not to,
> we miss a TLB flush. But fortunately, that's the only part of the
> protocol that does not nest: with the initialization removed, cpumask
> collects bits from upper and lower levels, and flushes TLB when needed.
>
> Fixes: 72b252aed506 ("mm: send one IPI per CPU to TLB flush all entries after
> unmapping pages")
> Signed-off-by: Hugh Dickins <hu...@google.com>
> Cc: sta...@vger.kernel.org # 4.3+
Acked-by: Mel Gorman <mgor...@techsingularity.net>
--
Mel Gorman
SUSE Labs
