Re: [PATCH v2 4/9] remoteproc: mediatek: Support probing for the 2nd core of dual-core SCP

From: TingHan Shen <tinghan.shen@mediatek.com>
To: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: <Project_Global_Chrome_Upstream_Group@mediatek.com>,
	<bjorn.andersson@linaro.org>, <bleung@chromium.org>,
	<chrome-platform@lists.linux.dev>, <devicetree@vger.kernel.org>,
	<dnojiri@chromium.org>, <enric.balletbo@collabora.com>,
	<groeck@chromium.org>, <gustavoars@kernel.org>,
	<keescook@chromium.org>, <krzk+dt@kernel.org>,
	<lee.jones@linaro.org>, <linux-arm-kernel@lists.infradead.org>,
	<linux-kernel@vger.kernel.org>,
	<linux-mediatek@lists.infradead.org>,
	<linux-remoteproc@vger.kernel.org>, <matthias.bgg@gmail.com>,
	<pmalani@chromium.org>, <robh+dt@kernel.org>,
	<sebastian.reichel@collabora.com>, <weishunc@google.com>
Subject: Re: [PATCH v2 4/9] remoteproc: mediatek: Support probing for the 2nd core of dual-core SCP
Date: Fri, 16 Sep 2022 19:59:50 +0800	[thread overview]
Message-ID: <a5ecd1dd567ca58807b289f2488d933f27e087dd.camel@mediatek.com> (raw)
In-Reply-To: <CANLsYkx6kXk8u_ajFbnhdWTkZBLtrq_z02jryLBSVH0x--_ZFw@mail.gmail.com>

On Thu, 2022-09-08 at 14:58 -0600, Mathieu Poirier wrote:
> On Thu, 8 Sept 2022 at 05:21, Tinghan Shen <tinghan.shen@mediatek.com>
> wrote:
> 
> > Hi Mathieu,
> > 
> > > > The mtk_scp.c driver only supports the single core SCP and the
> > > > 1st core of a dual-core SCP. This patch extends it for the 2nd core.
> > > > 
> > > > MT8195 SCP is a dual-core MCU. Both cores are housed in the same
> > 
> > subsys.
> > > 
> > > s/subsys/subsystem
> > > 
> > > > They have the same viewpoint of registers and memory.
> > > > 
> > > > Core 1 of the SCP features its own set of core configuration registers,
> > > > interrupt controller, timers, and DMAs. The rest of the peripherals
> > > > in this subsystem are shared by core 0 and core 1.
> > > > 
> > > > As for memory, core 1 has its own cache memory. the SCP SRAM is shared
> > > 
> > > /the/The
> > > 
> > > > by core 0 and core 1.
> > > > 
> > > > Signed-off-by: Tinghan Shen <tinghan.shen@mediatek.com>
> > > > ---
> > > >  drivers/remoteproc/mtk_scp.c | 22 ++++++++++++++++++++--
> > > >  1 file changed, 20 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/remoteproc/mtk_scp.c
> > 
> > b/drivers/remoteproc/mtk_scp.c
> > > > index 3510c6d0bbc8..91b4aefde4ac 100644
> > > > --- a/drivers/remoteproc/mtk_scp.c
> > > > +++ b/drivers/remoteproc/mtk_scp.c
> > > > @@ -23,6 +23,10 @@
> > > >  #define MAX_CODE_SIZE 0x500000
> > > >  #define SECTION_NAME_IPI_BUFFER ".ipi_buffer"
> > > > 
> > > > +#define SCP_CORE_0 0
> > > > +#define SCP_CORE_1 1
> > > > +#define SCP_CORE_SINGLE 0xF
> > > > +
> > > >  /**
> > > >   * scp_get() - get a reference to SCP.
> > > >   *
> > > > @@ -836,6 +840,7 @@ static int scp_probe(struct platform_device *pdev)
> > > >     struct resource *res;
> > > >     const char *fw_name = "scp.img";
> > > >     int ret, i;
> > > > +   u32 core_id = SCP_CORE_SINGLE;
> > > > 
> > > >     ret = rproc_of_parse_firmware(dev, 0, &fw_name);
> > > >     if (ret < 0 && ret != -EINVAL)
> > > > @@ -851,8 +856,16 @@ static int scp_probe(struct platform_device *pdev)
> > > >     scp->data = of_device_get_match_data(dev);
> > > >     platform_set_drvdata(pdev, scp);
> > > > 
> > > > +   ret = of_property_read_u32_index(dev->of_node,
> > 
> > "mediatek,scp-core", 1, &core_id);
> > > > +   if (ret == 0)
> > > > +           dev_info(dev, "Boot SCP dual core %u\n", core_id);
> > > 
> > > Why is the DT property "mediatek,scp-core" needed at all?  Since the
> > 
> > compatible
> > > "mediatek,mt8195-scp-dual" has already been defined previously in this
> > 
> > patchset,
> > > initialising the second core, if present, is a matter of looking at the
> > > compatile string.
> > 
> > This idea of identify cores by the compatible looks workable.
> > I'll update this series at next version.
> > Thanks!
> > 
> > > 
> > > > +
> > > >     res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "sram");
> > > > -   scp->sram_base = devm_ioremap_resource(dev, res);
> > > > +   if (core_id == SCP_CORE_1)
> > > > +           scp->sram_base = devm_ioremap(dev, res->start,
> > 
> > resource_size(res));
> > > > +   else
> > > > +           scp->sram_base = devm_ioremap_resource(dev, res);
> > > > +
> > > 
> > > This looks very broken...  For this to work you would need to have two DT
> > > entries with the "mediatek,mt8195-scp-dual" compatible properly, one with
> > > "mediatek,scp-core = <&scp_dual1 0>;" and another one with
> > 
> > "mediatek,scp-core = <&scp_dual0 1>;".
> > > 
> > > Which is also very broken...  Here you have a binding whose first
> > 
> > argument is a
> > > reference to the core sibling while the second argument is a
> > 
> > characteristic of
> > > the current core, which is highly confusing.
> > > 
> > > I suggest what when you see the compatible binding
> > 
> > "mediatek,mt8195-scp", a
> > > single core is initialized.  If you see "mediatek,mt8195-scp-dual", both
> > 
> > cores
> > > are initialized as part of the _same_ probe.
> > > 
> > > If the above analysis is not correct it means I misinterpreted your
> > > work and if so, a serious amount of comments is needed _and_ a very
> > 
> > detailed
> > > example in "mtk,scp.yaml" that leaves no room for interpretation.
> > > 
> > > I will stop reviewing this patchset until you have clarified how this
> > 
> > works.
> > > 
> > > Thanks,
> > > Mathieu
> > 
> > There's one problem of initializng the CORE1 using the same probe flow.
> > The register space of CORE0 and CORE1 are overlapped in the device node.
> > Both cores need to use the 'cfg' registers defined in scp yaml.
> > The devm_ioremap_resource catches address overlapping and returns error
> > when
> > probing CORE1 driver.
> > 
> 
> That is exactly why I suggest to initialise both cores within the same
> probe() function.
> 

Hi Mathieu,

I'm thinking about how to initialise in the same probe() function.
I'm wondering if this implies that using one scp driver to initialize 2 cores?
If it is, I assume the dts descriptions for both cores should be contained in one node.

When there's one node for both cores, it looks like that there is a problem of 
using dma_allocate_coherent(). Each core has its own reserved memory region. 
When there's only one device for both cores, it's not able to identify the memory region 
by the device parameter of dma_allocate_coherent().

Is it acceptable to consider manually allocating core 1 device in the probe() when probing core 0?

Best regards,
TingHan