mm, memcg: throttle allocators when failing reclaim over memory.high

We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup.  This happens especially with swap space -- either
when none is configured, or swap is full.  These failures often also don't
have enough warning to allow one to react, whether for a human or for a
daemon monitoring PSI.

Here is output from a simple program showing how long it takes in usec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:

    [root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
    > --wait -t timeout 300 /root/mdf
    [...]
    95  1035
    96  1038
    97  1000
    98  1036
    99  1048
    100 1590
    101 1968
    102 1776
    103 1863
    104 1757
    105 1921
    106 1893
    107 1760
    108 1748
    109 1843
    110 1716
    111 1924
    112 1776
    113 1831
    114 1766
    115 1836
    116 1588
    117 1912
    118 1802
    119 1857
    120 1731
    [...]
    [System OOM in 2-3 seconds]

The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode, but
it's nowhere near enough to contain growth.  It also doesn't get worse the
more pages you have, since it only considers nr_pages.

The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers.  In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached.  Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads as
they exceed their high threshold.  However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with no
swap, or where swap has become exhausted, we can end up with memory.high
being breached and us having no weapons left in our arsenal to combat
runaway growth with, since reclaim is futile.

It's also hard for system monitoring software or users to tell how bad the
situation is, as "high" events for the memcg may in some cases be benign,
and in others be catastrophic.  The current status quo is that we fail
containment in a way that doesn't provide any advance warning that things
are about to go horribly wrong (for example, we are about to invoke the
kernel OOM killer).

This patch introduces explicit throttling when reclaim is failing to keep
memcg size contained at the memory.high setting.  It does so by applying
an exponential delay curve derived from the memcg's overage compared to
memory.high.  In the normal case where the memcg is either below or only
marginally over its memory.high setting, no throttling will be performed.

This composes well with system health monitoring and remediation, as these
allocator delays are factored into PSI's memory pressure calculations.
This both creates a mechanism system administrators or applications
consuming the PSI interface to trivially see that the memcg in question is
struggling and use that to make more reasonable decisions, and permits
them enough time to act.  Either of these can act with significantly more
nuance than that we can provide using the system OOM killer.

This is a similar idea to memory.oom_control in cgroup v1 which would put
the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and also
doesn't schedule indefinitely, which previously made tracing and other
introspection difficult (ie.  it's clamped at 2*HZ per allocation through
MEMCG_MAX_HIGH_DELAY_JIFFIES).

Contrast the previous results with a kernel with this patch:

    [root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
    > --wait -t timeout 300 /root/mdf
    [...]
    95  1002
    96  1000
    97  1002
    98  1003
    99  1000
    100 1043
    101 84724
    102 330628
    103 610511
    104 1016265
    105 1503969
    106 2391692
    107 2872061
    108 3248003
    109 4791904
    110 5759832
    111 6912509
    112 8127818
    113 9472203
    114 12287622
    115 12480079
    116 14144008
    117 15808029
    118 16384500
    119 16383242
    120 16384979
    [...]

As you can see, in the normal case, memory allocation takes around 1000
usec.  However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first.  Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then the
next is almost an entire second.  This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.

We use an exponential curve for our delay penalty for a few reasons:

1. We run mem_cgroup_handle_over_high to potentially do reclaim after
   we've already performed allocations, which means that temporarily
   going over memory.high by a small amount may be perfectly legitimate,
   even for compliant workloads. We don't want to unduly penalise such
   cases.
2. An exponential curve (as opposed to a static or linear delay) allows
   ramping up memory pressure stats more gradually, which can be useful
   to work out that you have set memory.high too low, without destroying
   application performance entirely.

This patch expands on earlier work by Johannes Weiner. Thanks!

[akpm@linux-foundation.org: fix max() warning]
[akpm@linux-foundation.org: fix __udivdi3 ref on 32-bit]
[akpm@linux-foundation.org: fix it even more]
[chris@chrisdown.name: fix 64-bit divide even more]
Link: http://lkml.kernel.org/r/20190723180700.GA29459@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
1 file changed