Re: [RFT][PATCH v2 1/2] PM / domains: Rework governor code to be more consistent

From: Ulf Hansson
Date: Mon Nov 06 2017 - 07:10:10 EST


On 3 November 2017 at 12:47, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
>
> The genpd governor currently uses negative PM QoS values to indicate
> the "no suspend" condition and 0 as "no restriction", but it doesn't
> use them consistently. Moreover, it tries to refresh QoS values for
> already suspended devices in a quite questionable way.
>
> For the above reasons, rework it to be a bit more consistent.
>
> First off, note that dev_pm_qos_read_value() in
> dev_update_qos_constraint() and __default_power_down_ok() is
> evaluated for devices in suspend. Moreover, that only happens if the
> effective_constraint_ns value for them is negative (meaning "no
> suspend"). It is not evaluated in any other cases, so effectively
> the QoS values are only updated for devices in suspend that should
> not have been suspended in the first place. In all of the other
> cases, the QoS values taken into account are the effective ones from
> the time before the device has been suspended, so generally devices
> need to be resumed and suspended again for new QoS values to take
> effect anyway. Thus evaluating dev_update_qos_constraint() in
> those two places doesn't make sense at all, so drop it.
>
> Second, initialize effective_constraint_ns to 0 ("no constraint")
> rather than to (-1) ("no suspend"), which makes more sense in
> general and in case effective_constraint_ns is never updated
> (the device is in suspend all the time or it is never suspended)
> it doesn't affect the device's parent and so on.
>
> Finally, rework default_suspend_ok() to explicitly handle the
> "no restriction" special case.
>
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> ---
> drivers/base/power/domain.c | 2 -
> drivers/base/power/domain_governor.c | 61 +++++++++++++++++++++--------------
> 2 files changed, 38 insertions(+), 25 deletions(-)
>
> Index: linux-pm/drivers/base/power/domain.c
> ===================================================================
> --- linux-pm.orig/drivers/base/power/domain.c
> +++ linux-pm/drivers/base/power/domain.c
> @@ -1331,7 +1331,7 @@ static struct generic_pm_domain_data *ge
>
> gpd_data->base.dev = dev;
> gpd_data->td.constraint_changed = true;
> - gpd_data->td.effective_constraint_ns = -1;
> + gpd_data->td.effective_constraint_ns = 0;
> gpd_data->nb.notifier_call = genpd_dev_pm_qos_notifier;
>
> spin_lock_irq(&dev->power.lock);
> Index: linux-pm/drivers/base/power/domain_governor.c
> ===================================================================
> --- linux-pm.orig/drivers/base/power/domain_governor.c
> +++ linux-pm/drivers/base/power/domain_governor.c
> @@ -14,22 +14,22 @@
> static int dev_update_qos_constraint(struct device *dev, void *data)
> {
> s64 *constraint_ns_p = data;
> - s32 constraint_ns = -1;
> -
> - if (dev->power.subsys_data && dev->power.subsys_data->domain_data)
> - constraint_ns = dev_gpd_data(dev)->td.effective_constraint_ns;
> + s64 constraint_ns;
>
> - if (constraint_ns < 0) {
> - constraint_ns = dev_pm_qos_read_value(dev);
> - constraint_ns *= NSEC_PER_USEC;
> - }
> - if (constraint_ns == 0)
> + if (!dev->power.subsys_data || !dev->power.subsys_data->domain_data)
> return 0;
>
> /*
> - * constraint_ns cannot be negative here, because the device has been
> - * suspended.
> + * Only take suspend-time QoS constraints of devices into account,
> + * because constraints updated after the device has been suspended are
> + * not guaranteed to be taken into account anyway. In order for them
> + * to take effect, the device has to be resumed and suspended again.
> */

This means a change in behavior, because earlier we took into account
QoS values for child devices that were not attached to a genpd.

Is there any reason you think we should change this, or is it just
something you overlooked here?

I understand, that if the QoS constraint has been updated after such
child device has been suspended, it's tricky to take a correct
decision.

To really solve this, we would either have to register a QoS notifier
for all children devices that has its parent attached to a genpd, or
always runtime resume devices before updating QoS constraints.

Non of these options is perfect, so perhaps we should consider a "best
effort" approach instead? Whatever that may be.

> + constraint_ns = dev_gpd_data(dev)->td.effective_constraint_ns;
> + /* 0 means "no constraint" */
> + if (constraint_ns == 0)
> + return 0;
> +
> if (constraint_ns < *constraint_ns_p || *constraint_ns_p == 0)
> *constraint_ns_p = constraint_ns;
>
> @@ -76,14 +76,29 @@ static bool default_suspend_ok(struct de
> device_for_each_child(dev, &constraint_ns,
> dev_update_qos_constraint);
>
> - if (constraint_ns > 0) {
> + if (constraint_ns == 0) {
> + /* "No restriction", so the device is allowed to suspend. */
> + td->effective_constraint_ns = 0;
> + td->cached_suspend_ok = true;
> + } else {
> + /*
> + * constraint_ns must be positive here, because the children
> + * walked above are all suspended, so effective_constraint_ns
> + * cannot be negative for them.
> + */
> constraint_ns -= td->suspend_latency_ns +
> td->resume_latency_ns;
> - if (constraint_ns == 0)
> + /*
> + * effective_constraint_ns is negative already and
> + * cached_suspend_ok is false, so if the computed value is not
> + * positive, return right away.
> + */
> + if (constraint_ns <= 0)
> return false;
> +
> + td->effective_constraint_ns = constraint_ns;
> + td->cached_suspend_ok = true;
> }
> - td->effective_constraint_ns = constraint_ns;
> - td->cached_suspend_ok = constraint_ns >= 0;
>
> /*
> * The children have been suspended already, so we don't need to take
> @@ -144,18 +159,16 @@ static bool __default_power_down_ok(stru
> */
> td = &to_gpd_data(pdd)->td;
> constraint_ns = td->effective_constraint_ns;
> - /* default_suspend_ok() need not be called before us. */
> - if (constraint_ns < 0) {
> - constraint_ns = dev_pm_qos_read_value(pdd->dev);
> - constraint_ns *= NSEC_PER_USEC;
> - }
> + /*
> + * Negative values mean "no suspend at all" and this runs only
> + * when all devices in the domain are suspended, so it must be
> + * 0 at least.
> + *
> + * 0 means "no constraint"
> + */
> if (constraint_ns == 0)
> continue;
>
> - /*
> - * constraint_ns cannot be negative here, because the device has
> - * been suspended.
> - */
> if (constraint_ns <= off_on_time_ns)
> return false;
>
>

Kind regards
Uffe