Fabio, On Sat, Sep 04 2021 at 18:10, Fabio Estevam wrote: > On Fri, Sep 3, 2021 at 5:07 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > I did as suggested and here is trace.txt: > https://pastebin.com/VUfLRJ8a Lacks a stack trace, but yes this one is the culprit: kworker/u4:2-70 [000] d..1 87.940929: softirq_raise: vec=3 [action=NET_RX] It has only interrupts and preemption disabled and it's in task context. So if there is no interrupt raised and no local_bh_disable / enable() pair invoked before the CPU goes idle nothing will handle the softirq and the raised bit stays pending which makes the NOHZ idle code complain. > Also, while investigating this problem I saw a commit that fixed a > similar issue: > e63052a5dd3c ("mlx5e: add add missing BH locking around napi_schdule()"). > > I then tried the same approach on the ath10k sdio driver: > > diff --git a/drivers/net/wireless/ath/ath10k/sdio.c > b/drivers/net/wireless/ath/ath10k/sdio.c > index b746052737e0..eb705214f3f0 100644 > --- a/drivers/net/wireless/ath/ath10k/sdio.c > +++ b/drivers/net/wireless/ath/ath10k/sdio.c > @@ -1363,8 +1363,11 @@ static void > ath10k_rx_indication_async_work(struct work_struct *work) > ep->ep_ops.ep_rx_complete(ar, skb); > } > > - if (test_bit(ATH10K_FLAG_CORE_REGISTERED, &ar->dev_flags)) > + if (test_bit(ATH10K_FLAG_CORE_REGISTERED, &ar->dev_flags)) { > + local_bh_disable(); > napi_schedule(&ar->napi); > + local_bh_enable(); > + } > } > > and no longer get the "NOHZ tick-stop error: Non-RCU local softirq work is > pending, handler #08!!!" error messages after launching hostapd. > > Is this a proper fix? Yes. This is correct. See above. Thanks, tglx
