From: Mauro Carvalho Chehab <mchehab@redhat.com>
Date: Mon, 5 Oct 2009 17:52:34 -0400
Subject: [misc] don't call printk while crashing
Message-id: <20091005145234.5de5a65e@pedra.chehab.org>
Patchwork-id: 21046
O-Subject: [PATCH RHEL5.5] BZ#497195 Don't call printk while crashing
Bugzilla: 497195
RH-Acked-by: Eduardo Habkost <ehabkost@redhat.com>
RH-Acked-by: Neil Horman <nhorman@redhat.com>

Upstream code for mdelay() is:

#define mdelay(n) (								\
        (__builtin_constant_p(n) && (n)<=MAX_UDELAY_MS) ? udelay((n)*1000) :	\
        ({unsigned long __ms=(n); while (__ms--) udelay(1000);}))

However, RHEL5 changeset 3a4d4483d4cb7d0514c8d0d061de87da97c03425 (From Aug,
2005!) changed it to:

#define mdelay(n) (                          \
{                                            \
 static int warned=0;                        \
 unsigned long __ms=(n);                     \
 WARN_ON(in_irq() && !(warned++));	     \
 while (__ms--) udelay(1000);                \
})

With the RHEL variant, WARN_ON() is called if a code with mdelay() is used, while at
IRQ time. This caused a regression, as reported on some bugzillas like #458368 and #497195.

While this is generally ok, when kernel crash is called, we need to wait up to 1 second
in order to be sure that all CPU's but one were disabled. However, due to WARN_ON, it
will print a message like:

	BUG: warning at arch/i386/kernel/crash.c:148/nmi_shootdown_cpus() (Not tainted)

The previous approach were to patch printk to reduce the risk of having it called during
CPU shutdown princess, but this didn't solve the issue completely.

On some cases, calling a printk during the crash is still causing failures.

So, let's use a different approach: just replace all occurences of mdelay(1) to
udelay(1000) to avoid the WARN_ON() call during nmi_shootdown_cpus().

>From comment https://bugzilla.redhat.com/show_bug.cgi?id=497195#c6, this approach seems
to solve the issue.

Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>

diff --git a/arch/i386/kernel/crash.c b/arch/i386/kernel/crash.c
index a203412..c7cc162 100644
--- a/arch/i386/kernel/crash.c
+++ b/arch/i386/kernel/crash.c
@@ -144,7 +144,7 @@ static void nmi_shootdown_cpus(void)
 
 	msecs = 1000; /* Wait at most a second for the other cpus to stop */
 	while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
-		mdelay(1);
+		udelay(1000);
 		msecs--;
 	}
 
diff --git a/arch/powerpc/kernel/crash.c b/arch/powerpc/kernel/crash.c
index f35cb59..99fbabe 100644
--- a/arch/powerpc/kernel/crash.c
+++ b/arch/powerpc/kernel/crash.c
@@ -189,7 +189,7 @@ static void crash_kexec_prepare_cpus(int cpu)
 	msecs = 10000;
 	while ((cpus_weight(cpus_in_crash) < ncpus) && (--msecs > 0)) {
 		cpu_relax();
-		mdelay(1);
+		udelay(1000);
 	}
 
 	/* Would it be better to replace the trap vector here? */
@@ -241,7 +241,7 @@ void crash_kexec_secondary(struct pt_regs *regs)
 			local_irq_restore(flags);
 			return;
 		}
-		mdelay(1);
+		udelay(1000);
 		cpu_relax();
 	}
 	if (cpu == crashing_cpu) {
diff --git a/arch/x86_64/kernel/crash.c b/arch/x86_64/kernel/crash.c
index baebf74..de84891 100644
--- a/arch/x86_64/kernel/crash.c
+++ b/arch/x86_64/kernel/crash.c
@@ -160,7 +160,7 @@ static void nmi_shootdown_cpus(void)
 
 	msecs = 1000; /* Wait at most a second for the other cpus to stop */
 	while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
-		mdelay(1);
+		udelay(1000);
 		msecs--;
 	}
 	/* Leave the nmi callback set */