Subject: [PATCH 0/3] Make count list_lru_one::nr_items lockless
From: Kirill Tkhai <ktkhai@virtuozzo.com>
To: apolyakov@beget.ru, linux-kernel@vger.kernel.org, linux-mm@kvack.org,
        ktkhai@virtuozzo.com, vdavydov.dev@gmail.com, aryabinin@virtuozzo.com,
        akpm@linux-foundation.org
Date: Tue, 22 Aug 2017 15:29:08 +0300
Message-ID: <150340381428.3845.6099251634440472539.stgit@localhost.localdomain>
User-Agent: StGit/0.17.1-dirty
MIME-Version: 1.0
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
SpamDiagnosticOutput: 1:99
SpamDiagnosticMetadata: NSPM
X-MS-Exchange-CrossTenant-OriginalArrivalTime: 22 Aug 2017 12:29:03.2138 (UTC)
X-MS-Exchange-CrossTenant-FromEntityHeader: Hosted
X-MS-Exchange-Transport-CrossTenantHeadersStamped: DB6PR0801MB1336
Sender: linux-kernel-owner@vger.kernel.org
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org
Content-Length: 1930
Lines: 41

This series aims to improve scalability of list_lru shrinking
and to make list_lru_count_one() working more effective.

On RHEL7 3.10 kernel I observe high system time usage and time
spent in super_cache_count() during slab shrinking:

0,94%  mysqld         [kernel.vmlinux]  [k] _raw_spin_lock                [k] _raw_spin_lock
0,57%  mysqld         [kernel.vmlinux]  [k] shrink_slab                   [k] shrink_slab
0,51%  mysqld         [kernel.vmlinux]  [k] super_cache_count             [k] super_cache_count
0,32%  mysqld         [kernel.vmlinux]  [k] __list_lru_count_one.isra.2   [k] _raw_spin_lock
0,32%  mysqld         [kernel.vmlinux]  [k] list_lru_count_one            [k] __list_lru_count_one.isra.2

(percentage of all node time; collected via $perf record --call-graph fp -j k -a).
It's an example, how the processes traces look like. And many processes spend time
in the above.

There is a node with many containers (more, than 200), and (as it's usually happen)
containers have no free memory (cache is actively used). Since shrink_slab() iterates
all superblocks, and it happens frequently, the shrink scales badly, and node spends
in sys more than 90% of time.

The patchset makes list_lru_count_one() lockless via RCU technics. Patch [1/3]
adds a new rcu field to struct list_lru_memcg and makes functions account its
size during allocations. Patch [2/3] makes list_lru_node::memcg_lrus RCU-protected
and RCU-accessible. Patch [3/3] removes the lock and adds rcu read protection
into __list_lru_count_one().

---

Kirill Tkhai (3):
      mm: Add rcu field to struct list_lru_memcg
      mm: Make list_lru_node::memcg_lrus RCU protected
      mm: Count list_lru_one::nr_items lockless


 include/linux/list_lru.h |    3 +-
 mm/list_lru.c            |   77 ++++++++++++++++++++++++++++++----------------
 2 files changed, 53 insertions(+), 27 deletions(-)

--
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>