Re: [PATCH 7/7] mm/page_alloc: Replace local_lock with normal spinlock

[Mel Gorman]

On Wed, Jun 15, 2022 at 04:04:46PM -0700, Andrew Morton wrote:
> On Thu, 16 Jun 2022 00:48:55 +0200 Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote:
> 
> > In the logs I see lots of errors like:
> > 
> > BUG: sleeping function called from invalid context at 
> > ./include/linux/sched/mm.h:274
> > 
> > BUG: scheduling while atomic: systemd-udevd/288/0x00000002
> > 
> > BUG: sleeping function called from invalid context at mm/filemap.c:2647
> > 
> > however there are also a fatal ones like:
> > 
> > Unable to handle kernel paging request at virtual address 00000000017a87b4
> > 
> > 
> > The issues seems to be a bit random. Looks like memory trashing. 
> > Reverting $subject on top of current linux-next fixes all those issues.
> > 
> > 
> 
> This?
> 
> --- a/mm/page_alloc.c~mm-page_alloc-replace-local_lock-with-normal-spinlock-fix
> +++ a/mm/page_alloc.c
> @@ -183,8 +183,10 @@ static DEFINE_MUTEX(pcp_batch_high_lock)
>  	type *_ret;							\
>  	pcpu_task_pin();						\
>  	_ret = this_cpu_ptr(ptr);					\
> -	if (!spin_trylock_irqsave(&_ret->member, flags))		\
> +	if (!spin_trylock_irqsave(&_ret->member, flags)) {		\
> +		pcpu_task_unpin();					\
>  		_ret = NULL;						\
> +	}								\
>  	_ret;								\
>  })
>  

This is the correct fix. I *had* a fix for this but in a patch that was
not posted that drops irqsave :(

-- 
Mel Gorman
SUSE Labs