1 // SPDX-License-Identifier: GPL-2.0
3 * drivers/base/power/domain_governor.c - Governors for device PM domains.
7 #include <linux/kernel.h>
8 #include <linux/pm_domain.h>
9 #include <linux/pm_qos.h>
10 #include <linux/hrtimer.h>
11 #include <linux/cpuidle.h>
12 #include <linux/cpumask.h>
13 #include <linux/ktime.h>
15 static int dev_update_qos_constraint(struct device *dev, void *data)
17 s64 *constraint_ns_p = data;
20 if (dev->power.subsys_data && dev->power.subsys_data->domain_data) {
21 struct gpd_timing_data *td = dev_gpd_data(dev)->td;
24 * Only take suspend-time QoS constraints of devices into
25 * account, because constraints updated after the device has
26 * been suspended are not guaranteed to be taken into account
27 * anyway. In order for them to take effect, the device has to
28 * be resumed and suspended again.
30 constraint_ns = td ? td->effective_constraint_ns :
31 PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
34 * The child is not in a domain and there's no info on its
35 * suspend/resume latencies, so assume them to be negligible and
36 * take its current PM QoS constraint (that's the only thing
37 * known at this point anyway).
39 constraint_ns = dev_pm_qos_read_value(dev, DEV_PM_QOS_RESUME_LATENCY);
40 constraint_ns *= NSEC_PER_USEC;
43 if (constraint_ns < *constraint_ns_p)
44 *constraint_ns_p = constraint_ns;
50 * default_suspend_ok - Default PM domain governor routine to suspend devices.
51 * @dev: Device to check.
53 static bool default_suspend_ok(struct device *dev)
55 struct gpd_timing_data *td = dev_gpd_data(dev)->td;
59 dev_dbg(dev, "%s()\n", __func__);
61 spin_lock_irqsave(&dev->power.lock, flags);
63 if (!td->constraint_changed) {
64 bool ret = td->cached_suspend_ok;
66 spin_unlock_irqrestore(&dev->power.lock, flags);
69 td->constraint_changed = false;
70 td->cached_suspend_ok = false;
71 td->effective_constraint_ns = 0;
72 constraint_ns = __dev_pm_qos_resume_latency(dev);
74 spin_unlock_irqrestore(&dev->power.lock, flags);
76 if (constraint_ns == 0)
79 constraint_ns *= NSEC_PER_USEC;
81 * We can walk the children without any additional locking, because
82 * they all have been suspended at this point and their
83 * effective_constraint_ns fields won't be modified in parallel with us.
85 if (!dev->power.ignore_children)
86 device_for_each_child(dev, &constraint_ns,
87 dev_update_qos_constraint);
89 if (constraint_ns == PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS) {
90 /* "No restriction", so the device is allowed to suspend. */
91 td->effective_constraint_ns = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS;
92 td->cached_suspend_ok = true;
93 } else if (constraint_ns == 0) {
95 * This triggers if one of the children that don't belong to a
96 * domain has a zero PM QoS constraint and it's better not to
97 * suspend then. effective_constraint_ns is zero already and
98 * cached_suspend_ok is false, so bail out.
102 constraint_ns -= td->suspend_latency_ns +
103 td->resume_latency_ns;
105 * effective_constraint_ns is zero already and cached_suspend_ok
106 * is false, so if the computed value is not positive, return
109 if (constraint_ns <= 0)
112 td->effective_constraint_ns = constraint_ns;
113 td->cached_suspend_ok = true;
117 * The children have been suspended already, so we don't need to take
118 * their suspend latencies into account here.
120 return td->cached_suspend_ok;
123 static void update_domain_next_wakeup(struct generic_pm_domain *genpd, ktime_t now)
125 ktime_t domain_wakeup = KTIME_MAX;
127 struct pm_domain_data *pdd;
128 struct gpd_link *link;
130 if (!(genpd->flags & GENPD_FLAG_MIN_RESIDENCY))
134 * Devices that have a predictable wakeup pattern, may specify
135 * their next wakeup. Let's find the next wakeup from all the
136 * devices attached to this domain and from all the sub-domains.
137 * It is possible that component's a next wakeup may have become
138 * stale when we read that here. We will ignore to ensure the domain
139 * is able to enter its optimal idle state.
