hi Philipp: On 2023/12/4 21:30, Philipp Zabel wrote: > Hi Keith, > > On Mo, 2023-12-04 at 20:33 +0800, Keith Zhao wrote: >> Implement drm device registration interface >> >> Signed-off-by: Keith Zhao <keith.zhao@xxxxxxxxxxxxxxxx> >> --- > [...] >> diff --git a/drivers/gpu/drm/verisilicon/Kconfig b/drivers/gpu/drm/verisilicon/Kconfig >> new file mode 100644 >> index 000000000000..e10fa97635aa >> --- /dev/null >> +++ b/drivers/gpu/drm/verisilicon/Kconfig >> @@ -0,0 +1,13 @@ >> +# SPDX-License-Identifier: GPL-2.0 >> +config DRM_VERISILICON >> + tristate "DRM Support for VeriSilicon" >> + depends on DRM >> + select DRM_KMS_HELPER >> + select DRM_GEM_DMA_HELPER >> + select CMA >> + select DMA_CMA >> + help >> + Choose this option if you have a VeriSilicon soc chipset. > > This seems a bit generic. Doesn't the VeriSilicon display controller IP > used on JH7110 have a product name? yes , there is a product name "dc8200", I will match it. > > [...] >> diff --git a/drivers/gpu/drm/verisilicon/vs_drv.c b/drivers/gpu/drm/verisilicon/vs_drv.c >> new file mode 100644 >> index 000000000000..4fb1f29ef84b >> --- /dev/null >> +++ b/drivers/gpu/drm/verisilicon/vs_drv.c >> @@ -0,0 +1,316 @@ >> +// SPDX-License-Identifier: GPL-2.0 > [...] >> +static void vs_drm_device_release_clocks(void *res) >> +{ >> + struct vs_drm_device *priv = res; >> + unsigned int i; >> + >> + reset_control_bulk_assert(priv->nrsts, priv->rst_vout); >> + >> + for (i = 0; i < priv->clk_count; ++i) { >> + if (priv->clks[i]) { >> + clk_disable_unprepare(priv->clks[i]); >> + clk_put(priv->clks[i]); >> + } >> + } > > Why not use the bulk API for clk as well? ok , will do it next version > > [...] >> +static int vs_drm_device_init_clocks(struct vs_drm_device *priv) >> +{ >> + struct drm_device *dev = &priv->base; >> + struct platform_device *pdev = to_platform_device(dev->dev); >> + struct device_node *of_node = pdev->dev.of_node; >> + struct clk *clock; >> + unsigned int i; >> + int ret; >> + >> + if (dev_get_platdata(&pdev->dev) || !of_node) >> + return 0; >> + >> + priv->nrsts = ARRAY_SIZE(priv->rst_vout); >> + for (int i = 0; i < priv->nrsts; ++i) >> + priv->rst_vout[i].id = vout_resets[i]; >> + ret = devm_reset_control_bulk_get_shared(dev->dev, priv->nrsts, >> + priv->rst_vout); > > I would request resets and clocks in _probe(). > > If component_bind_all() returns -EPROBE_DEFER because of a still > missing DSI panel backlight or similar, this doesn't have to be done > multiple times. I got what you mean. component_bind_all should be done multiple times to prevent the dsi panel driver from lagging load. in my drm subsystem , there are 2 pipeline +------------------------------+ | | | | +----+ | +-------------------+ | +-------+ +------+ +------+ | +----->+ dc controller 0 +--->----->+HDMICtl| ->+ PHY +-->+PANEL0+ |AXI | | +-------------------+ | +-------+ +------+ +------+ | | | | | | | | | | | | | | | | |APB | | +-------------------+ +---------+ +------+ +-------+ | +----->+ dc controller 1 +--->---->+ dsiTx +--->+DPHY +->+ PANEL1+ | | | +-------------------+ +---------+ +------+ +-------+ +----+ | | +------------------------------+ component_bind_all will bind the hdmi encoder and dsi encoder . binding the hdmi encoder will always return ok . binging the dsi encoder has a question : I used the panel-raspberrypi-touchscreen.c as panel driver , this driver is a i2c device and it use a i2c command to read reg ID if read success , it will do drm_panel_add. if I disconnect the panel ,it will not do drm_panel_add. dsiTx will fail to find panel , The consequence is that the inputbridge