Message-ID: <00a212ee-1433-937f-1f15-f82e3137778c@bytedance.com>
Date:   Fri, 24 Feb 2023 12:16:26 +0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:102.0)
 Gecko/20100101 Thunderbird/102.7.2
Subject: Re: [PATCH v2 2/7] mm: vmscan: make global slab shrink lockless
Content-Language: en-US
From:   Qi Zheng <zhengqi.arch@bytedance.com>
To:     Sultan Alsawaf <sultan@kerneltoast.com>, paulmck@kernel.org
Cc:     akpm@linux-foundation.org, tkhai@ya.ru, hannes@cmpxchg.org,
        shakeelb@google.com, mhocko@kernel.org, roman.gushchin@linux.dev,
        muchun.song@linux.dev, david@redhat.com, shy828301@gmail.com,
        dave@stgolabs.net, penguin-kernel@i-love.sakura.ne.jp,
        linux-mm@kvack.org, linux-kernel@vger.kernel.org
References: <20230223132725.11685-1-zhengqi.arch@bytedance.com>
 <20230223132725.11685-3-zhengqi.arch@bytedance.com>
 <Y/evb+PBeaahx9Os@sultan-box.localdomain>
 <8049b6ed-435f-b518-f947-5516a514aec2@bytedance.com>
In-Reply-To: <8049b6ed-435f-b518-f947-5516a514aec2@bytedance.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 2023/2/24 12:00, Qi Zheng wrote:
> 
> 
> On 2023/2/24 02:24, Sultan Alsawaf wrote:
>> On Thu, Feb 23, 2023 at 09:27:20PM +0800, Qi Zheng wrote:
>>> The shrinker_rwsem is a global lock in shrinkers subsystem,
>>> it is easy to cause blocking in the following cases:
>>>
>>> a. the write lock of shrinker_rwsem was held for too long.
>>>     For example, there are many memcgs in the system, which
>>>     causes some paths to hold locks and traverse it for too
>>>     long. (e.g. expand_shrinker_info())
>>> b. the read lock of shrinker_rwsem was held for too long,
>>>     and a writer came at this time. Then this writer will be
>>>     forced to wait and block all subsequent readers.
>>>     For example:
>>>     - be scheduled when the read lock of shrinker_rwsem is
>>>       held in do_shrink_slab()
>>>     - some shrinker are blocked for too long. Like the case
>>>       mentioned in the patchset[1].
>>>
>>> Therefore, many times in history ([2],[3],[4],[5]), some
>>> people wanted to replace shrinker_rwsem reader with SRCU,
>>> but they all gave up because SRCU was not unconditionally
>>> enabled.
>>>
>>> But now, since commit 1cd0bd06093c ("rcu: Remove CONFIG_SRCU"),
>>> the SRCU is unconditionally enabled. So it's time to use
>>> SRCU to protect readers who previously held shrinker_rwsem.
>>>
>>> [1]. 
>>> https://lore.kernel.org/lkml/20191129214541.3110-1-ptikhomirov@virtuozzo.com/
>>> [2]. https://lore.kernel.org/all/1437080113.3596.2.camel@stgolabs.net/
>>> [3]. 
>>> https://lore.kernel.org/lkml/1510609063-3327-1-git-send-email-penguin-kernel@I-love.SAKURA.ne.jp/
>>> [4]. 
>>> https://lore.kernel.org/lkml/153365347929.19074.12509495712735843805.stgit@localhost.localdomain/
>>> [5]. 
>>> https://lore.kernel.org/lkml/20210927074823.5825-1-sultan@kerneltoast.com/
>>>
>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>> ---
>>>   mm/vmscan.c | 27 +++++++++++----------------
>>>   1 file changed, 11 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>>> index 9f895ca6216c..02987a6f95d1 100644
>>> --- a/mm/vmscan.c
>>> +++ b/mm/vmscan.c
>>> @@ -202,6 +202,7 @@ static void set_task_reclaim_state(struct 
>>> task_struct *task,
>>>   LIST_HEAD(shrinker_list);
>>>   DECLARE_RWSEM(shrinker_rwsem);
>>> +DEFINE_SRCU(shrinker_srcu);
>>>   #ifdef CONFIG_MEMCG
>>>   static int shrinker_nr_max;
>>> @@ -706,7 +707,7 @@ void free_prealloced_shrinker(struct shrinker 
>>> *shrinker)
>>>   void register_shrinker_prepared(struct shrinker *shrinker)
>>>   {
>>>       down_write(&shrinker_rwsem);
>>> -    list_add_tail(&shrinker->list, &shrinker_list);
>>> +    list_add_tail_rcu(&shrinker->list, &shrinker_list);
>>>       shrinker->flags |= SHRINKER_REGISTERED;
>>>       shrinker_debugfs_add(shrinker);
>>>       up_write(&shrinker_rwsem);
>>> @@ -760,13 +761,15 @@ void unregister_shrinker(struct shrinker 
>>> *shrinker)
>>>           return;
>>>       down_write(&shrinker_rwsem);
>>> -    list_del(&shrinker->list);
>>> +    list_del_rcu(&shrinker->list);
>>>       shrinker->flags &= ~SHRINKER_REGISTERED;
>>>       if (shrinker->flags & SHRINKER_MEMCG_AWARE)
>>>           unregister_memcg_shrinker(shrinker);
>>>       debugfs_entry = shrinker_debugfs_remove(shrinker);
>>>       up_write(&shrinker_rwsem);
>>> +    synchronize_srcu(&shrinker_srcu);
>>> +
>>>       debugfs_remove_recursive(debugfs_entry);
>>>       kfree(shrinker->nr_deferred);
>>> @@ -786,6 +789,7 @@ void synchronize_shrinkers(void)
>>>   {
>>>       down_write(&shrinker_rwsem);
>>>       up_write(&shrinker_rwsem);
>>> +    synchronize_srcu(&shrinker_srcu);
>>>   }
>>>   EXPORT_SYMBOL(synchronize_shrinkers);
>>> @@ -996,6 +1000,7 @@ static unsigned long shrink_slab(gfp_t gfp_mask, 
>>> int nid,
>>>   {
>>>       unsigned long ret, freed = 0;
>>>       struct shrinker *shrinker;
>>> +    int srcu_idx;
>>>       /*
>>>        * The root memcg might be allocated even though memcg is disabled
>>> @@ -1007,10 +1012,10 @@ static unsigned long shrink_slab(gfp_t 
>>> gfp_mask, int nid,
>>>       if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
>>>           return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
>>> -    if (!down_read_trylock(&shrinker_rwsem))
>>> -        goto out;
>>> +    srcu_idx = srcu_read_lock(&shrinker_srcu);
>>> -    list_for_each_entry(shrinker, &shrinker_list, list) {
>>> +    list_for_each_entry_srcu(shrinker, &shrinker_list, list,
>>> +                 srcu_read_lock_held(&shrinker_srcu)) {
>>>           struct shrink_control sc = {
>>>               .gfp_mask = gfp_mask,
>>>               .nid = nid,
>>> @@ -1021,19 +1026,9 @@ static unsigned long shrink_slab(gfp_t 
>>> gfp_mask, int nid,
>>>           if (ret == SHRINK_EMPTY)
>>>               ret = 0;
>>>           freed += ret;
>>> -        /*
>>> -         * Bail out if someone want to register a new shrinker to
>>> -         * prevent the registration from being stalled for long periods
>>> -         * by parallel ongoing shrinking.
>>> -         */
>>> -        if (rwsem_is_contended(&shrinker_rwsem)) {
>>> -            freed = freed ? : 1;
>>> -            break;
>>> -        }
>>>       }
>>> -    up_read(&shrinker_rwsem);
>>> -out:
>>> +    srcu_read_unlock(&shrinker_srcu, srcu_idx);
>>>       cond_resched();
>>>       return freed;
>>>   }
>>> -- 
>>> 2.20.1
>>>
>>>
>>
>> Hi Qi,
>>
>> A different problem I realized after my old attempt to use SRCU was 
>> that the
>> unregister_shrinker() path became quite slow due to the heavy 
>> synchronize_srcu()
>> call. Both register_shrinker() *and* unregister_shrinker() are called 
>> frequently
>> these days, and SRCU is too unfair to the unregister path IMO.
> 
> Hi Sultan,
> 
> IIUC, for unregister_shrinker(), the wait time is hardly longer with
> SRCU than with shrinker_rwsem before.
> 
> And I just did a simple test. After using the script in cover letter to
> increase the shrink_slab hotspot, I did umount 1k times at the same
> time, and then I used bpftrace to measure the time consumption of
> unregister_shrinker() as follows:
> 
> bpftrace -e 'kprobe:unregister_shrinker { @start[tid] = nsecs; } 
> kretprobe:unregister_shrinker /@start[tid]/ { @ns[comm] = hist(nsecs - 
> @start[tid]); delete(@start[tid]); }'
> 
> @ns[umount]:
> [16K, 32K)             3 |      |
> [32K, 64K)            66 |@@@@@@@@@@      |
> [64K, 128K)           32 |@@@@@      |
> [128K, 256K)          22 |@@@      |
> [256K, 512K)          48 |@@@@@@@      |
> [512K, 1M)            19 |@@@      |
> [1M, 2M)             131 |@@@@@@@@@@@@@@@@@@@@@      |
> [2M, 4M)             313 
> |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
> [4M, 8M)             302 
> |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |
> [8M, 16M)             55 |@@@@@@@@@
> 
> I see that the highest time-consuming of unregister_shrinker() is 
> between 8ms and 16ms, which feels tolerable?

