Skip to contentSkip to conversation

sched/psi: Allow unprivileged PSI polling

PATCH Superseded Domenico Cerasuolo 2 participants 6 messages Mar 30, 2023

Subsystems (4)

v5 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v5 cover sched/psi: Allow unprivileged PSI polling

PSI offers 2 mechanisms to get information about a specific resource pressure. One is reading from /proc/pressure/<resource>, which gives average pressures aggregated every 2s. The other is creating a pollable fd for a specific resource and cgroup.

The trigger creation requires CAP_SYS_RESOURCE, and gives the possibility to pick specific time window and threshold, spawing an RT thread to aggregate the data.

Systemd would like to provide containers the option to monitor pressure on their own cgroup and sub-cgroups. For example, if systemd launches a container that itself then launches services, the container should have the ability to poll() for pressure in individual services. But neither the container nor the services are privileged.

The series is implemented in 4 steps in order to reduce the noise of the change.

V5:
- few suggested code style changes in psi.c and psy_types.h
- fix trigger destruction logic now using proper locks

V4:
- fixed psi_open leftover usage in IRQ accounting (patch 4/4)

V3:
- restored renaming patch (#2 of 4) as suggested in review
- rebased #3 and #4 on the renaming commit

V2:
- removed renaming patch (previous 2/4) and applied suggested solution
- changed update_triggers side effect removal as suggested in review
- rebased core patch on other V2 changes

Domenico Cerasuolo (4):
sched/psi: rearrange polling code in preparation sched/psi: rename existing poll members in preparation sched/psi: extract update_triggers side effect sched/psi: allow unprivileged polling of N*2s period

 Documentation/accounting/psi.rst |   4 +
 include/linux/psi.h              |   2 +-
 include/linux/psi_types.h        |  43 +--
 kernel/cgroup/cgroup.c           |   2 +-
 kernel/sched/psi.c               | 458 +++++++++++++++++--------------
 5 files changed, 280 insertions(+), 229 deletions(-)
Signature · 1 line
2.34.1
#1 lore raw

PATCH v5 1/4 sched/psi: rearrange polling code in preparation

Move a few functions up in the file to avoid forward declaration needed in the patch implementing unprivileged PSI triggers.

---
 kernel/sched/psi.c | 196 ++++++++++++++++++++++-----------------------
 1 file changed, 98 insertions(+), 98 deletions(-)
kernel/sched/psi.c +98 -98
index 02e011cabe91..fe9269f1d2a4 100644
@@ -384,92 +384,6 @@ static void collect_percpu_times(struct psi_group *group,
 		*pchanged_states = changed_states;
 }
 
-static u64 update_averages(struct psi_group *group, u64 now)
-{
-	unsigned long missed_periods = 0;
-	u64 expires, period;
-	u64 avg_next_update;
-	int s;
-
-	/* avgX= */
-	expires = group->avg_next_update;
-	if (now - expires >= psi_period)
-		missed_periods = div_u64(now - expires, psi_period);
-
-	/*
-	 * The periodic clock tick can get delayed for various
-	 * reasons, especially on loaded systems. To avoid clock
-	 * drift, we schedule the clock in fixed psi_period intervals.
-	 * But the deltas we sample out of the per-cpu buckets above
-	 * are based on the actual time elapsing between clock ticks.
-	 */
-	avg_next_update = expires + ((1 + missed_periods) * psi_period);
-	period = now - (group->avg_last_update + (missed_periods * psi_period));
-	group->avg_last_update = now;
-
-	for (s = 0; s < NR_PSI_STATES - 1; s++) {
-		u32 sample;
-
-		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
-		/*
-		 * Due to the lockless sampling of the time buckets,
-		 * recorded time deltas can slip into the next period,
-		 * which under full pressure can result in samples in
-		 * excess of the period length.
-		 *
-		 * We don't want to report non-sensical pressures in
-		 * excess of 100%, nor do we want to drop such events
-		 * on the floor. Instead we punt any overage into the
-		 * future until pressure subsides. By doing this we
-		 * don't underreport the occurring pressure curve, we
-		 * just report it delayed by one period length.
-		 *
-		 * The error isn't cumulative. As soon as another
-		 * delta slips from a period P to P+1, by definition
-		 * it frees up its time T in P.
-		 */
-		if (sample > period)
-			sample = period;
-		group->avg_total[s] += sample;
-		calc_avgs(group->avg[s], missed_periods, sample, period);
-	}
-
-	return avg_next_update;
-}
-
-static void psi_avgs_work(struct work_struct *work)
-{
-	struct delayed_work *dwork;
-	struct psi_group *group;
-	u32 changed_states;
-	u64 now;
-
-	dwork = to_delayed_work(work);
-	group = container_of(dwork, struct psi_group, avgs_work);
-
-	mutex_lock(&group->avgs_lock);
-
-	now = sched_clock();
-
-	collect_percpu_times(group, PSI_AVGS, &changed_states);
-	/*
-	 * If there is task activity, periodically fold the per-cpu
-	 * times and feed samples into the running averages. If things
-	 * are idle and there is no data to process, stop the clock.
-	 * Once restarted, we'll catch up the running averages in one
-	 * go - see calc_avgs() and missed_periods.
-	 */
-	if (now >= group->avg_next_update)
-		group->avg_next_update = update_averages(group, now);
-
-	if (changed_states & PSI_STATE_RESCHEDULE) {
-		schedule_delayed_work(dwork, nsecs_to_jiffies(
-				group->avg_next_update - now) + 1);
-	}
-
-	mutex_unlock(&group->avgs_lock);
-}
-
 /* Trigger tracking window manipulations */
 static void window_reset(struct psi_window *win, u64 now, u64 value,
 			 u64 prev_growth)
@@ -516,18 +430,6 @@ static u64 window_update(struct psi_window *win, u64 now, u64 value)
 	return growth;
 }
 
