| Message ID | 20220531072757.1.Ie7f6d4bf8cce28131da31a43354727e417cae98d@changeid (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | drm/msm/dpu: Move min BW request and full BW disable back to mdss | expand |
Hi Doug On 5/31/2022 7:28 AM, Douglas Anderson wrote: > In commit a670ff578f1f ("drm/msm/dpu: always use mdp device to scale > bandwidth") we fully moved interconnect stuff to the DPU driver. This > had no change for sc7180 but _did_ have an impact for other SoCs. It > made them match the sc7180 scheme. > > Unfortunately, the sc7180 scheme seems like it was a bit broken. > Specifically the interconnect needs to be on for more than just the > DPU driver's AXI bus. In the very least it also needs to be on for the > DSI driver's AXI bus. This can be seen fairly easily by doing this on > a ChromeOS sc7180-trogdor class device: > > set_power_policy --ac_screen_dim_delay=5 --ac_screen_off_delay=10 > sleep 10 > cd /sys/bus/platform/devices/ae94000.dsi/power > echo on > control > > When you do that, you'll get a warning splat in the logs about > "gcc_disp_hf_axi_clk status stuck at 'off'". > > One could argue that perhaps what I have done above is "illegal" and > that it can't happen naturally in the system because in normal system > usage the DPU is pretty much always on when DSI is on. That being > said: > * In official ChromeOS builds (admittedly a 5.4 kernel with backports) > we have seen that splat at bootup. > * Even though we don't use "autosuspend" for these components, we > don't use the "put_sync" variants. Thus plausibly the DSI could stay > "runtime enabled" past when the DPU is enabled. Techncially we > shouldn't do that if the DPU's suspend ends up yanking our clock. > > Let's change things such that the "bare minimum" request for the > interconnect happens in the mdss driver again. That means that all of > the children can assume that the interconnect is on at the minimum > bandwidth. We'll then let the DPU request the higher amount that it > wants. > > It should be noted that this isn't as hacky of a solution as it might > initially appear. Specifically: > * Since MDSS and DPU individually get their own references to the > interconnect then the framework will actually handle aggregating > them. The two drivers are _not_ clobbering each other. > * When the Qualcomm interconnect driver aggregates it takes the max of > all the peaks. Thus having MDSS request a peak, as we're doing here, > won't actually change the total interconnect bandwidth (it won't be > added to the request for the DPU). This perhaps explains why the > "average" requested in MDSS was historically 0 since that one > _would_ be added in. > > NOTE also that in the downstream ChromeOS 5.4 and 5.15 kernels, we're > also seeing some RPMH hangs that are addressed by this fix. These > hangs are showing up in the field and on _some_ devices with enough > stress testing of suspend/resume. Specifically right at suspend time > with a stack crawl that looks like this (from chromeos-5.15 tree): > rpmh_write_batch+0x19c/0x240 > qcom_icc_bcm_voter_commit+0x210/0x420 > qcom_icc_set+0x28/0x38 > apply_constraints+0x70/0xa4 > icc_set_bw+0x150/0x24c > dpu_runtime_resume+0x50/0x1c4 > pm_generic_runtime_resume+0x30/0x44 > __genpd_runtime_resume+0x68/0x7c > genpd_runtime_resume+0x12c/0x20c > __rpm_callback+0x98/0x138 > rpm_callback+0x30/0x88 > rpm_resume+0x370/0x4a0 > __pm_runtime_resume+0x80/0xb0 > dpu_kms_enable_commit+0x24/0x30 > msm_atomic_commit_tail+0x12c/0x630 > commit_tail+0xac/0x150 > drm_atomic_helper_commit+0x114/0x11c > drm_atomic_commit+0x68/0x78 > drm_atomic_helper_disable_all+0x158/0x1c8 > drm_atomic_helper_suspend+0xc0/0x1c0 > drm_mode_config_helper_suspend+0x2c/0x60 > msm_pm_prepare+0x2c/0x40 > pm_generic_prepare+0x30/0x44 > genpd_prepare+0x80/0xd0 > device_prepare+0x78/0x17c > dpm_prepare+0xb0/0x384 > dpm_suspend_start+0x34/0xc0 > > We don't completely understand all the mechanisms in play, but the > hang seemed to come and go with random factors. It's not terribly > surprising that the hang is gone after this patch since the line of > code that was failing is no longer present in the kernel. > > Fixes: a670ff578f1f ("drm/msm/dpu: always use mdp device to scale bandwidth") > Fixes: c33b7c0389e1 ("drm/msm/dpu: add support for clk and bw scaling for display") > Signed-off-by: Douglas Anderson <dianders@chromium.org> > --- > > drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c | 8 ---- > drivers/gpu/drm/msm/msm_mdss.c | 58 +++++++++++++++++++++++++ > 2 files changed, 58 insertions(+), 8 deletions(-) > > diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c > index 2b9d931474e0..3025184053e0 100644 > --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c > +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c > @@ -49,8 +49,6 @@ > #define DPU_DEBUGFS_DIR "msm_dpu" > #define DPU_DEBUGFS_HWMASKNAME "hw_log_mask" > > -#define MIN_IB_BW 400000000ULL /* Min ib vote 400MB */ > - > static int dpu_kms_hw_init(struct msm_kms *kms); > static void _dpu_kms_mmu_destroy(struct dpu_kms *dpu_kms); > > @@ -1303,15 +1301,9 @@ static int __maybe_unused dpu_runtime_resume(struct device *dev) > struct dpu_kms *dpu_kms = to_dpu_kms(priv->kms); > struct drm_encoder *encoder; > struct drm_device *ddev; > - int i; > > ddev = dpu_kms->dev; > > - WARN_ON(!(dpu_kms->num_paths)); > - /* Min vote of BW is required before turning on AXI clk */ > - for (i = 0; i < dpu_kms->num_paths; i++) > - icc_set_bw(dpu_kms->path[i], 0, Bps_to_icc(MIN_IB_BW)); > - > rc = clk_bulk_prepare_enable(dpu_kms->num_clocks, dpu_kms->clocks); > if (rc) { > DPU_ERROR("clock enable failed rc:%d\n", rc); > diff --git a/drivers/gpu/drm/msm/msm_mdss.c b/drivers/gpu/drm/msm/msm_mdss.c > index 0454a571adf7..fe2c398d99b7 100644 > --- a/drivers/gpu/drm/msm/msm_mdss.c > +++ b/drivers/gpu/drm/msm/msm_mdss.c > @@ -5,6 +5,7 @@ > > #include <linux/clk.h> > #include <linux/delay.h> > +#include <linux/interconnect.h> > #include <linux/irq.h> > #include <linux/irqchip.h> > #include <linux/irqdesc.h> > @@ -25,6 +26,8 @@ > #define UBWC_CTRL_2 0x150 > #define UBWC_PREDICTION_MODE 0x154 > > +#define MIN_IB_BW 400000000UL /* Min ib vote 400MB */ > + > struct msm_mdss { > struct device *dev; > > @@ -36,8 +39,47 @@ struct msm_mdss { > unsigned long enabled_mask; > struct irq_domain *domain; > } irq_controller; > + struct icc_path *path[2]; > + u32 num_paths; > }; > > +static int msm_mdss_parse_data_bus_icc_path(struct device *dev, > + struct msm_mdss *msm_mdss) > +{ > + struct icc_path *path0 = of_icc_get(dev, "mdp0-mem"); > + struct