Commit 9e62b885 authored by Steven Price's avatar Steven Price Committed by Rob Herring

drm/panfrost: Simplify devfreq utilisation tracking

Instead of tracking per-slot utilisation track a single value for the
entire GPU. Ultimately it doesn't matter if the GPU is busy with only
vertex or a combination of vertex and fragment processing - if it's busy
then it's busy and devfreq should be scaling appropriately.

This also makes way for being able to submit multiple jobs per slot
which requires more values than the original boolean per slot.
Reviewed-by: default avatarTomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: default avatarAlyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Signed-off-by: default avatarSteven Price <steven.price@arm.com>
Signed-off-by: default avatarRob Herring <robh@kernel.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20191025134143.14324-3-steven.price@arm.com
parent 221bc779
...@@ -13,7 +13,7 @@ ...@@ -13,7 +13,7 @@
#include "panfrost_gpu.h" #include "panfrost_gpu.h"
#include "panfrost_regs.h" #include "panfrost_regs.h"
static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev, int slot); static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev);
static int panfrost_devfreq_target(struct device *dev, unsigned long *freq, static int panfrost_devfreq_target(struct device *dev, unsigned long *freq,
u32 flags) u32 flags)
...@@ -32,37 +32,23 @@ static int panfrost_devfreq_target(struct device *dev, unsigned long *freq, ...@@ -32,37 +32,23 @@ static int panfrost_devfreq_target(struct device *dev, unsigned long *freq,
static void panfrost_devfreq_reset(struct panfrost_device *pfdev) static void panfrost_devfreq_reset(struct panfrost_device *pfdev)
{ {
ktime_t now = ktime_get(); pfdev->devfreq.busy_time = 0;
int i; pfdev->devfreq.idle_time = 0;
pfdev->devfreq.time_last_update = ktime_get();
for (i = 0; i < NUM_JOB_SLOTS; i++) {
pfdev->devfreq.slot[i].busy_time = 0;
pfdev->devfreq.slot[i].idle_time = 0;
pfdev->devfreq.slot[i].time_last_update = now;
}
} }
static int panfrost_devfreq_get_dev_status(struct device *dev, static int panfrost_devfreq_get_dev_status(struct device *dev,
struct devfreq_dev_status *status) struct devfreq_dev_status *status)
{ {
struct panfrost_device *pfdev = dev_get_drvdata(dev); struct panfrost_device *pfdev = dev_get_drvdata(dev);
int i;
for (i = 0; i < NUM_JOB_SLOTS; i++) { panfrost_devfreq_update_utilization(pfdev);
panfrost_devfreq_update_utilization(pfdev, i);
}
status->current_frequency = clk_get_rate(pfdev->clock); status->current_frequency = clk_get_rate(pfdev->clock);
status->total_time = ktime_to_ns(ktime_add(pfdev->devfreq.slot[0].busy_time, status->total_time = ktime_to_ns(ktime_add(pfdev->devfreq.busy_time,
pfdev->devfreq.slot[0].idle_time)); pfdev->devfreq.idle_time));
status->busy_time = 0;
for (i = 0; i < NUM_JOB_SLOTS; i++) {
status->busy_time += ktime_to_ns(pfdev->devfreq.slot[i].busy_time);
}
/* We're scheduling only to one core atm, so don't divide for now */ status->busy_time = ktime_to_ns(pfdev->devfreq.busy_time);
/* status->busy_time /= NUM_JOB_SLOTS; */
panfrost_devfreq_reset(pfdev); panfrost_devfreq_reset(pfdev);
...@@ -134,14 +120,10 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev) ...@@ -134,14 +120,10 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev)
void panfrost_devfreq_resume(struct panfrost_device *pfdev) void panfrost_devfreq_resume(struct panfrost_device *pfdev)
{ {
int i;
if (!pfdev->devfreq.devfreq) if (!pfdev->devfreq.devfreq)
return; return;
panfrost_devfreq_reset(pfdev); panfrost_devfreq_reset(pfdev);
for (i = 0; i < NUM_JOB_SLOTS; i++)
pfdev->devfreq.slot[i].busy = false;