-static void init_triggers(struct psi_group *group, u64 now)
-{
-	struct psi_trigger *t;
-
-	list_for_each_entry(t, &group->triggers, node)
-		window_reset(&t->win, now,
-				group->total[PSI_POLL][t->state], 0);
-	memcpy(group->polling_total, group->total[PSI_POLL],
-		   sizeof(group->polling_total));
-	group->polling_next_update = now + group->poll_min_period;
-}
-
 static u64 update_triggers(struct psi_group *group, u64 now)
 {
 	struct psi_trigger *t;
@@ -590,6 +492,104 @@ static u64 update_triggers(struct psi_group *group, u64 now)
 	return now + group->poll_min_period;
 }
 
+static u64 update_averages(struct psi_group *group, u64 now)
+{
+	unsigned long missed_periods = 0;
+	u64 expires, period;
+	u64 avg_next_update;
+	int s;
+
+	/* avgX= */
+	expires = group->avg_next_update;
+	if (now - expires >= psi_period)
+		missed_periods = div_u64(now - expires, psi_period);
+
+	/*
+	 * The periodic clock tick can get delayed for various
+	 * reasons, especially on loaded systems. To avoid clock
+	 * drift, we schedule the clock in fixed psi_period intervals.
+	 * But the deltas we sample out of the per-cpu buckets above
+	 * are based on the actual time elapsing between clock ticks.
+	 */
+	avg_next_update = expires + ((1 + missed_periods) * psi_period);
+	period = now - (group->avg_last_update + (missed_periods * psi_period));
+	group->avg_last_update = now;
+
+	for (s = 0; s < NR_PSI_STATES - 1; s++) {
+		u32 sample;
+
+		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
+		/*
+		 * Due to the lockless sampling of the time buckets,
+		 * recorded time deltas can slip into the next period,
+		 * which under full pressure can result in samples in
+		 * excess of the period length.
+		 *
+		 * We don't want to report non-sensical pressures in
+		 * excess of 100%, nor do we want to drop such events
+		 * on the floor. Instead we punt any overage into the
+		 * future until pressure subsides. By doing this we
+		 * don't underreport the occurring pressure curve, we
+		 * just report it delayed by one period length.
+		 *
+		 * The error isn't cumulative. As soon as another
+		 * delta slips from a period P to P+1, by definition
+		 * it frees up its time T in P.
+		 */
+		if (sample > period)
+			sample = period;
+		group->avg_total[s] += sample;
+		calc_avgs(group->avg[s], missed_periods, sample, period);
+	}
+
+	return avg_next_update;
+}
+
+static void psi_avgs_work(struct work_struct *work)
+{
+	struct delayed_work *dwork;
+	struct psi_group *group;
+	u32 changed_states;
+	u64 now;
+
+	dwork = to_delayed_work(work);
+	group = container_of(dwork, struct psi_group, avgs_work);
+
+	mutex_lock(&group->avgs_lock);
+
+	now = sched_clock();
+
+	collect_percpu_times(group, PSI_AVGS, &changed_states);
+	/*
+	 * If there is task activity, periodically fold the per-cpu
+	 * times and feed samples into the running averages. If things
+	 * are idle and there is no data to process, stop the clock.
+	 * Once restarted, we'll catch up the running averages in one
+	 * go - see calc_avgs() and missed_periods.
+	 */
+	if (now >= group->avg_next_update)
+		group->avg_next_update = update_averages(group, now);
+
+	if (changed_states & PSI_STATE_RESCHEDULE) {
+		schedule_delayed_work(dwork, nsecs_to_jiffies(
+				group->avg_next_update - now) + 1);
+	}
+
+	mutex_unlock(&group->avgs_lock);
+}
+
+static void init_triggers(struct psi_group *group, u64 now)
+{
+	struct psi_trigger *t;
+
+	list_for_each_entry(t, &group->triggers, node)
+		window_reset(&t->win, now,
+				group->total[PSI_POLL][t->state], 0);
+	memcpy(group->polling_total, group->total[PSI_POLL],
+		   sizeof(group->polling_total));
+	group->polling_next_update = now + group->poll_min_period;
+}
+
 /* Schedule polling if it's not already scheduled or forced. */
 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay,
 				   bool force)
Signature · 1 line
2.34.1
#2 lore raw

