| Message ID | 20191227173707.20413-2-martin.blumenstingl@googlemail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm: lima: devfreq and cooling device support | expand |
Hi Martin, On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: > Most platforms with a Mali-400 or Mali-450 GPU also have support for > changing the GPU clock frequency. Add devfreq support so the GPU clock > rate is updated based on the actual GPU usage when the > "operating-points-v2" property is present in the board.dts. > > The actual devfreq code is taken from panfrost_devfreq.c and modified so > it matches what the lima hardware needs: > - a call to dev_pm_opp_set_clkname() during initialization because there > are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > the GPU so we need to control it using devfreq. > - locking when reading or writing the devfreq statistics because (unlike > than panfrost) we have multiple PP and GP IRQs which may finish jobs > concurrently. I gave this a quick try on my RK3328, and the clock scaling indeed kicks in nicely on the glmark2 scenes that struggle, however something appears to be missing in terms of regulator association, as the appropriate OPP voltages aren't reflected in the GPU supply (fortunately the initial voltage seems close enough to that of the highest OPP not to cause major problems, on my box at least). With panfrost on RK3399 I do see the supply voltage scaling accordingly, but I don't know my way around devfreq well enough to know what matters in the difference :/ Thanks, Robin. > Signed-off-by: Martin Blumenstingl <martin.blumenstingl@googlemail.com> > --- > drivers/gpu/drm/lima/Kconfig | 1 + > drivers/gpu/drm/lima/Makefile | 3 +- > drivers/gpu/drm/lima/lima_devfreq.c | 183 ++++++++++++++++++++++++++++ > drivers/gpu/drm/lima/lima_devfreq.h | 15 +++ > drivers/gpu/drm/lima/lima_device.c | 4 + > drivers/gpu/drm/lima/lima_device.h | 17 +++ > drivers/gpu/drm/lima/lima_drv.c | 14 ++- > drivers/gpu/drm/lima/lima_sched.c | 7 ++ > drivers/gpu/drm/lima/lima_sched.h | 3 + > 9 files changed, 244 insertions(+), 3 deletions(-) > create mode 100644 drivers/gpu/drm/lima/lima_devfreq.c > create mode 100644 drivers/gpu/drm/lima/lima_devfreq.h > > diff --git a/drivers/gpu/drm/lima/Kconfig b/drivers/gpu/drm/lima/Kconfig > index 571dc369a7e9..cdd24b68b5d4 100644 > --- a/drivers/gpu/drm/lima/Kconfig > +++ b/drivers/gpu/drm/lima/Kconfig > @@ -10,5 +10,6 @@ config DRM_LIMA > depends on OF > select DRM_SCHED > select DRM_GEM_SHMEM_HELPER > + select PM_DEVFREQ > help > DRM driver for ARM Mali 400/450 GPUs. > diff --git a/drivers/gpu/drm/lima/Makefile b/drivers/gpu/drm/lima/Makefile > index a85444b0a1d4..5e5c29875e9c 100644 > --- a/drivers/gpu/drm/lima/Makefile > +++ b/drivers/gpu/drm/lima/Makefile > @@ -14,6 +14,7 @@ lima-y := \ > lima_sched.o \ > lima_ctx.o \ > lima_dlbu.o \ > - lima_bcast.o > + lima_bcast.o \ > + lima_devfreq.o > > obj-$(CONFIG_DRM_LIMA) += lima.o > diff --git a/drivers/gpu/drm/lima/lima_devfreq.c b/drivers/gpu/drm/lima/lima_devfreq.c > new file mode 100644 > index 000000000000..a5fd6b8faa77 > --- /dev/null > +++ b/drivers/gpu/drm/lima/lima_devfreq.c > @@ -0,0 +1,183 @@ > +// SPDX-License-Identifier: GPL-2.0 > +/* > + * Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> > + * > + * Based on panfrost_devfreq.c: > + * Copyright 2019 Collabora ltd. > + */ > +#include <linux/clk.h> > +#include <linux/devfreq.h> > +#include <linux/devfreq_cooling.h> > +#include <linux/device.h> > +#include <linux/platform_device.h> > +#include <linux/pm_opp.h> > + > +#include "lima_device.h" > +#include "lima_devfreq.h" > + > +static void lima_devfreq_update_utilization(struct lima_device *ldev) > +{ > + unsigned long irqflags; > + ktime_t now, last; > + > + if (!ldev->devfreq.devfreq) > + return; > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + now = ktime_get(); > + last = ldev->devfreq.time_last_update; > + > + if (atomic_read(&ldev->devfreq.busy_count) > 0) > + ldev->devfreq.busy_time += ktime_sub(now, last); > + else > + ldev->devfreq.idle_time += ktime_sub(now, last); > + > + ldev->devfreq.time_last_update = now; > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > +} > + > +static int lima_devfreq_target(struct device *dev, unsigned long *freq, > + u32 flags) > +{ > + struct dev_pm_opp *opp; > + int err; > + > + opp = devfreq_recommended_opp(dev, freq, flags); > + if (IS_ERR(opp)) > + return PTR_ERR(opp); > + dev_pm_opp_put(opp); > + > + err = dev_pm_opp_set_rate(dev, *freq); > + if (err) > + return err; > + > + return 0; > +} > + > +static void lima_devfreq_reset(struct lima_device *ldev) > +{ > + unsigned long irqflags; > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + ldev->devfreq.busy_time = 0; > + ldev->devfreq.idle_time = 0; > + ldev->devfreq.time_last_update = ktime_get(); > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > +} > + > +static int lima_devfreq_get_dev_status(struct device *dev, > + struct devfreq_dev_status *status) > +{ > + struct lima_device *ldev = dev_get_drvdata(dev); > + unsigned long irqflags; > + > + lima_devfreq_update_utilization(ldev); > + > + status->current_frequency = clk_get_rate(ldev->clk_gpu); > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + status->total_time = ktime_to_ns(ktime_add(ldev->devfreq.busy_time, > + ldev->devfreq.idle_time)); > + status->busy_time = ktime_to_ns(ldev->devfreq.busy_time); > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > + > + lima_devfreq_reset(ldev); > + > + dev_dbg(ldev->dev, "busy %lu total %lu %lu %% freq %lu MHz\n", > + status->busy_time, status->total_time, > + status->busy_time / (status->total_time / 100), > + status->current_frequency / 1000 / 1000); > + > + return 0; > +} > + > +static struct devfreq_dev_profile lima_devfreq_profile = { > + .polling_ms = 50, /* ~3 frames */ > + .target = lima_devfreq_target, > + .get_dev_status = lima_devfreq_get_dev_status, > +}; > + > +int lima_devfreq_init(struct lima_device *ldev) > +{ > + struct thermal_cooling_device *cooling; > + struct device *dev = &ldev->pdev->dev; > + struct devfreq *devfreq; > + struct dev_pm_opp *opp; > + unsigned long cur_freq; > + int ret; > + > + spin_lock_init(&ldev->devfreq.lock); > + > + ldev->devfreq.opp_table = dev_pm_opp_set_clkname(dev, "core"); > + if (IS_ERR(ldev->devfreq.opp_table)) > + return PTR_ERR(ldev->devfreq.opp_table); > + > + ret = dev_pm_opp_of_add_table(dev); > + if (ret == -ENODEV) /* Optional, continue without devfreq */ > + return 0; > + else if (ret) > + return ret; > + > + lima_devfreq_reset(ldev); > + > + cur_freq = clk_get_rate(ldev->clk_gpu); > + > + opp = devfreq_recommended_opp(dev, &cur_freq, 0); > + if (IS_ERR(opp)) > + return PTR_ERR(opp); > + > + lima_devfreq_profile.initial_freq = cur_freq; > + dev_pm_opp_put(opp); > + > + devfreq = devm_devfreq_add_device(dev, &lima_devfreq_profile, > + DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL); > + if (IS_ERR(devfreq)) { > + dev_err(dev, "Couldn't initialize GPU devfreq\n"); > + dev_pm_opp_of_remove_table(dev); > + return PTR_ERR(devfreq); > + } > + > + ldev->devfreq.devfreq = devfreq; > + > + cooling = of_devfreq_cooling_register(dev->of_node, devfreq); > + if (IS_ERR(cooling)) > + dev_info(dev, "Failed to register cooling device\n"); > + else > + ldev->devfreq.cooling = cooling; > + > + return 0; > +} > + > +void lima_devfreq_fini(struct lima_device *ldev) > +{ > + if (ldev->devfreq.cooling) > + devfreq_cooling_unregister(ldev->devfreq.cooling); > + > + if (ldev->devfreq.opp_table) { > + dev_pm_opp_put_clkname(ldev->devfreq.opp_table); > + ldev->devfreq.opp_table = NULL; > + } > + > + dev_pm_opp_of_remove_table(&ldev->pdev->dev); > +} > + > +void