On Fri, Jan 22, 2021 at 01:43:41PM -0500, Johannes Weiner wrote:
> This reverts commit 536d3bf261a2fc3b05b3e91e7eef7383443015cf, as it
> can cause writers to memory.high to get stuck in the kernel forever,
> performing page reclaim and consuming excessive amounts of CPU cycles.
>
> Before the patch, a write to memory.high would first put the new limit
> in place for the workload, and then reclaim the requested delta. After
> the patch, the kernel tries to reclaim the delta before putting the
> new limit into place, in order to not overwhelm the workload with a
> sudden, large excess over the limit. However, if reclaim is actively
> racing with new allocations from the uncurbed workload, it can keep
> the write() working inside the kernel indefinitely.
>
> This is causing problems in Facebook production. A privileged
> system-level daemon that adjusts memory.high for various workloads
> running on a host can get unexpectedly stuck in the kernel and
> essentially turn into a sort of involuntary kswapd for one of the
> workloads. We've observed that daemon busy-spin in a write() for
> minutes at a time, neglecting its other duties on the system, and
> expending privileged system resources on behalf of a workload.
>
> To remedy this, we have first considered changing the reclaim logic to
> break out after a couple of loops - whether the workload has converged
> to the new limit or not - and bound the write() call this way.
> However, the root cause that inspired the sequence change in the first
> place has been fixed through other means, and so a revert back to the
> proven limit-setting sequence, also used by memory.max, is preferable.
>
> The sequence was changed to avoid extreme latencies in the workload
> when the limit was lowered: the sudden, large excess created by the
> limit lowering would erroneously trigger the penalty sleeping code
> that is meant to throttle excessive growth from below. Allocating
> threads could end up sleeping long after the write() had already
> reclaimed the delta for which they were being punished.
>
> However, erroneous throttling also caused problems in other scenarios
> at around the same time. This resulted in commit b3ff92916af3 ("mm,
> memcg: reclaim more aggressively before high allocator throttling"),
> included in the same release as the offending commit. When allocating
> threads now encounter large excess caused by a racing write() to
> memory.high, instead of entering punitive sleeps, they will simply be
> tasked with helping reclaim down the excess, and will be held no
> longer than it takes to accomplish that. This is in line with regular
> limit enforcement - i.e. if the workload allocates up against or over
> an otherwise unchanged limit from below.
>
> With the patch breaking userspace, and the root cause addressed by
> other means already, revert it again.
>
> Fixes: 536d3bf261a2 ("mm: memcontrol: avoid workload stalls when lowering memory.high")
> Cc: <stable@xxxxxxxxxxxxxxx> # 5.8+
> Reported-by: Tejun Heo <tj@xxxxxxxxxx>
> Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
Acked-by: Roman Gushchin <guro@xxxxxx>
