Re: Linux 4.15-rc6

[Andres Freund]

On 2018-01-03 13:57:25 +0100, Willy Tarreau wrote:
> On Tue, Jan 02, 2018 at 01:09:13PM -0800, Linus Torvalds wrote:
> > On Tue, Jan 2, 2018 at 12:28 PM, Andres Freund <andres@xxxxxxxxxxx> wrote:
> > >
> > > I thought it'd be interesting to run a short benchmark to be able to
> > > estimate the impact of the PTI work on postgres workloads (which I work
> > > on). On my skylake laptop, a memory resident, OLTP workload with 16
> > > connections results in:
> > 
> > Yeah, that's actually pretty much in line with expectations.
> > 
> > Something around 5% performance impact of the isolation is what people
> > are looking at.
> > 
> > Obviously it depends on just exactly what you do. Some loads will
> > hardly be affected at all, if they just spend all their time in user
> > space. And if you do a lot of small system calls, you might see
> > double-digit slowdowns.
> 
> I can confirm, I've just run some tests on haproxy on a core i7-4790K
> and I'm observing a performance loss of ~17%, making the connection
> rate go down from 245k/s to 204k/s. It's indeed quite significant for
> such use cases, eventhough I think it might reasonably be absorbed by
> usual noise in most use cases.

Yea, I've expanded the postgres benchmarks a bit, and it's not hard to
construct cases with significantly increased slowdowns:
https://www.postgresql.org/message-id/20180102222354.qikjmf7dvnjgbkxe@xxxxxxxxxxxxxxxxx

and that's on a laptop, not a large system. I'd assume at least the
nopcid cases gets considerably worse on larger sysstems.


> With that said, I think we should start to think about an option to
> disable this per process. We could imagine for example a prctl()
> requiring CAP_SYS_ADMIN to disable it. This would at least allow
> processes started as root to disable it when they consider themselves
> irrelevant to this kind of protection (mostly I/O intensive or network
> intensive applications).

That might not be a bad idea. If so, it'd be a good idea to keep it
separate from CAP_SYS_ADMIN. E.g. postgres refuses to run as root, but
setcap'ing to allow CAP_SYS_LIVE_AND_LET_LIVE_SYSCALL or such would
work.

But I suspect this isn't something easily done on a capability/prctl
level? Seems not uncomplicated to change this after a process has
already been created - so maybe it'd be easier to force this via
personality()?


> > > This isn't a complaint, I just thought it might be useful
> > > information. If it helps for anything/anybody, I'm happy to run
> > > additional benchmarks / provide additional information.
> > 
> > Note that it will depend heavily on the hardware too. Older CPU's
> > without PCID will be impacted more by the isolation.
> 
> Interesting. This CPU has PCID, so it's possible that older hardware
> may indeed be hit a bit more.

The post linked above has numbers with nopcid disabling pcid use, and
indeed, the difference is quite measurable.

Greetings,

Andres Freund