lima_devfreq_record_busy(struct lima_device *ldev) > +{ > + lima_devfreq_update_utilization(ldev); > + atomic_inc(&ldev->devfreq.busy_count); > +} > + > +void lima_devfreq_record_idle(struct lima_device *ldev) > +{ > + int count; > + > + lima_devfreq_update_utilization(ldev); > + count = atomic_dec_if_positive(&ldev->devfreq.busy_count); > + WARN_ON(count < 0); > +} > diff --git a/drivers/gpu/drm/lima/lima_devfreq.h b/drivers/gpu/drm/lima/lima_devfreq.h > new file mode 100644 > index 000000000000..fe4f8a437033 > --- /dev/null > +++ b/drivers/gpu/drm/lima/lima_devfreq.h > @@ -0,0 +1,15 @@ > +/* SPDX-License-Identifier: GPL-2.0 */ > +/* Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> */ > + > +#ifndef __LIMA_DEVFREQ_H__ > +#define __LIMA_DEVFREQ_H__ > + > +struct lima_device; > + > +int lima_devfreq_init(struct lima_device *ldev); > +void lima_devfreq_fini(struct lima_device *ldev); > + > +void lima_devfreq_record_busy(struct lima_device *ldev); > +void lima_devfreq_record_idle(struct lima_device *ldev); > + > +#endif > diff --git a/drivers/gpu/drm/lima/lima_device.c b/drivers/gpu/drm/lima/lima_device.c > index 19829b543024..7f1f7a1c03e5 100644 > --- a/drivers/gpu/drm/lima/lima_device.c > +++ b/drivers/gpu/drm/lima/lima_device.c > @@ -214,6 +214,8 @@ static int lima_init_gp_pipe(struct lima_device *dev) > struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_gp; > int err; > > + pipe->ldev = dev; > + > err = lima_sched_pipe_init(pipe, "gp"); > if (err) > return err; > @@ -244,6 +246,8 @@ static int lima_init_pp_pipe(struct lima_device *dev) > struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_pp; > int err, i; > > + pipe->ldev = dev; > + > err = lima_sched_pipe_init(pipe, "pp"); > if (err) > return err; > diff --git a/drivers/gpu/drm/lima/lima_device.h b/drivers/gpu/drm/lima/lima_device.h > index 31158d86271c..26f0efdd17f1 100644 > --- a/drivers/gpu/drm/lima/lima_device.h > +++ b/drivers/gpu/drm/lima/lima_device.h > @@ -5,6 +5,7 @@ > #define __LIMA_DEVICE_H__ > > #include <drm/drm_device.h> > +#include <linux/atomic.h> > #include <linux/delay.h> > > #include "lima_sched.h" > @@ -94,6 +95,22 @@ struct lima_device { > > u32 *dlbu_cpu; > dma_addr_t dlbu_dma; > + > + struct { > + struct devfreq *devfreq; > + struct opp_table *opp_table; > + struct thermal_cooling_device *cooling; > + ktime_t busy_time; > + ktime_t idle_time; > + ktime_t time_last_update; > + atomic_t busy_count; > + /* > + * Protect busy_time, idle_time and time_last_update because > + * these can be updated concurrently - for example by the GP > + * and PP interrupts. > + */ > + spinlock_t lock; > + } devfreq; > }; > > static inline struct lima_device * > diff --git a/drivers/gpu/drm/lima/lima_drv.c b/drivers/gpu/drm/lima/lima_drv.c > index 124efe4fa97b..b64b1777f220 100644 > --- a/drivers/gpu/drm/lima/lima_drv.c > +++ b/drivers/gpu/drm/lima/lima_drv.c > @@ -10,6 +10,7 @@ > #include <drm/drm_prime.h> > #include <drm/lima_drm.h> > > +#include "lima_devfreq.h" > #include "lima_drv.h" > #include "lima_gem.h" > #include "lima_vm.h" > @@ -296,18 +297,26 @@ static int lima_pdev_probe(struct platform_device *pdev) > if (err) > goto err_out1; > > + err = lima_devfreq_init(ldev); > + if (err) { > + dev_err(&pdev->dev, "Fatal error during devfreq init\n"); > + goto err_out2; > + } > + > /* > * Register the DRM device with the core and the connectors with > * sysfs. > */ > err = drm_dev_register(ddev, 0); > if (err < 0) > - goto err_out2; > + goto err_out3; > > return 0; > > -err_out2: > +err_out3: > lima_device_fini(ldev); > +err_out2: > + lima_devfreq_fini(ldev); > err_out1: > drm_dev_put(ddev); > err_out0: > @@ -321,6 +330,7 @@ static int lima_pdev_remove(struct platform_device *pdev) > struct drm_device *ddev = ldev->ddev; > > drm_dev_unregister(ddev); > + lima_devfreq_fini(ldev); > lima_device_fini(ldev); > drm_dev_put(ddev); > lima_sched_slab_fini(); > diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c > index f522c5f99729..851c496a168b 100644 > --- a/drivers/gpu/drm/lima/lima_sched.c > +++ b/drivers/gpu/drm/lima/lima_sched.c > @@ -5,6 +5,7 @@ > #include <linux/slab.h> > #include <linux/xarray.h> > > +#include "lima_devfreq.h" > #include "lima_drv.h" > #include "lima_sched.h" > #include "lima_vm.h" > @@ -213,6 +214,8 @@ static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job) > */ > ret = dma_fence_get(task->fence); > > + lima_devfreq_record_busy(pipe->ldev); > + > pipe->current_task = task; > > /* this is needed for MMU to work correctly, otherwise GP/PP > @@ -280,6 +283,8 @@ static void lima_sched_handle_error_task(struct lima_sched_pipe *pipe, > pipe->current_vm = NULL; > pipe->current_task = NULL; > > + lima_devfreq_record_idle(pipe->ldev); > + > drm_sched_resubmit_jobs(&pipe->base); > drm_sched_start(&pipe->base, true); > } > @@ -348,6 +353,8 @@ void lima_sched_pipe_fini(struct lima_sched_pipe *pipe) > > void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe) > { > + lima_devfreq_record_idle(pipe->ldev); > + > if (pipe->error) > schedule_work(&pipe->error_work); > else { > diff --git a/drivers/gpu/drm/lima/lima_sched.h b/drivers/gpu/drm/lima/lima_sched.h > index 928af91c1118..9ae7df7d7fbb 100644 > --- a/drivers/gpu/drm/lima/lima_sched.h > +++ b/drivers/gpu/drm/lima/lima_sched.h > @@ -6,6 +6,7 @@ > > #include <drm/gpu_scheduler.h> > > +struct lima_device; > struct lima_vm; > > struct lima_sched_task { > @@ -41,6 +42,8 @@ struct lima_sched_pipe { > u32 fence_seqno; > spinlock_t fence_lock; > > + struct lima_device *ldev; > + > struct lima_sched_task *current_task; > struct lima_vm *current_vm; > >
Hi Robin, On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: > > Hi Martin, > > On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: > > Most platforms with a Mali-400 or Mali-450 GPU also have support for > > changing the GPU clock frequency. Add devfreq support so the GPU clock > > rate is updated based on the actual GPU usage when the > > "operating-points-v2" property is present in the board.dts. > > > > The actual devfreq code is taken from panfrost_devfreq.c and modified so > > it matches what the lima hardware needs: > > - a call to dev_pm_opp_set_clkname() during initialization because there > > are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > > the GPU so we need to control it using devfreq. > > - locking when reading or writing the devfreq statistics because (unlike > > than panfrost) we have multiple PP and GP IRQs which may finish jobs > > concurrently. > > I gave this a quick try on my RK3328, and the clock scaling indeed kicks > in nicely on the glmark2 scenes that struggle, however something appears > to be missing in terms of regulator association, as the appropriate OPP > voltages aren't reflected in the GPU supply (fortunately the initial > voltage seems close enough to that of the highest OPP not to cause major > problems, on my box at least). With panfrost on RK3399 I do see the > supply voltage scaling accordingly, but I don't know my way around > devfreq well enough to know what matters in the difference :/ first of all: thank you for trying this out! :-) does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use generic code for devfreq") for your panfrost test? if I understand the devfreq API correct then I suspect with that commit panfrost also won't change the voltage anymore. this is probably due to a missing call to dev_pm_opp_set_regulators() which is supposed to attach the regulator to the devfreq instance. I didn't notice this yet because on Amlogic SoCs the voltage is the same for all OPPs. I'll debug this in the next days and send an updated patch (and drop the RFC prefix if there are no more comments). Regards Martin
