Subject: 32-on-64 blkif protocol negotiation fallback for old guests.
From: kraxel@suse.de
References: 244055
Patch-mainline: never.

See the comment below.  Oh well.

--- head.orig/drivers/xen/Kconfig	2012-02-10 11:56:27.000000000 +0100
+++ head/drivers/xen/Kconfig	2012-02-17 14:38:40.000000000 +0100
@@ -23,6 +23,9 @@ config XEN_UNPRIVILEGED_GUEST
 	select PM
 	select SUSPEND
 
+config XEN_DOMCTL
+	tristate
+
 config XEN_XENBUS_DEV
 	def_bool y
 	depends on PROC_FS
@@ -42,6 +45,7 @@ config XEN_BLKDEV_BACKEND
 	tristate "Block-device backend driver"
 	depends on BLOCK && XEN_BACKEND
 	default XEN_BACKEND
+	select XEN_DOMCTL
 	help
 	  The block-device backend driver allows the kernel to export its
 	  block devices to other guests via a high-performance shared-memory
@@ -51,6 +55,7 @@ config XEN_BLKDEV_TAP
 	tristate "Block-device tap backend driver"
 	depends on BLOCK && XEN_BACKEND
 	default XEN_BACKEND
+	select XEN_DOMCTL
 	help
 	  The block tap driver is an alternative to the block back driver
 	  and allows VM block requests to be redirected to userspace through
--- head.orig/drivers/xen/blkback/xenbus.c	2012-03-26 15:49:36.000000000 +0200
+++ head/drivers/xen/blkback/xenbus.c	2012-03-26 15:54:30.000000000 +0200
@@ -20,6 +20,7 @@
 #include <stdarg.h>
 #include <linux/kthread.h>
 #include "common.h"
+#include "../core/domctl.h"
 
 #undef DPRINTK
 #define DPRINTK(fmt, args...)				\
@@ -546,24 +547,25 @@ static int connect_ring(struct backend_i
 
 	be->blkif->blk_protocol = BLKIF_PROTOCOL_NATIVE;
 	protocol = xenbus_read(XBT_NIL, dev->otherend, "protocol", NULL);
-	if (IS_ERR(protocol))
+	if (IS_ERR(protocol)) {
 		protocol = NULL;
+		be->blkif->blk_protocol = xen_guest_blkif_protocol(be->blkif->domid);
 #ifndef CONFIG_X86_32
-	else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_32))
+	} else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_32)) {
 		be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_32;
 #endif
 #ifndef CONFIG_X86_64
-	else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_64))
+	} else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_64)) {
 		be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_64;
 #endif
-	else if (0 != strcmp(protocol, XEN_IO_PROTO_ABI_NATIVE)) {
+	} else if (0 != strcmp(protocol, XEN_IO_PROTO_ABI_NATIVE)) {
 		xenbus_dev_fatal(dev, err, "unknown fe protocol %s", protocol);
 		kfree(protocol);
 		return -1;
 	}
 	pr_info("blkback: ring-ref %u, event-channel %u, protocol %d (%s)\n",
 		ring_ref, evtchn, be->blkif->blk_protocol,
-		protocol ?: "unspecified, assuming native");
+		protocol ?: "unspecified");
 	kfree(protocol);
 
 	/* Map the shared frame, irq etc. */
--- head.orig/drivers/xen/blktap/xenbus.c	2012-03-26 15:50:12.000000000 +0200
+++ head/drivers/xen/blktap/xenbus.c	2012-03-26 15:54:07.000000000 +0200
@@ -38,6 +38,7 @@
 #include <linux/kthread.h>
 #include <xen/xenbus.h>
 #include "common.h"
+#include "../core/domctl.h"
 
 
 struct backend_info
@@ -441,24 +442,25 @@ static int connect_ring(struct backend_i
 
 	be->blkif->blk_protocol = BLKIF_PROTOCOL_NATIVE;
 	protocol = xenbus_read(XBT_NIL, dev->otherend, "protocol", NULL);
-	if (IS_ERR(protocol))
+	if (IS_ERR(protocol)) {
 		protocol = NULL;
+		be->blkif->blk_protocol = xen_guest_blkif_protocol(be->blkif->domid);
 #ifndef CONFIG_X86_32
-	else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_32))
+	} else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_32)) {
 		be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_32;
 #endif
 #ifndef CONFIG_X86_64
-	else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_64))
+	} else if (0 == strcmp(protocol, XEN_IO_PROTO_ABI_X86_64)) {
 		be->blkif->blk_protocol = BLKIF_PROTOCOL_X86_64;
 #endif
-	else if (0 != strcmp(protocol, XEN_IO_PROTO_ABI_NATIVE)) {
+	} else if (0 != strcmp(protocol, XEN_IO_PROTO_ABI_NATIVE)) {
 		xenbus_dev_fatal(dev, err, "unknown fe protocol %s", protocol);
 		kfree(protocol);
 		return -1;
 	}
 	pr_info("blktap: ring-ref %u, event-channel %u, protocol %d (%s)\n",
 		ring_ref, evtchn, be->blkif->blk_protocol,
-		protocol ?: "unspecified, assuming native");
+		protocol ?: "unspecified");
 	kfree(protocol);
 