icc_path *path1 = of_icc_get(dev, "mdp1-mem"); > + > + if (IS_ERR_OR_NULL(path0)) > + return PTR_ERR_OR_ZERO(path0); > + > + msm_mdss->path[0] = path0; > + msm_mdss->num_paths = 1; > + > + if (!IS_ERR_OR_NULL(path1)) { > + msm_mdss->path[1] = path1; > + msm_mdss->num_paths++; > + } > + > + return 0; > +} > + > +static void msm_mdss_put_icc_path(void *data) > +{ > + struct msm_mdss *msm_mdss = data; > + int i; > + > + for (i = 0; i < msm_mdss->num_paths; i++) > + icc_put(msm_mdss->path[i]); > +} > + > +static void msm_mdss_icc_request_bw(struct msm_mdss *msm_mdss, unsigned long bw) > +{ > + int i; > + > + for (i = 0; i < msm_mdss->num_paths; i++) > + icc_set_bw(msm_mdss->path[i], 0, Bps_to_icc(bw)); > +} > + > static void msm_mdss_irq(struct irq_desc *desc) > { > struct msm_mdss *msm_mdss = irq_desc_get_handler_data(desc); > @@ -136,6 +178,13 @@ static int msm_mdss_enable(struct msm_mdss *msm_mdss) > { > int ret; > > + /* > + * Several components have AXI clocks that can only be turned on if > + * the interconnect is enabled (non-zero bandwidth). Let's make sure > + * that the interconnects are at least at a minimum amount. > + */ > + msm_mdss_icc_request_bw(msm_mdss, MIN_IB_BW); > + This patch does two things : 1) Move the min icc vote from the dpu_runtime_resume() path to the msm_mdss_enable() so that while resuming, we add the min vote to the parent device. We do need a min vote before turning on the AXI clk as per this comment mentioned in c33b7c0389e1 (drm/msm/dpu: add support for clk and bw scaling for display) @@ -1101,8 +1129,15 @@ static int __maybe_unused dpu_runtime_resume(struct device *dev) struct drm_encoder *encoder; struct drm_device *ddev; struct dss_module_power *mp = &dpu_kms->mp; + int i; ddev = dpu_kms->dev; + + /* Min vote of BW is required before turning on AXI clk */ + for (i = 0; i < dpu_kms->num_paths; i++) + icc_set_bw(dpu_kms->path[i], 0, + dpu_kms->catalog->perf.min_dram_ib); So I understand and I am fine with this part. 2) Add a min vote in msm_mdss_init(). This is the part I am not able to fully follow. If we only need to add the min vote before voting for the clocks, adding it in mdss_mdss_enable() should be enough. Do we need this part of adding the min vote to the msm_mdss_init()? If so, why? > ret = clk_bulk_prepare_enable(msm_mdss->num_clocks, msm_mdss->clocks); > if (ret) { > dev_err(msm_mdss->dev, "clock enable failed, ret:%d\n", ret); > @@ -178,6 +227,7 @@ static int msm_mdss_enable(struct msm_mdss *msm_mdss) > static int msm_mdss_disable(struct msm_mdss *msm_mdss) > { > clk_bulk_disable_unprepare(msm_mdss->num_clocks, msm_mdss->clocks); > + msm_mdss_icc_request_bw(msm_mdss, 0); > > return 0; > } > @@ -271,6 +321,14 @@ static struct msm_mdss *msm_mdss_init(struct platform_device *pdev, bool is_mdp5 > > dev_dbg(&pdev->dev, "mapped mdss address space @%pK\n", msm_mdss->mmio); > > + ret = msm_mdss_parse_data_bus_icc_path(&pdev->dev, msm_mdss); > + if (ret) > + return ERR_PTR(ret); > + ret = devm_add_action_or_reset(&pdev->dev, msm_mdss_put_icc_path, msm_mdss); > + if (ret) > + return ERR_PTR(ret); > + msm_mdss_icc_request_bw(msm_mdss, MIN_IB_BW); > + > if (is_mdp5) > ret = mdp5_mdss_parse_clock(pdev, &msm_mdss->clocks); > else