cannot be created , also outputbridge cannot be created. for encoder bind , it will fail to find the input bridge of dsi. Under this premise, although returning -EPROBE_DEFER allows bind to be executed multiple times, the final result is that the entire bind fails. returning -EPROBE_DEFER can solve panel driver from lagging load , but for no panel case , it will destory all pipeline (include hdmi and dsi). I did two things: late_initcall_sync(vs_drm_init); to make sure the panel drive has been probed; dsi encoder bind always return ok to make sure hdmi pipeline ok at lease. component_bind_all do once . > >> + if (ret) { >> + drm_err(dev, "Failed to get reset controls\n"); >> + return ret; >> + } >> + >> + priv->clk_count = of_clk_get_parent_count(of_node); >> + if (!priv->clk_count) >> + return 0; >> + >> + priv->clks = drmm_kzalloc(dev, priv->clk_count * sizeof(priv->clks[0]), >> + GFP_KERNEL); >> + if (!priv->clks) >> + return -ENOMEM; >> + >> + for (i = 0; i < priv->clk_count; ++i) { >> + clock = of_clk_get(of_node, i); >> + if (IS_ERR(clock)) { >> + ret = PTR_ERR(clock); >> + if (ret == -EPROBE_DEFER) >> + goto err; >> + drm_err(dev, "clock %u not found: %d\n", i, ret); >> + continue; >> + } >> + ret = clk_prepare_enable(clock); >> + if (ret) { >> + drm_err(dev, "failed to enable clock %u: %d\n", >> + i, ret); >> + clk_put(clock); >> + continue; >> + } >> + priv->clks[i] = clock; >> + } >> + >> + ret = reset_control_bulk_deassert(priv->nrsts, priv->rst_vout); >> + if (ret) >> + return ret; > > This should goto err, otherwise clocks are left enabled. > >> + >> + return devm_add_action_or_reset(&pdev->dev, >> + vs_drm_device_release_clocks, >> + priv); >> + >> +err: >> + while (i) { >> + --i; >> + if (priv->clks[i]) { >> + clk_disable_unprepare(priv->clks[i]); >> + clk_put(priv->clks[i]); >> + } >> + } >> + return ret; >> +} >> + >> +static int vs_drm_bind(struct device *dev) >> +{ >> + struct platform_device *pdev = to_platform_device(dev); >> + struct vs_drm_device *priv; >> + int ret; >> + struct drm_device *drm_dev; >> + >> + /* Remove existing drivers that may own the framebuffer memory. */ >> + ret = drm_aperture_remove_framebuffers(&vs_drm_driver); >> + if (ret) >> + return ret; >> + >> + priv = devm_drm_dev_alloc(dev, &vs_drm_driver, struct vs_drm_device, base); >> + if (IS_ERR(priv)) >> + return PTR_ERR(priv); >> + >> + priv->pitch_alignment = 64; > > Why is this a variable instead of a constant? the dc controllers of VeriSilicon have different series, like dc8200 , dc9200... Their pitch values may vary from series to series, like 64 , 128 > >> + ret = dma_set_coherent_mask(priv->base.dev, DMA_BIT_MASK(40)); >> + if (ret) >> + return ret; >> + >> + drm_dev = &priv->base; >> + platform_set_drvdata(pdev, drm_dev); >> + >> + ret = vs_drm_device_init_clocks(priv); >> + if (ret) >> + return ret; >> + >> + vs_mode_config_init(drm_dev); >> + >> + /* Now try and bind all our sub-components */ >> + ret = component_bind_all(dev, drm_dev); >> + if (ret) >> + return ret; >> + >> + ret = drm_vblank_init(drm_dev, drm_dev->mode_config.num_crtc); >> + if (ret) >> + return ret; > > Missing component_unbind_all(), see below. > >> + >> + drm_mode_config_reset(drm_dev); >> + >> + drm_kms_helper_poll_init(drm_dev); >> + >> + ret = drm_dev_register(drm_dev, 0); >> + if (ret) >> + return ret; > > Missing drm_kms_helper_poll_fini(), see below. > >> + >> + drm_fbdev_generic_setup(drm_dev, 32); >> + >> + return 0; > > Here I'd expect an error path calling drm_kms_helper_poll_fini() and > component_unbind_all() as appropriate. > ok > > regards > Philipp