Hi Peiyong,
On Mon, Nov 30, 2020 at 02:33:59PM -0800, Peiyong Lin wrote:
> On Tue, Nov 17, 2020 at 1:31 PM Peiyong Lin wrote:
> >
> > On Thu, Oct 22, 2020 at 10:34 AM Peiyong Lin wrote:
> > >
> > > Historically there is no common trace event for GPU frequency, in
> > > downstream Android each different hardware vendor implements their own
> > > way to expose GPU frequency, for example as a debugfs node. This patch
> > > standardize it as a common trace event in upstream linux kernel to help
> > > the ecosystem have a common implementation across hardware vendors.
> > > Toolings in the Linux ecosystem will benefit from this especially in the
> > > downstream Android, where this information is critical to graphics
> > > developers.
> > >
> > > Signed-off-by: Peiyong Lin
> > > ---
> > >
> > > Changelog since v3:
> > > - Correct copyright title.
> > >
> > > Changelog since v2:
> > > - Add more comments to indicate when the event should be emitted.
> > > - Change state to frequency.
> > >
> > > Changelog since v1:
> > > - Use %u in TP_printk
> > >
> > > drivers/gpu/Makefile| 1 +
> > > drivers/gpu/trace/Kconfig | 3 +++
> > > drivers/gpu/trace/Makefile | 1 +
> > > drivers/gpu/trace/trace_gpu_frequency.c | 13 ++
> > > include/trace/events/power.h| 33 +
> > > 5 files changed, 51 insertions(+)
> > > create mode 100644 drivers/gpu/trace/trace_gpu_frequency.c
> > >
> > > diff --git a/drivers/gpu/Makefile b/drivers/gpu/Makefile
> > > index 835c88318cec..f289a47eb031 100644
> > > --- a/drivers/gpu/Makefile
> > > +++ b/drivers/gpu/Makefile
> > > @@ -6,3 +6,4 @@ obj-$(CONFIG_TEGRA_HOST1X) += host1x/
> > > obj-y += drm/ vga/
> > > obj-$(CONFIG_IMX_IPUV3_CORE) += ipu-v3/
> > > obj-$(CONFIG_TRACE_GPU_MEM)+= trace/
> > > +obj-$(CONFIG_TRACE_GPU_FREQUENCY) += trace/
> > > diff --git a/drivers/gpu/trace/Kconfig b/drivers/gpu/trace/Kconfig
> > > index c24e9edd022e..ac4aec8d5845 100644
> > > --- a/drivers/gpu/trace/Kconfig
> > > +++ b/drivers/gpu/trace/Kconfig
> > > @@ -2,3 +2,6 @@
> > >
> > > config TRACE_GPU_MEM
> > > bool
> > > +
> > > +config TRACE_GPU_FREQUENCY
> > > + bool
> > > diff --git a/drivers/gpu/trace/Makefile b/drivers/gpu/trace/Makefile
> > > index b70fbdc5847f..2b7ae69327d6 100644
> > > --- a/drivers/gpu/trace/Makefile
> > > +++ b/drivers/gpu/trace/Makefile
> > > @@ -1,3 +1,4 @@
> > > # SPDX-License-Identifier: GPL-2.0
> > >
> > > obj-$(CONFIG_TRACE_GPU_MEM) += trace_gpu_mem.o
> > > +obj-$(CONFIG_TRACE_GPU_FREQUENCY) += trace_gpu_frequency.o
> > > diff --git a/drivers/gpu/trace/trace_gpu_frequency.c
> > > b/drivers/gpu/trace/trace_gpu_frequency.c
> > > new file mode 100644
> > > index ..668fabd6b77a
> > > --- /dev/null
> > > +++ b/drivers/gpu/trace/trace_gpu_frequency.c
> > > @@ -0,0 +1,13 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +/*
> > > + * GPU frequency trace points
> > > + *
> > > + * Copyright (C) 2020 Google LLC
> > > + */
> > > +
> > > +#include
> > > +
> > > +#define CREATE_TRACE_POINTS
> > > +#include
> > > +
> > > +EXPORT_TRACEPOINT_SYMBOL(gpu_frequency);
> > > diff --git a/include/trace/events/power.h b/include/trace/events/power.h
> > > index af5018aa9517..343825a76953 100644
> > > --- a/include/trace/events/power.h
> > > +++ b/include/trace/events/power.h
> > > @@ -500,6 +500,39 @@ DEFINE_EVENT(dev_pm_qos_request,
> > > dev_pm_qos_remove_request,
> > >
> > > TP_ARGS(name, type, new_value)
> > > );
> > > +
> > > +/**
> > > + * gpu_frequency - Reports the GPU frequency in GPU clock domains.
> > > + *
> > > + * This event should be emitted whenever there's a GPU frequency change
> > > happens,
> > > + * or a GPU goes from idle state to active state, or vice versa.
> > > + *
> > > + * When the GPU goes from idle state to active state, this event should
> > > report
> > > + * the GPU frequency of the active state. When the GPU goes from active
> > > state to
> > > + * idle state, this event should report a zero frequency value.
> > > + *
> > > + * @frequency: New frequency (in KHz)
> > > + * @gpu_id: Id for each GPU clock domain
Thinking about options for how to assign this, there are a few
existing tracepoints (I see block device and jb2) which use dev_t, so
perhaps that would be an option. We'd still want to be able to expose
multiple clocks per device though, either with a separate field or
with a defined packing into this one.
dev_t is a little bit tricky because the packing can (is?) different
between the kernel and C library. major + minor + clock domain as
3 separate fields would be the most explicit.
As the intended use-case is tools, I imagine userspace will want to
link this to data exposed via other APIs, so gpu_id needs to somehow
allow that. That probably means some opaque integer assigned within
the kernel isn't OK.
Hopefully some other vendors have thoughts on what
