releases/3.4.77/sched-fix-race-on-toggling-cfs_bandwidth_used.patch

   1 From 1ee14e6c8cddeeb8a490d7b54cd9016e4bb900b4 Mon Sep 17 00:00:00 2001
   2 From: Ben Segall <bsegall@google.com>
   3 Date: Wed, 16 Oct 2013 11:16:12 -0700
   4 Subject: sched: Fix race on toggling cfs_bandwidth_used
   5
   6 From: Ben Segall <bsegall@google.com>
   7
   8 commit 1ee14e6c8cddeeb8a490d7b54cd9016e4bb900b4 upstream.
   9
  10 When we transition cfs_bandwidth_used to false, any currently
  11 throttled groups will incorrectly return false from cfs_rq_throttled.
  12 While tg_set_cfs_bandwidth will unthrottle them eventually, currently
  13 running code (including at least dequeue_task_fair and
  14 distribute_cfs_runtime) will cause errors.
  15
  16 Fix this by turning off cfs_bandwidth_used only after unthrottling all
  17 cfs_rqs.
  18
  19 Tested: toggle bandwidth back and forth on a loaded cgroup. Caused
  20 crashes in minutes without the patch, hasn't crashed with it.
  21
  22 Signed-off-by: Ben Segall <bsegall@google.com>
  23 Signed-off-by: Peter Zijlstra <peterz@infradead.org>
  24 Cc: pjt@google.com
  25 Link: http://lkml.kernel.org/r/20131016181611.22647.80365.stgit@sword-of-the-dawn.mtv.corp.google.com
  26 Signed-off-by: Ingo Molnar <mingo@kernel.org>
  27 Cc: Chris J Arges <chris.j.arges@canonical.com>
  28 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  29
  30 ---
  31  kernel/sched/core.c  |    9 ++++++++-
  32  kernel/sched/fair.c  |   16 +++++++++-------
  33  kernel/sched/sched.h |    3 ++-
  34  3 files changed, 19 insertions(+), 9 deletions(-)
  35
  36 --- a/kernel/sched/core.c
  37 +++ b/kernel/sched/core.c
  38 @@ -7906,7 +7906,12 @@ static int tg_set_cfs_bandwidth(struct t
  39
  40         runtime_enabled = quota != RUNTIME_INF;
  41         runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
  42 -       account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
  43 +       /*
  44 +        * If we need to toggle cfs_bandwidth_used, off->on must occur
  45 +        * before making related changes, and on->off must occur afterwards
  46 +        */
  47 +       if (runtime_enabled && !runtime_was_enabled)
  48 +               cfs_bandwidth_usage_inc();
  49         raw_spin_lock_irq(&cfs_b->lock);
  50         cfs_b->period = ns_to_ktime(period);
  51         cfs_b->quota = quota;
  52 @@ -7932,6 +7937,8 @@ static int tg_set_cfs_bandwidth(struct t
  53                         unthrottle_cfs_rq(cfs_rq);
  54                 raw_spin_unlock_irq(&rq->lock);
  55         }
  56 +       if (runtime_was_enabled && !runtime_enabled)
  57 +               cfs_bandwidth_usage_dec();
  58  out_unlock:
  59         mutex_unlock(&cfs_constraints_mutex);
  60
  61 --- a/kernel/sched/fair.c
  62 +++ b/kernel/sched/fair.c
  63 @@ -1393,13 +1393,14 @@ static inline bool cfs_bandwidth_used(vo
  64         return static_key_false(&__cfs_bandwidth_used);
  65  }
  66
  67 -void account_cfs_bandwidth_used(int enabled, int was_enabled)
  68 +void cfs_bandwidth_usage_inc(void)
  69  {
  70 -       /* only need to count groups transitioning between enabled/!enabled */
  71 -       if (enabled && !was_enabled)
  72 -               static_key_slow_inc(&__cfs_bandwidth_used);
  73 -       else if (!enabled && was_enabled)
  74 -               static_key_slow_dec(&__cfs_bandwidth_used);
  75 +       static_key_slow_inc(&__cfs_bandwidth_used);
  76 +}
  77 +
  78 +void cfs_bandwidth_usage_dec(void)
  79 +{
  80 +       static_key_slow_dec(&__cfs_bandwidth_used);
  81  }
  82  #else /* HAVE_JUMP_LABEL */
  83  static bool cfs_bandwidth_used(void)
  84 @@ -1407,7 +1408,8 @@ static bool cfs_bandwidth_used(void)
  85         return true;
  86  }
  87
  88 -void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
  89 +void cfs_bandwidth_usage_inc(void) {}
  90 +void cfs_bandwidth_usage_dec(void) {}
  91  #endif /* HAVE_JUMP_LABEL */
  92
  93  /*
  94 --- a/kernel/sched/sched.h
  95 +++ b/kernel/sched/sched.h
  96 @@ -1140,7 +1140,8 @@ extern void init_cfs_rq(struct cfs_rq *c
  97  extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq);
  98  extern void unthrottle_offline_cfs_rqs(struct rq *rq);
  99
 100 -extern void account_cfs_bandwidth_used(int enabled, int was_enabled);
 101 +extern void cfs_bandwidth_usage_inc(void);
 102 +extern void cfs_bandwidth_usage_dec(void);
 103
 104  #ifdef CONFIG_NO_HZ
 105  enum rq_nohz_flag_bits {