 	/* Map the shared frame, irq etc. */
--- head.orig/drivers/xen/core/Makefile	2012-02-17 14:37:50.000000000 +0100
+++ head/drivers/xen/core/Makefile	2012-02-17 14:38:43.000000000 +0100
@@ -11,3 +11,6 @@ obj-$(CONFIG_XEN_SMPBOOT)	+= smpboot.o
 obj-$(CONFIG_SMP)		+= spinlock.o
 obj-$(CONFIG_KEXEC)		+= machine_kexec.o
 obj-$(CONFIG_GENERIC_CLOCKEVENTS) += clockevents.o
+obj-$(CONFIG_XEN_DOMCTL)	+= domctl.o
+CFLAGS_domctl.o			:= -D__XEN_PUBLIC_XEN_H__ -D__XEN_PUBLIC_GRANT_TABLE_H__
+CFLAGS_domctl.o			+= -D__XEN_TOOLS__ -imacros xen/interface/domctl.h
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ head/drivers/xen/core/domctl.c	2011-10-04 17:01:11.000000000 +0200
@@ -0,0 +1,131 @@
+/*
+ * !!!  dirty hack alert  !!!
+ *
+ * Problem: old guests kernels don't have a "protocol" node
+ *          in the frontend xenstore directory, so mixing
+ *          32 and 64bit domains doesn't work.
+ *
+ * Upstream plans to solve this in the tools, by letting them
+ * create a protocol node.  Which certainly makes sense.
+ * But it isn't trivial and isn't done yet.  Too bad.
+ *
+ * So for the time being we use the get_address_size domctl
+ * hypercall for a pretty good guess.  Not nice as the domctl
+ * hypercall isn't supposed to be used by the kernel.  Because
+ * we don't want to have dependencies between dom0 kernel and
+ * xen kernel versions.  Now we have one.  Ouch.
+ */
+#undef __XEN_PUBLIC_XEN_H__
+#undef __XEN_PUBLIC_GRANT_TABLE_H__
+#undef __XEN_TOOLS__
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <asm/hypervisor.h>
+#include <xen/blkif.h>
+
+#include "domctl.h"
+
+/* stuff copied from xen/interface/domctl.h, which we can't
+ * include directly for the reasons outlined above .... */
+
+typedef struct xen_domctl_address_size {
+	uint32_t size;
+} xen_domctl_address_size_t;
+
+typedef __attribute__((aligned(8))) uint64_t uint64_aligned_t;
+
+union xen_domctl {
+	/* v4: sle10 sp1: xen 3.0.4 + 32-on-64 patches */
+	struct {
+		uint32_t cmd;
+		uint32_t interface_version;
+		domid_t  domain;
+		union {
+			/* left out lots of other struct xen_domctl_foobar */
+			struct xen_domctl_address_size       address_size;
+			uint64_t                             dummy_align;
+			uint8_t                              dummy_pad[128];
+		};
+	} v4;
+
+	/*
+	 * v5: upstream: xen 3.1
+	 * v6: upstream: xen 4.0
+	 * v7: upstream: xen 4.1; sle11 sp1: xen 4.0 + cpupools patches
+	 * v8: upstream: xen 4.2
+	 */
+	struct {
+		uint32_t cmd;
+		uint32_t interface_version;
+		domid_t  domain;
+		union {
+			struct xen_domctl_address_size       address_size;
+			uint64_aligned_t                     dummy_align;
+			uint8_t                              dummy_pad[128];
+		};
+	} v5, v6, v7, v8;
+};
+
+/* The actual code comes here */
+
+static inline int hypervisor_domctl(void *domctl)
+{
+	return _hypercall1(int, domctl, domctl);
+}
+
+int xen_guest_address_size(int domid)
+{
+	union xen_domctl domctl;
+	int low, ret;
+
+#define guest_address_size(ver) do {					\
+	memset(&domctl, 0, sizeof(domctl));				\
+	domctl.v##ver.cmd = XEN_DOMCTL_get_address_size;		\
+	domctl.v##ver.interface_version = low = ver;			\
+	domctl.v##ver.domain = domid;					\
+	ret = hypervisor_domctl(&domctl) ?: domctl.v##ver.address_size.size; \
+	if (ret == 32 || ret == 64) {					\
+		pr_info("v" #ver " domctl worked ok: dom%d is %d-bit\n",\
+			domid, ret);					\
+		return ret;						\
+	}								\
+} while (0)
+
+	BUILD_BUG_ON(XEN_DOMCTL_INTERFACE_VERSION > 8);
+	guest_address_size(8);
+#if CONFIG_XEN_COMPAT < 0x040200
+	guest_address_size(7);
+#endif
+#if CONFIG_XEN_COMPAT < 0x040100
+	guest_address_size(6);
+#endif
+#if CONFIG_XEN_COMPAT < 0x040000
+	guest_address_size(5);
+#endif
+#if CONFIG_XEN_COMPAT < 0x030100
+	guest_address_size(4);
+#endif
+
+	ret = BITS_PER_LONG;
+	pr_warn("v%d...%d domctls failed, assuming dom%d is native: %d\n",
+		low, XEN_DOMCTL_INTERFACE_VERSION, domid, ret);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(xen_guest_address_size);
+
+int xen_guest_blkif_protocol(int domid)
+{
+	int address_size = xen_guest_address_size(domid);
+
+	if (address_size == BITS_PER_LONG)
+		return BLKIF_PROTOCOL_NATIVE;
+	if (address_size == 32)
+		return BLKIF_PROTOCOL_X86_32;
+	if (address_size == 64)
+		return BLKIF_PROTOCOL_X86_64;
+	return BLKIF_PROTOCOL_NATIVE;
+}
+EXPORT_SYMBOL_GPL(xen_guest_blkif_protocol);
+
+MODULE_LICENSE("GPL");
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ head/drivers/xen/core/domctl.h	2010-03-25 14:37:59.000000000 +0100
@@ -0,0 +1,2 @@
+int xen_guest_address_size(int domid);
+int xen_guest_blkif_protocol(int domid);