Re: [PATCH net] net: Use this_cpu_inc() to increment net->core_stats

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2022-04-21 08:32:30 [-0700], Eric Dumazet wrote:
> On Thu, Apr 21, 2022 at 7:00 AM Sebastian Andrzej Siewior
> <bigeasy@linutronix.de> wrote:
> >
> > The macro dev_core_stats_##FIELD##_inc() disables preemption and invokes
> > netdev_core_stats_alloc() to return a per-CPU pointer.
> > netdev_core_stats_alloc() will allocate memory on its first invocation
> > which breaks on PREEMPT_RT because it requires non-atomic context for
> > memory allocation.
> 
> Can you elaborate on this, I am confused ?
> 
> You are saying that on PREEMPT_RT, we can not call
> alloc_percpu_gfp(XXX, GFP_ATOMIC | __GFP_NOWARN);
> under some contexts ?

Correct. On PREEMPT_RT you must not explicitly create an atomic context
by
- using preempt_disable()
- acquiring a raw_spinlock_t lock
- using local_irq_disable()

while allocating memory. GFP_ATOMIC won't save you. The internal locks
within mm (kmalloc() and per-CPU memory) are sleeping locks and can not
be acquired in atomic context.

> preemption might be disabled by callers of net->core_stats anyways...

It won't be disabled by
- acquiring a spinlock_t lock
- running in softirq or interrupt handler

I haven't seen any splats (with RT enabled) other than this
preempt_disable() section so far. However only the first caller
allocates memory so maybe I add a check later on to be sure.

Sebastian