On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: > Hi Robin, > > On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: >> >> Hi Martin, >> >> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: >>> Most platforms with a Mali-400 or Mali-450 GPU also have support for >>> changing the GPU clock frequency. Add devfreq support so the GPU clock >>> rate is updated based on the actual GPU usage when the >>> "operating-points-v2" property is present in the board.dts. >>> >>> The actual devfreq code is taken from panfrost_devfreq.c and modified so >>> it matches what the lima hardware needs: >>> - a call to dev_pm_opp_set_clkname() during initialization because there >>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks >>> the GPU so we need to control it using devfreq. >>> - locking when reading or writing the devfreq statistics because (unlike >>> than panfrost) we have multiple PP and GP IRQs which may finish jobs >>> concurrently. >> >> I gave this a quick try on my RK3328, and the clock scaling indeed kicks >> in nicely on the glmark2 scenes that struggle, however something appears >> to be missing in terms of regulator association, as the appropriate OPP >> voltages aren't reflected in the GPU supply (fortunately the initial >> voltage seems close enough to that of the highest OPP not to cause major >> problems, on my box at least). With panfrost on RK3399 I do see the >> supply voltage scaling accordingly, but I don't know my way around >> devfreq well enough to know what matters in the difference :/ > first of all: thank you for trying this out! :-) > > does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use > generic code for devfreq") for your panfrost test? > if I understand the devfreq API correct then I suspect with that > commit panfrost also won't change the voltage anymore. Oh, you're quite right - I was already considering that change as ancient history, but indeed it's only in 5.5-rc, while that board is still on 5.4.y release kernels. No wonder I couldn't make sense of how the (current) code could possibly be working :) I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is hopefully fixed), but I'm already fairly confident you've called it correctly. Cheers, Robin. > this is probably due to a missing call to dev_pm_opp_set_regulators() > which is supposed to attach the regulator to the devfreq instance. > I didn't notice this yet because on Amlogic SoCs the voltage is the > same for all OPPs. > > I'll debug this in the next days and send an updated patch (and drop > the RFC prefix if there are no more comments). > > > Regards > Martin >
On Sat, Dec 28, 2019 at 1:37 AM Martin Blumenstingl <martin.blumenstingl@googlemail.com> wrote: > > Most platforms with a Mali-400 or Mali-450 GPU also have support for > changing the GPU clock frequency. Add devfreq support so the GPU clock > rate is updated based on the actual GPU usage when the > "operating-points-v2" property is present in the board.dts. > > The actual devfreq code is taken from panfrost_devfreq.c and modified so > it matches what the lima hardware needs: > - a call to dev_pm_opp_set_clkname() during initialization because there > are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > the GPU so we need to control it using devfreq. > - locking when reading or writing the devfreq statistics because (unlike > than panfrost) we have multiple PP and GP IRQs which may finish jobs > concurrently. > > Signed-off-by: Martin Blumenstingl <martin.blumenstingl@googlemail.com> > --- > drivers/gpu/drm/lima/Kconfig | 1 + > drivers/gpu/drm/lima/Makefile | 3 +- > drivers/gpu/drm/lima/lima_devfreq.c | 183 ++++++++++++++++++++++++++++ > drivers/gpu/drm/lima/lima_devfreq.h | 15 +++ > drivers/gpu/drm/lima/lima_device.c | 4 + > drivers/gpu/drm/lima/lima_device.h | 17 +++ > drivers/gpu/drm/lima/lima_drv.c | 14 ++- > drivers/gpu/drm/lima/lima_sched.c | 7 ++ > drivers/gpu/drm/lima/lima_sched.h | 3 + > 9 files changed, 244 insertions(+), 3 deletions(-) > create mode 100644 drivers/gpu/drm/lima/lima_devfreq.c > create mode 100644 drivers/gpu/drm/lima/lima_devfreq.h > > diff --git a/drivers/gpu/drm/lima/Kconfig b/drivers/gpu/drm/lima/Kconfig > index 571dc369a7e9..cdd24b68b5d4 100644 > --- a/drivers/gpu/drm/lima/Kconfig > +++ b/drivers/gpu/drm/lima/Kconfig > @@ -10,5 +10,6 @@ config DRM_LIMA > depends on OF > select DRM_SCHED > select DRM_GEM_SHMEM_HELPER > + select PM_DEVFREQ > help > DRM driver for ARM Mali 400/450 GPUs. > diff --git a/drivers/gpu/drm/lima/Makefile b/drivers/gpu/drm/lima/Makefile > index a85444b0a1d4..5e5c29875e9c 100644 > --- a/drivers/gpu/drm/lima/Makefile > +++ b/drivers/gpu/drm/lima/Makefile > @@ -14,6 +14,7 @@ lima-y := \ > lima_sched.o \ > lima_ctx.o \ > lima_dlbu.o \ > - lima_bcast.o > + lima_bcast.o \ > + lima_devfreq.o > > obj-$(CONFIG_DRM_LIMA) += lima.o > diff --git a/drivers/gpu/drm/lima/lima_devfreq.c b/drivers/gpu/drm/lima/lima_devfreq.c > new file mode 100644 > index 000000000000..a5fd6b8faa77 > --- /dev/null > +++ b/drivers/gpu/drm/lima/lima_devfreq.c > @@ -0,0 +1,183 @@ > +// SPDX-License-Identifier: GPL-2.0 > +/* > + * Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> > + * > + * Based on panfrost_devfreq.c: > + * Copyright 2019 Collabora ltd. > + */ > +#include <linux/clk.h> > +#include <linux/devfreq.h> > +#include <linux/devfreq_cooling.h> > +#include <linux/device.h> > +#include <linux/platform_device.h> > +#include <linux/pm_opp.h> > + > +#include "lima_device.h" > +#include "lima_devfreq.h" > + > +static void lima_devfreq_update_utilization(struct lima_device *ldev) > +{ > + unsigned long irqflags; > + ktime_t now, last; > + > + if (!ldev->devfreq.devfreq) > + return; > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + now = ktime_get(); > + last = ldev->devfreq.time_last_update; > + > + if (atomic_read(&ldev->devfreq.busy_count) > 0) > + ldev->devfreq.busy_time += ktime_sub(now, last); > + else > + ldev->devfreq.idle_time += ktime_sub(now, last); > + > + ldev->devfreq.time_last_update = now; > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > +} > + > +static int lima_devfreq_target(struct device *dev, unsigned long *freq, > + u32 flags) > +{ > + struct dev_pm_opp *opp; > + int err; > + > + opp = devfreq_recommended_opp(dev, freq, flags); > + if (IS_ERR(opp)) > + return PTR_ERR(opp); > + dev_pm_opp_put(opp); > + > + err = dev_pm_opp_set_rate(dev, *freq); > + if (err) > + return err; > + > + return 0; > +} > + > +static void lima_devfreq_reset(struct lima_device *ldev) > +{ > + unsigned long irqflags; > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + ldev->devfreq.busy_time = 0; > + ldev->devfreq.idle_time = 0; > + ldev->devfreq.time_last_update = ktime_get(); > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > +} > + > +static int lima_devfreq_get_dev_status(struct device *dev, > + struct devfreq_dev_status *status) > +{ > + struct lima_device *ldev = dev_get_drvdata(dev); > + unsigned long irqflags; > + > + lima_devfreq_update_utilization(ldev); > + > + status->current_frequency = clk_get_rate(ldev->clk_gpu); > + > + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); > + > + status->total_time = ktime_to_ns(ktime_add(ldev->devfreq.busy_time, > + ldev->devfreq.idle_time)); > + status->busy_time = ktime_to_ns(ldev->devfreq.busy_time); > + > + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); > + > + lima_devfreq_reset(ldev); > + > + dev_dbg(ldev->dev, "busy %lu total %lu %lu %% freq %lu MHz\n", > + status->busy_time, status->total_time, > + status->busy_time / (status->total_time / 100), > + status->current_frequency / 1000 / 1000); > + > + return 0; > +} > + > +static struct