mbox series

[RFC,0/7] Make rstat flushing IRQ and sleep friendly

Message ID 20230323040037.2389095-1-yosryahmed@google.com
Headers
Series Make rstat flushing IRQ and sleep friendly |

Message

Yosry Ahmed March 23, 2023, 4 a.m. UTC 
  Currently, if rstat flushing is invoked using the irqsafe variant
cgroup_rstat_flush_irqsafe(), we keep interrupts disabled and do not
sleep for the entire flush operation, which is O(# cpus * # cgroups).
This can be rather dangerous.

Not all contexts that use cgroup_rstat_flush_irqsafe() actually cannot
sleep, and among those that cannot sleep, not all contexts require
interrupts to be disabled. This patch series breaks down the
O(# cpus * # cgroups) duration that we disable interrupts for into a
series of O(# cgroups) durations. Disabling interrupts is deferred to
the caller if needed.

Patch 1 mainly addresses this by not requiring interrupts to be
disabled for the global rstat lock to be acquired. As a side effect of
that, the we disable rstat flushing in interrupt context. See patch 1
for more details.

One thing I am not sure about is whether the only caller of
cgroup_rstat_flush_hold() -- cgroup_base_stat_cputime_show(),
currently has any dependency on that call disabling interrupts.

Patch 2 follows suit for stats_flush_lock in the memcg code, allowing it
to be acquired without disabling interrupts.

Patch 3 removes cgroup_rstat_flush_irqsafe() and updates
cgroup_rstat_flush() to be more explicit about sleeping.

Patch 4 changes memcg code paths that invoke rstat flushing to sleep
where possible. The patch changes code paths where it is naturally saef
to sleep: userspace reads and the background periodic flusher.

Patches 5 & 6 allow sleeping while rstat flushing in reclaim context and
refault context. I am not sure if this is okay, especially the latter,
so I placed them in separate patches for ease of revert/drop.

Patch 7 is a slightly tangential optimization that limits the work done
by rstat flushing in some scenarios.

Yosry Ahmed (7):
  cgroup: rstat: only disable interrupts for the percpu lock
  memcg: do not disable interrupts when holding stats_flush_lock
  cgroup: rstat: remove cgroup_rstat_flush_irqsafe()
  memcg: sleep during flushing stats in safe contexts
  vmscan: memcg: sleep when flushing stats during reclaim
  workingset: memcg: sleep when flushing stats in workingset_refault()
  memcg: do not modify rstat tree for zero updates

 block/blk-cgroup.c         |  2 +-
 include/linux/cgroup.h     |  3 +--
 include/linux/memcontrol.h |  8 +++---
 kernel/cgroup/cgroup.c     |  4 +--
 kernel/cgroup/rstat.c      | 54 ++++++++++++++++++++------------------
 mm/memcontrol.c            | 52 ++++++++++++++++++++++--------------
 mm/vmscan.c                |  2 +-
 mm/workingset.c            |  4 +--
 8 files changed, 73 insertions(+), 56 deletions(-)

Comments

Yosry Ahmed March 23, 2023, 5:07 a.m. UTC | #1 
On Wed, Mar 22, 2023 at 9:10 PM Shakeel Butt <shakeelb@google.com> wrote:
>
> On Wed, Mar 22, 2023 at 9:00 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > Currently, if rstat flushing is invoked using the irqsafe variant
> > cgroup_rstat_flush_irqsafe(), we keep interrupts disabled and do not
> > sleep for the entire flush operation, which is O(# cpus * # cgroups).
> > This can be rather dangerous.
> >
> > Not all contexts that use cgroup_rstat_flush_irqsafe() actually cannot
> > sleep, and among those that cannot sleep, not all contexts require
> > interrupts to be disabled.
>
> Too many negations in the above sentence is making it very confusing.

Sorry, this is indeed very confusing. I guess a better rephrasing is:

Multiple code paths use cgroup_rstat_flush_irqsafe(), but many of them
can sleep. Even among the code paths that actually cannot sleep,
multiple ones are interruptible.