[PATCH] time/sched_clock: Call enable_sched_clock_irqtime() with IRQs enabled

["Benoît Monin" <benoit.monin@bootlin.com>]

sched_clock_register() disables interrupts with local_irq_save() and
calls enable_sched_clock_irqtime() before restoring them.

With CONFIG_IRQ_TIME_ACCOUNTING, enable_sched_clock_irqtime() calls
static_branch_enable(). With CONFIG_JUMP_LABEL and CONFIG_HOTPLUG_CPU,
this takes cpus_read_lock(), which eventually calls might_sleep() from
percpu_down_read_internal(). So with CONFIG_DEBUG_ATOMIC_SLEEP the
following is reported:

  BUG: sleeping function called from invalid context at include/linux/percpu-rwsem.h:51
  ...
  Call Trace:
  ...
  [<a8000001008ad9c4>] __might_resched+0x3ac/0x3d0
  [<a8000001008598b0>] percpu_down_read_internal.constprop.0+0x30/0x1b8
  [<a800000100af5ddc>] static_key_enable+0x14/0x30
  [<a80000010099b600>] sched_clock_register+0x2a0/0x370
  ...

Most callers of sched_clock_register() are early init functions,
for which might_sleep() stays silent because system_state is still
SYSTEM_BOOTING. However, platform drivers such as timer-nxp-stm,
timer-ti-dm, timer-mediatek and timer-rzn1, as well as the mips-gic-timer
promotion work, call it later, where the warning is reported.

Only the sched_clock data update requires interrupts to be disabled.
Enabling IRQ time accounting only flips a static branch, so restore
interrupts before calling enable_sched_clock_irqtime().

Fixes: 2707745533d6 ("time/sched_clock: Disable interrupts in sched_clock_register()")
Reported-by: Chen-Yu Tsai <wenst@chromium.org>
Closes: https://lore.kernel.org/lkml/20261109095036.GA722431@google.com/
Signed-off-by: Benoît Monin <benoit.monin@bootlin.com>
---
Found the sleeping-in-atomic warning while working on the GIC
synchronization series [1] and testing configuration options on the
Mobileye EyeQ6H SoC, but the issue is not limited to that clock source.

Chen-Yu Tsai reported the same issue with timer-mediatek.

I also discussed the issue with my colleague Hervé, who reproduced the
BUG on a board using timer-rzn1.

[1]: https://lore.kernel.org/lkml/20260929-sync-gic-counters-v4-0-ec70c4b60434@bootlin.com/
---
 kernel/time/sched_clock.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c
index f3aaef695b8c..b4e317b8a585 100644
--- a/kernel/time/sched_clock.c
+++ b/kernel/time/sched_clock.c
@@ -238,12 +238,13 @@ void sched_clock_register(u64 (*read)(void), int bits, unsigned long rate)
 	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 		bits, r, r_unit, res, wrap);
 
+	/* Restore interrupts now as enable_sched_clock_irqtime() might sleep */
+	local_irq_restore(flags);
+
 	/* Enable IRQ time accounting if we have a fast enough sched_clock() */
 	if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 		enable_sched_clock_irqtime();
 
-	local_irq_restore(flags);
-
 	pr_debug("Registered %pS as sched_clock source\n", read);
 }
 EXPORT_SYMBOL_GPL(sched_clock_register);

---
base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
change-id: 20261012-sched-clock-might-sleep-48af26c8911b

Best regards,
--  
Benoît Monin, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com