devfreq_dev_profile lima_devfreq_profile = { > + .polling_ms = 50, /* ~3 frames */ > + .target = lima_devfreq_target, > + .get_dev_status = lima_devfreq_get_dev_status, > +}; > + > +int lima_devfreq_init(struct lima_device *ldev) > +{ > + struct thermal_cooling_device *cooling; > + struct device *dev = &ldev->pdev->dev; > + struct devfreq *devfreq; > + struct dev_pm_opp *opp; > + unsigned long cur_freq; > + int ret; > + > + spin_lock_init(&ldev->devfreq.lock); > + > + ldev->devfreq.opp_table = dev_pm_opp_set_clkname(dev, "core"); > + if (IS_ERR(ldev->devfreq.opp_table)) > + return PTR_ERR(ldev->devfreq.opp_table); > + > + ret = dev_pm_opp_of_add_table(dev); > + if (ret == -ENODEV) /* Optional, continue without devfreq */ > + return 0; > + else if (ret) > + return ret; > + > + lima_devfreq_reset(ldev); > + > + cur_freq = clk_get_rate(ldev->clk_gpu); > + > + opp = devfreq_recommended_opp(dev, &cur_freq, 0); > + if (IS_ERR(opp)) > + return PTR_ERR(opp); > + > + lima_devfreq_profile.initial_freq = cur_freq; > + dev_pm_opp_put(opp); > + > + devfreq = devm_devfreq_add_device(dev, &lima_devfreq_profile, > + DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL); > + if (IS_ERR(devfreq)) { > + dev_err(dev, "Couldn't initialize GPU devfreq\n"); > + dev_pm_opp_of_remove_table(dev); > + return PTR_ERR(devfreq); > + } > + > + ldev->devfreq.devfreq = devfreq; > + > + cooling = of_devfreq_cooling_register(dev->of_node, devfreq); > + if (IS_ERR(cooling)) > + dev_info(dev, "Failed to register cooling device\n"); > + else > + ldev->devfreq.cooling = cooling; > + > + return 0; > +} > + > +void lima_devfreq_fini(struct lima_device *ldev) > +{ > + if (ldev->devfreq.cooling) > + devfreq_cooling_unregister(ldev->devfreq.cooling); > + > + if (ldev->devfreq.opp_table) { > + dev_pm_opp_put_clkname(ldev->devfreq.opp_table); > + ldev->devfreq.opp_table = NULL; > + } > + > + dev_pm_opp_of_remove_table(&ldev->pdev->dev); > +} > + > +void lima_devfreq_record_busy(struct lima_device *ldev) > +{ > + lima_devfreq_update_utilization(ldev); > + atomic_inc(&ldev->devfreq.busy_count); > +} > + > +void lima_devfreq_record_idle(struct lima_device *ldev) > +{ > + int count; > + > + lima_devfreq_update_utilization(ldev); > + count = atomic_dec_if_positive(&ldev->devfreq.busy_count); > + WARN_ON(count < 0); > +} > diff --git a/drivers/gpu/drm/lima/lima_devfreq.h b/drivers/gpu/drm/lima/lima_devfreq.h > new file mode 100644 > index 000000000000..fe4f8a437033 > --- /dev/null > +++ b/drivers/gpu/drm/lima/lima_devfreq.h > @@ -0,0 +1,15 @@ > +/* SPDX-License-Identifier: GPL-2.0 */ > +/* Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> */ > + > +#ifndef __LIMA_DEVFREQ_H__ > +#define __LIMA_DEVFREQ_H__ > + > +struct lima_device; > + > +int lima_devfreq_init(struct lima_device *ldev); > +void lima_devfreq_fini(struct lima_device *ldev); > + > +void lima_devfreq_record_busy(struct lima_device *ldev); > +void lima_devfreq_record_idle(struct lima_device *ldev); > + > +#endif > diff --git a/drivers/gpu/drm/lima/lima_device.c b/drivers/gpu/drm/lima/lima_device.c > index 19829b543024..7f1f7a1c03e5 100644 > --- a/drivers/gpu/drm/lima/lima_device.c > +++ b/drivers/gpu/drm/lima/lima_device.c > @@ -214,6 +214,8 @@ static int lima_init_gp_pipe(struct lima_device *dev) > struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_gp; > int err; > > + pipe->ldev = dev; > + > err = lima_sched_pipe_init(pipe, "gp"); > if (err) > return err; > @@ -244,6 +246,8 @@ static int lima_init_pp_pipe(struct lima_device *dev) > struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_pp; > int err, i; > > + pipe->ldev = dev; > + > err = lima_sched_pipe_init(pipe, "pp"); > if (err) > return err; > diff --git a/drivers/gpu/drm/lima/lima_device.h b/drivers/gpu/drm/lima/lima_device.h > index 31158d86271c..26f0efdd17f1 100644 > --- a/drivers/gpu/drm/lima/lima_device.h > +++ b/drivers/gpu/drm/lima/lima_device.h > @@ -5,6 +5,7 @@ > #define __LIMA_DEVICE_H__ > > #include <drm/drm_device.h> > +#include <linux/atomic.h> > #include <linux/delay.h> > > #include "lima_sched.h" > @@ -94,6 +95,22 @@ struct lima_device { > > u32 *dlbu_cpu; > dma_addr_t dlbu_dma; > + > + struct { > + struct devfreq *devfreq; > + struct opp_table *opp_table; > + struct thermal_cooling_device *cooling; > + ktime_t busy_time; > + ktime_t idle_time; > + ktime_t time_last_update; > + atomic_t busy_count; > + /* > + * Protect busy_time, idle_time and time_last_update because > + * these can be updated concurrently - for example by the GP > + * and PP interrupts. > + */ > + spinlock_t lock; > + } devfreq; > }; > > static inline struct lima_device * > diff --git a/drivers/gpu/drm/lima/lima_drv.c b/drivers/gpu/drm/lima/lima_drv.c > index 124efe4fa97b..b64b1777f220 100644 > --- a/drivers/gpu/drm/lima/lima_drv.c > +++ b/drivers/gpu/drm/lima/lima_drv.c > @@ -10,6 +10,7 @@ > #include <drm/drm_prime.h> > #include <drm/lima_drm.h> > > +#include "lima_devfreq.h" > #include "lima_drv.h" > #include "lima_gem.h" > #include "lima_vm.h" > @@ -296,18 +297,26 @@ static int lima_pdev_probe(struct platform_device *pdev) > if (err) > goto err_out1; > > + err = lima_devfreq_init(ldev); > + if (err) { > + dev_err(&pdev->dev, "Fatal error during devfreq init\n"); > + goto err_out2; > + } > + > /* > * Register the DRM device with the core and the connectors with > * sysfs. > */ > err = drm_dev_register(ddev, 0); > if (err < 0) > - goto err_out2; > + goto err_out3; > > return 0; > > -err_out2: > +err_out3: > lima_device_fini(ldev); > +err_out2: > + lima_devfreq_fini(ldev); > err_out1: > drm_dev_put(ddev); > err_out0: > @@ -321,6 +330,7 @@ static int lima_pdev_remove(struct platform_device *pdev) > struct drm_device *ddev = ldev->ddev; > > drm_dev_unregister(ddev); > + lima_devfreq_fini(ldev); > lima_device_fini(ldev); > drm_dev_put(ddev); > lima_sched_slab_fini(); > diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c > index f522c5f99729..851c496a168b 100644 > --- a/drivers/gpu/drm/lima/lima_sched.c > +++ b/drivers/gpu/drm/lima/lima_sched.c > @@ -5,6 +5,7 @@ > #include <linux/slab.h> > #include <linux/xarray.h> > > +#include "lima_devfreq.h" > #include "lima_drv.h" > #include "lima_sched.h" > #include "lima_vm.h" > @@ -213,6 +214,8 @@ static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job) > */ > ret = dma_fence_get(task->fence); > > + lima_devfreq_record_busy(pipe->ldev); > + > pipe->current_task = task; > > /* this is needed for MMU to work correctly, otherwise GP/PP > @@ -280,6 +283,8 @@ static void lima_sched_handle_error_task(struct lima_sched_pipe *pipe, > pipe->current_vm = NULL; > pipe->current_task = NULL; > > + lima_devfreq_record_idle(pipe->ldev); > + > drm_sched_resubmit_jobs(&pipe->base); > drm_sched_start(&pipe->base, true); > } > @@ -348,6 +353,8 @@ void lima_sched_pipe_fini(struct lima_sched_pipe *pipe) > > void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe) > { > + lima_devfreq_record_idle(pipe->ldev); This should be moved to the else {} part of the following code. As you have added the save idle record to the lima_sched_handle_error_task which is just the if {} part of the following code, so no need to call it twice for error tasks. BTW. lima_sched_handle_error_task is also used for timeout tasks, so add idle record in it is fine. Regards, Qiang > + > if (pipe->error) > schedule_work(&pipe->error_work); > else { > diff --git a/drivers/gpu/drm/lima/lima_sched.h b/drivers/gpu/drm/lima/lima_sched.h > index 928af91c1118..9ae7df7d7fbb 100644 > --- a/drivers/gpu/drm/lima/lima_sched.h > +++ b/drivers/gpu/drm/lima/lima_sched.h > @@ -6,6 +6,7 @@ > > #include <drm/gpu_scheduler.h> > > +struct lima_device; > struct lima_vm; > > struct lima_sched_task { > @@ -41,6 +42,8 @@ struct lima_sched_pipe { > u32 fence_seqno; > spinlock_t fence_lock; > > + struct lima_device *ldev; > + > struct lima_sched_task *current_task; > struct lima_vm *current_vm; > > -- > 2.24.1 >