PATCH v5 2/4 sched/psi: rename existing poll members in preparation

In reply to Domenico Cerasuolo

Renaming in PSI implementation to make a clear distinction between privileged and unprivileged triggers code to be implemented in the next patch.

---
 include/linux/psi_types.h |  36 ++++-----
 kernel/sched/psi.c        | 163 +++++++++++++++++++-------------------
 2 files changed, 100 insertions(+), 99 deletions(-)
include/linux/psi_types.h +18 -18
index 1e0a0d7ace3a..1819afa8b198 100644
@@ -175,26 +175,26 @@ struct psi_group {
 	u64 total[NR_PSI_AGGREGATORS][NR_PSI_STATES - 1];
 	unsigned long avg[NR_PSI_STATES - 1][3];
 
-	/* Monitor work control */
-	struct task_struct __rcu *poll_task;
-	struct timer_list poll_timer;
-	wait_queue_head_t poll_wait;
-	atomic_t poll_wakeup;
-	atomic_t poll_scheduled;
+	/* Monitor RT polling work control */
+	struct task_struct __rcu *rtpoll_task;
+	struct timer_list rtpoll_timer;
+	wait_queue_head_t rtpoll_wait;
+	atomic_t rtpoll_wakeup;
+	atomic_t rtpoll_scheduled;
 
 	/* Protects data used by the monitor */
-	struct mutex trigger_lock;
-
-	/* Configured polling triggers */
-	struct list_head triggers;
-	u32 nr_triggers[NR_PSI_STATES - 1];
-	u32 poll_states;
-	u64 poll_min_period;
-
-	/* Total stall times at the start of monitor activation */
-	u64 polling_total[NR_PSI_STATES - 1];
-	u64 polling_next_update;
-	u64 polling_until;
+	struct mutex rtpoll_trigger_lock;
+
+	/* Configured RT polling triggers */
+	struct list_head rtpoll_triggers;
+	u32 rtpoll_nr_triggers[NR_PSI_STATES - 1];
+	u32 rtpoll_states;
+	u64 rtpoll_min_period;
+
+	/* Total stall times at the start of RT polling monitor activation */
+	u64 rtpoll_total[NR_PSI_STATES - 1];
+	u64 rtpoll_next_update;
+	u64 rtpoll_until;
 };
 
 #else /* CONFIG_PSI */
kernel/sched/psi.c +82 -81

339 lines not on this page · raw message

Signature · 1 line
2.34.1
#3 lore raw

PATCH v5 3/4 sched/psi: extract update_triggers side effect

In reply to Domenico Cerasuolo

This change moves update_total flag out of update_triggers function, currently called only in psi_poll_work.
In the next patch, update_triggers will be called also in psi_avgs_work, but the total update information is specific to psi_poll_work. Returning update_total value to the caller let us avoid differentiating the implementation of update_triggers for different aggregators.