And when I use the synchronize_srcu_expedited() mentioned by Paul,
the measured time consumption has a more obvious decrease:

@ns[umount]:
[16K, 32K)           105 |@@@@@@@@@@ 
      |
[32K, 64K)           521 
|@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[64K, 128K)          119 |@@@@@@@@@@@ 
      |
[128K, 256K)          32 |@@@ 
      |
[256K, 512K)          70 |@@@@@@ 
      |
[512K, 1M)            49 |@@@@ 
      |
[1M, 2M)              34 |@@@ 
      |
[2M, 4M)              18 |@ 
      |
[4M, 8M)               4 |

> 
> Thanks,
> Qi
> 
>>
>> Although I never got around to submitting it, I made a non-SRCU 
>> solution [1]
>> that uses fine-grained locking instead, which is fair to both the 
>> register path
>> and unregister path. (The patch I've linked is a version of this 
>> adapted to an
>> older 4.14 kernel FYI, but it can be reworked for the current kernel.)
>>
>> What do you think about the fine-grained locking approach?
>>
>> Thanks,
>> Sultan
>>
>> [1] 
>> https://github.com/kerneltoast/android_kernel_google_floral/commit/012378f3173a82d2333d3ae7326691544301e76a
>>
> 

-- 
Thanks,
Qi