Hi Robin, On Mon, Dec 30, 2019 at 1:47 AM Robin Murphy <robin.murphy@arm.com> wrote: > > On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: > > Hi Robin, > > > > On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: > >> > >> Hi Martin, > >> > >> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: > >>> Most platforms with a Mali-400 or Mali-450 GPU also have support for > >>> changing the GPU clock frequency. Add devfreq support so the GPU clock > >>> rate is updated based on the actual GPU usage when the > >>> "operating-points-v2" property is present in the board.dts. > >>> > >>> The actual devfreq code is taken from panfrost_devfreq.c and modified so > >>> it matches what the lima hardware needs: > >>> - a call to dev_pm_opp_set_clkname() during initialization because there > >>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > >>> the GPU so we need to control it using devfreq. > >>> - locking when reading or writing the devfreq statistics because (unlike > >>> than panfrost) we have multiple PP and GP IRQs which may finish jobs > >>> concurrently. > >> > >> I gave this a quick try on my RK3328, and the clock scaling indeed kicks > >> in nicely on the glmark2 scenes that struggle, however something appears > >> to be missing in terms of regulator association, as the appropriate OPP > >> voltages aren't reflected in the GPU supply (fortunately the initial > >> voltage seems close enough to that of the highest OPP not to cause major > >> problems, on my box at least). With panfrost on RK3399 I do see the > >> supply voltage scaling accordingly, but I don't know my way around > >> devfreq well enough to know what matters in the difference :/ > > first of all: thank you for trying this out! :-) > > > > does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use > > generic code for devfreq") for your panfrost test? > > if I understand the devfreq API correct then I suspect with that > > commit panfrost also won't change the voltage anymore. > > Oh, you're quite right - I was already considering that change as > ancient history, but indeed it's only in 5.5-rc, while that board is > still on 5.4.y release kernels. No wonder I couldn't make sense of how > the (current) code could possibly be working :) > > I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is > hopefully fixed), but I'm already fairly confident you've called it > correctly. I just tested it with the lima driver (by undervolting the GPU by 0.05V) and it seems that dev_pm_opp_set_regulators is really needed. I'll fix this in the next version of this patch and also submit a fix for panfrost (I won't be able to test that though, so help is appreciated in terms of testing :)) Martin
Hi Qiang, On Tue, Dec 31, 2019 at 3:54 AM Qiang Yu <yuq825@gmail.com> wrote: [...] > > diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c > > index f522c5f99729..851c496a168b 100644 > > --- a/drivers/gpu/drm/lima/lima_sched.c > > +++ b/drivers/gpu/drm/lima/lima_sched.c > > @@ -5,6 +5,7 @@ > > #include <linux/slab.h> > > #include <linux/xarray.h> > > > > +#include "lima_devfreq.h" > > #include "lima_drv.h" > > #include "lima_sched.h" > > #include "lima_vm.h" > > @@ -213,6 +214,8 @@ static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job) > > */ > > ret = dma_fence_get(task->fence); > > > > + lima_devfreq_record_busy(pipe->ldev); > > + > > pipe->current_task = task; > > > > /* this is needed for MMU to work correctly, otherwise GP/PP > > @@ -280,6 +283,8 @@ static void lima_sched_handle_error_task(struct lima_sched_pipe *pipe, > > pipe->current_vm = NULL; > > pipe->current_task = NULL; > > > > + lima_devfreq_record_idle(pipe->ldev); > > + > > drm_sched_resubmit_jobs(&pipe->base); > > drm_sched_start(&pipe->base, true); > > } > > @@ -348,6 +353,8 @@ void lima_sched_pipe_fini(struct lima_sched_pipe *pipe) > > > > void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe) > > { > > + lima_devfreq_record_idle(pipe->ldev); > > This should be moved to the else {} part of the following code. As you > have added the save > idle record to the lima_sched_handle_error_task which is just the if > {} part of the following > code, so no need to call it twice for error tasks. BTW. > lima_sched_handle_error_task is also > used for timeout tasks, so add idle record in it is fine. oh, that is a good catch - thank you! I will fix that in the next version Martin
On 2019-12-31 2:17 pm, Martin Blumenstingl wrote: > Hi Robin, > > On Mon, Dec 30, 2019 at 1:47 AM Robin Murphy <robin.murphy@arm.com> wrote: >> >> On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: >>> Hi Robin, >>> >>> On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: >>>> >>>> Hi Martin, >>>> >>>> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: >>>>> Most platforms with a Mali-400 or Mali-450 GPU also have support for >>>>> changing the GPU clock frequency. Add devfreq support so the GPU clock >>>>> rate is updated based on the actual GPU usage when the >>>>> "operating-points-v2" property is present in the board.dts. >>>>> >>>>> The actual devfreq code is taken from panfrost_devfreq.c and modified so >>>>> it matches what the lima hardware needs: >>>>> - a call to dev_pm_opp_set_clkname() during initialization because there >>>>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks >>>>> the GPU so we need to control it using devfreq. >>>>> - locking when reading or writing the devfreq statistics because (unlike >>>>> than panfrost) we have multiple PP and GP IRQs which may finish jobs >>>>> concurrently. >>>> >>>> I gave this a quick try on my RK3328, and the clock scaling indeed kicks >>>> in nicely on the glmark2 scenes that struggle, however something appears >>>> to be missing in terms of regulator association, as the appropriate OPP >>>> voltages aren't reflected in the GPU supply (fortunately the initial >>>> voltage seems close enough to that of the highest OPP not to cause major >>>> problems, on my box at least). With panfrost on RK3399 I do see the >>>> supply voltage scaling accordingly, but I don't know my way around >>>> devfreq well enough to know what matters in the difference :/ >>> first of all: thank you for trying this out! :-) >>> >>> does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use >>> generic code for devfreq") for your panfrost test? >>> if I understand the devfreq API correct then I suspect with that >>> commit panfrost also won't change the voltage anymore. >> >> Oh, you're quite right - I was already considering that change as >> ancient history, but indeed it's only in 5.5-rc, while that board is >> still on 5.4.y release kernels. No wonder I couldn't make sense of how >> the (current) code could possibly be working :) >> >> I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is >> hopefully fixed), but I'm already fairly confident you've called it >> correctly. > I just tested it with the lima driver (by undervolting the GPU by > 0.05V) and it seems that dev_pm_opp_set_regulators is really needed. > I'll fix this in the next version of this patch and also submit a fix > for panfrost (I won't be able to test that though, so help is > appreciated in terms of testing :)) Yeah, I started hacking something up for panfrost yesterday, but at the point of realising the core OPP code wants refactoring to actually handle optional regulators without spewing errors, decided that was crossing the line into "work" and thus could wait until next week :D Cheers, Robin.
