From: Jonathan Lim <jolim@redhat.com>
Date: Wed, 23 Jan 2008 18:22:05 -0500
Subject: [ia64] kdump: slave CPUs drop to POD
Message-id: 20080123232205.GA29185@sgi-desktop.boston.redhat.com
O-Subject: [RHEL 5.2 PATCH] BZ 429956: slave CPU's drop to POD when booting kdump kernel on IA64
Bugzilla: 429956

The current implementation of kdump on INIT events would enter kdump processing
on DIE_INIT_MONARCH_ENTER and DIE_INIT_SLAVE_ENTER events.  Thus, the monarch
CPU would go ahead and boot up the kdump kernel.

On SN shub2 systems, this out-of-sync situation causes some slave CPU's on
different nodes to enter POD.

This patch moves kdump entry points to DIE_INIT_MONARCH_LEAVE and
DIE_INIT_SLAVE_LEAVE.  It also sets the kdump_in_progress variable in the
DIE_INIT_MONARCH_PROCESS event to not dump all active stack traces to the
console in the case of kdump.

The patch has been committed upstream:

  http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=311f594dec9b0c8693ec7df75b82c251b6b0e7c2

Following is a diff against the 2.6.18-62.el5 kernel source:

Acked-by: Neil Horman <nhorman@redhat.com>

diff --git a/arch/ia64/kernel/crash.c b/arch/ia64/kernel/crash.c
index 6aa564b..cf2c1b9 100644
--- a/arch/ia64/kernel/crash.c
+++ b/arch/ia64/kernel/crash.c
@@ -24,7 +24,7 @@
 int kdump_status[NR_CPUS];
 atomic_t kdump_cpu_freezed;
 int kdump_on_init = 1;
-extern int kdump_in_progress;
+atomic_t kdump_in_progress;
 
 ssize_t
 copy_oldmem_page(unsigned long pfn, char *buf,
@@ -181,30 +181,36 @@ kdump_init_notifier(struct notifier_block *self, unsigned long val, void *data)
 		return NOTIFY_DONE;
 	}
 
-	if (val != DIE_INIT_MONARCH_ENTER &&
-	    val != DIE_INIT_SLAVE_ENTER &&
+	if (val != DIE_INIT_MONARCH_LEAVE &&
+	    val != DIE_INIT_SLAVE_LEAVE &&
+	    val != DIE_INIT_MONARCH_PROCESS &&
 	    val != DIE_MCA_RENDZVOUS_LEAVE &&
 	    val != DIE_MCA_MONARCH_LEAVE)
 		return NOTIFY_DONE;
 
 	nd = (struct ia64_mca_notify_die *)args->err;
 	/* Reason code 1 means machine check rendezous*/
-	if ((val==DIE_INIT_MONARCH_ENTER || val==DIE_INIT_SLAVE_ENTER) &&
-	    nd->sos->rv_rc == 1)
+	if ((val == DIE_INIT_MONARCH_LEAVE || val == DIE_INIT_SLAVE_LEAVE
+	    || val == DIE_INIT_MONARCH_PROCESS) && nd->sos->rv_rc == 1)
 		return NOTIFY_DONE;
 
 	if (kdump_sending_init)
 		unw_init_running(kdump_cpu_freeze, NULL);
 
 	switch (val) {
-		case DIE_INIT_MONARCH_ENTER:
+		case DIE_INIT_MONARCH_PROCESS:
+			atomic_set(&kdump_in_progress, 1);
+			*(nd->monarch_cpu) = -1;
+			break;
+		case DIE_INIT_MONARCH_LEAVE:
 			machine_kdump_on_init();
 			break;
-		case DIE_INIT_SLAVE_ENTER:
-			unw_init_running(kdump_cpu_freeze, NULL);
+		case DIE_INIT_SLAVE_LEAVE:
+			if (atomic_read(&kdump_in_progress))
+				unw_init_running(kdump_cpu_freeze, NULL);
 			break;
 		case DIE_MCA_RENDZVOUS_LEAVE:
-			if (kdump_in_progress)
+			if (atomic_read(&kdump_in_progress))
 				unw_init_running(kdump_cpu_freeze, NULL);
 			break;
 		case DIE_MCA_MONARCH_LEAVE:
@@ -244,8 +250,10 @@ static int
 machine_crash_setup(void)
 {
 	char *from = strstr(saved_command_line, "elfcorehdr=");
+	/* be notified before default_monarch_init_process */
 	static struct notifier_block kdump_init_notifier_nb = {
 		.notifier_call = kdump_init_notifier,
+		.priority = 1,
 	};
 	int ret;
 	if (from)
diff --git a/arch/ia64/kernel/mca.c b/arch/ia64/kernel/mca.c
index e0bcc7f..a4f8a96 100644
--- a/arch/ia64/kernel/mca.c
+++ b/arch/ia64/kernel/mca.c
@@ -83,6 +83,7 @@
 #include <asm/system.h>
 #include <asm/sal.h>
 #include <asm/mca.h>
+#include <asm/kexec.h>
 
 #include <asm/irq.h>
 #include <asm/hw_irq.h>
@@ -96,11 +97,6 @@
 # define IA64_MCA_DEBUG(fmt...)
 #endif
 
-#ifdef CONFIG_KEXEC
-/* Used by arch/ia64/kernel/crash.c */
-int kdump_in_progress;
-#endif
-
 /* Used by mca_asm.S */
 u32				ia64_mca_serialize;
 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
@@ -1244,7 +1240,7 @@ ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
 		/* Dump buffered message to console */
 		ia64_mlogbuf_finish(1);
 #ifdef CONFIG_KEXEC
-		kdump_in_progress = 1;
+		atomic_set(&kdump_in_progress, 1);
 		/* In the case of (!recover), notify_die(DIE_MCA_MONARCH_LEAVE)
 		   will not return. A dump kernel will be booted. Need to set
 		   nonarch_cpu here to get slave cpus out of looping in OS.
@@ -1490,6 +1486,10 @@ default_monarch_init_process(struct notifier_block *self, unsigned long val, voi
 	struct task_struct *g, *t;
 	if (val != DIE_INIT_MONARCH_PROCESS)
 		return NOTIFY_DONE;
+#ifdef CONFIG_KEXEC
+	if (atomic_read(&kdump_in_progress))
+		return NOTIFY_DONE;
+#endif
 
 	/*
 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
diff --git a/include/asm-ia64/kexec.h b/include/asm-ia64/kexec.h
index 1c789b1..7bd0963 100644
--- a/include/asm-ia64/kexec.h
+++ b/include/asm-ia64/kexec.h
@@ -49,6 +49,7 @@ extern unsigned long kdump_find_rsvd_region(unsigned long size,
 extern void kdump_cpu_freeze(struct unw_frame_info *info, void *arg);
 extern int kdump_status[];
 extern atomic_t kdump_cpu_freezed;
+extern atomic_t kdump_in_progress;
 extern int kdump_kernel;
 
 #endif /* _ASM_IA64_KEXEC_H */