141 list_for_each_entry(pdd, &genpd->dev_list, list_node) {
142 next_wakeup = to_gpd_data(pdd)->td->next_wakeup;
143 if (next_wakeup != KTIME_MAX && !ktime_before(next_wakeup, now))
144 if (ktime_before(next_wakeup, domain_wakeup))
145 domain_wakeup = next_wakeup;
148 list_for_each_entry(link, &genpd->parent_links, parent_node) {
149 struct genpd_governor_data *cgd = link->child->gd;
151 next_wakeup = cgd ? cgd->next_wakeup : KTIME_MAX;
152 if (next_wakeup != KTIME_MAX && !ktime_before(next_wakeup, now))
153 if (ktime_before(next_wakeup, domain_wakeup))
154 domain_wakeup = next_wakeup;
157 genpd->gd->next_wakeup = domain_wakeup;
160 static bool next_wakeup_allows_state(struct generic_pm_domain *genpd,
161 unsigned int state, ktime_t now)
163 ktime_t domain_wakeup = genpd->gd->next_wakeup;
164 s64 idle_time_ns, min_sleep_ns;
166 min_sleep_ns = genpd->states[state].power_off_latency_ns +
167 genpd->states[state].residency_ns;
169 idle_time_ns = ktime_to_ns(ktime_sub(domain_wakeup, now));
171 return idle_time_ns >= min_sleep_ns;
174 static bool __default_power_down_ok(struct dev_pm_domain *pd,
177 struct generic_pm_domain *genpd = pd_to_genpd(pd);
178 struct gpd_link *link;
179 struct pm_domain_data *pdd;
183 off_on_time_ns = genpd->states[state].power_off_latency_ns +
184 genpd->states[state].power_on_latency_ns;
186 min_off_time_ns = -1;
188 * Check if subdomains can be off for enough time.
190 * All subdomains have been powered off already at this point.
192 list_for_each_entry(link, &genpd->parent_links, parent_node) {
193 struct genpd_governor_data *cgd = link->child->gd;
195 s64 sd_max_off_ns = cgd ? cgd->max_off_time_ns : -1;
197 if (sd_max_off_ns < 0)
201 * Check if the subdomain is allowed to be off long enough for
202 * the current domain to turn off and on (that's how much time
203 * it will have to wait worst case).
205 if (sd_max_off_ns <= off_on_time_ns)
208 if (min_off_time_ns > sd_max_off_ns || min_off_time_ns < 0)
209 min_off_time_ns = sd_max_off_ns;
213 * Check if the devices in the domain can be off enough time.
215 list_for_each_entry(pdd, &genpd->dev_list, list_node) {
216 struct gpd_timing_data *td;
220 * Check if the device is allowed to be off long enough for the
221 * domain to turn off and on (that's how much time it will
222 * have to wait worst case).
224 td = to_gpd_data(pdd)->td;
225 constraint_ns = td->effective_constraint_ns;
227 * Zero means "no suspend at all" and this runs only when all
228 * devices in the domain are suspended, so it must be positive.
230 if (constraint_ns == PM_QOS_RESUME_LATENCY_NO_CONSTRAINT_NS)
233 if (constraint_ns <= off_on_time_ns)
236 if (min_off_time_ns > constraint_ns || min_off_time_ns < 0)
237 min_off_time_ns = constraint_ns;
241 * If the computed minimum device off time is negative, there are no
242 * latency constraints, so the domain can spend arbitrary time in the
245 if (min_off_time_ns < 0)
249 * The difference between the computed minimum subdomain or device off
250 * time and the time needed to turn the domain on is the maximum
251 * theoretical time this domain can spend in the "off" state.
253 genpd->gd->max_off_time_ns = min_off_time_ns -
254 genpd->states[state].power_on_latency_ns;
259 * _default_power_down_ok - Default generic PM domain power off governor routine.
260 * @pd: PM domain to check.
261 * @now: current ktime.
263 * This routine must be executed under the PM domain's lock.
