[PATCH 2/2] printk: Reinitialize klogd percpu state for secondary CPUs

[Jann Horn <jannh@google.com>]

When printk_deferred() is used before percpu initialization, it will queue
up wake_up_klogd_work and set printk_pending on the boot CPU; percpu
initialization then copies the the pending work's state and the
printk_pending flag to all secondary CPUs. The end result is that e.g.
if you run "dmesg -w" when the system is up, it won't notice new printk
messages.
To ensure that the secondary CPUs don't think that they already have
pending printk work, reset the secondary CPUs' work items and
printk_pending state.

Signed-off-by: Jann Horn <jannh@google.com>
---
 kernel/printk/printk.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
index fada22dc4ab6..e531407188f1 100644
--- a/kernel/printk/printk.c
+++ b/kernel/printk/printk.c
@@ -2964,6 +2964,29 @@ static DEFINE_PER_CPU(struct irq_work, wake_up_klogd_work) = {
 	.flags = ATOMIC_INIT(IRQ_WORK_LAZY),
 };
 
+/*
+ * If we called defer_console_output() before fully initializing the percpu
+ * subsystem, e.g. via printk_deferred(), the irq_work flags and the
+ * printk_pending flag may have been copied from the boot CPU to the others
+ * while work was pending.
+ * To fix that, manually reset the percpu data for all secondary processors
+ * before bringing up SMP.
+ */
+static int __init klogd_percpu_fixup(void)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		if (cpu == smp_processor_id())
+			continue;
+		atomic_set(&per_cpu(wake_up_klogd_work.flags, cpu),
+			   IRQ_WORK_LAZY);
+		per_cpu(printk_pending, cpu) = 0;
+	}
+	return 0;
+}
+early_initcall(klogd_percpu_fixup)
+
 void wake_up_klogd(void)
 {
 	preempt_disable();
-- 
2.25.1.696.g5e7596f4ac-goog