devfreq_resume_device(pfdev->devfreq.devfreq); devfreq_resume_device(pfdev->devfreq.devfreq);
} }
...@@ -154,9 +136,8 @@ void panfrost_devfreq_suspend(struct panfrost_device *pfdev) ...@@ -154,9 +136,8 @@ void panfrost_devfreq_suspend(struct panfrost_device *pfdev)
devfreq_suspend_device(pfdev->devfreq.devfreq); devfreq_suspend_device(pfdev->devfreq.devfreq);
} }
static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev, int slot) static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev)
{ {
struct panfrost_devfreq_slot *devfreq_slot = &pfdev->devfreq.slot[slot];
ktime_t now; ktime_t now;
ktime_t last; ktime_t last;
...@@ -164,22 +145,27 @@ static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev, i ...@@ -164,22 +145,27 @@ static void panfrost_devfreq_update_utilization(struct panfrost_device *pfdev, i
return; return;
now = ktime_get(); now = ktime_get();
last = pfdev->devfreq.slot[slot].time_last_update; last = pfdev->devfreq.time_last_update;
/* If we last recorded a transition to busy, we have been idle since */ if (atomic_read(&pfdev->devfreq.busy_count) > 0)
if (devfreq_slot->busy) pfdev->devfreq.busy_time += ktime_sub(now, last);
pfdev->devfreq.slot[slot].busy_time += ktime_sub(now, last);
else else
pfdev->devfreq.slot[slot].idle_time += ktime_sub(now, last); pfdev->devfreq.idle_time += ktime_sub(now, last);
pfdev->devfreq.time_last_update = now;
}
pfdev->devfreq.slot[slot].time_last_update = now; void panfrost_devfreq_record_busy(struct panfrost_device *pfdev)
{
panfrost_devfreq_update_utilization(pfdev);
atomic_inc(&pfdev->devfreq.busy_count);
} }
/* The job scheduler is expected to call this at every transition busy <-> idle */ void panfrost_devfreq_record_idle(struct panfrost_device *pfdev)
void panfrost_devfreq_record_transition(struct panfrost_device *pfdev, int slot)
{ {
struct panfrost_devfreq_slot *devfreq_slot = &pfdev->devfreq.slot[slot]; int count;
panfrost_devfreq_update_utilization(pfdev, slot); panfrost_devfreq_update_utilization(pfdev);
devfreq_slot->busy = !devfreq_slot->busy; count = atomic_dec_if_positive(&pfdev->devfreq.busy_count);
WARN_ON(count < 0);
} }
...@@ -10,6 +10,7 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev); ...@@ -10,6 +10,7 @@ void panfrost_devfreq_fini(struct panfrost_device *pfdev);
void panfrost_devfreq_resume(struct panfrost_device *pfdev); void panfrost_devfreq_resume(struct panfrost_device *pfdev);
void panfrost_devfreq_suspend(struct panfrost_device *pfdev); void panfrost_devfreq_suspend(struct panfrost_device *pfdev);
void panfrost_devfreq_record_transition(struct panfrost_device *pfdev, int slot); void panfrost_devfreq_record_busy(struct panfrost_device *pfdev);
void panfrost_devfreq_record_idle(struct panfrost_device *pfdev);
#endif /* __PANFROST_DEVFREQ_H__ */ #endif /* __PANFROST_DEVFREQ_H__ */
...@@ -51,13 +51,6 @@ struct panfrost_features { ...@@ -51,13 +51,6 @@ struct panfrost_features {
unsigned long hw_issues[64 / BITS_PER_LONG]; unsigned long hw_issues[64 / BITS_PER_LONG];
}; };
struct panfrost_devfreq_slot {
ktime_t busy_time;
ktime_t idle_time;
ktime_t time_last_update;
bool busy;
};
struct panfrost_device { struct panfrost_device {
struct device *dev; struct device *dev;
struct drm_device *ddev; struct drm_device *ddev;
...@@ -93,7 +86,10 @@ struct panfrost_device { ...@@ -93,7 +86,10 @@ struct panfrost_device {
struct { struct {
struct devfreq *devfreq; struct devfreq *devfreq;
struct thermal_cooling_device *cooling; struct thermal_cooling_device *cooling;
struct panfrost_devfreq_slot slot[NUM_JOB_SLOTS]; ktime_t busy_time;
ktime_t idle_time;
ktime_t time_last_update;
atomic_t busy_count;
} devfreq; } devfreq;
}; };
......
