Re: [PATCH 1/3] perf/core: Make sure the ring-buffer is mapped in all page-tables

[Andy Lutomirski]

> On Jul 20, 2018, at 11:37 AM, Joerg Roedel <jroedel@xxxxxxx> wrote:
> 
>> On Fri, Jul 20, 2018 at 12:32:10PM -0700, Andy Lutomirski wrote:
>> I'm just reading your changelog, and you said the PMDs are no longer
>> shared between the page tables.  So this presumably means that
>> vmalloc_fault() no longer actually works correctly on PTI systems.  I
>> didn't read the code to figure out *why* it doesn't work, but throwing
>> random vmalloc_sync_all() calls around is wrong.
> 
> Hmm, so the whole point of vmalloc_fault() fault is to sync changes from
> swapper_pg_dir to process page-tables when the relevant parts of the
> kernel page-table are not shared, no?
> 
> That is also the reason we don't see this on 64 bit, because there these
> parts *are* shared.
> 
> So with that reasoning vmalloc_fault() works as designed, except that
> a warning is issued when it's happens in the NMI path. That warning comes
> from
> 
>    ebc8827f75954 x86: Barf when vmalloc and kmemcheck faults happen in NMI
> 
> which went into 2.6.37 and was added because the NMI handler were not
> nesting-safe back then. Reason probably was that the handler on 64 bit
> has to use an IST stack and a nested NMI would overwrite the stack of
> the upper handler.  We don't have this problem on 32 bit as a nested NMI
> will not do another stack-switch there.
> 

Thanks for digging!  The problem was presumably that vmalloc_fault() will IRET and re-enable NMIs on the way out.  But weâve supported page faults on user memory in NMI handlers on 32-bit and 64-bit for quite a while, and itâs fine now.

I would remove the warning, re-test, and revert the other patch.

The one case we canât handle in vmalloc_fault() is a fault on a stack access. I donât expect this to be a problem for PTI. It was a problem for CONFIG_VMAP_STACK, though.

> I am not sure about 64 bit, but there is a lot of assembly magic to make
> NMIs nesting-safe, so I guess the problem should be gone there too.
> 
> 
> Regards,
> 
>    Joerg