Hi Robin, On Tue, Dec 31, 2019 at 5:40 PM Robin Murphy <robin.murphy@arm.com> wrote: > > On 2019-12-31 2:17 pm, Martin Blumenstingl wrote: > > Hi Robin, > > > > On Mon, Dec 30, 2019 at 1:47 AM Robin Murphy <robin.murphy@arm.com> wrote: > >> > >> On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: > >>> Hi Robin, > >>> > >>> On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: > >>>> > >>>> Hi Martin, > >>>> > >>>> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: > >>>>> Most platforms with a Mali-400 or Mali-450 GPU also have support for > >>>>> changing the GPU clock frequency. Add devfreq support so the GPU clock > >>>>> rate is updated based on the actual GPU usage when the > >>>>> "operating-points-v2" property is present in the board.dts. > >>>>> > >>>>> The actual devfreq code is taken from panfrost_devfreq.c and modified so > >>>>> it matches what the lima hardware needs: > >>>>> - a call to dev_pm_opp_set_clkname() during initialization because there > >>>>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > >>>>> the GPU so we need to control it using devfreq. > >>>>> - locking when reading or writing the devfreq statistics because (unlike > >>>>> than panfrost) we have multiple PP and GP IRQs which may finish jobs > >>>>> concurrently. > >>>> > >>>> I gave this a quick try on my RK3328, and the clock scaling indeed kicks > >>>> in nicely on the glmark2 scenes that struggle, however something appears > >>>> to be missing in terms of regulator association, as the appropriate OPP > >>>> voltages aren't reflected in the GPU supply (fortunately the initial > >>>> voltage seems close enough to that of the highest OPP not to cause major > >>>> problems, on my box at least). With panfrost on RK3399 I do see the > >>>> supply voltage scaling accordingly, but I don't know my way around > >>>> devfreq well enough to know what matters in the difference :/ > >>> first of all: thank you for trying this out! :-) > >>> > >>> does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use > >>> generic code for devfreq") for your panfrost test? > >>> if I understand the devfreq API correct then I suspect with that > >>> commit panfrost also won't change the voltage anymore. > >> > >> Oh, you're quite right - I was already considering that change as > >> ancient history, but indeed it's only in 5.5-rc, while that board is > >> still on 5.4.y release kernels. No wonder I couldn't make sense of how > >> the (current) code could possibly be working :) > >> > >> I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is > >> hopefully fixed), but I'm already fairly confident you've called it > >> correctly. > > I just tested it with the lima driver (by undervolting the GPU by > > 0.05V) and it seems that dev_pm_opp_set_regulators is really needed. > > I'll fix this in the next version of this patch and also submit a fix > > for panfrost (I won't be able to test that though, so help is > > appreciated in terms of testing :)) > > Yeah, I started hacking something up for panfrost yesterday, but at the > point of realising the core OPP code wants refactoring to actually > handle optional regulators without spewing errors, decided that was > crossing the line into "work" and thus could wait until next week :D I'm not sure what you mean, dev_pm_opp_set_regulators uses regulator_get_optional. doesn't that mean that it is optional already? Martin
On 2019-12-31 4:47 pm, Martin Blumenstingl wrote: > Hi Robin, > > On Tue, Dec 31, 2019 at 5:40 PM Robin Murphy <robin.murphy@arm.com> wrote: >> >> On 2019-12-31 2:17 pm, Martin Blumenstingl wrote: >>> Hi Robin, >>> >>> On Mon, Dec 30, 2019 at 1:47 AM Robin Murphy <robin.murphy@arm.com> wrote: >>>> >>>> On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: >>>>> Hi Robin, >>>>> >>>>> On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: >>>>>> >>>>>> Hi Martin, >>>>>> >>>>>> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: >>>>>>> Most platforms with a Mali-400 or Mali-450 GPU also have support for >>>>>>> changing the GPU clock frequency. Add devfreq support so the GPU clock >>>>>>> rate is updated based on the actual GPU usage when the >>>>>>> "operating-points-v2" property is present in the board.dts. >>>>>>> >>>>>>> The actual devfreq code is taken from panfrost_devfreq.c and modified so >>>>>>> it matches what the lima hardware needs: >>>>>>> - a call to dev_pm_opp_set_clkname() during initialization because there >>>>>>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks >>>>>>> the GPU so we need to control it using devfreq. >>>>>>> - locking when reading or writing the devfreq statistics because (unlike >>>>>>> than panfrost) we have multiple PP and GP IRQs which may finish jobs >>>>>>> concurrently. >>>>>> >>>>>> I gave this a quick try on my RK3328, and the clock scaling indeed kicks >>>>>> in nicely on the glmark2 scenes that struggle, however something appears >>>>>> to be missing in terms of regulator association, as the appropriate OPP >>>>>> voltages aren't reflected in the GPU supply (fortunately the initial >>>>>> voltage seems close enough to that of the highest OPP not to cause major >>>>>> problems, on my box at least). With panfrost on RK3399 I do see the >>>>>> supply voltage scaling accordingly, but I don't know my way around >>>>>> devfreq well enough to know what matters in the difference :/ >>>>> first of all: thank you for trying this out! :-) >>>>> >>>>> does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use >>>>> generic code for devfreq") for your panfrost test? >>>>> if I understand the devfreq API correct then I suspect with that >>>>> commit panfrost also won't change the voltage anymore. >>>> >>>> Oh, you're quite right - I was already considering that change as >>>> ancient history, but indeed it's only in 5.5-rc, while that board is >>>> still on 5.4.y release kernels. No wonder I couldn't make sense of how >>>> the (current) code could possibly be working :) >>>> >>>> I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is >>>> hopefully fixed), but I'm already fairly confident you've called it >>>> correctly. >>> I just tested it with the lima driver (by undervolting the GPU by >>> 0.05V) and it seems that dev_pm_opp_set_regulators is really needed. >>> I'll fix this in the next version of this patch and also submit a fix >>> for panfrost (I won't be able to test that though, so help is >>> appreciated in terms of testing :)) >> >> Yeah, I started hacking something up for panfrost yesterday, but at the >> point of realising the core OPP code wants refactoring to actually >> handle optional regulators without spewing errors, decided that was >> crossing the line into "work" and thus could wait until next week :D > I'm not sure what you mean, dev_pm_opp_set_regulators uses > regulator_get_optional. > doesn't that mean that it is optional already? Indeed it does call regulator_get_optional(), but it then goes on to treat the absence of a supposedly-optional regulator as a hard failure. It doesn't seem very useful having a nice abstracted interface if users still end up have to dance around and duplicate half the parsing in order to work out whether it's worth calling or not - far better IMO if it could just successfully set/put zero regulators in the cases where the OPPs are behind a firmware/mailbox DVFS interface rather than explicit in-kernel clock/regulator control. That said, given that I think the current lima/panfrost users should all be relatively simple with either 0 or 1 regulator, you could probably just special-case -ENODEV and accept a spurious error message sometimes for the sake of an immediate fix, then we can make general improvements to the interface separately afterwards. Robin.
