* [PATCH 1/2] mm: Avoid full RCU lookup of memcg for statistics updates
2014-07-23 11:24 [PATCH 0/2] Avoid unnecessary overhead in fault paths due to memcg and rss stats Mel Gorman
@ 2014-07-23 11:24 ` Mel Gorman
2014-07-23 11:44 ` Johannes Weiner
2014-07-23 11:24 ` [PATCH 2/2] mm: Move RSS stat event count synchronisation out of the fast path Mel Gorman
1 sibling, 1 reply; 5+ messages in thread
From: Mel Gorman @ 2014-07-23 11:24 UTC (permalink / raw)
To: Andrew Morton
Cc: Johannes Weiner, Michal Hocko, Mel Gorman, Linux Kernel, Linux-MM
When updating memcg VM statistics like PGFAULT we take the rcu read
lock and lookup the memcg. For statistic updates this is overkill
when the process may not belong to a memcg. This patch adds a light
check to check if a memcg potentially exists. It's race-prone in that
some VM stats may be missed when a process first joins a memcg but
that is not serious enough to justify a constant performance penalty.
The exact impact of this is difficult to quantify because it's timing
sensitive, workload sensitive and sensitive to the RCU options set. However,
broadly speaking there should be less interference due to page fault
activity in both the number of RCU grace periods and their age.
Signed-off-by: Mel Gorman <mgorman@suse.de>
---
include/linux/memcontrol.h | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index eb65d29..76fa97d 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -220,6 +220,14 @@ static inline void mem_cgroup_count_vm_event(struct mm_struct *mm,
{
if (mem_cgroup_disabled())
return;
+ /*
+ * For statistic updates it's overkill to take the RCU lock and do
+ * a fully safe lookup of an associated memcg. Do a simple check
+ * first. At worst, we miss a few stat updates when a process is
+ * moved to a memcg for the first time.
+ */
+ if (!rcu_access_pointer(mm->owner))
+ return;
__mem_cgroup_count_vm_event(mm, idx);
}
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
--
1.8.4.5
^ permalink raw reply related [flat|nested] 5+ messages in thread* [PATCH 2/2] mm: Move RSS stat event count synchronisation out of the fast path
2014-07-23 11:24 [PATCH 0/2] Avoid unnecessary overhead in fault paths due to memcg and rss stats Mel Gorman
2014-07-23 11:24 ` [PATCH 1/2] mm: Avoid full RCU lookup of memcg for statistics updates Mel Gorman
@ 2014-07-23 11:24 ` Mel Gorman
1 sibling, 0 replies; 5+ messages in thread
From: Mel Gorman @ 2014-07-23 11:24 UTC (permalink / raw)
To: Andrew Morton
Cc: Johannes Weiner, Michal Hocko, Mel Gorman, Linux Kernel, Linux-MM
With split RSS counters there is a per-task RSS counter that is synced with
the mm counters every 64 page faults in the fast path. Not all faults result
in modifications to these stats and the sync is potentially a waste. This
patch synchronises the counts to synchronise when the counter overflows
the threshold. It may mean that drift can be long-lived if the number of
events was limited but the amount of drift will be bounded.
Unlike the previous patch this is easier to quantify by anything that is
fault intensive by monitoring CPU usage.
kernbench
3.16.0-rc5 3.16.0-rc5
memcgstat-preempt-v1r1 rssthresh-preempt-v1r1
User min 579.44 ( 0.00%) 580.04 ( -0.10%)
User mean 579.92 ( 0.00%) 580.70 ( -0.13%)
User stddev 0.27 ( 0.00%) 0.49 (-84.55%)
User max 580.25 ( 0.00%) 581.33 ( -0.19%)
User range 0.81 ( 0.00%) 1.29 (-59.26%)
System min 35.94 ( 0.00%) 35.93 ( 0.03%)
System mean 36.22 ( 0.00%) 36.05 ( 0.46%)
System stddev 0.21 ( 0.00%) 0.10 ( 50.88%)
System max 36.53 ( 0.00%) 36.24 ( 0.79%)
System range 0.59 ( 0.00%) 0.31 ( 47.46%)
Elapsed min 83.99 ( 0.00%) 83.98 ( 0.01%)
Elapsed mean 84.17 ( 0.00%) 84.31 ( -0.17%)
Elapsed stddev 0.21 ( 0.00%) 0.22 ( -3.37%)
Elapsed max 84.55 ( 0.00%) 84.51 ( 0.05%)
Elapsed range 0.56 ( 0.00%) 0.53 ( 5.36%)
