On 11/17/2016 11:25 AM, Ying Xue wrote:
> On 11/16/2016 10:14 PM, Parthasarathy Bhuvaragan wrote:
>> Commit 333f796235a527 ("tipc: fix a race condition leading to
>> subscriber refcnt bug") reveals a soft lockup while acquiring
>> nametbl_lock.
>
> Please explain why commit 333f796235a527 introduces the soft lockup.
[partha]
Before commit 333f796235a527, the tipc_conn_shutdown() was always called
from tipc_close_conn() i.e in the context of tipc_topsrv_stop().
In that context, we are allowed to grab the nametbl_lock.
In commit 333f796235a527, i moved the tipc_conn_release (renamed from
tipc_conn_shutdown) to the connection refcount cleanup. This allows
either tipc_nametbl_withdraw() or tipc_topsrv_stop() to perform
tipc_sock_release().
The fact that tipc_exit_net() first calls tipc_topsrv_stop() and then
tipc_net_stop()->tipc_nametble_withdraw() increases the chances for the
later to perform the connection cleanup.
>
>>
>> The soft lock is the call chain of tipc_nametbl_withdraw(), when it
>> performs the tipc_conn_kref_release() as it tries to grab nametbl_lock
>> again while holding it already.
>> tipc_nametbl_withdraw
>> spin_lock_bh(&tn->nametbl_lock);
>> tipc_nametbl_remove_publ
>> tipc_nameseq_remove_publ
>> tipc_subscrp_report_overlap
>> tipc_subscrp_send_event
>> tipc_conn_sendmsg
>> Here, the (test_bit(CF_CONNECTED, &con->flags)) fails, leading to the
>> else case where do do a conn_put() and that triggers the cleanup as
>> refcount reached 0. Leading the call chain below:
>> tipc_conn_kref_release
>> tipc_sock_release
>> tipc_conn_release
>> tipc_subscrb_delete
>> tipc_subscrp_delete
>> tipc_nametbl_unsubscribe
>> spin_lock_bh(&tn->nametbl_lock); << Soft Lockup >>
>>
>> The above lockup is caused due to two reasons:
>> 1. During module exit tipc_exit_net() first calls tipc_topsrv_stop()
>> followed by tipc_nametbl_withdraw() in scope of tipc_net_stop().
>> 2. tipc_server_stop() grabs and releases the idr_lock twice for every
>> connection as it needs to execute functions that might sleep.
>> The first lock is to find the connection and the second sets the
>> connection flag, removes it from list and decrements the refcount.
>> There is a possibility for another thread to grab the connection
>> in between the two, thus owning the connection and triggering the
>> cleanup while decrementing the refcount later.
>>
>> In this commit, we perform:
>> - call tipc_nametbl_withdraw() before tipc_topsrv_stop()
>> - In tipc_server_stop():
>> we remove the connection id's of all connections from connection
>> list in the scope of the idr_lock.
>> copy s->idr_in_use to stack, as this is being updated in the loop.
>> Thus no other thread will find a connection which has unset
>> CF_CONNECTED in its flags.
>>
>> Fixes: 333f796235a52727 ("tipc: fix a race condition leading to
>> subscriber refcnt bug")
>> Signed-off-by: Parthasarathy Bhuvaragan
>> <[email protected]>
>> ---
>> net/tipc/core.c | 4 ++++
>> net/tipc/net.c | 2 --
>> net/tipc/server.c | 19 +++++++++++++++----
>> 3 files changed, 19 insertions(+), 6 deletions(-)
>>
>> diff --git a/net/tipc/core.c b/net/tipc/core.c
>> index 236b043a4156..3ea1452e2f06 100644
>> --- a/net/tipc/core.c
>> +++ b/net/tipc/core.c
>> @@ -93,6 +93,10 @@ static int __net_init tipc_init_net(struct net *net)
>>
>> static void __net_exit tipc_exit_net(struct net *net)
>> {
>> + struct tipc_net *tn = net_generic(net, tipc_net_id);
>> +
>> + tipc_nametbl_withdraw(net, TIPC_CFG_SRV, tn->own_addr, 0,
>> + tn->own_addr);
>
> I could not understand why we move tipc_nametbl_withdraw to here.
