[Intel-gfx] [PATCH v6 0/5] drm/i915: Expose more GPU properties through sysfs
Chris Wilson
chris at chris-wilson.co.uk
Wed Dec 13 13:35:40 UTC 2017
Quoting Daniel Vetter (2017-12-13 08:17:17)
> On Tue, Dec 12, 2017 at 02:33:38PM +0000, Lionel Landwerlin wrote:
> > On 12/12/17 11:19, Tvrtko Ursulin wrote:
> > >
> > > On 11/12/2017 21:05, Daniel Vetter wrote:
> > > > On Mon, Dec 11, 2017 at 02:38:53PM +0000, Tvrtko Ursulin wrote:
> > > > > On 11/12/2017 10:50, Joonas Lahtinen wrote:
> > > > > > + Daniel, Chris
> > > > > >
> > > > > > On Thu, 2017-12-07 at 09:21 +0000, Tvrtko Ursulin wrote:
> > > > > > > On 04/12/2017 15:02, Lionel Landwerlin wrote:
> > > > > > > > Hi,
> > > > > > > >
> > > > > > > > After discussion with Chris, Joonas & Tvrtko, this series adds an
> > > > > > > > additional commit to link the render node back to the card through a
> > > > > > > > symlink. Making it obvious from an application using
> > > > > > > > a render node to
> > > > > > > > know where to get the information it needs.
> > > > > > >
> > > > > > > Important thing to mention as well is that it is trivial
> > > > > > > to get from the
> > > > > > > master drm fd to the sysfs root, via fstat and opendir
> > > > > > > /sys/dev/char/<major>:<minor>. With the addition of the
> > > > > > > card symlink to
> > > > > > > render nodes it is trivial for render node fd as well.
> > > > > > >
> > > > > > > I am happy with this approach - it is extensible, flexible and avoids
> > > > > > > issues with ioctl versioning or whatnot. With one value
> > > > > > > per file it is
> > > > > > > trivial for userspace to access.
> > > > > > >
> > > > > > > So for what I'm concerned, given how gputop would use
> > > > > > > all of this and so
> > > > > > > be the userspace, if everyone else is happy, I think we could do a
> > > > > > > detailed review and prehaps also think about including gputop in some
> > > > > > > distribution to make the case 100% straightforward.
> > > > > >
> > > > > > For the GPU topology I agree this is the right choice, it's
> > > > > > going to be
> > > > > > about topology after all, and directory tree is the perfect candidate.
> > > > > > And if a new platform appears, then it's a new platform and may change
> > > > > > the topology well the hardware topology has changed.
> > > > > >
> > > > > > For the engine enumeration, I'm not equally sold for sysfs
> > > > > > exposing it.
> > > > > > It's a "linear list of engine instances with flags" how the userspace
> > > > > > is going to be looking at them. And it's also information
> > > > > > about what to
> > > > > > pass to an IOCTL as arguments after decision has been made, and then
> > > > > > you already have the FD you know you'll be dealing with, at hand. So
> > > > > > another IOCTL for that seems more convenient.
> > > > >
> > > > > Apart from more flexibility and easier to extend, sysfs might be
> > > > > a better
> > > > > fit for applications which do not otherwise need a drm fd. Say a
> > > > > top-like
> > > > > tool which shows engine utilization, or those patches I RFC-ed recently
> > > > > which do the same but per DRM client.
> > > > >
> > > > > Okay, these stats are now available also via PMU so the argument
> > > > > is not the
> > > > > strongest I admit, but I still find it quite neat. It also might
> > > > > allow us to
> > > > > define our own policy with regards to needed privilege to access these
> > > > > stats, and not be governed by the perf API rules.
> > > >
> > > > How exactly is sysfs easier to extend than ioctl? There's lots of
> > >
> > > Easier as in no need to version, add has_this/has_that markers, try to
> > > guess today how big a field for something might be needed in the future
> > > and similar.
> > >
> > > > serializing and deserializing going on, ime that's more boilerplate. Imo
> > > > the only reason for sysfs is when you _must_ access it without having an
> > > > fd to the gpu. The inverse is generally not true (i.e. using sysfs when
> > > > you have the fd already), and as soon as you add a writeable field here
> > > > you're screwed (because sysfs can't be passed to anyone else but root,
> > > > compared to drm fd - viz the backlight fiasco).
> > >
> > > I would perhaps expand the "must access without having a drm fd" to
> > > "more convenient to access without a drm fd". My use case here was the
> > > per-client engine usage stats. Equivalence with /proc/<pid>/stat, or
> > > even /proc/stat if you want. So I was interested in creating a foothold
> > > in sysfs for that purpose.
> > >
> > > No writable fields were imagined in all these two to three use cases.
> > >
> > > > But even without writeable fields: Think of highly contained
> > > > containers/clients which only get the drm fd to render. If sysfs is
> > > > gone,
> > > > will they fall on their faces? Atm "drm fd is all you need" is very
> > > > deeply
> > > > ingrained into our various OS stacks.
> > >
> > > Okay, this is something which was mentioned but I think the answer was
> > > unclear. If current stacks do work without sysfs then this is a good
> > > argument to keep that ability.
> > >
> > > As I said I am OK to drop the engine info bits from this series.
> > > Question for Lionel, gpu-top and Mesa then is whether sysfs works for
> > > them, for the remaining topology information. Attractiveness of sysfs
> > > there was that it looked easier to be future proof for more or less
> > > hypothetical future topologies.
> >
> > We did a couple of versions with ioctls :
> >
> > https://patchwork.freedesktop.org/patch/185959/ (through GET_PARAM)
> > https://patchwork.freedesktop.org/series/33436/ (though a new discovery uAPI
> > initially targeted at the engine discovery)
> >
> > Eventually Chris suggested sysfs (which I find kind of convenient), even
> > though you Daniel raised a valid point with sandboxed processes.
> >
> > I personally don't care much in which way this should be implemented.
> > Just give me some direction :)
> >
> > Daniel: Would something in the style of
> > https://patchwork.freedesktop.org/series/33436/ work? If yes, what would you
> > recommend to change?
>
> tbh I feel bad derailing this even further, but I think if you want a more
> flexible query ioctl that's easier to extend we could maybe look at the
> chunk list approach amdgpu_cs_ioctl uses: All the massive pile of execbuf
> metadata is supplied in a chunk list, and if you want to extend new stuff,
> or add a new type, you add a new chunk_id. The chunks themselves are all
> linearized into one allocation.
>
> We could use the exact same trick for output, and require that userspace
> simply jumps over chunks it doesn't understand. Pronto, you have your
> insta-extensibility.
That maps to my suggestion to use json (or one of the other compact
readable formats) as the interchange. (Presumably with a (seq)file
approach, so you pass in the read offset to avoid having to allocate
everything up front.) A useful suggestion is that the same output is
provided via debugfs for dev convenience, and including in the error
state.
-Chris
More information about the Intel-gfx
mailing list