Hi, On Tue, May 31, 2022 at 2:29 PM Abhinav Kumar <quic_abhinavk@quicinc.com> wrote: > > > @@ -136,6 +178,13 @@ static int msm_mdss_enable(struct msm_mdss *msm_mdss) > > { > > int ret; > > > > + /* > > + * Several components have AXI clocks that can only be turned on if > > + * the interconnect is enabled (non-zero bandwidth). Let's make sure > > + * that the interconnects are at least at a minimum amount. > > + */ > > + msm_mdss_icc_request_bw(msm_mdss, MIN_IB_BW); > > + > > This patch does two things : > > 1) Move the min icc vote from the dpu_runtime_resume() path to the > msm_mdss_enable() so that while resuming, we add the min vote to the > parent device. We do need a min vote before turning on the AXI clk as > per this comment mentioned in c33b7c0389e1 (drm/msm/dpu: add support for > clk and bw scaling for display) > > > @@ -1101,8 +1129,15 @@ static int __maybe_unused > dpu_runtime_resume(struct device *dev) > struct drm_encoder *encoder; > struct drm_device *ddev; > struct dss_module_power *mp = &dpu_kms->mp; > + int i; > > ddev = dpu_kms->dev; > + > + /* Min vote of BW is required before turning on AXI clk */ > + for (i = 0; i < dpu_kms->num_paths; i++) > + icc_set_bw(dpu_kms->path[i], 0, > + dpu_kms->catalog->perf.min_dram_ib); > > So I understand and I am fine with this part. > > 2) Add a min vote in msm_mdss_init(). > > This is the part I am not able to fully follow. > > If we only need to add the min vote before voting for the clocks, adding > it in mdss_mdss_enable() should be enough. > > Do we need this part of adding the min vote to the msm_mdss_init()? > > If so, why? Ah, good question. Mostly I was matching how things were done before commit a670ff578f1f ("drm/msm/dpu: always use mdp device to scale bandwidth"). Before that commit we used to put the min vote in the init path. ...but you're right, I don't see any good reason for it. I'll send a v2 removing that line from msm_mdss_init(). -Doug
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c index 2b9d931474e0..3025184053e0 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c @@ -49,8 +49,6 @@ #define DPU_DEBUGFS_DIR "msm_dpu" #define DPU_DEBUGFS_HWMASKNAME "hw_log_mask" -#define MIN_IB_BW 400000000ULL /* Min ib vote 400MB */ - static int dpu_kms_hw_init(struct msm_kms *kms); static void _dpu_kms_mmu_destroy(struct dpu_kms *dpu_kms); @@ -1303,15 +1301,9 @@ static int __maybe_unused dpu_runtime_resume(struct device *dev) struct dpu_kms *dpu_kms = to_dpu_kms(priv->kms); struct drm_encoder *encoder; struct drm_device *ddev; - int i; ddev = dpu_kms->dev; - WARN_ON(!(dpu_kms->num_paths)); - /* Min vote of BW is required before turning on AXI clk */ - for (i = 0; i < dpu_kms->num_paths; i++) - icc_set_bw(dpu_kms->path[i], 0, Bps_to_icc(MIN_IB_BW)); - rc = clk_bulk_prepare_enable(dpu_kms->num_clocks, dpu_kms->clocks); if (rc) { DPU_ERROR("clock enable failed rc:%d\n", rc); diff --git a/drivers/gpu/drm/msm/msm_mdss.c b/drivers/gpu/drm/msm/msm_mdss.c index 0454a571adf7..fe2c398d99b7 100644 --- a/drivers/gpu/drm/msm/msm_mdss.c +++ b/drivers/gpu/drm/msm/msm_mdss.c @@ -5,6 +5,7 @@ #include <linux/clk.h> #include <linux/delay.h> +#include <linux/interconnect.h> #include <linux/irq.h> #include <linux/irqchip.h> #include <linux/irqdesc.h> @@ -25,6 +26,8 @@ #define UBWC_CTRL_2 0x150 #define