[PATCH 2/2] drm: Measure Self Refresh Entry/Exit times to avoid thrashing
Daniel Vetter
daniel at ffwll.ch
Wed Sep 18 08:17:40 UTC 2019
On Tue, Sep 17, 2019 at 04:04:33PM -0400, Sean Paul wrote:
> From: Sean Paul <seanpaul at chromium.org>
>
> Currently the self refresh idle timer is a const set by the crtc. This
> is fine if the self refresh entry/exit times are well-known for all
> panels used on that crtc. However panels and workloads can vary quite a
> bit, and a timeout which works well for one doesn't work well for
> another.
>
> In the extreme, if the timeout is too short we could get in a situation
> where the self refresh exits are taking so long we queue up a self refresh
> entry before the exit commit is even finished.
>
> This patch changes the idle timeout to a moving average of the entry
> times + a moving average of exit times + the crtc constant.
>
> This patch was tested on rockchip, with a "kevin" CrOS panel the idle
> delay averages out to about ~235ms (35 entry + 100 exit + 100 const). On
> the same board, the "bob" panel idle delay lands around ~340ms (90 entry
> + 150 exit + 100 const).
>
> Signed-off-by: Sean Paul <seanpaul at chromium.org>
> ---
> drivers/gpu/drm/drm_atomic_helper.c | 20 +++++++
> drivers/gpu/drm/drm_self_refresh_helper.c | 71 ++++++++++++++++++++++-
> include/drm/drm_self_refresh_helper.h | 2 +
> 3 files changed, 92 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
> index 9d7e4da6c292..3f13fa9a9e24 100644
> --- a/drivers/gpu/drm/drm_atomic_helper.c
> +++ b/drivers/gpu/drm/drm_atomic_helper.c
> @@ -26,6 +26,7 @@
> */
>
> #include <linux/dma-fence.h>
> +#include <linux/ktime.h>
>
> #include <drm/drm_atomic.h>
> #include <drm/drm_atomic_helper.h>
> @@ -1570,9 +1571,23 @@ static void commit_tail(struct drm_atomic_state *old_state)
> {
> struct drm_device *dev = old_state->dev;
> const struct drm_mode_config_helper_funcs *funcs;
> + ktime_t start;
> + s64 commit_time_ms;
>
> funcs = dev->mode_config.helper_private;
>
> + /*
> + * We're measuring the _entire_ commit, so the time will vary depending
> + * on how many fences and objects are involved. For the purposes of self
> + * refresh, this is desirable since it'll give us an idea of how
> + * congested things are. This will inform our decision on how often we
> + * should enter self refresh after idle.
> + *
> + * These times will be averaged out in the self refresh helpers to avoid
> + * overreacting over one outlier frame
> + */
> + start = ktime_get();
> +
> drm_atomic_helper_wait_for_fences(dev, old_state, false);
>
> drm_atomic_helper_wait_for_dependencies(old_state);
> @@ -1582,6 +1597,11 @@ static void commit_tail(struct drm_atomic_state *old_state)
> else
> drm_atomic_helper_commit_tail(old_state);
>
> + commit_time_ms = ktime_ms_delta(ktime_get(), start);
> + if (commit_time_ms > 0)
> + drm_self_refresh_helper_update_avg_times(old_state,
> + (unsigned long)commit_time_ms);
> +
> drm_atomic_helper_commit_cleanup_done(old_state);
>
> drm_atomic_state_put(old_state);
> diff --git a/drivers/gpu/drm/drm_self_refresh_helper.c b/drivers/gpu/drm/drm_self_refresh_helper.c
> index 9095cebf2147..522430f8eef1 100644
> --- a/drivers/gpu/drm/drm_self_refresh_helper.c
> +++ b/drivers/gpu/drm/drm_self_refresh_helper.c
> @@ -5,6 +5,7 @@
> * Authors:
> * Sean Paul <seanpaul at chromium.org>
> */
> +#include <linux/average.h>
> #include <linux/bitops.h>
> #include <linux/slab.h>
> #include <linux/workqueue.h>
> @@ -50,10 +51,16 @@
> * atomic_check when &drm_crtc_state.self_refresh_active is true.
> */
>
> +DECLARE_EWMA(psr_time, 4, 4)
> +
> struct drm_self_refresh_data {
> struct drm_crtc *crtc;
> struct delayed_work entry_work;
> unsigned int entry_delay_ms;
> +
> + struct mutex avg_mutex;
> + struct ewma_psr_time entry_avg_ms;
> + struct ewma_psr_time exit_avg_ms;
> };
>
> static void drm_self_refresh_helper_entry_work(struct work_struct *work)
> @@ -121,6 +128,59 @@ static void drm_self_refresh_helper_entry_work(struct work_struct *work)
> drm_modeset_acquire_fini(&ctx);
> }
>
> +/**
> + * drm_self_refresh_helper_update_avg_times - Updates a crtc's SR time averages
> + * @state: the state which has just been applied to hardware
> + * @commit_time_ms: the amount of time in ms that this commit took to complete
> + *
> + * Called after &drm_mode_config_funcs.atomic_commit_tail, this function will
> + * update the average entry/exit self refresh times on self refresh transitions.