CPU min 728.00 ( 0.00%) 729.00 ( -0.14%)
CPU mean 731.60 ( 0.00%) 731.00 ( 0.08%)
CPU stddev 1.96 ( 0.00%) 2.10 ( -7.04%)
CPU max 733.00 ( 0.00%) 734.00 ( -0.14%)
CPU range 5.00 ( 0.00%) 5.00 ( 0.00%)
memcgstat- rssthresh
User 7313.65 7210.70
System 484.33 466.93
Elapsed 1154.23 1133.01
page fault test
3.16.0-rc5 3.16.0-rc5
memcgstat-preempt-v1r1rssthresh-preempt-v1r1
System 1 0.4790 ( 0.00%) 0.4980 ( -3.97%)
System 2 0.5120 ( 0.00%) 0.5035 ( 1.66%)
System 3 0.5410 ( 0.00%) 0.5355 ( 1.02%)
System 4 0.6805 ( 0.00%) 0.6735 ( 1.03%)
System 5 0.8530 ( 0.00%) 0.8560 ( -0.35%)
System 6 1.0425 ( 0.00%) 1.0360 ( 0.62%)
System 7 1.2645 ( 0.00%) 1.2525 ( 0.95%)
System 8 1.5070 ( 0.00%) 1.5040 ( 0.20%)
Elapsed 1 0.5630 ( 0.00%) 0.5835 ( -3.64%)
Elapsed 2 0.3010 ( 0.00%) 0.2955 ( 1.83%)
Elapsed 3 0.2105 ( 0.00%) 0.2105 ( 0.00%)
Elapsed 4 0.1965 ( 0.00%) 0.1920 ( 2.29%)
Elapsed 5 0.2145 ( 0.00%) 0.2190 ( -2.10%)
Elapsed 6 0.2055 ( 0.00%) 0.2040 ( 0.73%)
Elapsed 7 0.2080 ( 0.00%) 0.2020 ( 2.88%)
Elapsed 8 0.2100 ( 0.00%) 0.2100 ( 0.00%)
3.16.0-rc5 3.16.0-rc5
memcgstat-preempt-v1r1rssthresh-preempt-v1r1
User 109.91 104.41
System 370.77 360.09
Elapsed 274.71 266.62
The difference is marginal in the overall cost of a page fault but there
is no point having unnecessary overhead either.
Signed-off-by: Mel Gorman <mgorman@suse.de>
---
include/linux/mm_types.h | 1 -
mm/memory.c | 32 +++++++++++++-------------------
2 files changed, 13 insertions(+), 20 deletions(-)
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 96c5750..c9404e4 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -332,7 +332,6 @@ enum {
#define SPLIT_RSS_COUNTING
/* per-thread cached information, */
struct task_rss_stat {
- int events; /* for synchronization threshold */
int count[NR_MM_COUNTERS];
};
#endif /* USE_SPLIT_PTE_PTLOCKS */
diff --git a/mm/memory.c b/mm/memory.c
index d67fd9f..26b41be 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -131,6 +131,9 @@ core_initcall(init_zero_pfn);
#if defined(SPLIT_RSS_COUNTING)
+/* sync counter once per 64 rss stat update events */
+#define TASK_RSS_EVENTS_THRESH (64)
+
void sync_mm_rss(struct mm_struct *mm)
{
int i;
@@ -141,39 +144,33 @@ void sync_mm_rss(struct mm_struct *mm)
current->rss_stat.count[i] = 0;
}
}
- current->rss_stat.events = 0;
}
static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
{
struct task_struct *task = current;
- if (likely(task->mm == mm))
+ if (likely(task->mm == mm)) {
task->rss_stat.count[member] += val;
- else
+ if (task != current)
+ return;
+
+ if (task->rss_stat.count[member] > TASK_RSS_EVENTS_THRESH ||
+ task->rss_stat.count[member] < -TASK_RSS_EVENTS_THRESH) {
+ add_mm_counter(mm, member, current->rss_stat.count[member]);
+ current->rss_stat.count[member] = 0;
+ }
+ } else
add_mm_counter(mm, member, val);
}
#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
-/* sync counter once per 64 page faults */
-#define TASK_RSS_EVENTS_THRESH (64)
-static void check_sync_rss_stat(struct task_struct *task)
-{
- if (unlikely(task != current))
- return;
- if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
- sync_mm_rss(task->mm);
-}
#else /* SPLIT_RSS_COUNTING */
#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
-static void check_sync_rss_stat(struct task_struct *task)
-{
-}
-
#endif /* SPLIT_RSS_COUNTING */
#ifdef HAVE_GENERIC_MMU_GATHER
@@ -3319,9 +3316,6 @@ int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
count_vm_event(PGFAULT);
mem_cgroup_count_vm_event(mm, PGFAULT);
- /* do counter updates before entering really critical section. */
- check_sync_rss_stat(current);
-
/*
* Enable the memcg OOM handling for faults triggered in user
* space. Kernel faults are handled more gracefully.
--
1.8.4.5
^ permalink raw reply related [flat|nested] 5+ messages in thread