UBWC_PREDICTION_MODE 0x154 +#define MIN_IB_BW 400000000UL /* Min ib vote 400MB */ + struct msm_mdss { struct device *dev; @@ -36,8 +39,47 @@ struct msm_mdss { unsigned long enabled_mask; struct irq_domain *domain; } irq_controller; + struct icc_path *path[2]; + u32 num_paths; }; +static int msm_mdss_parse_data_bus_icc_path(struct device *dev, + struct msm_mdss *msm_mdss) +{ + struct icc_path *path0 = of_icc_get(dev, "mdp0-mem"); + struct icc_path *path1 = of_icc_get(dev, "mdp1-mem"); + + if (IS_ERR_OR_NULL(path0)) + return PTR_ERR_OR_ZERO(path0); + + msm_mdss->path[0] = path0; + msm_mdss->num_paths = 1; + + if (!IS_ERR_OR_NULL(path1)) { + msm_mdss->path[1] = path1; + msm_mdss->num_paths++; + } + + return 0; +} + +static void msm_mdss_put_icc_path(void *data) +{ + struct msm_mdss *msm_mdss = data; + int i; + + for (i = 0; i < msm_mdss->num_paths; i++) + icc_put(msm_mdss->path[i]); +} + +static void msm_mdss_icc_request_bw(struct msm_mdss *msm_mdss, unsigned long bw) +{ + int i; + + for (i = 0; i < msm_mdss->num_paths; i++) + icc_set_bw(msm_mdss->path[i], 0, Bps_to_icc(bw)); +} + static void msm_mdss_irq(struct irq_desc *desc) { struct msm_mdss *msm_mdss = irq_desc_get_handler_data(desc); @@ -136,6 +178,13 @@ static int msm_mdss_enable(struct msm_mdss *msm_mdss) { int ret; + /* + * Several components have AXI clocks that can only be turned on if + * the interconnect is enabled (non-zero bandwidth). Let's make sure + * that the interconnects are at least at a minimum amount. + */ + msm_mdss_icc_request_bw(msm_mdss, MIN_IB_BW); + ret = clk_bulk_prepare_enable(msm_mdss->num_clocks, msm_mdss->clocks); if (ret) { dev_err(msm_mdss->dev, "clock enable failed, ret:%d\n", ret); @@ -178,6 +227,7 @@ static int msm_mdss_enable(struct msm_mdss *msm_mdss) static int msm_mdss_disable(struct msm_mdss *msm_mdss) { clk_bulk_disable_unprepare(msm_mdss->num_clocks, msm_mdss->clocks); + msm_mdss_icc_request_bw(msm_mdss, 0); return 0; } @@ -271,6 +321,14 @@ static struct msm_mdss *msm_mdss_init(struct platform_device *pdev, bool is_mdp5 dev_dbg(&pdev->dev, "mapped mdss address space @%pK\n", msm_mdss->mmio); + ret = msm_mdss_parse_data_bus_icc_path(&pdev->dev, msm_mdss); + if (ret) + return ERR_PTR(ret); + ret = devm_add_action_or_reset(&pdev->dev, msm_mdss_put_icc_path, msm_mdss); + if (ret) + return ERR_PTR(ret); + msm_mdss_icc_request_bw(msm_mdss, MIN_IB_BW); + if (is_mdp5) ret = mdp5_mdss_parse_clock(pdev, &msm_mdss->clocks); else
In commit a670ff578f1f ("drm/msm/dpu: always use mdp device to scale bandwidth") we fully moved interconnect stuff to the DPU driver. This had no change for sc7180 but _did_ have an impact for other SoCs. It made them match the sc7180 scheme. Unfortunately, the sc7180 scheme seems like it was a bit broken. Specifically the interconnect needs to be on for more than just the DPU driver's AXI bus. In the very least it also needs to be on for the DSI driver's AXI bus. This can be seen fairly easily by doing this on a ChromeOS sc7180-trogdor class device: set_power_policy --ac_screen_dim_delay=5 --ac_screen_off_delay=10 sleep 10 cd /sys/bus/platform/devices/ae94000.dsi/power echo on > control When you do that, you'll get a warning splat in the logs about "gcc_disp_hf_axi_clk status stuck at 'off'". One could argue that perhaps