Re: [PATCH v5 3/9] x86/head: Move early exception panic code into early_fixup_exception

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Apr 04, 2016 at 01:52:06PM +0200, Jan Kara wrote:
> Sounds like a good idea to me. I've also consulted this with Petr Mladek
> (added to CC) who is using printk_func per-cpu variable in his
> printk-from-NMI patches and he also doesn't see a problem with this.

There's a few printk() variants that do not go through this; which means
they're broken for a number of cases, including the kdb printk
redirection, this NMI stuff etc.

> I was just wondering about one thing - this way we add more early printks
> if I understand your intention right. Are we guaranteed that they happen
> only from a single CPU? Because currently there is no locking in
> early_printk() and thus we can end up writing to early console several
> messages in parallel from different CPUs. Not sure what's going to happen
> in that case...

You get lovely per char interleaving on you serial line ;-)

What I've done in the past was something like the below; that way you
only get the normal task->softirq->irq->nmi nesting, which is mostly
decipherable.

diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
index bfbf284e4218..c4c3269ff104 100644
--- a/kernel/printk/printk.c
+++ b/kernel/printk/printk.c
@@ -1907,17 +1907,36 @@ struct console *early_console;
 asmlinkage __visible void early_printk(const char *fmt, ...)
 {
 	va_list ap;
+	static int print_cpu = -1;
 	char buf[512];
-	int n;
+	int n, cpu;
 
 	if (!early_console)
 		return;
 
+	preempt_disable();
+	cpu = raw_smp_processor_id();
+	for (;;) {
+		if (READ_ONCE(print_cpu) == cpu)
+			break;
+
+		if (READ_ONCE(print_cpu) == -1 &&
+		    cmpxchg(&print_cpu, -1, cpu) == -1) {
+			cpu = -1;
+			break;
+		}
+
+		cpu_relax();
+	}
+
 	va_start(ap, fmt);
 	n = vscnprintf(buf, sizeof(buf), fmt, ap);
 	va_end(ap);
 
 	early_console->write(early_console, buf, n);
+
+	smp_store_release(&print_cpu, cpu);
+	preempt_enable();
 }
 #endif