Hi Robin, On Wed, Jan 1, 2020 at 1:55 PM Robin Murphy <robin.murphy@arm.com> wrote: > > On 2019-12-31 4:47 pm, Martin Blumenstingl wrote: > > Hi Robin, > > > > On Tue, Dec 31, 2019 at 5:40 PM Robin Murphy <robin.murphy@arm.com> wrote: > >> > >> On 2019-12-31 2:17 pm, Martin Blumenstingl wrote: > >>> Hi Robin, > >>> > >>> On Mon, Dec 30, 2019 at 1:47 AM Robin Murphy <robin.murphy@arm.com> wrote: > >>>> > >>>> On 2019-12-29 11:19 pm, Martin Blumenstingl wrote: > >>>>> Hi Robin, > >>>>> > >>>>> On Sun, Dec 29, 2019 at 11:58 PM Robin Murphy <robin.murphy@arm.com> wrote: > >>>>>> > >>>>>> Hi Martin, > >>>>>> > >>>>>> On 2019-12-27 5:37 pm, Martin Blumenstingl wrote: > >>>>>>> Most platforms with a Mali-400 or Mali-450 GPU also have support for > >>>>>>> changing the GPU clock frequency. Add devfreq support so the GPU clock > >>>>>>> rate is updated based on the actual GPU usage when the > >>>>>>> "operating-points-v2" property is present in the board.dts. > >>>>>>> > >>>>>>> The actual devfreq code is taken from panfrost_devfreq.c and modified so > >>>>>>> it matches what the lima hardware needs: > >>>>>>> - a call to dev_pm_opp_set_clkname() during initialization because there > >>>>>>> are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks > >>>>>>> the GPU so we need to control it using devfreq. > >>>>>>> - locking when reading or writing the devfreq statistics because (unlike > >>>>>>> than panfrost) we have multiple PP and GP IRQs which may finish jobs > >>>>>>> concurrently. > >>>>>> > >>>>>> I gave this a quick try on my RK3328, and the clock scaling indeed kicks > >>>>>> in nicely on the glmark2 scenes that struggle, however something appears > >>>>>> to be missing in terms of regulator association, as the appropriate OPP > >>>>>> voltages aren't reflected in the GPU supply (fortunately the initial > >>>>>> voltage seems close enough to that of the highest OPP not to cause major > >>>>>> problems, on my box at least). With panfrost on RK3399 I do see the > >>>>>> supply voltage scaling accordingly, but I don't know my way around > >>>>>> devfreq well enough to know what matters in the difference :/ > >>>>> first of all: thank you for trying this out! :-) > >>>>> > >>>>> does your kernel include commit 221bc77914cbcc ("drm/panfrost: Use > >>>>> generic code for devfreq") for your panfrost test? > >>>>> if I understand the devfreq API correct then I suspect with that > >>>>> commit panfrost also won't change the voltage anymore. > >>>> > >>>> Oh, you're quite right - I was already considering that change as > >>>> ancient history, but indeed it's only in 5.5-rc, while that board is > >>>> still on 5.4.y release kernels. No wonder I couldn't make sense of how > >>>> the (current) code could possibly be working :) > >>>> > >>>> I'll try the latest -rc kernel tomorrow to confirm (now that PCIe is > >>>> hopefully fixed), but I'm already fairly confident you've called it > >>>> correctly. > >>> I just tested it with the lima driver (by undervolting the GPU by > >>> 0.05V) and it seems that dev_pm_opp_set_regulators is really needed. > >>> I'll fix this in the next version of this patch and also submit a fix > >>> for panfrost (I won't be able to test that though, so help is > >>> appreciated in terms of testing :)) > >> > >> Yeah, I started hacking something up for panfrost yesterday, but at the > >> point of realising the core OPP code wants refactoring to actually > >> handle optional regulators without spewing errors, decided that was > >> crossing the line into "work" and thus could wait until next week :D > > I'm not sure what you mean, dev_pm_opp_set_regulators uses > > regulator_get_optional. > > doesn't that mean that it is optional already? > > Indeed it does call regulator_get_optional(), but it then goes on to > treat the absence of a supposedly-optional regulator as a hard failure. > It doesn't seem very useful having a nice abstracted interface if users > still end up have to dance around and duplicate half the parsing in > order to work out whether it's worth calling or not - far better IMO if > it could just successfully set/put zero regulators in the cases where > the OPPs are behind a firmware/mailbox DVFS interface rather than > explicit in-kernel clock/regulator control. thank you for the explanation I'll experience this case on newer Amlogic SoCs where regulators are hidden behind SCPI firmware. so far I have only tested the case without OPP-table on those SoCs. > That said, given that I think the current lima/panfrost users should all > be relatively simple with either 0 or 1 regulator, you could probably > just special-case -ENODEV and accept a spurious error message sometimes > for the sake of an immediate fix, then we can make general improvements > to the interface separately afterwards. OK, I'll be working on this next week again if you get to fix panfrost earlier then please Cc me so we don't duplicate work Martin
diff --git a/drivers/gpu/drm/lima/Kconfig b/drivers/gpu/drm/lima/Kconfig index 571dc369a7e9..cdd24b68b5d4 100644 --- a/drivers/gpu/drm/lima/Kconfig +++ b/drivers/gpu/drm/lima/Kconfig @@ -10,5 +10,6 @@ config DRM_LIMA depends on OF select DRM_SCHED select DRM_GEM_SHMEM_HELPER + select PM_DEVFREQ help DRM driver for ARM Mali 400/450 GPUs. diff --git a/drivers/gpu/drm/lima/Makefile b/drivers/gpu/drm/lima/Makefile index a85444b0a1d4..5e5c29875e9c 100644 --- a/drivers/gpu/drm/lima/Makefile +++ b/drivers/gpu/drm/lima/Makefile @@ -14,6 +14,7 @@ lima-y := \ lima_sched.o \ lima_ctx.o \ lima_dlbu.o \ - lima_bcast.o + lima_bcast.o \ + lima_devfreq.o obj-$(CONFIG_DRM_LIMA) += lima.o diff --git a/drivers/gpu/drm/lima/lima_devfreq.c b/drivers/gpu/drm/lima/lima_devfreq.c new file mode 100644 index 000000000000..a5fd6b8faa77 --- /dev/null +++ b/drivers/gpu/drm/lima/lima_devfreq.c @@ -0,0 +1,183 @@ +// SPDX-License-Identifier: GPL-2.0 +/* + * Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> + * + * Based on panfrost_devfreq.c: + * Copyright 2019 Collabora ltd. + */ +#include <linux/clk.h> +#include <linux/devfreq.h> +#include <linux/devfreq_cooling.h> +#include <linux/device.h> +#include <linux/platform_device.h> +#include <linux/pm_opp.h> + +#include "lima_device.h" +#include "lima_devfreq.h" + +static void lima_devfreq_update_utilization(struct lima_device *ldev) +{ + unsigned long irqflags; + ktime_t now, last; + + if (!ldev->devfreq.devfreq) + return; + + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); + + now = ktime_get(); + last = ldev->devfreq.time_last_update; + + if (atomic_read(&ldev->devfreq.busy_count) > 0) + ldev->devfreq.busy_time += ktime_sub(now, last); + else + ldev->devfreq.idle_time += ktime_sub(now, last); + + ldev->devfreq.time_last_update = now; + + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); +} + +static int lima_devfreq_target(struct device *dev, unsigned long *freq, + u32 flags) +{ + struct dev_pm_opp *opp; + int err; + + opp = devfreq_recommended_opp(dev, freq, flags); + if (IS_ERR(opp)) + return PTR_ERR(opp); + dev_pm_opp_put(opp); + + err = dev_pm_opp_set_rate(dev, *freq); + if (err) + return err; + + return 0; +} + +static void lima_devfreq_reset(struct lima_device *ldev) +{ + unsigned long irqflags; + + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); + + ldev->devfreq.busy_time = 0; + ldev->devfreq.idle_time = 0; + ldev->devfreq.time_last_update = ktime_get(); + + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); +} + +static int lima_devfreq_get_dev_status(struct device *dev, + struct devfreq_dev_status *status) +{ + struct lima_device *ldev = dev_get_drvdata(dev); + unsigned long irqflags; + + lima_devfreq_update_utilization(ldev); + + status->current_frequency = clk_get_rate(ldev->clk_gpu); + + spin_lock_irqsave(&ldev->devfreq.lock, irqflags); + + status->total_time = ktime_to_ns(ktime_add(ldev->devfreq.busy_time, + ldev->devfreq.idle_time)); + status->busy_time = ktime_to_ns(ldev->devfreq.busy_time); + + spin_unlock_irqrestore(&ldev->devfreq.lock, irqflags); + + lima_devfreq_reset(ldev); + + dev_dbg(ldev->dev, "busy %lu total %lu %lu %% freq %lu MHz\n", + status->busy_time, status->total_time, + status->busy_time / (status->total_time / 100), + status->current_frequency / 1000 / 1000); + + return 0; +} + +static struct devfreq_dev_profile lima_devfreq_profile = { + .polling_ms = 50, /* ~3 frames */ + .target = lima_devfreq_target, + .get_dev_status = lima_devfreq_get_dev_status, +}; + +int lima_devfreq_init(struct lima_device *ldev) +{ + struct thermal_cooling_device *cooling; + struct device *dev = &ldev->pdev->dev; + struct devfreq *devfreq; + struct dev_pm_opp *opp; + unsigned long cur_freq; + int ret; + + spin_lock_init(&ldev->devfreq.lock); + + ldev->devfreq.opp_table = dev_pm_opp_set_clkname(dev, "core"); + if (IS_ERR(ldev->devfreq.opp_table)) + return PTR_ERR(ldev->devfreq.opp_table); + + ret = dev_pm_opp_of_add_table(dev); + if (ret == -ENODEV) /* Optional, continue without devfreq */ + return 0; + else if (ret) + return ret; + + lima_devfreq_reset(ldev); + + cur_freq = clk_get_rate(ldev->clk_gpu); + + opp = devfreq_recommended_opp(dev, &cur_freq, 0); + if (IS_ERR(opp)) + return PTR_ERR(opp); + + lima_devfreq_profile.initial_freq = cur_freq; + dev_pm_opp_put(opp); + + devfreq = devm_devfreq_add_device(dev, &lima_devfreq_profile, + DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL); + if (IS_ERR(devfreq)) { + dev_err(dev, "Couldn't initialize GPU devfreq\n"); + dev_pm_opp_of_remove_table(dev); + return PTR_ERR(devfreq); + } + + ldev->devfreq.devfreq = devfreq; + + cooling = of_devfreq_cooling_register(dev->of_node, devfreq); + if (IS_ERR(cooling)) + dev_info(dev, "Failed to register cooling device\n"); + else + ldev->devfreq.cooling = cooling; + + return 0; +} + +void lima_devfreq_fini(struct lima_device *ldev) +{ + if (ldev->devfreq.cooling) + devfreq_cooling_unregister(ldev->devfreq.cooling); + + if (ldev->devfreq.opp_table) { + dev_pm_opp_put_clkname(ldev->devfreq.opp_table); + ldev->devfreq.opp_table = NULL; + } + + dev_pm_opp_of_remove_table(&ldev->pdev->dev); +} + +void lima_devfreq_record_busy(struct lima_device *ldev) +{ + lima_devfreq_update_utilization(ldev); + atomic_inc(&ldev->devfreq.busy_count); +} + +void lima_devfreq_record_idle(struct lima_device *ldev) +{ + int count; + + lima_devfreq_update_utilization(ldev); + count = atomic_dec_if_positive(&ldev->devfreq.busy_count); + WARN_ON(count < 0); +} diff --git a/drivers/gpu/drm/lima/lima_devfreq.h b/drivers/gpu/drm/lima/lima_devfreq.h new file mode 100644 index 000000000000..fe4f8a437033 --- /dev/null +++ b/drivers/gpu/drm/lima/lima_devfreq.h @@ -0,0 +1,15 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +/* Copyright 2019 Martin Blumenstingl <martin.blumenstingl@googlemail.com> */ + +#ifndef __LIMA_DEVFREQ_H__ +#define __LIMA_DEVFREQ_H__ + +struct lima_device; + +int lima_devfreq_init(struct lima_device *ldev); +void lima_devfreq_fini(struct lima_device *ldev); + +void lima_devfreq_record_busy(struct lima_device *ldev); +void lima_devfreq_record_idle(struct lima_device *ldev); + +#endif diff --git a/drivers/gpu/drm/lima/lima_device.c b/drivers/gpu/drm/lima/lima_device.c index 19829b543024..7f1f7a1c03e5 100644 --- a/drivers/gpu/drm/lima/lima_device.c +++ b/drivers/gpu/drm/lima/lima_device.c @@ -214,6 +214,8 @@ static int lima_init_gp_pipe(struct lima_device *dev) struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_gp; int err; + pipe->ldev = dev; + err = lima_sched_pipe_init(pipe, "gp"); if (err) return err; @@ -244,6 +246,8 @@ static int lima_init_pp_pipe(struct lima_device *dev) struct lima_sched_pipe *pipe = dev->pipe + lima_pipe_pp; int err, i; + pipe->ldev = dev; + err = lima_sched_pipe_init(pipe, "pp"); if (err) return err; diff --git a/drivers/gpu/drm/lima/lima_device.h b/drivers/gpu/drm/lima/lima_device.h index 31158d86271c..26f0efdd17f1 100644 --- a/drivers/gpu/drm/lima/lima_device.h +++ b/drivers/gpu/drm/lima/lima_device.h @@ -5,6 +5,7 @@ #define __LIMA_DEVICE_H__ #include <drm/drm_device.h> +#include <linux/atomic.h> #include <linux/delay.h> #include "lima_sched.h" @@ -94,6 +95,22 @@ struct lima_device { u32 *dlbu_cpu; dma_addr_t dlbu_dma; + + struct { + struct devfreq *devfreq; + struct opp_table *opp_table; + struct thermal_cooling_device *cooling; + ktime_t busy_time; + ktime_t idle_time; + ktime_t time_last_update; + atomic_t busy_count; + /* + * Protect busy_time, idle_time and time_last_update because + * these can be updated concurrently - for example by the GP + * and PP interrupts. + */ + spinlock_t lock; + } devfreq; }; static inline struct lima_device * diff --git a/drivers/gpu/drm/lima/lima_drv.c b/drivers/gpu/drm/lima/lima_drv.c index 124efe4fa97b..b64b1777f220 100644 --- a/drivers/gpu/drm/lima/lima_drv.c +++ b/drivers/gpu/drm/lima/lima_drv.c @@ -10,6 +10,7 @@ #include <drm/drm_prime.h> #include <drm/lima_drm.h> +#include "lima_devfreq.h" #include "lima_drv.h" #include "lima_gem.h" #include "lima_vm.h" @@ -296,18 +297,26 @@ static int lima_pdev_probe(struct platform_device *pdev) if (err) goto err_out1; + err = lima_devfreq_init(ldev); + if (err) { + dev_err(&pdev->dev, "Fatal error during devfreq init\n"); + goto err_out2; + } + /* * Register the DRM device with the core and the connectors with * sysfs. */ err = drm_dev_register(ddev, 0); if (err < 0) - goto err_out2; + goto err_out3; return 0; -err_out2: +err_out3: lima_device_fini(ldev); +err_out2: + lima_devfreq_fini(ldev); err_out1: drm_dev_put(ddev); err_out0: @@ -321,6 +330,7 @@ static int lima_pdev_remove(struct platform_device *pdev) struct drm_device *ddev = ldev->ddev; drm_dev_unregister(ddev); + lima_devfreq_fini(ldev); lima_device_fini(ldev); drm_dev_put(ddev); lima_sched_slab_fini(); diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c index f522c5f99729..851c496a168b 100644 --- a/drivers/gpu/drm/lima/lima_sched.c +++ b/drivers/gpu/drm/lima/lima_sched.c @@ -5,6 +5,7 @@ #include <linux/slab.h> #include <linux/xarray.h> +#include "lima_devfreq.h" #include "lima_drv.h" #include "lima_sched.h" #include "lima_vm.h" @@ -213,6 +214,8 @@ static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job) */ ret = dma_fence_get(task->fence); + lima_devfreq_record_busy(pipe->ldev); + pipe->current_task = task; /* this is needed for MMU to work correctly, otherwise GP/PP @@ -280,6 +283,8 @@ static void lima_sched_handle_error_task(struct lima_sched_pipe *pipe, pipe->current_vm = NULL; pipe->current_task = NULL; + lima_devfreq_record_idle(pipe->ldev); + drm_sched_resubmit_jobs(&pipe->base); drm_sched_start(&pipe->base, true); } @@ -348,6 +353,8 @@ void lima_sched_pipe_fini(struct lima_sched_pipe *pipe) void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe) { + lima_devfreq_record_idle(pipe->ldev); + if (pipe->error) schedule_work(&pipe->error_work); else { diff --git a/drivers/gpu/drm/lima/lima_sched.h b/drivers/gpu/drm/lima/lima_sched.h index 928af91c1118..9ae7df7d7fbb 100644 --- a/drivers/gpu/drm/lima/lima_sched.h +++ b/drivers/gpu/drm/lima/lima_sched.h @@ -6,6 +6,7 @@ #include <drm/gpu_scheduler.h> +struct lima_device; struct lima_vm; struct lima_sched_task { @@ -41,6 +42,8 @@ struct lima_sched_pipe { u32 fence_seqno; spinlock_t fence_lock; + struct lima_device *ldev; + struct lima_sched_task *current_task; struct lima_vm *current_vm;
Most platforms with a Mali-400 or Mali-450 GPU also have support for changing the GPU clock frequency. Add devfreq support so the GPU clock rate is updated based on the actual GPU usage when the "operating-points-v2" property is present in the board.dts. The actual devfreq code is taken from panfrost_devfreq.c and modified so it matches what the lima hardware needs: - a call to dev_pm_opp_set_clkname() during initialization because there are two clocks on Mali-4x0 IPs. "core" is the one that actually clocks the GPU so we need to control it using devfreq. - locking when reading or writing the devfreq statistics because (unlike than panfrost) we have multiple PP and GP IRQs which may finish jobs concurrently. Signed-off-by: Martin Blumenstingl <martin.blumenstingl@googlemail.com> --- drivers/gpu/drm/lima/Kconfig | 1 + drivers/gpu/drm/lima/Makefile | 3 +- drivers/gpu/drm/lima/lima_devfreq.c | 183 ++++++++++++++++++++++++++++ drivers/gpu/drm/lima/lima_devfreq.h | 15 +++ drivers/gpu/drm/lima/lima_device.c | 4 + drivers/gpu/drm/lima/lima_device.h | 17 +++ drivers/gpu/drm/lima/lima_drv.c | 14 ++- drivers/gpu/drm/lima/lima_sched.c | 7 ++ drivers/gpu/drm/lima/lima_sched.h | 3 + 9 files changed, 244 insertions(+), 3 deletions(-) create mode 100644 drivers/gpu/drm/lima/lima_devfreq.c create mode 100644 drivers/gpu/drm/lima/lima_devfreq.h