...@@ -155,8 +155,7 @@ static void panfrost_job_hw_submit(struct panfrost_job *job, int js) ...@@ -155,8 +155,7 @@ static void panfrost_job_hw_submit(struct panfrost_job *job, int js)
} }
cfg = panfrost_mmu_as_get(pfdev, &job->file_priv->mmu); cfg = panfrost_mmu_as_get(pfdev, &job->file_priv->mmu);
panfrost_devfreq_record_busy(pfdev);
panfrost_devfreq_record_transition(pfdev, js);
job_write(pfdev, JS_HEAD_NEXT_LO(js), jc_head & 0xFFFFFFFF); job_write(pfdev, JS_HEAD_NEXT_LO(js), jc_head & 0xFFFFFFFF);
job_write(pfdev, JS_HEAD_NEXT_HI(js), jc_head >> 32); job_write(pfdev, JS_HEAD_NEXT_HI(js), jc_head >> 32);
...@@ -404,7 +403,7 @@ static void panfrost_job_timedout(struct drm_sched_job *sched_job) ...@@ -404,7 +403,7 @@ static void panfrost_job_timedout(struct drm_sched_job *sched_job)
} }
spin_unlock_irqrestore(&pfdev->js->job_lock, flags); spin_unlock_irqrestore(&pfdev->js->job_lock, flags);
panfrost_devfreq_record_transition(pfdev, js); panfrost_devfreq_record_idle(pfdev);
panfrost_device_reset(pfdev); panfrost_device_reset(pfdev);
for (i = 0; i < NUM_JOB_SLOTS; i++) for (i = 0; i < NUM_JOB_SLOTS; i++)
...@@ -467,7 +466,7 @@ static irqreturn_t panfrost_job_irq_handler(int irq, void *data) ...@@ -467,7 +466,7 @@ static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
pfdev->jobs[j] = NULL; pfdev->jobs[j] = NULL;
panfrost_mmu_as_put(pfdev, &job->file_priv->mmu); panfrost_mmu_as_put(pfdev, &job->file_priv->mmu);
panfrost_devfreq_record_transition(pfdev, j); panfrost_devfreq_record_idle(pfdev);
dma_fence_signal_locked(job->done_fence); dma_fence_signal_locked(job->done_fence);
pm_runtime_put_autosuspend(pfdev->dev); pm_runtime_put_autosuspend(pfdev->dev);
...@@ -568,14 +567,14 @@ int panfrost_job_is_idle(struct panfrost_device *pfdev) ...@@ -568,14 +567,14 @@ int panfrost_job_is_idle(struct panfrost_device *pfdev)
struct panfrost_job_slot *js = pfdev->js; struct panfrost_job_slot *js = pfdev->js;
int i; int i;
/* Check whether the hardware is idle */
if (atomic_read(&pfdev->devfreq.busy_count))
return false;
for (i = 0; i < NUM_JOB_SLOTS; i++) { for (i = 0; i < NUM_JOB_SLOTS; i++) {
/* If there are any jobs in the HW queue, we're not idle */ /* If there are any jobs in the HW queue, we're not idle */
if (atomic_read(&js->queue[i].sched.hw_rq_count)) if (atomic_read(&js->queue[i].sched.hw_rq_count))
return false; return false;
/* Check whether the hardware is idle */
if (pfdev->devfreq.slot[i].busy)
return false;
} }
return true; return true;
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment