Re: [PATCH 1/1] mm/page_alloc: add scheduling point to free_unref_page_list

[wangjianxing]

spin_lock will preempt_disable(), interrupt context will 
__irq_enter/local_bh_disable and also add preempt count with offset.

cond_resched check whether if preempt_count == 0 in first and won't 
schedule in previous context.

Is this right?


With another way, could we add some condition to avoid call cond_resched 
in interrupt context or spin_lock()?

+ if (preemptible())
+       cond_resched();

On 03/10/2022 09:05 AM, Andrew Morton wrote:
> On Tue, 8 Mar 2022 16:19:33 +0000 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> > On Tue, Mar 01, 2022 at 08:38:25PM -0500, wangjianxing wrote:
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 3589febc6..1b96421c8 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -3479,6 +3479,9 @@ void free_unref_page_list(struct list_head *list)
> > >   		 */
> > >   		if (++batch_count == SWAP_CLUSTER_MAX) {
> > >   			local_unlock_irqrestore(&pagesets.lock, flags);
> > > +
> > > +			cond_resched();
> > This isn't safe.  This path can be called from interrupt context
> > (otherwise we'd be using local_unlock_irq() instead of irqrestore()).
> What a shame it is that we don't document our interfaces :(
>
> I can't immediately find such callers, but I could imagine
> put_pages_list() (which didn't document its interface this way either)
> being called from IRQ.
>
> And drivers/iommu/dma-iommu.c:fq_ring_free() calls put_pages_list()
> from inside spin_lock().