tipc: fix a race condition leading to subscriber refcnt bug
authorParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Tue, 12 Apr 2016 11:05:21 +0000 (13:05 +0200)
committerDavid S. Miller <davem@davemloft.net>
Thu, 14 Apr 2016 20:46:46 +0000 (16:46 -0400)
commit333f796235a52727db7e0a13888045f3aa3d5335
tree7e1230a9d803af9b5cef3492f5482ba01cc382ff
parentedd93cd72741b80633969c247a2f86f177242cfb
tipc: fix a race condition leading to subscriber refcnt bug

Until now, the requests sent to topology server are queued
to a workqueue by the generic server framework.
These messages are processed by worker threads and trigger the
registered callbacks.
To reduce latency on uniprocessor systems, explicit rescheduling
is performed using cond_resched() after MAX_RECV_MSG_COUNT(25)
messages.

This implementation on SMP systems leads to an subscriber refcnt
error as described below:
When a worker thread yields by calling cond_resched() in a SMP
system, a new worker is created on another CPU to process the
pending workitem. Sometimes the sleeping thread wakes up before
the new thread finishes execution.
This breaks the assumption on ordering and being single threaded.
The fault is more frequent when MAX_RECV_MSG_COUNT is lowered.

If the first thread was processing subscription create and the
second thread processing close(), the close request will free
the subscriber and the create request oops as follows:

[31.224137] WARNING: CPU: 2 PID: 266 at include/linux/kref.h:46 tipc_subscrb_rcv_cb+0x317/0x380         [tipc]
[31.228143] CPU: 2 PID: 266 Comm: kworker/u8:1 Not tainted 4.5.0+ #97
[31.228377] Workqueue: tipc_rcv tipc_recv_work [tipc]
[...]
[31.228377] Call Trace:
[31.228377]  [<ffffffff812fbb6b>] dump_stack+0x4d/0x72
[31.228377]  [<ffffffff8105a311>] __warn+0xd1/0xf0
[31.228377]  [<ffffffff8105a3fd>] warn_slowpath_null+0x1d/0x20
[31.228377]  [<ffffffffa0098067>] tipc_subscrb_rcv_cb+0x317/0x380 [tipc]
[31.228377]  [<ffffffffa00a4984>] tipc_receive_from_sock+0xd4/0x130 [tipc]
[31.228377]  [<ffffffffa00a439b>] tipc_recv_work+0x2b/0x50 [tipc]
[31.228377]  [<ffffffff81071925>] process_one_work+0x145/0x3d0
[31.246554] ---[ end trace c3882c9baa05a4fd ]---
[31.248327] BUG: spinlock bad magic on CPU#2, kworker/u8:1/266
[31.249119] BUG: unable to handle kernel NULL pointer dereference at 0000000000000428
[31.249323] IP: [<ffffffff81099d0c>] spin_dump+0x5c/0xe0
[31.249323] PGD 0
[31.249323] Oops: 0000 [#1] SMP

In this commit, we
- rename tipc_conn_shutdown() to tipc_conn_release().
- move connection release callback execution from tipc_close_conn()
  to a new function tipc_sock_release(), which is executed before
  we free the connection.
Thus we release the subscriber during connection release procedure
rather than connection shutdown procedure.

Signed-off-by: Parthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/tipc/server.c
net/tipc/server.h
net/tipc/subscr.c