In tests with two Lima boards from 8devices (QCA4531 based) on OpenWrt
19.07 we could force a silent restart of a device with no serial
output when we were sending a high amount of UDP traffic (iperf3 at 80
MBit/s in both directions from external hosts, saturating the wifi and
causing a load of about 4.5 to 6) and were then triggering an
ath9k_queue_reset().
Further debugging showed that the restart was caused by the ath79
watchdog. With disabled watchdog we could observe that the device was
constantly going into ath_isr() interrupt handler and was returning
early after the ATH_OP_HW_RESET flag test, without clearing any
interrupts. Even though ath9k_queue_reset() calls
ath9k_hw_kill_interrupts().
With JTAG we could observe the following race condition:
1) ath9k_queue_reset()
...
-> ath9k_hw_kill_interrupts()
-> set_bit(ATH_OP_HW_RESET, &common->op_flags);
...
<- returns
2) ath9k_tasklet()
...
-> ath9k_hw_resume_interrupts()
...
<- returns
3) loops around:
...
handle_int()
-> ath_isr()
...
-> if (test_bit(ATH_OP_HW_RESET,
&common->op_flags))
return IRQ_HANDLED;
x) ath_reset_internal():
=> never reached <=
And in ath_isr() we would typically see the following interrupts /
interrupt causes:
* status: 0x00111030 or 0x00110030
* async_cause: 2 (AR_INTR_MAC_IPQ)
* sync_cause: 0
So the ath9k_tasklet() reenables the ath9k interrupts
through ath9k_hw_resume_interrupts() which ath9k_queue_reset() had just
disabled. And ath_isr() then keeps firing because it returns IRQ_HANDLED
without actually clearing the interrupt.
To fix this IRQ storm also clear/disable the interrupts again when we
are in reset state.
Cc: Sven Eckelmann <sven@narfation.org>
Cc: Simon Wunderlich <sw@simonwunderlich.de>
Cc: Linus Lüssing <linus.luessing@c0d3.blue>
Fixes:
872b5d814f99 ("ath9k: do not access hardware on IRQs during reset")
Signed-off-by: Linus Lüssing <ll@simonwunderlich.de>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Link: https://lore.kernel.org/r/20210914192515.9273-3-linus.luessing@c0d3.blue