> + * These averages will be used when calculating how long to delay before
> + * entering self refresh mode after activity.
> + */
> +void drm_self_refresh_helper_update_avg_times(struct drm_atomic_state *state,
> + unsigned int commit_time_ms)
> +{
> + struct drm_crtc *crtc;
> + struct drm_crtc_state *old_crtc_state, *new_crtc_state;
> + int i;
> +
> + for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state,
> + new_crtc_state, i) {
> + struct drm_self_refresh_data *sr_data = crtc->self_refresh_data;
> + struct ewma_psr_time *time;
> +
> + if (old_crtc_state->self_refresh_active ==
> + new_crtc_state->self_refresh_active)
> + continue;
> +
> + if (new_crtc_state->self_refresh_active)
> + time = &sr_data->entry_avg_ms;
> + else
> + time = &sr_data->exit_avg_ms;
> +
> + /*
> + * It might be nice if we could rely on &drm_crtc.mutex to
> + * protect &drm_self_refresh_data.exit_avg_ms, as we do with
> + * &drm_self_refresh_data.entry_avg_ms, but there are a few
> + * reasons why a separate lock is a better choice:
> + * - We can't rely on &drm_crtc.mutex being held here if we're
> + * doing a nonblocking commit
> + * - We can't grab &drm_crtc.mutex here since drm_modeset_lock()
> + * doesn't tell us whether the lock was already held in the
> + * acquire context (it eats -EALREADY), so we can't tell if we
> + * should drop it or not
> + * - We don't need such a heavy-handed lock for what we're
> + * trying to do here, commit ordering doesn't matter, so a
> + * point-of-use lock will be less contentious
> + */
This comment here feels rather misplaced, I think better to put that into
the commit message. Elaborate locking analysis in code comments tends to
not age well ime.
> + mutex_lock(&sr_data->avg_mutex);
> + ewma_psr_time_add(time, commit_time_ms);
> + mutex_unlock(&sr_data->avg_mutex);
> + }
> +}
> +EXPORT_SYMBOL(drm_self_refresh_helper_update_avg_times);
> +
> /**
> * drm_self_refresh_helper_alter_state - Alters the atomic state for SR exit
> * @state: the state currently being checked
> @@ -152,6 +212,7 @@ void drm_self_refresh_helper_alter_state(struct drm_atomic_state *state)
>
> for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
> struct drm_self_refresh_data *sr_data;
> + unsigned int delay;
>
> /* Don't trigger the entry timer when we're already in SR */
> if (crtc_state->self_refresh_active)
> @@ -161,8 +222,13 @@ void drm_self_refresh_helper_alter_state(struct drm_atomic_state *state)
> if (!sr_data)
> continue;
>
> + mutex_lock(&sr_data->avg_mutex);
> + delay = ewma_psr_time_read(&sr_data->entry_avg_ms) +
> + ewma_psr_time_read(&sr_data->exit_avg_ms) +
> + sr_data->entry_delay_ms;
Since you auto-tune now, I'd remove the entry_delay_ms thing outright, and
just use 2x the entry+exit times for this. That would scale lot better
from real quick panels that take only 1 frame to enter/exit (i.e. no real
delay) ot real horrrors that might take even longer than the panels you
have. Adding a constant 100ms still assumes that entry+exit aren't too far
away from that 100ms value you hardcoded in drivers.
Should we have a debug print somewhere that tells us the self-refresh
delay? I'd expect a "why am I not entering sr?" moment with this
otherwise.
> + mutex_unlock(&sr_data->avg_mutex);
> mod_delayed_work(system_wq, &sr_data->entry_work,
> - msecs_to_jiffies(sr_data->entry_delay_ms));
> + msecs_to_jiffies(delay));
> }
> }
> EXPORT_SYMBOL(drm_self_refresh_helper_alter_state);
> @@ -191,6 +257,9 @@ int drm_self_refresh_helper_init(struct drm_crtc *crtc,
> drm_self_refresh_helper_entry_work);
> sr_data->entry_delay_ms = entry_delay_ms;
> sr_data->crtc = crtc;
> + mutex_init(&sr_data->avg_mutex);
> + ewma_psr_time_init(&sr_data->entry_avg_ms);
> + ewma_psr_time_init(&sr_data->exit_avg_ms);
>
> crtc->self_refresh_data = sr_data;
> return 0;
> diff --git a/include/drm/drm_self_refresh_helper.h b/include/drm/drm_self_refresh_helper.h
> index 397a583ccca7..ff777690c564 100644
> --- a/include/drm/drm_self_refresh_helper.h
> +++ b/include/drm/drm_self_refresh_helper.h
> @@ -12,6 +12,8 @@ struct drm_atomic_state;
> struct drm_crtc;
>
> void drm_self_refresh_helper_alter_state(struct drm_atomic_state *state);
> +void drm_self_refresh_helper_update_avg_times(struct drm_atomic_state *state,
> + unsigned int commit_time_ms);
>
> int drm_self_refresh_helper_init(struct drm_crtc *crtc,
> unsigned int entry_delay_ms);
With the bikesheds addressed somehow:
Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> --
> Sean Paul, Software Engineer, Google / Chromium OS
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the dri-devel
mailing list