migrate PV EOI MSR

In preparation to adding PV EOI migration for 1.2, trivially refactor some some compat code to make it easier to add version specific cpuid tweaks.

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/pc_piix.c | 44 ++++++++++++++++++++++++++++++++++++--------
 1 file changed, 36 insertions(+), 8 deletions(-)

hw/pc_piix.c +36 -8

index a771d79..008d42f 100644
@@ -369,6 +369,22 @@ static QEMUMachine pc_machine_v1_2 = {
     .default_machine_opts = KVM_MACHINE_OPTIONS,
 };
 
+static void pc_machine_v1_1_compat(void)
+{
+}
+
+static void pc_init_pci_v1_1(ram_addr_t ram_size,
+                             const char *boot_device,
+                             const char *kernel_filename,
+                             const char *kernel_cmdline,
+                             const char *initrd_filename,
+                             const char *cpu_model)
+{
+    pc_machine_v1_1_compat();
+    pc_init_pci(ram_size, boot_device, kernel_filename,
+                kernel_cmdline, initrd_filename, cpu_model);
+}
+
 #define PC_COMPAT_1_1 \
         {\
             .driver   = "virtio-scsi-pci",\

@@ -403,7 +419,7 @@ static QEMUMachine pc_machine_v1_2 = {
 static QEMUMachine pc_machine_v1_1 = {
     .name = "pc-1.1",
     .desc = "Standard PC",
-    .init = pc_init_pci,
+    .init = pc_init_pci_v1_1,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -439,7 +455,7 @@ static QEMUMachine pc_machine_v1_1 = {
 static QEMUMachine pc_machine_v1_0 = {
     .name = "pc-1.0",
     .desc = "Standard PC",
-    .init = pc_init_pci,
+    .init = pc_init_pci_v1_1,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -455,7 +471,7 @@ static QEMUMachine pc_machine_v1_0 = {
 static QEMUMachine pc_machine_v0_15 = {
     .name = "pc-0.15",
     .desc = "Standard PC",
-    .init = pc_init_pci,
+    .init = pc_init_pci_v1_1,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -488,7 +504,7 @@ static QEMUMachine pc_machine_v0_15 = {
 static QEMUMachine pc_machine_v0_14 = {
     .name = "pc-0.14",
     .desc = "Standard PC",
-    .init = pc_init_pci,
+    .init = pc_init_pci_v1_1,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -519,10 +535,22 @@ static QEMUMachine pc_machine_v0_14 = {
             .value    = stringify(1),\
         }
 
+static void pc_init_pci_v0_13(ram_addr_t ram_size,
+                             const char *boot_device,
+                             const char *kernel_filename,
+                             const char *kernel_cmdline,
+                             const char *initrd_filename,
+                             const char *cpu_model)
+{
+    pc_machine_v1_1_compat();
+    pc_init_pci_no_kvmclock(ram_size, boot_device, kernel_filename,
+                            kernel_cmdline, initrd_filename, cpu_model);
+}
+
 static QEMUMachine pc_machine_v0_13 = {
     .name = "pc-0.13",
     .desc = "Standard PC",
-    .init = pc_init_pci_no_kvmclock,
+    .init = pc_init_pci_v0_13,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -560,7 +588,7 @@ static QEMUMachine pc_machine_v0_13 = {
 static QEMUMachine pc_machine_v0_12 = {
     .name = "pc-0.12",
     .desc = "Standard PC",
-    .init = pc_init_pci_no_kvmclock,
+    .init = pc_init_pci_v0_13,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -594,7 +622,7 @@ static QEMUMachine pc_machine_v0_12 = {
 static QEMUMachine pc_machine_v0_11 = {
     .name = "pc-0.11",
     .desc = "Standard PC, qemu 0.11",
-    .init = pc_init_pci_no_kvmclock,
+    .init = pc_init_pci_v0_13,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

@@ -616,7 +644,7 @@ static QEMUMachine pc_machine_v0_11 = {
 static QEMUMachine pc_machine_v0_10 = {
     .name = "pc-0.10",
     .desc = "Standard PC, qemu 0.10",
-    .init = pc_init_pci_no_kvmclock,
+    .init = pc_init_pci_v0_13,
     .max_cpus = 255,
     .default_machine_opts = KVM_MACHINE_OPTIONS,
     .compat_props = (GlobalProperty[]) {

Signature · 1 line

MST

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Quoted 133 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

Too late for 1.2?