[partha]
Otherwise we call tipc_topsrv_stop() before tipc_nametbl_withdraw(),
which might lead to:
- soft lockup
- its too late to actually notify the subscribers, as the topology
server might already have started shutting down.
I couldn't just change the existing order as tipc_net_stop() does a lot
more than just tipc_nametbl_withdraw().
>
>> tipc_topsrv_stop(net);
>> tipc_net_stop(net);
>> tipc_bcast_stop(net);
>> diff --git a/net/tipc/net.c b/net/tipc/net.c
>> index 28bf4feeb81c..92696cc6e763 100644
>> --- a/net/tipc/net.c
>> +++ b/net/tipc/net.c
>> @@ -130,8 +130,6 @@ void tipc_net_stop(struct net *net)
>> if (!tn->own_addr)
>> return;
>>
>> - tipc_nametbl_withdraw(net, TIPC_CFG_SRV, tn->own_addr, 0,
>> - tn->own_addr);
>> rtnl_lock();
>> tipc_bearer_stop(net);
>> tipc_node_stop(net);
>> diff --git a/net/tipc/server.c b/net/tipc/server.c
>> index 215849ce453d..22a4caf66065 100644
>> --- a/net/tipc/server.c
>> +++ b/net/tipc/server.c
>> @@ -622,22 +622,33 @@ int tipc_server_start(struct tipc_server *s)
>>
>> void tipc_server_stop(struct tipc_server *s)
>> {
>> + int max_in_use = s->idr_in_use;
>> struct tipc_conn *con;
>> int total = 0;
>> int id;
>>
>> spin_lock_bh(&s->idr_lock);
>> - for (id = 0; total < s->idr_in_use; id++) {
>> + for (id = 0; total < max_in_use; id++) {
>> con = idr_find(&s->conn_idr, id);
>> if (con) {
>> total++;
>> - spin_unlock_bh(&s->idr_lock);
>> - tipc_close_conn(con);
>> - spin_lock_bh(&s->idr_lock);
>
> In fact the change is unnecessarily so complex at all, instead we can do
> this like:
>
> spin_lock_bh(&s->idr_lock);
> for (id = 0; total < s->idr_in_use; id++) {
> con = idr_find(&s->conn_idr, id);
> if (con) {
> total++;
> if (test_and_clear_bit(CF_CONNECTED, &con->flags))
> idr_remove(&s->conn_idr, con->conid);
> s->idr_in_use--;
> spin_unlock_bh(&s->idr_lock);
> kernel_sock_shutdown(con->sock, SHUT_RDWR);
> conn_put(con);
> spin_lock_bh(&s->idr_lock);
> }
> }
> }
> spin_unlock_bh(&s->idr_lock);
[partha]
Thanks for the good suggestion. In addition, i can remove the variable
total and use s->idr_in_use as exit criteria.
like: for (id = 0; s->idr_in_use; id++)
>
> Regards,
> Ying
>
>> + if (test_and_clear_bit(CF_CONNECTED, &con->flags)) {
>> + idr_remove(&s->conn_idr, con->conid);
>> + s->idr_in_use--;
>> + }
>> }
>> }
>> spin_unlock_bh(&s->idr_lock);
>>
>> + for (id = 0; total < max_in_use; id++) {
>> + con = idr_find(&s->conn_idr, id);
>> + if (con) {
>> + total++;
>> + kernel_sock_shutdown(con->sock, SHUT_RDWR);
>> + conn_put(con);
>> + }
>> + }
>> +
>> tipc_work_stop(s);
>> kmem_cache_destroy(s->rcvbuf_cache);
>> idr_destroy(&s->conn_idr);
>>
>
------------------------------------------------------------------------------
_______________________________________________
tipc-discussion mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/tipc-discussion
