Re: [Intel-xe] [PATCH 1/1] drm/xe: Report tile count from device query api.

["Dixit, Ashutosh" <ashutosh.dixit@intel.com>]

On Wed, 07 Jun 2023 11:31:15 -0700, Matt Roper wrote:
>
> On Wed, Jun 07, 2023 at 04:49:09AM +0000, Ghimiray, Himal Prasad wrote:
> >
> >
> > > -----Original Message-----
> > > From: Roper, Matthew D <matthew.d.roper@intel.com>
> > > Sent: 07 June 2023 01:26
> > > To: Ghimiray, Himal Prasad <himal.prasad.ghimiray@intel.com>
> > > Cc: intel-xe@lists.freedesktop.org; Dugast, Francois
> > > <francois.dugast@intel.com>
> > > Subject: Re: [PATCH 1/1] drm/xe: Report tile count from device query api.
> > >
> > > On Tue, Jun 06, 2023 at 03:19:29PM +0530, Himal Prasad Ghimiray wrote:
> > > > Cc: Matt Roper <matthew.d.roper@intel.com>
> > > > Cc: Francois Dugast <francois.dugast@intel.com>
> > > >
> > > > Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> > >
> > > Does userspace actually need this reported in the device query?  It seems more
> > > likely that they're going to need a tile-based equivalent of 'drm_xe_query_gt'
> > > that returns not only the tile count, but also further details about each tile.  If
> > > that's the case, then also adding this count as an extra field to the device-level
> > > query seems redundant (and the existing GT count seems like it's probably
> > > redundant as well).
> > Hi Matt,
> >
> > From userspace  perspective the requirement is to know number of tiles supported by device
> > and agreed 'drm_xe_query_tile' should provide this info along with other details.
> > Can you please provide your inputs on what other details can be part of tile query ?
>
> It's really something that would need to be discussed and agreed upon
> with the userspace teams, but I imagine details about VRAM might be
> something they'd want to receive via such a query.
>
> > >
> > > Are there patches for any of the userspace components yet that show
> > > how/when they'd use this?
> > https://patchwork.freedesktop.org/series/118927/ exposes per tile vram addr_range via sysfs.
> > While writing IGT to test this sysfs entry, I came across the requirement. IGT patches are WIP.
> >
> > Just a thought, will it makes sense to expose number of tiles device supports by sysfs ?
> > Something like: <device-properties>/number_tiles

In i915 e.g. the number of gt's was never exposed, userland would just
count the number of gt/gtN directories in sysfs to get number of gt's (see
how this is done in IGT). So something like that is possible for tiles too
(see email thread "Re: [PATCH 0/3] drm/xe: Add sysfs entry to report per gt
memory size"), though like Matt says needs discussion with UMD teams to see
which particular option to go with.

Sometimes I think we should put whatever we can in sysfs and minimize
things exposed via the query ioctl. Though again not sure about the
tradeoff's of one way vs another.

Ashutosh

> > IMO this info should be available to userspace even without making any code changes .
> > There may be other such potential candidates which can fall under device-properties.
>
> That's really something that we need guidance on from the userspace
> teams (mesa, media, compute, etc.).  It's hard for us kernel developers
> to really know exactly what the best uapi would be for userspace.
> That's part of why the DRM subsystem requires that any uapi change
> proposed for i915 or Xe can't be added to the driver until the
> corresponding userspace patches have also been developed and fully
> reviewed.  We don't want to write UAPI that winds up not really meeting
> userspace's needs and then be stuck supporting that broken uapi forever.
>
>
> Matt
>
> >
> > Himal
> > >
> > >
> > > Matt
> > >
> > > > ---
> > > >  drivers/gpu/drm/xe/xe_query.c | 1 +
> > > >  include/uapi/drm/xe_drm.h     | 3 ++-
> > > >  2 files changed, 3 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/xe/xe_query.c
> > > > b/drivers/gpu/drm/xe/xe_query.c index c4165fa3428e..54efbefd663f
> > > > 100644
> > > > --- a/drivers/gpu/drm/xe/xe_query.c
> > > > +++ b/drivers/gpu/drm/xe/xe_query.c
> > > > @@ -197,6 +197,7 @@ static int query_config(struct xe_device *xe, struct
> > > drm_xe_device_query *query)
> > > >			hweight_long(xe->info.mem_region_mask);
> > > >		config->info[XE_QUERY_CONFIG_MAX_ENGINE_PRIORITY] =
> > > >			xe_engine_device_get_max_priority(xe);
> > > > +	config->info[XE_QUERY_CONFIG_TILE_COUNT] = xe->info.tile_count;
> > > >
> > > >		if (copy_to_user(query_ptr, config, size)) {
> > > >			kfree(config);
> > > > diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> > > > index 0ebc50beb5e5..8e552aa55037 100644
> > > > --- a/include/uapi/drm/xe_drm.h
> > > > +++ b/include/uapi/drm/xe_drm.h
> > > > @@ -185,7 +185,8 @@ struct drm_xe_query_config {
> > > >  #define XE_QUERY_CONFIG_GT_COUNT		4
> > > >  #define XE_QUERY_CONFIG_MEM_REGION_COUNT	5
> > > >  #define XE_QUERY_CONFIG_MAX_ENGINE_PRIORITY	6
> > > > -#define XE_QUERY_CONFIG_NUM_PARAM
> > >	XE_QUERY_CONFIG_MAX_ENGINE_PRIORITY + 1
> > > > +#define XE_QUERY_CONFIG_TILE_COUNT		7
> > > > +#define XE_QUERY_CONFIG_NUM_PARAM
> > >	XE_QUERY_CONFIG_TILE_COUNT + 1
> > > >		__u64 info[];
> > > >  };
> > > >
> > > > --
> > > > 2.25.1
> > > >
> > >
> > > --
> > > Matt Roper
> > > Graphics Software Engineer
> > > Linux GPU Platform Enablement
> > > Intel Corporation
>
> --
> Matt Roper
> Graphics Software Engineer
> Linux GPU Platform Enablement
> Intel Corporation