Re: [PATCH] mm: vmscan: fix IO/refault regression in cache workingset transition

[Johannes Weiner]

On Tue, Apr 04, 2017 at 03:07:03PM -0700, Andrew Morton wrote:
> On Tue,  4 Apr 2017 18:00:52 -0400 Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> 
> > Since 59dc76b0d4df ("mm: vmscan: reduce size of inactive file list")
> > we noticed bigger IO spikes during changes in cache access patterns.
> > 
> > The patch in question shrunk the inactive list size to leave more room
> > for the current workingset in the presence of streaming IO. However,
> > workingset transitions that previously happened on the inactive list
> > are now pushed out of memory and incur more refaults to complete.
> > 
> > This patch disables active list protection when refaults are being
> > observed. This accelerates workingset transitions, and allows more of
> > the new set to establish itself from memory, without eating into the
> > ability to protect the established workingset during stable periods.
> > 
> > Fixes: 59dc76b0d4df ("mm: vmscan: reduce size of inactive file list")
> > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx> # 4.7+
> 
> That's a pretty large patch and the problem has been there for a year. 
> I'm not sure that it's 4.11 material, let alone -stable.  Care to
> explain further?

The problem statement is a little terse, my apologies.

The workloads that were measurably affected for us were hit pretty bad
by it, with refault/majfault rates doubling and tripling during cache
transitions, and the machines sustaining half-hour periods of 100% IO
utilization, where they'd previously have sub-minute peaks at 60-90%.

Stateful services that handle user data tend to be more conservative
with kernel upgrades. As a result we hit most page cache issues with
some delay, as was the case here.

The severity seemed to warrant a stable tag, but I agree that holding
out until 4.11.1 is probably better, given the invasiveness of this.