From: Sagi Grimberg sagi@grimberg.me
[ Upstream commit b5332a9f3f3d884a1b646ce155e664cc558c1722 ]
We are not changing anything in the TCP connection state so we should not take a write_lock but rather a read lock.
This caused a deadlock when running nvmet-tcp and nvme-tcp on the same system, where state_change callbacks on the host and on the controller side have causal relationship and made lockdep report on this with blktests:
================================ WARNING: inconsistent lock state 5.12.0-rc3 #1 Tainted: G I -------------------------------- inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-R} usage. nvme/1324 [HC0[0]:SC0[0]:HE1:SE1] takes: ffff888363151000 (clock-AF_INET){++-?}-{2:2}, at: nvme_tcp_state_change+0x21/0x150 [nvme_tcp] {IN-SOFTIRQ-W} state was registered at: __lock_acquire+0x79b/0x18d0 lock_acquire+0x1ca/0x480 _raw_write_lock_bh+0x39/0x80 nvmet_tcp_state_change+0x21/0x170 [nvmet_tcp] tcp_fin+0x2a8/0x780 tcp_data_queue+0xf94/0x1f20 tcp_rcv_established+0x6ba/0x1f00 tcp_v4_do_rcv+0x502/0x760 tcp_v4_rcv+0x257e/0x3430 ip_protocol_deliver_rcu+0x69/0x6a0 ip_local_deliver_finish+0x1e2/0x2f0 ip_local_deliver+0x1a2/0x420 ip_rcv+0x4fb/0x6b0 __netif_receive_skb_one_core+0x162/0x1b0 process_backlog+0x1ff/0x770 __napi_poll.constprop.0+0xa9/0x5c0 net_rx_action+0x7b3/0xb30 __do_softirq+0x1f0/0x940 do_softirq+0xa1/0xd0 __local_bh_enable_ip+0xd8/0x100 ip_finish_output2+0x6b7/0x18a0 __ip_queue_xmit+0x706/0x1aa0 __tcp_transmit_skb+0x2068/0x2e20 tcp_write_xmit+0xc9e/0x2bb0 __tcp_push_pending_frames+0x92/0x310 inet_shutdown+0x158/0x300 __nvme_tcp_stop_queue+0x36/0x270 [nvme_tcp] nvme_tcp_stop_queue+0x87/0xb0 [nvme_tcp] nvme_tcp_teardown_admin_queue+0x69/0xe0 [nvme_tcp] nvme_do_delete_ctrl+0x100/0x10c [nvme_core] nvme_sysfs_delete.cold+0x8/0xd [nvme_core] kernfs_fop_write_iter+0x2c7/0x460 new_sync_write+0x36c/0x610 vfs_write+0x5c0/0x870 ksys_write+0xf9/0x1d0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xae irq event stamp: 10687 hardirqs last enabled at (10687): [<ffffffff9ec376bd>] _raw_spin_unlock_irqrestore+0x2d/0x40 hardirqs last disabled at (10686): [<ffffffff9ec374d8>] _raw_spin_lock_irqsave+0x68/0x90 softirqs last enabled at (10684): [<ffffffff9f000608>] __do_softirq+0x608/0x940 softirqs last disabled at (10649): [<ffffffff9cdedd31>] do_softirq+0xa1/0xd0
other info that might help us debug this: Possible unsafe locking scenario:
CPU0 ---- lock(clock-AF_INET); <Interrupt> lock(clock-AF_INET);
*** DEADLOCK ***
5 locks held by nvme/1324: #0: ffff8884a01fe470 (sb_writers#4){.+.+}-{0:0}, at: ksys_write+0xf9/0x1d0 #1: ffff8886e435c090 (&of->mutex){+.+.}-{3:3}, at: kernfs_fop_write_iter+0x216/0x460 #2: ffff888104d90c38 (kn->active#255){++++}-{0:0}, at: kernfs_remove_self+0x22d/0x330 #3: ffff8884634538d0 (&queue->queue_lock){+.+.}-{3:3}, at: nvme_tcp_stop_queue+0x52/0xb0 [nvme_tcp] #4: ffff888363150d30 (sk_lock-AF_INET){+.+.}-{0:0}, at: inet_shutdown+0x59/0x300
