Re: [linus:master] [mm] f95bdb700b: stress-ng.ramfs.ops_per_sec -88.8% regression

From: Qi Zheng
Date: Thu Jun 01 2023 - 06:45:07 EST


Hi Kirill,

On 2023/6/1 18:06, Kirill Tkhai wrote:
01.06.2023, 11:34, "Qi Zheng" <qi.zheng@xxxxxxxxx <mailto:qi.zheng@xxxxxxxxx>>:



On 2023/6/1 08:57, Kirill Tkhai wrote:

 Hi,

 On 30.05.2023 06:07, Qi Zheng wrote:

 Hi,

 On 2023/5/29 20:51, Paul E. McKenney wrote:

 On Mon, May 29, 2023 at 10:39:21AM +0800, Qi Zheng wrote:


 [...]


 Thanks for such a detailed explanation.

 Now I think we can continue to try to complete the
idea[1] from
 Kirill Tkhai. The patch moves heavy
synchronize_srcu() to delayed
 work, so it doesn't affect on user-visible
unregistration speed.

 [1].
https://lore.kernel.org/lkml/153365636747.19074.12610817307548583381.stgit@localhost.localdomain/ <https://lore.kernel.org/lkml/153365636747.19074.12610817307548583381.stgit@localhost.localdomain/>


 A blast from the past!  ;-)

 But yes, moving the long-latency synchronize_srcu()
off the user-visible
 critical code path can be even better.


 Yeah, I applied these patches  ([PATCH RFC 04/10]~[PATCH
RFC 10/10],
 with few conflicts), the ops/s does get back to the
previous levels.

 I'll continue updating this patchset after doing more testing.


 You may also fix the issue using the below generic solution.

 In addition to this we need patch, which calls
unregister_shrinker_delayed_initiate()
 instead of unregister_shrinker() in deactivate_locked_super(),
and calls
 unregister_shrinker_delayed_finalize() from
destroy_super_work(). Compilation tested only.

 ---
   include/linux/shrinker.h | 2 ++
   mm/vmscan.c | 38 ++++++++++++++++++++++++++++++--------
   2 files changed, 32 insertions(+), 8 deletions(-)
 diff --git a/include/linux/shrinker.h b/include/linux/shrinker.h
 index 224293b2dd06..4ba2986716d3 100644
 --- a/include/linux/shrinker.h
 +++ b/include/linux/shrinker.h
 @@ -4,6 +4,7 @@

   #include <linux/atomic.h>
   #include <linux/types.h>
 +#include <linux/rwsem.h>

   /*
    * This struct is used to pass information from page reclaim
to the shrinkers.
 @@ -83,6 +84,7 @@ struct shrinker {
   #endif
           /* objs pending delete, per node */
           atomic_long_t *nr_deferred;
 + struct rw_semaphore rwsem;
   };
   #define DEFAULT_SEEKS 2 /* A good number if you don't know
better. */

 diff --git a/mm/vmscan.c b/mm/vmscan.c
 index eeca83e28c9b..19fc129771ce 100644
 --- a/mm/vmscan.c
 +++ b/mm/vmscan.c
 @@ -706,6 +706,7 @@ static int __prealloc_shrinker(struct
shrinker *shrinker)
           if (!shrinker->nr_deferred)
                   return -ENOMEM;

 + init_rwsem(&shrinker->rwsem);
           return 0;
   }

 @@ -757,7 +758,9 @@ void register_shrinker_prepared(struct
shrinker *shrinker)
   {
           mutex_lock(&shrinker_mutex);
           list_add_tail_rcu(&shrinker->list, &shrinker_list);
 + down_write(&shrinker->rwsem);
           shrinker->flags |= SHRINKER_REGISTERED;
 + up_write(&shrinker->rwsem);
           shrinker_debugfs_add(shrinker);
           mutex_unlock(&shrinker_mutex);
   }
 @@ -802,7 +805,7 @@ EXPORT_SYMBOL(register_shrinker);
   /*
    * Remove one
    */
 -void unregister_shrinker(struct shrinker *shrinker)
 +void unregister_shrinker_delayed_initiate(struct shrinker
*shrinker)
   {
           struct dentry *debugfs_entry;
           int debugfs_id;
 @@ -812,20 +815,33 @@ void unregister_shrinker(struct shrinker
*shrinker)

           mutex_lock(&shrinker_mutex);
           list_del_rcu(&shrinker->list);
 + down_write(&shrinker->rwsem);
           shrinker->flags &= ~SHRINKER_REGISTERED;
 + up_write(&shrinker->rwsem);
           if (shrinker->flags & SHRINKER_MEMCG_AWARE)
                   unregister_memcg_shrinker(shrinker);
           debugfs_entry = shrinker_debugfs_detach(shrinker,
&debugfs_id);
           mutex_unlock(&shrinker_mutex);

