[igt-dev] [PATCH i-g-t 0/4] device selection && lsgpu

[Arkadiusz Hiler]

On Tue, Nov 12, 2019 at 10:14:13AM +0000, Tvrtko Ursulin wrote:
> 
> On 24/10/2019 12:05, Arkadiusz Hiler wrote:
> > Hey,
> > 
> > This series aims to make running IGT on hosts with multiple GPUs manageable
> > without the need to unload modules / unbind devices.
> > 
> > Suggested reviewing order:
> >   * lsgpu
> >   * igt_core && drm_test changes
> >   * implementation internals in igt_device_scan.c
> > 
> > Changes since Zbigniew's last revision:
> >   * rewritten most of the parts that were using glib
> >   * removed multiple filter support - this will be added back when the need
> >     arises
> >   * we don't second guess the "chipset" of the device and just let the underlying
> >     open to fail if it has to
> >   * extra looging around opening device when filter is set
> >   * sysfs filter now has it's own prefix
> >   * no "platform" filter - sysfs should suffice for now, it can be added by
> >     someone more knowledgeable if the need arises
> > 
> > TODO:
> >   * API for opening multiple devices in a single test (e.g. for prime) - I don't
> >     want to design this upfront
> > 
> > Example usage:
> >    $ build/tools/lsgpu
> >    sys:/sys/devices/pci0000:00/0000:00:02.0/drm/card0
> >        subsystem       : drm
> >        drm card        : /dev/dri/card0
> >        parent          : sys:/sys/devices/pci0000:00/0000:00:02.0
> > 
> >    sys:/sys/devices/pci0000:00/0000:00:02.0/drm/renderD128
> >        subsystem       : drm
> >        drm render      : /dev/dri/renderD128
> >        parent          : sys:/sys/devices/pci0000:00/0000:00:02.0
> > 
> >    sys:/sys/devices/platform/vgem/drm/card1
> >        subsystem       : drm
> >        drm card        : /dev/dri/card1
> >        parent          : sys:/sys/devices/platform/vgem
> > 
> >    sys:/sys/devices/platform/vgem/drm/renderD129
> >        subsystem       : drm
> >        drm render      : /dev/dri/renderD129
> >        parent          : sys:/sys/devices/platform/vgem
> > 
> >    sys:/sys/devices/pci0000:00/0000:00:02.0
> >        subsystem       : pci
> >        drm card        : /dev/dri/card0
> >        drm render      : /dev/dri/renderD128
> >        vendor          : 8086
> >        device          : 5927
> > 
> >    sys:/sys/devices/platform/vgem
> >        subsystem       : platform
> >        drm card        : /dev/dri/card1
> >        drm render      : /dev/dri/renderD129
> > 
> >    $ build/tools/lsgpu -d "sys:/sys/devices/pci0000:00/0000:00:02.0"
> >    Notice: Using --device filters
> >    === Device filter ===
> >    sys:/sys/devices/pci0000:00/0000:00:02.0
> > 
> >    === Testing device open ===
> >    Device detail:
> >    subsystem   : pci
> >    drm card    : /dev/dri/card0
> >    drm render  : /dev/dri/renderD128
> >    Device /dev/dri/card0 successfully opened
> >    Device /dev/dri/renderD128 successfully opened
> >    -------------------------------------------
> > 
> >    # build/tests/core_auth --run-subtest getclient-simple --device "pci:vendor=intel"
> >    IGT-Version: 1.24-g64068440 (x86_64) (Linux: 5.3.7-250.vanilla.knurd.1.fc30.x86_64 x86_64)
> >    Starting subtest: getclient-simple
> >    Looking for devices to open using filter: pci:vendor=intel
> >    Filter matched /dev/dri/card0 | /dev/dri/renderD128
> >    Looking for devices to open using filter: pci:vendor=intel
> >    Filter matched /dev/dri/card0 | /dev/dri/renderD128 >    Subtest getclient-simple: SUCCESS (0.007s)
> 
> Looks very usable!
> 
> Can I be cheeky and ask if you could also cover the existing tools
> (especially intel_gpu_top (VLK-5588))?

I am happy to do this but I would prefer to do that in the folowup
patches to not bloat this series any more.

https://gitlab.freedesktop.org/drm/igt-gpu-tools/issues/51

> Plus I would need a helper to get the selected device PCI string to use with
> i915 PMU? There I'd need to get a string like "0000:01:00.0" since the i915
> PMU registers the device as either:
> 
> 
> "/sys/bus/event_source/devices/i915" for integrated.
> 
> Or:
> 
> "/sys/bus/event_source/devices/i915-0000:01:00.0" for discrete.
> 
> So I am thinking something like igt_device_get_pci_string() could work for
> lib/igt_perf.c/i915_type_id().

same as above

https://gitlab.freedesktop.org/drm/igt-gpu-tools/issues/52