On Fri, 2022-09-16 at 11:15 -0600, Mathieu Poirier wrote: > On Fri, 16 Sept 2022 at 06:00, TingHan Shen <tinghan.shen@xxxxxxxxxxxx> wrote: > > > > On Thu, 2022-09-08 at 14:58 -0600, Mathieu Poirier wrote: > > > On Thu, 8 Sept 2022 at 05:21, Tinghan Shen <tinghan.shen@xxxxxxxxxxxx> > > > 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@xxxxxxxxxxxx> > > > > > > --- > > > > > > 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? > > Look at what Suman did for TI's K3 R5[1] and DSP[2] platforms. > Reviewing the bindings for both platforms will also give you a good > idea of how things work. > > [1]. https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v6.0-rc5/source/drivers/remoteproc/ti_k3_r5_remoteproc.c*L1683__;Iw!!CTRNKA9wMg0ARbw!zVcjdLSfKYGO5YQMNGqq339mle8u0VdULX30z0XV4vo3vCb9Wy-w5ixOTmzbv1akubM$ > [2]. https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v6.0-rc5/source/drivers/remoteproc/ti_k3_dsp_remoteproc.c*L673__;Iw!!CTRNKA9wMg0ARbw!zVcjdLSfKYGO5YQMNGqq339mle8u0VdULX30z0XV4vo3vCb9Wy-w5ixOTmzbfE2dtBg$ > Hi Mathieu, My plan is changing the dts as following, scp core 0 { // Keep current properties untouched. compatible = "mediatek,mt8195-scp"; // core 0 properties... // Add a new property for multi-core scp. // if not present, it's single core. // if present and core id = 0, it's the main core, otherwise the sub cores. mediatek,scp-core = <0>; // add sub cores as sub node. // sub nodes can find parent by OF API. scp core 1 { // use the same compatile name as core 0. compatible = "mediatek,mt8195-scp"; // assign id > 0 to sub cores. mediatek,scp-core = <1>; // core 1 properties... }; }; The driver probe/remove behavior will be modified as below, scp probe() { // common init... // check core id to have different memory mapping flow if (core id == 0) // mapping cfg, sram and others else // mapping sram // reuse the cfg paddr/vaddr from core 0 // common init... if (core id == 0) { ret = of_platform_populate(...) // boot core 0 and sub cores rproc_add(); } else { // add sub core as sub device to main core rproc_add_subdev() rproc->auto_boot = false; rpoc_add(); } } scp_remove() { if (core id == 0) of_platform_depopulate() else rproc_remove_subdev() // remove core } Best regards, TingHan