Re: [PATCH 1/3] x86/mm/pat: Disable preemption around __flush_tlb_all()

From: Andy Lutomirski
Date: Tue Oct 16 2018 - 17:25:24 EST


On Tue, Oct 16, 2018 at 1:25 PM Sebastian Andrzej Siewior
<bigeasy@xxxxxxxxxxxxx> wrote:
>
> I've seen the WARN_ON_ONCE(__read_cr3() != build_cr3()) in
> switch_mm_irqs_off() every once in a while during a snapshotted system
> upgrade.
> I also saw the warning early during which was introduced in commit
> decab0888e6e ("x86/mm: Remove preempt_disable/enable() from
> __native_flush_tlb()"). The callchain is
>
> get_page_from_freelist() -> post_alloc_hook() -> __kernel_map_pages()
>
> with CONFIG_DEBUG_PAGEALLOC enabled.
>
> Turns out, once I disable preemption around __flush_tlb_all() both
> warnings do not appear.
>
> Disable preemption during CR3 reset / __flush_tlb_all().
>
> Fixes: decab0888e6e ("x86/mm: Remove preempt_disable/enable() from __native_flush_tlb()")
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> ---
> arch/x86/mm/pageattr.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
> index 51a5a69ecac9f..fe6b21f0a6631 100644
> --- a/arch/x86/mm/pageattr.c
> +++ b/arch/x86/mm/pageattr.c
> @@ -2088,7 +2088,9 @@ void __kernel_map_pages(struct page *page, int numpages, int enable)
> * We should perform an IPI and flush all tlbs,
> * but that can deadlock->flush only current cpu:
> */
> + preempt_disable();
> __flush_tlb_all();
> + preempt_enable();
>

Depending on your CPU, __flush_tlb_all() is either
__native_flush_tlb_global() or __native_flush_tlb(). Only
__native_flush_tlb() could have any problem with preemption, but it
has a WARN_ON_ONCE(preemptible()); in it. Can you try to figure out
why that's not firing for you?

I suspect that a better fix would be to put preempt_disable() into
__native_flulsh_tlb(), but I'd still like to understand why the
warning isn't working.