From: Rik van Riel <riel@redhat.com>
Subject: [PATCH][RHEL5] fix iSCSI root oops on x86_64 xen domU
Date: Tue, 14 Nov 2006 15:46:28 -0500
Bugzilla: 215581
Message-Id: <455A2B24.1090002@redhat.com>
Changelog: Xen: fix iSCSI root oops on x86_64 xen domU


With iSCSI root on x86-64 domU, the kernel oopses in
scsi_calculate_bounce_limit() because dma_ops is NULL.

The attached patch (from upstream) initializes the iommu
operations pointer.  Even though the patch is to arch/i386/
it is x86-64 only, because this is inside the big x86-64
#ifdef in arch/i386/kernel/pci-dma-xen.c.

Thanks to Mark McLoughlin for noticing the issue and finding
the fix in the upstream tree.

Fixes bug #215581

-- 
Who do you trust?
The people with all the right answers?
Or the people with the right questions?

# HG changeset patch
# User chrisw@sous-sol.org
# Date Tue Oct 03 13:44:38 2006 -0400
# Node ID e5a7f30e1db3f1084f6789d21ea2a6fdaafdb96d
# parent: 6cd0fae5d84c4a4b15546ceaade74b7d7f044404
Make sure no_iommu_init is called when needed on x86_64.  Thanks
to Mark McLoughlin <markmc@redhat.com> for spotting the issue and
proposing a fix.

Index: latest/arch/i386/kernel/pci-dma-xen.c
===================================================================
--- latest.orig/arch/i386/kernel/pci-dma-xen.c
+++ latest/arch/i386/kernel/pci-dma-xen.c
@@ -20,6 +20,9 @@
 #include <asm/bug.h>
 
 #ifdef __x86_64__
+#include <asm/proto.h>
+#include <asm/calgary.h>
+
 int iommu_merge __read_mostly = 0;
 EXPORT_SYMBOL(iommu_merge);
 
@@ -70,6 +73,23 @@ __init int iommu_setup(char *p)
 {
     return 1;
 }
+
+static int __init pci_iommu_init(void)
+{
+#ifdef CONFIG_CALGARY_IOMMU
+	calgary_iommu_init();
+#endif
+
+#ifdef CONFIG_IOMMU
+	gart_iommu_init();
+#endif
+
+	no_iommu_init();
+	return 0;
+}
+
+/* Must execute after PCI subsystem */
+fs_initcall(pci_iommu_init);
 #endif
 
 struct dma_coherent_mem {