Re: [PATCH 14/22] x86/fpu: Eager switch PKRU state

From: Sebastian Andrzej Siewior
Date: Mon Mar 11 2019 - 07:06:21 EST

On 2019-03-08 11:01:25 [-0800], Dave Hansen wrote:
> On 3/8/19 10:08 AM, Sebastian Andrzej Siewior wrote:
> > On 2019-02-25 10:16:24 [-0800], Dave Hansen wrote:
> >>> + if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
> >>> + return;
> >>> +
> >>> + if (current->mm) {
> >>> + pk = get_xsave_addr(&new_fpu->state.xsave, XFEATURE_PKRU);
> >>> + WARN_ON_ONCE(!pk);
> Nothing will break, but the warning will trigger, which isn't nice.

the warning should trigger if something goes south, I was not expecting
it to happen.

> > My understanding is that the in-kernel XSAVE will always save everything
> > so we should never "lose" the XFEATURE_PKRU no matter what user space
> > does.
> >
> > So as test case you want
> > xsave (-1 & ~XFEATURE_PKRU)
> > xrestore (-1 & ~XFEATURE_PKRU)
> >
> > in userland and then a context switch to see if the warning above
> > triggers?
> I think you need an XRSTOR with RFBM=-1 (or at least with the PKRU bit
> set) and the PKRU bit in the XFEATURES field in the XSAVE buffer set to 0.

let me check that, write a test case in userland and I come back with
the results. I can remove that warning but I wasn't expecting it to
trigger so let me verify that first.

> >>> + if (pk)
> >>> + pkru_val = pk->pkru;
> >>> + }> + __write_pkru(pkru_val);
> >>> }
> >>
> >> A comment above __write_pkru() would be nice to say that it only
> >> actually does the slow instruction on changes to the value.
> >
> > Could we please not do this? It is a comment above one of the callers
> > function and we have two or three. And we have that comment already
> > within __write_pkru().
> I looked at this code and thought "writing PKRU is slow", and "this
> writes PKRU unconditionally", and "the __ version of the function
> shoudn't have much logic in it".
> I got 2/3 wrong. To me that means this site needs a 1-line comment.
> Feel free to move one of the other comments to here if you think it's
> over-commented, but this site needs one.

right because things changed as part of patch series.
You wanted to have in __write_pkru() the same semantic like in
__read_pkru() which is currently the case because __write_pkru() has the
check. It would be great if we could rename it to something else and
avoid the comment. (Because if this user gets a comment then other
should, too and I think this is an overkill).

> > Last time we talked about this we agreed (or this was my impression) that
> > 0 should be written so that the kernel thread should always be able to
> > write to user space in case it borrowed its mm (otherwise it has none
> > and it would fail anyway).
> We can't write to userspace when borrowing an mm. If the kernel borrows
> an mm, we might as well be on the init_mm which has no userspace mappings.

If a kernel thread borrows a mm from a user task via use_mm() then it
_can_ write to that task's user land memory from a kthread.

> > We didn't want to leave PKRU alone because the outcome (whether or not
> > the write by the kernel thread succeeds) should not depend on the last
> > running task (and be random) but deterministic.
> Right, so let's make it deterministically restrictive: either
> init_pkru_value, or -1 since kernel threads shouldn't be touching
> userspace in the first place.

I'm fine either way, just tell me what you want. Just consider the
use_mm() part above I wrote. (I remember you/luto suggest to have an API
for something like that so that the PKRU value can be