Please add 2dce224f469f ("netns: protect netns ID lookups with RCU") to LTS

[Haakon Bugge]

Hi Greg & Sasha,


tl;dr: Please add 2dce224f469f ("netns: protect netns ID lookups with RCU") to the stable releases from v5.4 and older. It fixes a spin_unlock_bh() in peernet2id() called with IRQs off. I think this neat side-effect of commit 2dce224f469f was quite un-intentional, hence no Fixes: tag or CC: stable.

The details:

From bugzilla.redhat.com/show_bug.cgi?id=1384179 (an ancient 4.9.0-0.rc0 kernel):

 dump_stack+0x86/0xc3
__warn+0xcb/0xf0
warn_slowpath_null+0x1d/0x20
__local_bh_enable_ip+0x9d/0xc0
_raw_spin_unlock_bh+0x35/0x40
peernet2id+0x54/0x80
netlink_broadcast_filtered+0x220/0x3c0
netlink_broadcast+0x1d/0x20
audit_log+0x6a/0x90
security_set_bools+0xee/0x200
[]

Note, security_set_bools() calls write_lock_irq(). peernet2id() calls spin_unlock_bh().


From an internal (UEK) stack trace based on the v4.14.35 kernel (LTS 4.14.231):

queued_spin_lock_slowpath+0xb/0xf
_raw_spin_lock_irqsave+0x46/0x48
send_mad+0x3d2/0x590 [ib_core]
ib_sa_path_rec_get+0x223/0x4d0 [ib_core]
path_rec_start+0xa3/0x140 [ib_ipoib]
ipoib_start_xmit+0x2b0/0x6a0 [ib_ipoib]
dev_hard_start_xmit+0xb2/0x237
sch_direct_xmit+0x114/0x1bf
__dev_queue_xmit+0x592/0x818
dev_queue_xmit+0x10/0x12
arp_xmit+0x38/0xa6
arp_send_dst.part.16+0x61/0x84
arp_process+0x825/0x889
arp_rcv+0x140/0x1c9
__netif_receive_skb_core+0x401/0xb39
__netif_receive_skb+0x18/0x59
netif_receive_skb_internal+0x45/0x119
napi_gro_receive+0xd8/0xf6
ipoib_ib_handle_rx_wc+0x1ca/0x520 [ib_ipoib]
ipoib_poll+0xcd/0x150 [ib_ipoib]
net_rx_action+0x289/0x3f4
__do_softirq+0xe1/0x2b5
do_softirq_own_stack+0x2a/0x35
</IRQ>
do_softirq+0x4d/0x6a
__local_bh_enable_ip+0x57/0x59
_raw_spin_unlock_bh+0x23/0x25
peernet2id+0x51/0x73
netlink_broadcast_filtered+0x223/0x41b
netlink_broadcast+0x1d/0x1f
rdma_nl_multicast+0x22/0x30 [ib_core]
send_mad+0x3e5/0x590 [ib_core]
ib_sa_path_rec_get+0x223/0x4d0 [ib_core]
rdma_resolve_route+0x287/0x810 [rdma_cm]
rds_rdma_cm_event_handler_cmn+0x311/0x7d0 [rds_rdma]
rds_rdma_cm_event_handler_worker+0x22/0x30 [rds_rdma]
process_one_work+0x169/0x3a6
worker_thread+0x4d/0x3e5
kthread+0x105/0x138
ret_from_fork+0x24/0x49

Here, pay attention to ib_nl_make_request() which calls spin_lock_irqsave() on a global lock just before calling rdma_nl_multicast(). Thereafter, peernet2id() enables SoftIRQs, and ipoib starts and calls the same path and end up trying to acquire the same global lock again.

I have tried to repro this with no luck. But, stack traces seldom lies ;-)


Thxs, Håkon