| Message ID | 20230330191801.1967435-5-yosryahmed@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | memcg: avoid flushing stats atomically where possible | expand |
Hello. On Thu, Mar 30, 2023 at 07:17:57PM +0000, Yosry Ahmed <yosryahmed@google.com> wrote: > static void __mem_cgroup_flush_stats(void) > { > - unsigned long flag; > - > - if (!spin_trylock_irqsave(&stats_flush_lock, flag)) > + /* > + * We always flush the entire tree, so concurrent flushers can just > + * skip. This avoids a thundering herd problem on the rstat global lock > + * from memcg flushers (e.g. reclaim, refault, etc). > + */ > + if (atomic_read(&stats_flush_ongoing) || > + atomic_xchg(&stats_flush_ongoing, 1)) > return; I'm curious about why this instead of if (atomic_xchg(&stats_flush_ongoing, 1)) return; Is that some microarchitectural cleverness? Thanks, Michal
On Tue, Apr 4, 2023 at 9:53 AM Michal Koutný <mkoutny@suse.com> wrote: > > Hello. > > On Thu, Mar 30, 2023 at 07:17:57PM +0000, Yosry Ahmed <yosryahmed@google.com> wrote: > > static void __mem_cgroup_flush_stats(void) > > { > > - unsigned long flag; > > - > > - if (!spin_trylock_irqsave(&stats_flush_lock, flag)) > > + /* > > + * We always flush the entire tree, so concurrent flushers can just > > + * skip. This avoids a thundering herd problem on the rstat global lock > > + * from memcg flushers (e.g. reclaim, refault, etc). > > + */ > > + if (atomic_read(&stats_flush_ongoing) || > > + atomic_xchg(&stats_flush_ongoing, 1)) > > return; > > I'm curious about why this instead of > > if (atomic_xchg(&stats_flush_ongoing, 1)) > return; > > Is that some microarchitectural cleverness? > Yes indeed it is. Basically we want to avoid unconditional cache dirtying. This pattern is also used at other places in the kernel like qspinlock.
On Tue, Apr 4, 2023 at 10:13 AM Shakeel Butt <shakeelb@google.com> wrote: > > On Tue, Apr 4, 2023 at 9:53 AM Michal Koutný <mkoutny@suse.com> wrote: > > > > Hello. > > > > On Thu, Mar 30, 2023 at 07:17:57PM +0000, Yosry Ahmed <yosryahmed@google.com> wrote: > > > static void __mem_cgroup_flush_stats(void) > > > { > > > - unsigned long flag; > > > - > > > - if (!spin_trylock_irqsave(&stats_flush_lock, flag)) > > > + /* > > > + * We always flush the entire tree, so concurrent flushers can just > > > + * skip. This avoids a thundering herd problem on the rstat global lock > > > + * from memcg flushers (e.g. reclaim, refault, etc). > > > + */ > > > + if (atomic_read(&stats_flush_ongoing) || > > > + atomic_xchg(&stats_flush_ongoing, 1)) > > > return; > > > > I'm curious about why this instead of > > > > if (atomic_xchg(&stats_flush_ongoing, 1)) > > return; > > > > Is that some microarchitectural cleverness? > > > > Yes indeed it is. Basically we want to avoid unconditional cache > dirtying. This pattern is also used at other places in the kernel like > qspinlock. Oh also take a look at https://lore.kernel.org/all/20230404052228.15788-1-feng.tang@intel.com/
On Tue, Apr 04, 2023 at 10:21:33AM -0700, Shakeel Butt <shakeelb@google.com> wrote: > > Yes indeed it is. Basically we want to avoid unconditional cache > > dirtying. This pattern is also used at other places in the kernel like > > qspinlock. Thanks for confirmation. (I remembered the commit 873f64b791a2 ("mm/memcontrol.c: remove the redundant updating of stats_flush_threshold"). But was slightly confused why would it be open-coded every time.) > Oh also take a look at > https://lore.kernel.org/all/20230404052228.15788-1-feng.tang@intel.com/ Thanks for the link. Michal
diff --git a/mm/memcontrol.c b/mm/memcontrol.c index ff39f78f962e..65750f8b8259 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -585,8 +585,8 @@ mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz) */ static void flush_memcg_stats_dwork(struct work_struct *w); static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork); -static DEFINE_SPINLOCK(stats_flush_lock); static DEFINE_PER_CPU(unsigned int, stats_updates); +static atomic_t stats_flush_ongoing = ATOMIC_INIT(0); static atomic_t stats_flush_threshold = ATOMIC_INIT(0); static u64 flush_next_time; @@ -636,15 +636,19 @@ static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val) static void __mem_cgroup_flush_stats(void) { - unsigned long flag; - - if (!spin_trylock_irqsave(&stats_flush_lock, flag)) + /* + * We always flush the entire tree, so concurrent flushers can just + * skip. This avoids a thundering herd problem on the rstat global lock + * from memcg flushers (e.g. reclaim, refault, etc). + */ + if (atomic_read(&stats_flush_ongoing) || + atomic_xchg(&stats_flush_ongoing, 1)) return; - flush_next_time = jiffies_64 + 2*FLUSH_TIME; + WRITE_ONCE(flush_next_time, jiffies_64 + 2*FLUSH_TIME); cgroup_rstat_flush_atomic(root_mem_cgroup->css.cgroup); atomic_set(&stats_flush_threshold, 0); - spin_unlock_irqrestore(&stats_flush_lock, flag); + atomic_set(&stats_flush_ongoing, 0); } void mem_cgroup_flush_stats(void) @@ -655,7 +659,7 @@ void mem_cgroup_flush_stats(void) void mem_cgroup_flush_stats_ratelimited(void) { - if (time_after64(jiffies_64, flush_next_time)) + if (time_after64(jiffies_64, READ_ONCE(flush_next_time))) mem_cgroup_flush_stats(); }
As Johannes notes in [1], stats_flush_lock is currently used to: (a) Protect updated to stats_flush_threshold. (b) Protect updates to flush_next_time. (c) Serializes calls to cgroup_rstat_flush() based on those ratelimits. However: 1. stats_flush_threshold is already an atomic 2. flush_next_time is not atomic. The writer is locked, but the reader is lockless. If the reader races with a flush, you could see this: if (time_after(jiffies, flush_next_time)) spin_trylock() flush_next_time = now + delay flush() spin_unlock() spin_trylock() flush_next_time = now + delay flush() spin_unlock() which means we already can get flushes at a higher frequency than FLUSH_TIME during races. But it isn't really a problem. The reader could also see garbled partial updates if the compiler decides to split the write, so it needs at least READ_ONCE and WRITE_ONCE protection. 3. Serializing cgroup_rstat_flush() calls against the ratelimit factors is currently broken because of the race in 2. But the race is actually harmless, all we might get is the occasional earlier flush. If there is no delta, the flush won't do much. And if there is, the flush is justified. So the lock can be removed all together. However, the lock also served the purpose of preventing a thundering herd problem for concurrent flushers, see [2]. Use an atomic instead to serve the purpose of unifying concurrent flushers. [1]https://lore.kernel.org/lkml/20230323172732.GE739026@cmpxchg.org/ [2]https://lore.kernel.org/lkml/20210716212137.1391164-2-shakeelb@google.com/ Signed-off-by: Yosry Ahmed <yosryahmed@google.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Shakeel Butt <shakeelb@google.com> Acked-by: Michal Hocko <mhocko@suse.com> --- mm/memcontrol.c | 18 +++++++++++------- 1 file changed, 11 insertions(+), 7 deletions(-)