Sorry. Add Håkon Bugge <haakon.bugge@xxxxxxxxxx>
He told me to notice the memory leak when caches are freed.
Zhu Yanjun
On 2019/6/3 20:48, Zhu Yanjun wrote:
When KASAN is enabled, after several rds connections are
created, then "rmmod rds_rdma" is run. The following will
appear.
"
BUG rds_ib_incoming (Not tainted): Objects remaining
in rds_ib_incoming on __kmem_cache_shutdown()
Call Trace:
dump_stack+0x71/0xab
slab_err+0xad/0xd0
__kmem_cache_shutdown+0x17d/0x370
shutdown_cache+0x17/0x130
kmem_cache_destroy+0x1df/0x210
rds_ib_recv_exit+0x11/0x20 [rds_rdma]
rds_ib_exit+0x7a/0x90 [rds_rdma]
__x64_sys_delete_module+0x224/0x2c0
? __ia32_sys_delete_module+0x2c0/0x2c0
do_syscall_64+0x73/0x190
entry_SYSCALL_64_after_hwframe+0x44/0xa9
"
This is rds connection memory leak. The root cause is:
When "rmmod rds_rdma" is run, rds_ib_remove_one will call
rds_ib_dev_shutdown to drop the rds connections.
rds_ib_dev_shutdown will call rds_conn_drop to drop rds
connections as below.
"
rds_conn_path_drop(&conn->c_path[0], false);
"
In the above, destroy is set to false.
void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
{
atomic_set(&cp->cp_state, RDS_CONN_ERROR);
rcu_read_lock();
if (!destroy && rds_destroy_pending(cp->cp_conn)) {
rcu_read_unlock();
return;
}
queue_work(rds_wq, &cp->cp_down_w);
rcu_read_unlock();
}
In the above function, destroy is set to false. rds_destroy_pending
is called. This does not move rds connections to ib_nodev_conns.
So destroy is set to true to move rds connections to ib_nodev_conns.
In rds_ib_unregister_client, flush_workqueue is called to make rds_wq
finsh shutdown rds connections. The function rds_ib_destroy_nodev_conns
is called to shutdown rds connections finally.
Then rds_ib_recv_exit is called to destroy slab.
void rds_ib_recv_exit(void)
{
kmem_cache_destroy(rds_ib_incoming_slab);
kmem_cache_destroy(rds_ib_frag_slab);
}
The above slab memory leak will not occur again.
>From tests,
256 rds connections
[root@ca-dev14 ~]# time rmmod rds_rdma
real 0m16.522s
user 0m0.000s
sys 0m8.152s
512 rds connections
[root@ca-dev14 ~]# time rmmod rds_rdma
real 0m32.054s
user 0m0.000s
sys 0m15.568s
To rmmod rds_rdma with 256 rds connections, about 16 seconds are needed.
And with 512 rds connections, about 32 seconds are needed.
>From ftrace, when one rds connection is destroyed,
"
19) | rds_conn_destroy [rds]() {
19) 7.782 us | rds_conn_path_drop [rds]();
15) | rds_shutdown_worker [rds]() {
15) | rds_conn_shutdown [rds]() {
15) 1.651 us | rds_send_path_reset [rds]();
15) 7.195 us | }
15) + 11.434 us | }
19) 2.285 us | rds_cong_remove_conn [rds]();
19) * 24062.76 us | }
"
So if many rds connections will be destroyed, this function
rds_ib_destroy_nodev_conns uses most of time.
Suggested-by: Håkon Bugge <haakon.bugge@xxxxxxxxxx>
Signed-off-by: Zhu Yanjun <yanjun.zhu@xxxxxxxxxx>
---
net/rds/ib.c | 2 +-
net/rds/ib_recv.c | 3 +++
2 files changed, 4 insertions(+), 1 deletion(-)
diff --git a/net/rds/ib.c b/net/rds/ib.c
index f9baf2d..ec05d91 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -87,7 +87,7 @@ static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
spin_lock_irqsave(&rds_ibdev->spinlock, flags);
list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
- rds_conn_drop(ic->conn);
+ rds_conn_path_drop(&ic->conn->c_path[0], true);
spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
}
diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index 8946c89..3cae88c 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -168,6 +168,7 @@ void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
list_del(&inc->ii_cache_entry);
WARN_ON(!list_empty(&inc->ii_frags));
kmem_cache_free(rds_ib_incoming_slab, inc);
+ atomic_dec(&rds_ib_allocation);
}
rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
@@ -1057,6 +1058,8 @@ int rds_ib_recv_init(void)
void rds_ib_recv_exit(void)
{
+ WARN_ON(atomic_read(&rds_ib_allocation));
+
kmem_cache_destroy(rds_ib_incoming_slab);
kmem_cache_destroy(rds_ib_frag_slab);
}
