Re: [Intel-xe] [18/21] drm/xe/uapi: More OA uapi fixes/additions

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

On Mon, 30 Oct 2023 03:06:23 -0700, Lionel Landwerlin wrote:
>

Hi Lionel,

> On 27/10/2023 23:28, Dixit, Ashutosh wrote:
> > On Fri, 20 Oct 2023 00:28:01 -0700, Lionel Landwerlin wrote:
> >>>      struct drm_xe_oa_open_param {
> >>> +	/** @extensions: Pointer to the first extension struct, if any */
> >>> +	__u64 extensions;
> >>> +
> >>> +	/**
> >>> +	 * @config_syncobj: (Output) handle to configuration syncobj
> >>> +	 *
> >>> +	 * Handle to a syncobj which the kernel will signal after stream
> >>> +	 * configuration or re-configuration is complete (after return from
> >>> +	 * the ioctl). This handle can be provided as a dependency to the
> >>> +	 * next XE exec ioctl.
> >>> +	 */
> >>> +	__u32 config_syncobj;
> >> So you're adding this, but there is no implementation for it?
> > Not yet, but there will be. Actually the plan is to do this:
> >
> > struct drm_xe_oa_open_param {
> >	/** @extensions: Pointer to the first extension struct, if any */
> >	__u64 extensions;
> >
> >	/** @num_syncs: Amount of struct drm_xe_sync in array. */
> >	__u32 num_syncs;
> >
> >	/** @syncs: Pointer to struct drm_xe_sync array. */
> >	__u64 syncs;
> >
> > Which is exactly the same as what we see in drm_xe_vm_bind or
> > drm_xe_exec. So the plan is to make perf_open behave exactly like an
> > xe_exec which can be pipelined with other operations.
>
> I don't think this is going to work because it means we have to close and
> reopen the stream every time we want to change the configuration.
>
> It would need to be a separate ioctl on the perf file descriptor to be
> usable.

I didn't mention it, but the plan is the provide the same sync mechanism
also on the perf fd stream reconfiguration ioctl.

> Otherwise we can't really pipeline anything since we have to drain the
> buffer until we can close it and reopen a new file descriptor.

This I don't understand, since the disable-reconfigure-enable sequence on
the perf fd also resets OA buffer head/tail pointers (lose contents of the
OA buffer). So not sure why we don't need to "drain the buffer" in this
case.

Of course, secretly I am hoping if it's possible to eliminate these
disable-reconfigure-enable ioctl's on the perf fd and just do stream_close
followed by reconfigure + stream_open. I don't quite understand why this is
not possible.

Thanks.
--
Ashutosh