stack backtrace: CPU: 26 PID: 1324 Comm: nvme Tainted: G I 5.12.0-rc3 #1 Hardware name: Dell Inc. PowerEdge R640/06NR82, BIOS 2.10.0 11/12/2020 Call Trace: dump_stack+0x93/0xc2 mark_lock_irq.cold+0x2c/0xb3 ? verify_lock_unused+0x390/0x390 ? stack_trace_consume_entry+0x160/0x160 ? lock_downgrade+0x100/0x100 ? save_trace+0x88/0x5e0 ? _raw_spin_unlock_irqrestore+0x2d/0x40 mark_lock+0x530/0x1470 ? mark_lock_irq+0x1d10/0x1d10 ? enqueue_timer+0x660/0x660 mark_usage+0x215/0x2a0 __lock_acquire+0x79b/0x18d0 ? tcp_schedule_loss_probe.part.0+0x38c/0x520 lock_acquire+0x1ca/0x480 ? nvme_tcp_state_change+0x21/0x150 [nvme_tcp] ? rcu_read_unlock+0x40/0x40 ? tcp_mtu_probe+0x1ae0/0x1ae0 ? kmalloc_reserve+0xa0/0xa0 ? sysfs_file_ops+0x170/0x170 _raw_read_lock+0x3d/0xa0 ? nvme_tcp_state_change+0x21/0x150 [nvme_tcp] nvme_tcp_state_change+0x21/0x150 [nvme_tcp] ? sysfs_file_ops+0x170/0x170 inet_shutdown+0x189/0x300 __nvme_tcp_stop_queue+0x36/0x270 [nvme_tcp] nvme_tcp_stop_queue+0x87/0xb0 [nvme_tcp] nvme_tcp_teardown_admin_queue+0x69/0xe0 [nvme_tcp] nvme_do_delete_ctrl+0x100/0x10c [nvme_core] nvme_sysfs_delete.cold+0x8/0xd [nvme_core] kernfs_fop_write_iter+0x2c7/0x460 new_sync_write+0x36c/0x610 ? new_sync_read+0x600/0x600 ? lock_acquire+0x1ca/0x480 ? rcu_read_unlock+0x40/0x40 ? lock_is_held_type+0x9a/0x110 vfs_write+0x5c0/0x870 ksys_write+0xf9/0x1d0 ? __ia32_sys_read+0xa0/0xa0 ? lockdep_hardirqs_on_prepare.part.0+0x198/0x340 ? syscall_enter_from_user_mode+0x27/0x70 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x44/0xae
Fixes: 872d26a391da ("nvmet-tcp: add NVMe over TCP target driver") Reported-by: Yi Zhang yi.zhang@redhat.com Signed-off-by: Sagi Grimberg sagi@grimberg.me Signed-off-by: Christoph Hellwig hch@lst.de Signed-off-by: Sasha Levin sashal@kernel.org --- drivers/nvme/target/tcp.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/drivers/nvme/target/tcp.c b/drivers/nvme/target/tcp.c index d658c6e8263a..218fd766dc74 100644 --- a/drivers/nvme/target/tcp.c +++ b/drivers/nvme/target/tcp.c @@ -1434,7 +1434,7 @@ static void nvmet_tcp_state_change(struct sock *sk) { struct nvmet_tcp_queue *queue;
- write_lock_bh(&sk->sk_callback_lock); + read_lock_bh(&sk->sk_callback_lock); queue = sk->sk_user_data; if (!queue) goto done; @@ -1452,7 +1452,7 @@ static void nvmet_tcp_state_change(struct sock *sk) queue->idx, sk->sk_state); } done: - write_unlock_bh(&sk->sk_callback_lock); + read_unlock_bh(&sk->sk_callback_lock); }
static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)