Re: [RFC PATCH] Add /proc/<pid>/numa_vamaps for numa node information

[Jann Horn <jannh@google.com>]

On Fri, Sep 14, 2018 at 8:08 PM Prakash Sangappa
<prakash.sangappa@oracle.com> wrote:
> On 9/14/18 5:49 AM, Jann Horn wrote:
> > On Fri, Sep 14, 2018 at 8:21 AM Michal Hocko <mhocko@kernel.org> wrote:
> >> On Fri 14-09-18 03:33:28, Jann Horn wrote:
> >>> On Wed, Sep 12, 2018 at 10:43 PM prakash.sangappa
> >>> <prakash.sangappa@oracle.com> wrote:
> >>>> On 05/09/2018 04:31 PM, Dave Hansen wrote:
> >>>>> On 05/07/2018 06:16 PM, prakash.sangappa wrote:
> >>>>>> It will be /proc/<pid>/numa_vamaps. Yes, the behavior will be
> >>>>>> different with respect to seeking. Output will still be text and
> >>>>>> the format will be same.
> >>>>>>
> >>>>>> I want to get feedback on this approach.
> >>>>> I think it would be really great if you can write down a list of the
> >>>>> things you actually want to accomplish.  Dare I say: you need a
> >>>>> requirements list.
> >>>>>
> >>>>> The numa_vamaps approach continues down the path of an ever-growing list
> >>>>> of highly-specialized /proc/<pid> files.  I don't think that is
> >>>>> sustainable, even if it has been our trajectory for many years.
> >>>>>
> >>>>> Pagemap wasn't exactly a shining example of us getting new ABIs right,
> >>>>> but it sounds like something along those is what we need.
> >>>> Just sent out a V2 patch.  This patch simplifies the file content. It
> >>>> only provides VA range to numa node id information.
> >>>>
> >>>> The requirement is basically observability for performance analysis.
> >>>>
> >>>> - Need to be able to determine VA range to numa node id information.
> >>>>     Which also gives an idea of which range has memory allocated.
> >>>>
> >>>> - The proc file /proc/<pid>/numa_vamaps is in text so it is easy to
> >>>>     directly view.
> >>>>
> >>>> The V2 patch supports seeking to a particular process VA from where
> >>>> the application could read the VA to  numa node id information.
> >>>>
> >>>> Also added the 'PTRACE_MODE_READ_REALCREDS' check when opening the
> >>>> file /proc file as was indicated by Michal Hacko
> >>> procfs files should use PTRACE_MODE_*_FSCREDS, not PTRACE_MODE_*_REALCREDS.
> >> Out of my curiosity, what is the semantic difference? At least
> >> kernel_move_pages uses PTRACE_MODE_READ_REALCREDS. Is this a bug?
> > No, that's fine. REALCREDS basically means "look at the caller's real
> > UID for the access check", while FSCREDS means "look at the caller's
> > filesystem UID". The ptrace access check has historically been using
> > the real UID, which is sorta weird, but normally works fine. Given
> > that this is documented, I didn't see any reason to change it for most
> > things that do ptrace access checks, even if the EUID would IMO be
> > more appropriate. But things that capture caller credentials at points
> > like open() really shouldn't look at the real UID; instead, they
> > should use the filesystem UID (which in practice is basically the same
> > as the EUID).
> >
> > So in short, it depends on the interface you're coming through: Direct
> > syscalls use REALCREDS, things that go through the VFS layer use
> > FSCREDS.
>
> So in this case can the REALCREDS check be done in the read() system call
> when reading the /proc file instead of the open call?

No, REALCREDS shouldn't be used in open() and shouldn't be used in read().
FSCREDS can be used in open(); in theory, using ptrace_may_access() in
any way in read() is currently unsafe, but in practice, it's used that
way anyway. I have plans to clean that up eventually...