265 static bool _default_power_down_ok(struct dev_pm_domain *pd, ktime_t now)
267 struct generic_pm_domain *genpd = pd_to_genpd(pd);
268 struct genpd_governor_data *gd = genpd->gd;
269 int state_idx = genpd->state_count - 1;
270 struct gpd_link *link;
273 * Find the next wakeup from devices that can determine their own wakeup
274 * to find when the domain would wakeup and do it for every device down
275 * the hierarchy. It is not worth while to sleep if the state's residency
278 update_domain_next_wakeup(genpd, now);
279 if ((genpd->flags & GENPD_FLAG_MIN_RESIDENCY) && (gd->next_wakeup != KTIME_MAX)) {
280 /* Let's find out the deepest domain idle state, the devices prefer */
281 while (state_idx >= 0) {
282 if (next_wakeup_allows_state(genpd, state_idx, now)) {
283 gd->max_off_time_changed = true;
291 gd->cached_power_down_ok = false;
296 if (!gd->max_off_time_changed) {
297 genpd->state_idx = gd->cached_power_down_state_idx;
298 return gd->cached_power_down_ok;
302 * We have to invalidate the cached results for the parents, so
303 * use the observation that default_power_down_ok() is not
304 * going to be called for any parent until this instance
307 list_for_each_entry(link, &genpd->child_links, child_node) {
308 struct genpd_governor_data *pgd = link->parent->gd;
311 pgd->max_off_time_changed = true;
314 gd->max_off_time_ns = -1;
315 gd->max_off_time_changed = false;
316 gd->cached_power_down_ok = true;
319 * Find a state to power down to, starting from the state
320 * determined by the next wakeup.
322 while (!__default_power_down_ok(pd, state_idx)) {
323 if (state_idx == 0) {
324 gd->cached_power_down_ok = false;
331 genpd->state_idx = state_idx;
332 gd->cached_power_down_state_idx = genpd->state_idx;
333 return gd->cached_power_down_ok;
336 static bool default_power_down_ok(struct dev_pm_domain *pd)
338 return _default_power_down_ok(pd, ktime_get());
341 #ifdef CONFIG_CPU_IDLE
342 static bool cpu_power_down_ok(struct dev_pm_domain *pd)
344 struct generic_pm_domain *genpd = pd_to_genpd(pd);
345 struct cpuidle_device *dev;
346 ktime_t domain_wakeup, next_hrtimer;
347 ktime_t now = ktime_get();
348 s64 idle_duration_ns;
351 /* Validate dev PM QoS constraints. */
352 if (!_default_power_down_ok(pd, now))
355 if (!(genpd->flags & GENPD_FLAG_CPU_DOMAIN))
359 * Find the next wakeup for any of the online CPUs within the PM domain
360 * and its subdomains. Note, we only need the genpd->cpus, as it already
361 * contains a mask of all CPUs from subdomains.
363 domain_wakeup = ktime_set(KTIME_SEC_MAX, 0);
364 for_each_cpu_and(cpu, genpd->cpus, cpu_online_mask) {
365 dev = per_cpu(cpuidle_devices, cpu);
367 next_hrtimer = READ_ONCE(dev->next_hrtimer);
368 if (ktime_before(next_hrtimer, domain_wakeup))
369 domain_wakeup = next_hrtimer;
373 /* The minimum idle duration is from now - until the next wakeup. */
374 idle_duration_ns = ktime_to_ns(ktime_sub(domain_wakeup, now));
375 if (idle_duration_ns <= 0)
379 * Find the deepest idle state that has its residency value satisfied
380 * and by also taking into account the power off latency for the state.
381 * Start at the state picked by the dev PM QoS constraint validation.
383 i = genpd->state_idx;
385 if (idle_duration_ns >= (genpd->states[i].residency_ns +
386 genpd->states[i].power_off_latency_ns)) {
387 genpd->state_idx = i;
395 struct dev_power_governor pm_domain_cpu_gov = {
396 .suspend_ok = default_suspend_ok,
397 .power_down_ok = cpu_power_down_ok,
401 struct dev_power_governor simple_qos_governor = {
402 .suspend_ok = default_suspend_ok,
403 .power_down_ok = default_power_down_ok,
407 * pm_genpd_gov_always_on - A governor implementing an always-on policy
409 struct dev_power_governor pm_domain_always_on_gov = {
410 .suspend_ok = default_suspend_ok,