---
 kernel/sched/psi.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)
kernel/sched/psi.c +10 -9
index a3d0b5cf797a..f3df6a8ff493 100644
@@ -430,11 +430,11 @@ static u64 window_update(struct psi_window *win, u64 now, u64 value)
 	return growth;
 }
 
-static u64 update_triggers(struct psi_group *group, u64 now)
+static u64 update_triggers(struct psi_group *group, u64 now, bool *update_total)
 {
 	struct psi_trigger *t;
-	bool update_total = false;
 	u64 *total = group->total[PSI_POLL];
+	*update_total = false;
 
 	/*
 	 * On subsequent updates, calculate growth deltas and let
@@ -462,7 +462,7 @@ static u64 update_triggers(struct psi_group *group, u64 now)
 			 * been through all of them. Also remember to extend the
 			 * polling time if we see new stall activity.
 			 */
-			update_total = true;
+			*update_total = true;
 
 			/* Calculate growth since last update */
 			growth = window_update(&t->win, now, total[t->state]);
@@ -485,10 +485,6 @@ static u64 update_triggers(struct psi_group *group, u64 now)
 		t->pending_event = false;
 	}
 
-	if (update_total)
-		memcpy(group->rtpoll_total, total,
-				sizeof(group->rtpoll_total));
-
 	return now + group->rtpoll_min_period;
 }
 
@@ -622,6 +618,7 @@ static void psi_rtpoll_work(struct psi_group *group)
 {
 	bool force_reschedule = false;
 	u32 changed_states;
+	bool update_total;
 	u64 now;
 
 	mutex_lock(&group->rtpoll_trigger_lock);
@@ -686,8 +683,12 @@ static void psi_rtpoll_work(struct psi_group *group)
 		goto out;
 	}
 
-	if (now >= group->rtpoll_next_update)
-		group->rtpoll_next_update = update_triggers(group, now);
+	if (now >= group->rtpoll_next_update) {
+		group->rtpoll_next_update = update_triggers(group, now, &update_total);
+		if (update_total)
+			memcpy(group->rtpoll_total, group->total[PSI_POLL],
+				   sizeof(group->rtpoll_total));
+	}
 
 	psi_schedule_rtpoll_work(group,
 		nsecs_to_jiffies(group->rtpoll_next_update - now) + 1,
Signature · 1 line
2.34.1
#4 lore raw

PATCH v5 4/4 sched/psi: allow unprivileged polling of N*2s period

In reply to Domenico Cerasuolo

PSI offers 2 mechanisms to get information about a specific resource pressure. One is reading from /proc/pressure/<resource>, which gives average pressures aggregated every 2s. The other is creating a pollable fd for a specific resource and cgroup.

The trigger creation requires CAP_SYS_RESOURCE, and gives the possibility to pick specific time window and threshold, spawing an RT thread to aggregate the data.

Systemd would like to provide containers the option to monitor pressure on their own cgroup and sub-cgroups. For example, if systemd launches a container that itself then launches services, the container should have the ability to poll() for pressure in individual services. But neither the container nor the services are privileged.

This patch implements a mechanism to allow unprivileged users to create pressure triggers. The difference with privileged triggers creation is that unprivileged ones must have a time window that's a multiple of 2s. This is so that we can avoid unrestricted spawning of rt threads, and use instead the same aggregation mechanism done for the averages, which runs independently of any triggers.

---
 Documentation/accounting/psi.rst |   4 +
 include/linux/psi.h              |   2 +-
 include/linux/psi_types.h        |   7 ++
 kernel/cgroup/cgroup.c           |   2 +-
 kernel/sched/psi.c               | 148 +++++++++++++++++++------------
 5 files changed, 106 insertions(+), 57 deletions(-)
Documentation/accounting/psi.rst +4 -0
index 5e40b3f437f9..df6062eb3abb 100644
@@ -105,6 +105,10 @@ prevent overly frequent polling. Max limit is chosen as a high enough number
 after which monitors are most likely not needed and psi averages can be used
 instead.
 
