On 2019-11-28, Thadeu Lima de Souza Cascardo <cascardo@canonical.com> wrote:
> On Thu, Nov 28, 2019 at 11:01:40PM +1100, Aleksa Sarai wrote:
> > Due to a userspace breakage, commit 1251201c0d34 ("sched/core: Fix
> > uclamp ABI bug, clean up and robustify sched_read_attr() ABI logic and
> > code") changed the semantics of sched_getattr(2) when the userspace
> > struct is smaller than the kernel struct. Now, any trailing non-zero
> > data in the kernel structure is ignored when copying to userspace.
> > 
> > Ref: 1251201c0d34 ("sched/core: Fix uclamp ABI bug, clean up and
> >                     robustify sched_read_attr() ABI logic and code")
> > Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
> > ---
> >  man2/sched_setattr.2 | 6 ++----
> >  1 file changed, 2 insertions(+), 4 deletions(-)
> > 
> > diff --git a/man2/sched_setattr.2 b/man2/sched_setattr.2
> > index 76ffa14eba85..fbb67b8eb98b 100644
> > --- a/man2/sched_setattr.2
> > +++ b/man2/sched_setattr.2
> > @@ -284,10 +284,8 @@ structure,
> >  the additional bytes in the user-space structure are not touched.
> >  If the caller-provided structure is smaller than the kernel
> >  .I sched_attr
> > -structure and the kernel needs to return values outside the provided space,
> > -.BR sched_getattr ()
> > -fails with the error
> > -.BR E2BIG .
> > +structure, the kernel will silently not return any values which would be stored
> > +outside the provided space.
> >  As with
> >  .BR sched_setattr (),
> >  these semantics allow for future extensibility of the interface.
> > -- 
> > 2.24.0
> > 
> 
> I was thinking about documenting the difference in behavior of older kernels,
> before uclamp support.
> 
> However, in practice, for sched_getattr, the kernel never returned E2BIG (the
> code uses EFBIG incorrectly, in fact). It does, however, return EINVAL for
> sizes smaller than SCHED_ATTR_SIZE_VER0.

I've been told the EFBIG was actually a typo and it was always meant to
be E2BIG. But yes, the precise problem with the old semantics was that
they weren't tested "in the wild" with a proper struct upgrade -- hence
all of the headaches.

If we ever do implement a copy_struct_to_user() we are almost certainly
going to implement it with the new sched_getattr() semantics. To be
honest, I'm not sure I can imagine a case where an old userspace program
would benefit from being given an error saying that the kernel has some
properties that it doesn't understand. (sched_getattr() is also weird
for other reasons, such as the fact it takes a separate size argument.)

> However, E2BIG is still mentioned below as a possible return value for
> sched_getattr. Can you remove that too?

Will do.

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>