 + shrinker_debugfs_remove(debugfs_entry, debugfs_id); // This
is moved in your patch
 +}
 +EXPORT_SYMBOL(unregister_shrinker_delayed_initiate);
 +
 +void unregister_shrinker_delayed_finalize(struct shrinker
*shrinker)
 +{
           atomic_inc(&shrinker_srcu_generation);
           synchronize_srcu(&shrinker_srcu);

 - shrinker_debugfs_remove(debugfs_entry, debugfs_id);
 -
           kfree(shrinker->nr_deferred);
           shrinker->nr_deferred = NULL;
   }
 +EXPORT_SYMBOL(unregister_shrinker_delayed_finalize);
 +
 +void unregister_shrinker(struct shrinker *shrinker)
 +{
 + unregister_shrinker_delayed_initiate(shrinker);
 + unregister_shrinker_delayed_finalize(shrinker);
 +}
   EXPORT_SYMBOL(unregister_shrinker);

   /**
 @@ -856,9 +872,14 @@ static unsigned long
do_shrink_slab(struct shrink_control *shrinkctl,
                                             : SHRINK_BATCH;
           long scanned = 0, next_deferred;

 + down_read(&shrinker->rwsem);
 + if (!(shrinker->flags & SHRINKER_REGISTERED))
 + goto unlock;
           freeable = shrinker->count_objects(shrinker, shrinkctl);
 - if (freeable == 0 || freeable == SHRINK_EMPTY)
 - return freeable;
 + if (freeable == 0 || freeable == SHRINK_EMPTY) {
 + freed = freeable;
 + goto unlock;
 + }

           /*
            * copy the current shrinker scan count into a local
variable
 @@ -935,6 +956,8 @@ static unsigned long do_shrink_slab(struct
shrink_control *shrinkctl,
            * manner that handles concurrent updates.
            */
           new_nr = add_nr_deferred(next_deferred, shrinker,
shrinkctl);
 +unlock:
 + up_read(&shrinker->rwsem);


It should be moved after trace_mm_shrink_slab_end().

Could you explain the reason? I don't see the variable it will protect.

We jump to unlock label before trace_mm_shrink_slab_start(), so I
think we should not go to call trace_mm_shrink_slab_end().



           trace_mm_shrink_slab_end(shrinker, shrinkctl->nid,
freed, nr, new_nr, total_scan);
           return freed;
 @@ -968,9 +991,8 @@ static unsigned long
shrink_slab_memcg(gfp_t gfp_mask, int nid,
                   struct shrinker *shrinker;

                   shrinker = idr_find(&shrinker_idr, i);
 - if (unlikely(!shrinker || !(shrinker->flags &
SHRINKER_REGISTERED))) {
 - if (!shrinker)
 - clear_bit(i, info->map);
 + if (unlikely(!shrinker)) {
 + clear_bit(i, info->map);
                           continue;
                   }


Keep this as a fast path?

Probably, yes. And also we need 1)down_trylock() instead of down_read()  and 2)rwsem_is_contended  in do_shrink_slab().

Agree, although the critical section of the writer of shrinker->rwsem is
very short, this prevents unnecessary sleeps.



After applying the above patch, the performance regression problem of
ops/s can be solved. And it can be guaranteed that the shrinker is not
running after unregister_shrinker_delayed_initiate(), so the previous
semantics are not broken.

Keeping old semantics or not is quite subjective, I think. It's possible to provide strong arguments for both cases. First is faster, second is easier to adopt for users. For me personally the faster approach looks better.

Agree. I also like completely lock-less slab shrink.

>nce the lock granularity of down_read() has changed to the granularity

of per shrinker, it seems that the down_read() perf hotspot will not be
very high. I'm not quite sure why.

(The test script used is the script in
https://lore.kernel.org/all/20230313112819.38938-4-zhengqi.arch@xxxxxxxxxxxxx/ <https://lore.kernel.org/all/20230313112819.38938-4-zhengqi.arch@xxxxxxxxxxxxx/>)

Hmm, possible original shrinker_rwsem had a lot of atomic intersections between cpus on down_read(), while with small locks it has not. Are

I guess so.

CONFIG_ for locks debug are same in original and this case?

Basically yes, I will do some more testing.

Thanks,
Qi



   25.28% [kernel] [k] do_shrink_slab
   21.91% [kernel] [k] pv_native_safe_halt
   10.81% [kernel] [k] _find_next_bit
   10.47% [kernel] [k] down_read
    8.75% [kernel] [k] shrink_slab
    4.03% [kernel] [k] up_read
    3.29% [kernel] [k] shrink_lruvec
    2.75% [kernel] [k] xa_load
    2.73% [kernel] [k] mem_cgroup_iter
    2.67% [kernel] [k] shrink_node
    1.30% [kernel] [k] list_lru_count_one

Thanks,
Qi



--
Thanks,
Qi