Re: [PATCH net 2/2] net: gro: Let the timeout timer expire in softirq context with `threadirqs'

[Eric Dumazet <edumazet@google.com>]

On Tue, Nov 26, 2019 at 2:20 PM Sebastian Andrzej Siewior
<bigeasy@linutronix.de> wrote:
>
> The timer callback (napi_watchdog()) invokes __napi_schedule_irqoff()
> with disabled interrupts. With the `threadirqs' commandline option all
> interrupt handler are threaded and using __napi_schedule_irqoff() is not
> an issue because everyone is using it in threaded context which is
> synchronised with local_bh_disable().
> The napi_watchdog() timer is still expiring in hardirq context and may
> interrupt a threaded handler which is in the middle of
> __napi_schedule_irqoff() leading to list corruption.

Sorry, I do not understand this changelog.

Where/how do you get list corruption  exactly ?

 __napi_schedule_irqoff() _has_ to be called with hard IRQ disabled.

Please post a stack trace.


>
> Let the napi_watchdog() expire in softirq context if `threadirqs' is
> used.
>
> Fixes: 3b47d30396bae ("net: gro: add a per device gro flush timer")

Are you sure this commit is the root cause of the problem you see ?



> Cc: Eric Dumazet <edumazet@google.com>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
>  net/core/dev.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 99ac84ff398f4..ec533d20931bc 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -5994,6 +5994,8 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
>                 napi_gro_flush(n, !!timeout);
>                 if (timeout)
>                         hrtimer_start(&n->timer, ns_to_ktime(timeout),
> +                                     force_irqthreads ?

Honestly something is wrong with this patch, force_irqthreads should
not be used in net/ territory,
that is some layering problem.

> +                                     HRTIMER_MODE_REL_PINNED_SOFT :
>                                       HRTIMER_MODE_REL_PINNED);
>         }
>         if (unlikely(!list_empty(&n->poll_list))) {
> @@ -6225,7 +6227,9 @@ void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
>                     int (*poll)(struct napi_struct *, int), int weight)
>  {
>         INIT_LIST_HEAD(&napi->poll_list);
> -       hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
> +       hrtimer_init(&napi->timer, CLOCK_MONOTONIC,
> +                    force_irqthreads ?
> +                    HRTIMER_MODE_REL_PINNED_SOFT : HRTIMER_MODE_REL_PINNED);
>         napi->timer.function = napi_watchdog;
>         init_gro_hash(napi);
>         napi->skb = NULL;
> --
> 2.24.0
>