[v3,0/8] memcg: avoid flushing stats atomically where possible

Message

Yosry Ahmed March 30, 2023, 7:17 p.m. UTC

rstat flushing is an expensive operation that scales with the number of
cpus and the number of cgroups in the system. The purpose of this series
is to minimize the contexts where we flush stats atomically.

Patches 1 and 2 are cleanups requested during reviews of prior versions
of this series.

Patch 3 makes sure we never try to flush from within an irq context.

Patches 4 to 7 introduce separate variants of mem_cgroup_flush_stats()
for atomic and non-atomic flushing, and make sure we only flush the
stats atomically when necessary.

Patch 8 is a slightly tangential optimization that limits the work done
by rstat flushing in some scenarios.

v2 -> v3:
- Collected more Acks (thanks everyone!).
- Dropped controversial patch 4 from v2.
- Improved commit logs and cover letter (Michal).
v2: https://lore.kernel.org/linux-mm/[e-mail]/

v1 -> v2:
- Added more context in patch 4's commit log.
- Added atomic_read() before atomic_xchg() in patch 5 to avoid
  needlessly locking the cache line (Shakeel).
- Refactored patch 6: added a common helper, do_flush_stats(), for
  mem_cgroup_flush_stats{_atomic}() (Johannes).
- Renamed mem_cgroup_flush_stats_ratelimited() to
  mem_cgroup_flush_stats_atomic_ratelimited() in patch 6. It is restored
  in patch 7, but this maintains consistency (Johannes).
- Added line break to keep the lock section visually separated in patch
  7 (Johannes).
v1: https://lore.kernel.org/lkml/[e-mail]/

RFC -> v1:
- Dropped patch 1 that attempted to make the global rstat lock a non-irq
  lock, will follow up on that separetly (Shakeel).
- Dropped stats_flush_lock entirely, replaced by an atomic (Johannes).
- Renamed cgroup_rstat_flush_irqsafe() to cgroup_rstat_flush_atomic()
  instead of removing it (Johannes).
- Added a patch to rename mem_cgroup_flush_stats_delayed() to
  mem_cgroup_flush_stats_ratelimited() (Johannes).
- Separate APIs for flushing memcg stats in atomic and non-atomic
  contexts instead of a boolean argument (Johannes).
- Added patches 3 & 4 to make sure we never flush from irq context
  (Shakeel & Johannes).
RFC: https://lore.kernel.org/lkml/[e-mail]/

Yosry Ahmed (8):
  cgroup: rename cgroup_rstat_flush_"irqsafe" to "atomic"
  memcg: rename mem_cgroup_flush_stats_"delayed" to "ratelimited"
  memcg: do not flush stats in irq context
  memcg: replace stats_flush_lock with an atomic
  memcg: sleep during flushing stats in safe contexts
  workingset: memcg: sleep when flushing stats in workingset_refault()
  vmscan: memcg: sleep when flushing stats during reclaim
  memcg: do not modify rstat tree for zero updates

 include/linux/cgroup.h     |  2 +-
 include/linux/memcontrol.h |  9 ++++-
 kernel/cgroup/rstat.c      |  4 +-
 mm/memcontrol.c            | 78 ++++++++++++++++++++++++++++++--------
 mm/workingset.c            |  5 ++-
 5 files changed, 76 insertions(+), 22 deletions(-)