Remi Pommarel <repk@xxxxxxxxxxxx> writes:
> On EDMA capable hardware, ath9k_txq_list_has_key() can enter infinite
> loop if it is called while all txq_fifos have packets that use different
> key that the one we are looking for. Fix it by exiting the loop if all
> txq_fifos have been checked already.
>
> Because this loop is called under spin_lock_bh() (see ath_txq_lock) it
> causes the following rcu stall:
>
> rcu: INFO: rcu_sched self-detected stall on CPU
> ath10k_pci 0000:01:00.0: failed to read temperature -11
> rcu: 1-....: (5254 ticks this GP) idle=189/1/0x4000000000000002 softirq=8442983/8442984 fqs=2579
> (t=5257 jiffies g=17983297 q=334)
> Task dump for CPU 1:
> task:hostapd state:R running task stack: 0 pid: 297 ppid: 289 flags:0x0000000a
> Call trace:
> dump_backtrace+0x0/0x170
> show_stack+0x1c/0x24
> sched_show_task+0x140/0x170
> dump_cpu_task+0x48/0x54
> rcu_dump_cpu_stacks+0xf0/0x134
> rcu_sched_clock_irq+0x8d8/0x9fc
> update_process_times+0xa0/0xec
> tick_sched_timer+0x5c/0xd0
> __hrtimer_run_queues+0x154/0x320
> hrtimer_interrupt+0x120/0x2f0
> arch_timer_handler_virt+0x38/0x44
> handle_percpu_devid_irq+0x9c/0x1e0
> handle_domain_irq+0x64/0x90
> gic_handle_irq+0x78/0xb0
> call_on_irq_stack+0x28/0x38
> do_interrupt_handler+0x54/0x5c
> el1_interrupt+0x2c/0x4c
> el1h_64_irq_handler+0x14/0x1c
> el1h_64_irq+0x74/0x78
> ath9k_txq_has_key+0x1bc/0x250 [ath9k]
> ath9k_set_key+0x1cc/0x3dc [ath9k]
> drv_set_key+0x78/0x170
> ieee80211_key_replace+0x564/0x6cc
> ieee80211_key_link+0x174/0x220
> ieee80211_add_key+0x11c/0x300
> nl80211_new_key+0x12c/0x330
> genl_family_rcv_msg_doit+0xbc/0x11c
> genl_rcv_msg+0xd8/0x1c4
> netlink_rcv_skb+0x40/0x100
> genl_rcv+0x3c/0x50
> netlink_unicast+0x1ec/0x2c0
> netlink_sendmsg+0x198/0x3c0
> ____sys_sendmsg+0x210/0x250
> ___sys_sendmsg+0x78/0xc4
> __sys_sendmsg+0x4c/0x90
> __arm64_sys_sendmsg+0x28/0x30
> invoke_syscall.constprop.0+0x60/0x100
> do_el0_svc+0x48/0xd0
> el0_svc+0x14/0x50
> el0t_64_sync_handler+0xa8/0xb0
> el0t_64_sync+0x158/0x15c
>
> This rcu stall is hard to reproduce as is, but changing ATH_TXFIFO_DEPTH
> from 8 to 2 makes it reasonably easy to reproduce.
>
> Fixes: ca2848022c12 ("ath9k: Postpone key cache entry deletion for TXQ frames reference it")
> Signed-off-by: Remi Pommarel <repk@xxxxxxxxxxxx>
Thank you for the fix!
Acked-by: Toke Høiland-Jørgensen <toke@xxxxxxx>
