Re: Redoing eXclusive Page Frame Ownership (XPFO) with isolated CPUs in mind (for KVM to isolate its guests per CPU)

[Juerg Haefliger <juergh@gmail.com>]

On Wed, Sep 12, 2018 at 5:37 PM, Julian Stecklina <jsteckli@amazon.de> wrote:
> Julian Stecklina <jsteckli@amazon.de> writes:
>
>> Linus Torvalds <torvalds@linux-foundation.org> writes:
>>
>>> On Fri, Aug 31, 2018 at 12:45 AM Julian Stecklina <jsteckli@amazon.de> wrote:
>>>>
>>>> I've been spending some cycles on the XPFO patch set this week. For the
>>>> patch set as it was posted for v4.13, the performance overhead of
>>>> compiling a Linux kernel is ~40% on x86_64[1]. The overhead comes almost
>>>> completely from TLB flushing. If we can live with stale TLB entries
>>>> allowing temporary access (which I think is reasonable), we can remove
>>>> all TLB flushing (on x86). This reduces the overhead to 2-3% for
>>>> kernel compile.
>>>
>>> I have to say, even 2-3% for a kernel compile sounds absolutely horrendous.
>>
>> Well, it's at least in a range where it doesn't look hopeless.
>>
>>> Kernel bullds are 90% user space at least for me, so a 2-3% slowdown
>>> from a kernel is not some small unnoticeable thing.
>>
>> The overhead seems to come from the hooks that XPFO adds to
>> alloc/free_pages. These hooks add a couple of atomic operations per
>> allocated (4K) page for book keeping. Some of these atomic ops are only
>> for debugging and could be removed. There is also some opportunity to
>> streamline the per-page space overhead of XPFO.
>
> I've updated my XPFO branch[1] to make some of the debugging optional
> and also integrated the XPFO bookkeeping with struct page, instead of
> requiring CONFIG_PAGE_EXTENSION, which removes some checks in the hot
> path.

FWIW, that was my original design but there was some resistance to
adding more to the page struct and page extension was suggested
instead.


> These changes push the overhead down to somewhere between 1.5 and
> 2% for my quad core box in kernel compile. This is close to the
> measurement noise, so I take suggestions for a better benchmark here.
>
> Of course, if you hit contention on the xpfo spinlock then performance
> will suffer. I guess this is what happened on Khalid's large box.
>
> I'll try to remove the spinlocks and add fixup code to the pagefault
> handler to see whether this improves the situation on large boxes. This
> might turn out to be ugly, though.

I'm wondering how much performance we're loosing by having to split
hugepages. Any chance this can be quantified somehow? Maybe we can
have a pool of some sorts reserved for userpages and group allocations
so that we can track the XPFO state at the hugepage level instead of
at the 4k level to prevent/reduce page splitting. Not sure if that
causes issues or has any unwanted side effects though...

...Juerg


> Julian
>
> [1] http://git.infradead.org/users/jsteckli/linux-xpfo.git/shortlog/refs/heads/xpfo-master
> --
> Amazon Development Center Germany GmbH
> Berlin - Dresden - Aachen
> main office: Krausenstr. 38, 10117 Berlin
> Geschaeftsfuehrer: Dr. Ralf Herbrich, Christian Schlaeger
> Ust-ID: DE289237879
> Eingetragen am Amtsgericht Charlottenburg HRB 149173 B
>