target/iscsi: don't corrupt bh_count in iscsit_stop_time2retain_timer()
authorJörn Engel <joern@logfs.org>
Thu, 30 May 2013 20:36:51 +0000 (16:36 -0400)
committerNicholas Bellinger <nab@linux-iscsi.org>
Fri, 14 Jun 2013 07:55:49 +0000 (00:55 -0700)
commit574780fd5e6ec52bd43e0bdb777a19e4c4c6aa9c
tree35fcb7d07dd7595c56c3cf3fa0d16c02456a599b
parent317ddd256b9c24b0d78fa8018f80f1e495481a10
target/iscsi: don't corrupt bh_count in iscsit_stop_time2retain_timer()

Here is a fun one.  Bug seems to have been introduced by commit 140854cb,
almost two years ago.  I have no idea why we only started seeing it now,
but we did.

Rough callgraph:
core_tpg_set_initiator_node_queue_depth()
`-> spin_lock_irqsave(&tpg->session_lock, flags);
`-> lio_tpg_shutdown_session()
    `-> iscsit_stop_time2retain_timer()
        `-> spin_unlock_bh(&se_tpg->session_lock);
        `-> spin_lock_bh(&se_tpg->session_lock);
`-> spin_unlock_irqrestore(&tpg->session_lock, flags);

core_tpg_set_initiator_node_queue_depth() used to call spin_lock_bh(),
but 140854cb changed that to spin_lock_irqsave().  However,
lio_tpg_shutdown_session() still claims to be called with spin_lock_bh()
held, as does iscsit_stop_time2retain_timer():
 *      Called with spin_lock_bh(&struct se_portal_group->session_lock) held

Stale documentation is mostly annoying, but in this case the dropping
the lock with the _bh variant is plain wrong.  It is also wrong to drop
locks two functions below the lock-holder, but I will ignore that bit
for now.

After some more locking and unlocking we eventually hit this backtrace:
------------[ cut here ]------------
WARNING: at kernel/softirq.c:159 local_bh_enable_ip+0xe8/0x100()
Pid: 24645, comm: lio_helper.py Tainted: G           O 3.6.11+
Call Trace:
 [<ffffffff8103e5ff>] warn_slowpath_common+0x7f/0xc0
 [<ffffffffa040ae37>] ? iscsit_inc_conn_usage_count+0x37/0x50 [iscsi_target_mod]
 [<ffffffff8103e65a>] warn_slowpath_null+0x1a/0x20
 [<ffffffff810472f8>] local_bh_enable_ip+0xe8/0x100
 [<ffffffff815b8365>] _raw_spin_unlock_bh+0x15/0x20
 [<ffffffffa040ae37>] iscsit_inc_conn_usage_count+0x37/0x50 [iscsi_target_mod]
 [<ffffffffa041149a>] iscsit_stop_session+0xfa/0x1c0 [iscsi_target_mod]
 [<ffffffffa0417fab>] lio_tpg_shutdown_session+0x7b/0x90 [iscsi_target_mod]
 [<ffffffffa033ede4>] core_tpg_set_initiator_node_queue_depth+0xe4/0x290 [target_core_mod]
 [<ffffffffa0409032>] iscsit_tpg_set_initiator_node_queue_depth+0x12/0x20 [iscsi_target_mod]
 [<ffffffffa0415c29>] lio_target_nacl_store_cmdsn_depth+0xa9/0x180 [iscsi_target_mod]
 [<ffffffffa0331b49>] target_fabric_nacl_base_attr_store+0x39/0x40 [target_core_mod]
 [<ffffffff811b857d>] configfs_write_file+0xbd/0x120
 [<ffffffff81148f36>] vfs_write+0xc6/0x180
 [<ffffffff81149251>] sys_write+0x51/0x90
 [<ffffffff815c0969>] system_call_fastpath+0x16/0x1b
---[ end trace 3747632b9b164652 ]---

As a pure band-aid, this patch drops the _bh.

Signed-off-by: Joern Engel <joern@logfs.org>
Cc: stable <stable@vger.kernel.org>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
drivers/target/iscsi/iscsi_target_erl0.c