+Unprivileged users can also create monitors, with the only limitation that the
+window size must be a multiple of 2s, in order to prevent excessive resource
+usage.
+
 When activated, psi monitor stays active for at least the duration of one
 tracking window to avoid repeated activations/deactivations when system is
 bouncing in and out of the stall state.
include/linux/psi.h +1 -1
index b029a847def1..ab26200c2803 100644
@@ -24,7 +24,7 @@ void psi_memstall_leave(unsigned long *flags);
 
 int psi_show(struct seq_file *s, struct psi_group *group, enum psi_res res);
 struct psi_trigger *psi_trigger_create(struct psi_group *group,
-			char *buf, enum psi_res res);
+			char *buf, enum psi_res res, struct file *file);
 void psi_trigger_destroy(struct psi_trigger *t);
 
 __poll_t psi_trigger_poll(void **trigger_ptr, struct file *file,
include/linux/psi_types.h +7 -0
index 1819afa8b198..040c089581c6 100644
@@ -151,6 +151,9 @@ struct psi_trigger {
 
 	/* Deferred event(s) from previous ratelimit window */
 	bool pending_event;
+
+	/* Trigger type - PSI_AVGS for unprivileged, PSI_POLL for RT */
+	enum psi_aggregators aggregator;
 };
 
 struct psi_group {
@@ -171,6 +174,10 @@ struct psi_group {
 	/* Aggregator work control */
 	struct delayed_work avgs_work;
 
+	/* Unprivileged triggers against N*PSI_FREQ windows */
+	struct list_head avg_triggers;
+	u32 avg_nr_triggers[NR_PSI_STATES - 1];
+
 	/* Total stall times and sampled pressure averages */
 	u64 total[NR_PSI_AGGREGATORS][NR_PSI_STATES - 1];
 	unsigned long avg[NR_PSI_STATES - 1][3];
kernel/cgroup/cgroup.c +1 -1
index 935e8121b21e..dead36969bba 100644
@@ -3761,7 +3761,7 @@ static ssize_t pressure_write(struct kernfs_open_file *of, char *buf,
 	}
 
 	psi = cgroup_psi(cgrp);
-	new = psi_trigger_create(psi, buf, res);
+	new = psi_trigger_create(psi, buf, res, of->file);
 	if (IS_ERR(new)) {
 		cgroup_put(cgrp);
 		return PTR_ERR(new);
kernel/sched/psi.c +93 -55

276 lines not on this page · raw message

Signature · 1 line
2.34.1
maintainer
#5 lore raw

Re: [PATCH v5 4/4] sched/psi: allow unprivileged polling of N*2s period

Hi Domenico,

thanks, the destruction path looks right to me now.

Quoted 11 lines · Domenico Cerasuolo
kernel/sched/psi.c +2 -5
@@ -1319,11 +1353,16 @@ struct psi_trigger *psi_trigger_create(struct psi_group *group,
 		rcu_assign_pointer(group->rtpoll_task, task);
 	}
 
-	list_add(&t->node, &group->rtpoll_triggers);
-	group->rtpoll_min_period = min(group->rtpoll_min_period,
-		div_u64(t->win.size, UPDATES_PER_WINDOW));
-	group->rtpoll_nr_triggers[t->state]++;
-	group->rtpoll_states |= (1 << t->state);
+	if (privileged) {
+		list_add(&t->node, &group->rtpoll_triggers);

kernel/sched/psi.c struct psi_trigger *psi_trigger_create(struct psi_group *group, · line 1367

Quoted 10 lines
kernel/sched/psi.c +8 -0
+		group->rtpoll_min_period = min(group->rtpoll_min_period,
+			div_u64(t->win.size, UPDATES_PER_WINDOW));
+		group->rtpoll_nr_triggers[t->state]++;
+		group->rtpoll_states |= (1 << t->state);
+	} else {
+		list_add(&t->node, &group->avg_triggers);
+		group->avg_nr_triggers[t->state]++;
+	}
 
 	mutex_unlock(&group->rtpoll_trigger_lock);

But creation needs the same treatment. group->avg_triggers and avg_nr_triggers are protected by the avgs_lock, not the rtpoll_trigger_lock.

I suppose the lock taking and the group->rtpoll_task creation (anything with rtpoll in the name, really) should be within a big

	if (privileged) {
		rtpoll_trigger_lock()
		kthread_create()
		init-and-link()
		rtpoll_trigger_unlock()
	} else {
		avgs_lock()
		...
	}