__get_user slower than get_user (was Re: [RFC PATCH V3 0/5] Hi:)

From: Michael S. Tsirkin
Date: Tue Jan 08 2019 - 23:31:19 EST


On Mon, Jan 07, 2019 at 02:44:24PM -0800, Dan Williams wrote:
> On Mon, Jan 7, 2019 at 2:25 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Mon, Jan 07, 2019 at 01:39:15PM -0800, Dan Williams wrote:
> > > On Mon, Jan 7, 2019 at 6:11 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > >
> > > > On Sun, Jan 06, 2019 at 11:15:20PM -0800, Dan Williams wrote:
> > > > > On Sun, Jan 6, 2019 at 8:17 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Mon, Jan 07, 2019 at 11:53:41AM +0800, Jason Wang wrote:
> > > > > > >
> > > > > > > On 2019/1/7 äå11:28, Michael S. Tsirkin wrote:
> > > > > > > > On Mon, Jan 07, 2019 at 10:19:03AM +0800, Jason Wang wrote:
> > > > > > > > > On 2019/1/3 äå4:47, Michael S. Tsirkin wrote:
> > > > > > > > > > On Sat, Dec 29, 2018 at 08:46:51PM +0800, Jason Wang wrote:
> > > > > > > > > > > This series tries to access virtqueue metadata through kernel virtual
> > > > > > > > > > > address instead of copy_user() friends since they had too much
> > > > > > > > > > > overheads like checks, spec barriers or even hardware feature
> > > > > > > > > > > toggling.
> > > > > > > > > > Will review, thanks!
> > > > > > > > > > One questions that comes to mind is whether it's all about bypassing
> > > > > > > > > > stac/clac. Could you please include a performance comparison with
> > > > > > > > > > nosmap?
> > > > > > > > > >
> > > > > > > > > On machine without SMAP (Sandy Bridge):
> > > > > > > > >
> > > > > > > > > Before: 4.8Mpps
> > > > > > > > >
> > > > > > > > > After: 5.2Mpps
> > > > > > > > OK so would you say it's really unsafe versus safe accesses?
> > > > > > > > Or would you say it's just a better written code?
> > > > > > >
> > > > > > >
> > > > > > > It's the effect of removing speculation barrier.
> > > > > >
> > > > > >
> > > > > > You mean __uaccess_begin_nospec introduced by
> > > > > > commit 304ec1b050310548db33063e567123fae8fd0301
> > > > > > ?
> > > > > >
> > > > > > So fundamentally we do access_ok checks when supplying
> > > > > > the memory table to the kernel thread, and we should
> > > > > > do the spec barrier there.
> > > > > >
> > > > > > Then we can just create and use a variant of uaccess macros that does
> > > > > > not include the barrier?
> > > > > >
> > > > > > Or, how about moving the barrier into access_ok?
> > > > > > This way repeated accesses with a single access_ok get a bit faster.
> > > > > > CC Dan Williams on this idea.
> > > > >
> > > > > It would be interesting to see how expensive re-doing the address
> > > > > limit check is compared to the speculation barrier. I.e. just switch
> > > > > vhost_get_user() to use get_user() rather than __get_user(). That will
> > > > > sanitize the pointer in the speculative path without a barrier.
> > > >
> > > > Hmm it's way cheaper even though IIRC it's measureable.
> > > > Jason, would you like to try?
> > > > Although frankly __get_user being slower than get_user feels very wrong.
> > > > Not yet sure what to do exactly but would you agree?
> > >
> > > Agree. __get_user() being faster than get_user() defeats the whole
> > > point of converting code paths to the access_ok() + __get_user()
> > > pattern.
> >
> > Did you mean the reverse?
>
> Hmm, no... I'll rephrase: __get_user() should have lower overhead than
> get_user().

Right ...

Linus, given that you just changed all users of access_ok anyway, do
you still think that the access_ok() conversion to return a speculation
sanitized pointer or NULL is too big a conversion?

It was previously discarded here:
https://lkml.org/lkml/2018/1/17/929
but at that point we didn't have numbers and there was an understandable
rush to ship something safe.

At this point I think that vhost can show very measureable gains from
this conversion.

Thanks,

--
MST