In the Linux kernel, the following vulnerability has been resolved:
smc91x: fix broken irq-context in PREEMPT_RT
When smc91x.c is built with PREEMPT_RT, the following splat occurs in FVP_RevC:
[ 13.055000] smc91x LNRO0003:00 eth0: link up, 10Mbps, half-duplex, lpa 0x0000 [ 13.062137] BUG: workqueue leaked atomic, lock or RCU: kworker/2:1[106] [ 13.062137] preempt=0x00000000 lock=0->0 RCU=0->1 workfn=mld_ifc_work [ 13.062266] C replaying previous printk message [ 13.062266] CPU: 2 UID: 0 PID: 106 Comm: kworker/2:1 Not tainted 6.18.0-dirty #179 PREEMPT_{RT,(full)} [ 13.062353] Hardware name: , BIOS [ 13.062382] Workqueue: mld mld_ifc_work [ 13.062469] Call trace: [ 13.062494] show_stack+0x24/0x40 (C) [ 13.062602] __dump_stack+0x28/0x48 [ 13.062710] dump_stack_lvl+0x7c/0xb0 [ 13.062818] dump_stack+0x18/0x34 [ 13.062926] process_scheduled_works+0x294/0x450 [ 13.063043] worker_thread+0x260/0x3d8 [ 13.063124] kthread+0x1c4/0x228 [ 13.063235] ret_from_fork+0x10/0x20
This happens because smc_special_trylock() disables IRQs even on PREEMPT_RT, but smc_special_unlock() does not restore IRQs on PREEMPT_RT. The reason is that smc_special_unlock() calls spin_unlock_irqrestore(), and rcu_read_unlock_bh() in __dev_queue_xmit() cannot invoke rcu_read_unlock() through __local_bh_enable_ip() when current->softirq_disable_cnt becomes zero.
To address this issue, replace smc_special_trylock() with spin_trylock_irqsave().