timers/migration: Fix another race between hotplug and idle entry/exit

Commit 10a0e6f3d3db ("timers/migration: Move hierarchy setup into
cpuhotplug prepare callback") fixed a race between idle exit and CPU
hotplug up leading to a wrong "0" value migrator assigned to the top
level. However there is still a situation that remains unhandled:

[GRP0:0]
migrator = TMIGR_NONE
active = NONE
groupmask = 0
/ \ \
0 1 2..7
idle idle idle

0) The system is fully idle.

[GRP0:0]
migrator = CPU 0
active = CPU 0
groupmask = 0
/ \ \
0 1 2..7
active idle idle

1) CPU 0 is activating. It has done the cmpxchg on the top's ->migr_state
but it hasn't yet returned to __walk_groups().

[GRP0:0]
migrator = CPU 0
active = CPU 0, CPU 1
groupmask = 0
/ \ \
0 1 2..7
active active idle

2) CPU 1 is activating. CPU 0 stays the migrator (still stuck in
__walk_groups(), delayed by #VMEXIT for example).

[GRP1:0]
migrator = TMIGR_NONE
active = NONE
groupmask = 0
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = NONE
groupmask = 2 groupmask = 1
/ \ \
0 1 2..7 8
active active idle !online

3) CPU 8 is preparing to boot. CPUHP_TMIGR_PREPARE is being ran by CPU 1
which has created the GRP0:1 and the new top GRP1:0 connected to GRP0:1
and GRP0:0. The groupmask of GRP0:0 is now 2. CPU 1 hasn't yet
propagated its activation up to GRP1:0.

[GRP1:0]
migrator = 0 (!!!)
active = NONE
groupmask = 0
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = NONE
groupmask = 2 groupmask = 1
/ \ \
0 1 2..7 8
active active idle !online

4) CPU 0 finally resumed after its #VMEXIT. It's in __walk_groups()
returning from tmigr_cpu_active(). The new top GRP1:0 is visible and
fetched but the freshly updated groupmask of GRP0:0 may not be visible
due to lack of ordering! As a result tmigr_active_up() is called to
GRP0:0 with a child's groupmask of "0". This buggy "0" groupmask then
becomes the migrator for GRP1:0 forever. As a result, timers on a fully
idle system get ignored.

One possible fix would be to define TMIGR_NONE as "0" so that such a
race would have no effect. And after all TMIGR_NONE doesn't need to be
anything else. However this would leave an uncomfortable state machine
where gears happen not to break by chance but are vulnerable to future
modifications.

Keep TMIGR_NONE as is instead and pre-initialize to "1" the groupmask of
any newly created top level. This groupmask is guaranteed to be visible
upon fetching the corresponding group for the 1st time:

_ By the upcoming CPU thanks to CPU hotplug synchronization between the
control CPU (BP) and the booting one (AP).

_ By the control CPU since the groupmask and parent pointers are
initialized locally.

_ By all CPUs belonging to the same group than the control CPU because
they must wait for it to ever become idle before needing to walk to
the new top. The cmpcxhg() on ->migr_state then makes sure its
groupmask is visible.

With this pre-initialization, it is guaranteed that if a future top level
is linked to an old one, it is walked through with a valid groupmask.

Fixes: 10a0e6f3d3db ("timers/migration: Move hierarchy setup into cpuhotplug prepare callback")
Signed-off-by: Frederic Weisbecker <[e-mail]>
Signed-off-by: Thomas Gleixner <[e-mail]>
Cc: [e-mail]
Link: https://lore.kernel.org/all/[e-mail]

kernel/time/timer_migration.c

···
 	s.seq = 0;
 	atomic_set(&group->migr_state, s.state);
 
+	/*
+	 * If this is a new top-level, prepare its groupmask in advance.
+	 * This avoids accidents where yet another new top-level is
+	 * created in the future and made visible before the current groupmask.
+	 */
+	if (list_empty(&tmigr_level_list[lvl])) {
+		group->groupmask = BIT(0);
+		/*
+		 * The previous top level has prepared its groupmask already,
+		 * simply account it as the first child.
+		 */
+		if (lvl > 0)
+			group->num_children = 1;
+	}
+
 	timerqueue_init_head(&group->events);
 	timerqueue_init(&group->groupevt.nextevt);
 	group->groupevt.nextevt.expires = KTIME_MAX;
···
 	raw_spin_lock_irq(&child->lock);
 	raw_spin_lock_nested(&parent->lock, SINGLE_DEPTH_NESTING);
 
+	if (activate) {
+		/*
+		 * @child is the old top and @parent the new one. In this
+		 * case groupmask is pre-initialized and @child already
+		 * accounted, along with its new sibling corresponding to the
+		 * CPU going up.
+		 */
+		WARN_ON_ONCE(child->groupmask != BIT(0) || parent->num_children != 2);
+	} else {
+		/* Adding @child for the CPU going up to @parent. */
+		child->groupmask = BIT(parent->num_children++);
+	}
+
 	child->parent = parent;
-	child->groupmask = BIT(parent->num_children++);
 
 	raw_spin_unlock(&parent->lock);
 	raw_spin_unlock_irq(&child->lock);