On 2025-02-25 13:42:55 [+0900], Sergey Senozhatsky wrote:
> On (25/02/24 10:10), Sebastian Andrzej Siewior wrote:
> > On 2025-02-22 07:25:45 [+0900], Sergey Senozhatsky wrote:
> > > static void *zstd_custom_alloc(void *opaque, size_t size)
> > > {
> > > - if (!preemptible())
> > > + /* Technically this should not happen */
> > > + if (WARN_ON_ONCE(!preemptible()))
> > > return kvzalloc(size, GFP_ATOMIC);
> >
> > This check works only on preemptible kernels.
>
> I'm not sure this is true.
>
> > If you run this on !PREEMPTIBLE kernels, preemptible() reports always 0
> > so that WARNING will always trigger there.
>
> I thought that preemptible() depends on PREEMPT_COUNT, not on
> PREEMPTIBLE, because even on !PREEMPTIBLE preempt-count still
> holds hard/soft irq counts, etc.
Yes. The preempt count is always there to hold NMI/ HARDIRQ/ SOFTIRQ.
However only on a preemptible (that is with PREEMPT_COUNT) kernel
preempt_disable() does something. So on !PREEMPTIBLE kernel you don't
see spin_lock() or preempt_disable() reflect in preempt_count. Unless
you enable debugging which force this option into a non-preemptible
kernel.
> I ran CONFIG_PREEMPT_NONE=y zram-zstd tests and didn't see any
> warnings.
Sebastian