what I have done above is "illegal" and that it can't happen naturally in the system because in normal system usage the DPU is pretty much always on when DSI is on. That being said: * In official ChromeOS builds (admittedly a 5.4 kernel with backports) we have seen that splat at bootup. * Even though we don't use "autosuspend" for these components, we don't use the "put_sync" variants. Thus plausibly the DSI could stay "runtime enabled" past when the DPU is enabled. Techncially we shouldn't do that if the DPU's suspend ends up yanking our clock. Let's change things such that the "bare minimum" request for the interconnect happens in the mdss driver again. That means that all of the children can assume that the interconnect is on at the minimum bandwidth. We'll then let the DPU request the higher amount that it wants. It should be noted that this isn't as hacky of a solution as it might initially appear. Specifically: * Since MDSS and DPU individually get their own references to the interconnect then the framework will actually handle aggregating them. The two drivers are _not_ clobbering each other. * When the Qualcomm interconnect driver aggregates it takes the max of all the peaks. Thus having MDSS request a peak, as we're doing here, won't actually change the total interconnect bandwidth (it won't be added to the request for the DPU). This perhaps explains why the "average" requested in MDSS was historically 0 since that one _would_ be added in. NOTE also that in the downstream ChromeOS 5.4 and 5.15 kernels, we're also seeing some RPMH hangs that are addressed by this fix. These hangs are showing up in the field and on _some_ devices with enough stress testing of suspend/resume. Specifically right at suspend time with a stack crawl that looks like this (from chromeos-5.15 tree): rpmh_write_batch+0x19c/0x240 qcom_icc_bcm_voter_commit+0x210/0x420 qcom_icc_set+0x28/0x38 apply_constraints+0x70/0xa4 icc_set_bw+0x150/0x24c dpu_runtime_resume+0x50/0x1c4 pm_generic_runtime_resume+0x30/0x44 __genpd_runtime_resume+0x68/0x7c genpd_runtime_resume+0x12c/0x20c __rpm_callback+0x98/0x138 rpm_callback+0x30/0x88 rpm_resume+0x370/0x4a0 __pm_runtime_resume+0x80/0xb0 dpu_kms_enable_commit+0x24/0x30 msm_atomic_commit_tail+0x12c/0x630 commit_tail+0xac/0x150 drm_atomic_helper_commit+0x114/0x11c drm_atomic_commit+0x68/0x78 drm_atomic_helper_disable_all+0x158/0x1c8 drm_atomic_helper_suspend+0xc0/0x1c0 drm_mode_config_helper_suspend+0x2c/0x60 msm_pm_prepare+0x2c/0x40 pm_generic_prepare+0x30/0x44 genpd_prepare+0x80/0xd0 device_prepare+0x78/0x17c dpm_prepare+0xb0/0x384 dpm_suspend_start+0x34/0xc0 We don't completely understand all the mechanisms in play, but the hang seemed to come and go with random factors. It's not terribly surprising that the hang is gone after this patch since the line of code that was failing is no longer present in the kernel. Fixes: a670ff578f1f ("drm/msm/dpu: always use mdp device to scale bandwidth") Fixes: c33b7c0389e1 ("drm/msm/dpu: add support for clk and bw scaling for display") Signed-off-by: Douglas Anderson <dianders@chromium.org> --- drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c | 8 ---- drivers/gpu/drm/msm/msm_mdss.c | 58 +++++++++++++++++++++++++ 2 files changed, 58 insertions(+), 8 deletions(-)