Re: [PATCH RFC] uaccess: user_access_begin_after_access_ok()

[Michael S. Tsirkin]

On Thu, Jun 04, 2020 at 03:59:24PM +0100, Al Viro wrote:
> On Thu, Jun 04, 2020 at 02:10:27PM +0800, Jason Wang wrote:
> 
> > > > get_user(flags, desc->flags)
> > > > smp_rmb()
> > > > if (flags & VALID)
> > > > copy_from_user(&adesc, desc, sizeof adesc);
> > > > 
> > > > this would be a good candidate I think.
> > > Perhaps, once we get stac/clac out of raw_copy_from_user() (coming cycle,
> > > probably).  BTW, how large is the structure and how is it aligned?
> > 
> > 
> > Each descriptor is 16 bytes, and 16 bytes aligned.
> 
> Won't it be cheaper to grap the entire thing unconditionally?

Yes but we must read the rest of descriptor after the flags are valid.
If it's read before then the value we get might be the invalid one -
the one it had before another thread gave up control.

>  And what does
> that rmb order, while we are at it - won't all coherency work in terms of
> entire cachelines anyway?

Would be great to know that, but it's hardly guaranteed on all architectures, is it?

> Confused...