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 30 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

Quoted 5 lines

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

We have a ton of code that reads in values then just ignores them, for compat with old qemu.
This will be exactly such a case:
we don't drop anything - protocol does not support this. We read and simply do not tell kvm about it. We also have tons of code that sends useless values again for compatibility.

Quoted 3 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

False, we often send internal device state which is not directly guest visible.

Quoted 6 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 169 lines

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 38 lines · Michael S. Tsirkin

Quoted 35 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

Quoted 2 lines

We have a ton of code that reads in values then just ignores them, for compat with old qemu.

Quoted 4 lines

This will be exactly such a case:
we don't drop anything - protocol does not
support this. We read and simply do not tell kvm
about it.

We also have tons of code that sends

Quoted 1 line

useless values again for compatibility.

Quoted 3 lines

Quoted 3 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

Quoted 4 lines

Quoted 1 line

False, we often send internal device state which is not directly guest visible.

Besides, all MSRs in KVM's MSR list are actually guest visible, even if guests normally do not invoke these MSRs - they can.

Quoted 11 lines

Quoted 7 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 171 lines

Quoted 171 lines

Quoted 3 lines

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 39 lines · Michael S. Tsirkin

Quoted 38 lines · Michael S. Tsirkin

Quoted 35 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

Quoted 2 lines

Quoted 2 lines

We have a ton of code that reads in values then just ignores them, for compat with old qemu.

The difference is that you ignore only the values you _know_ to be safe to be ignored. You can't ignore a value simply because you don't know what it means, and if it's important or not.

Quoted 4 lines

Quoted 4 lines

This will be exactly such a case:
we don't drop anything - protocol does not
support this. We read and simply do not tell kvm
about it.

This fits the definition of "dropping", to me.

Quoted 2 lines

We also have tons of code that sends

Quoted 1 line

useless values again for compatibility.

But these values should be ignored only if the receiver knows exactly what it means, and knows exactly why/when it can be ignored.

Quoted 3 lines

Quoted 3 lines

Quoted 3 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

It may be not directly visible, but if it's part of the machine state, it affects the guest in some way. If it didn't affect the guest in any way, we wouldn't even have to send it while migrating, in the first place.

Quoted 7 lines

Quoted 4 lines

Quoted 1 line

False, we often send internal device state which is not directly guest visible.

Besides, all MSRs in KVM's MSR list are actually guest visible, even if guests normally do not invoke these MSRs - they can.

The difference is that the machine doesn't guarantee a MSR to be migrated unless the corresponding feature bit reports the MSR as supported.

e.g. using the PV EOI MSR without having the PV EOI CPUID bit set is undefined behavior. The guest shouldn't do it.

On the other hand, if the PV EOI CPUID bit is set, the MSR should be always kept, no matter what.

This means that we need to migrate the MSR only if the corresponding CPUID bit is enabled.

Quoted 11 lines

Quoted 11 lines

Quoted 7 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

Quoted 171 lines

Quoted 171 lines

Quoted 171 lines

Quoted 3 lines

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 53 lines · Eduardo Habkost

Quoted 39 lines · Michael S. Tsirkin

Quoted 38 lines · Michael S. Tsirkin

Quoted 35 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

Quoted 2 lines

Quoted 2 lines

We have a ton of code that reads in values then just ignores them, for compat with old qemu.

The difference is that you ignore only the values you _know_ to be safe to be ignored. You can't ignore a value simply because you don't know what it means, and if it's important or not.

Quoted 4 lines

Quoted 4 lines

This will be exactly such a case:
we don't drop anything - protocol does not
support this. We read and simply do not tell kvm
about it.

This fits the definition of "dropping", to me.

Quoted 5 lines

Quoted 2 lines

We also have tons of code that sends

Quoted 1 line

useless values again for compatibility.

But these values should be ignored only if the receiver knows exactly what it means, and knows exactly why/when it can be ignored.

Sorry I just do not understand these meta arguments. Do you mean the example below? If yes let us focus on it please.

Quoted 22 lines

Quoted 4 lines

Quoted 4 lines

Quoted 4 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

It may be not directly visible, but if it's part of the machine state, it affects the guest in some way. If it didn't affect the guest in any way, we wouldn't even have to send it while migrating, in the first place.

Quoted 7 lines

Quoted 4 lines

Quoted 1 line

False, we often send internal device state which is not directly guest visible.

Besides, all MSRs in KVM's MSR list are actually guest visible, even if guests normally do not invoke these MSRs - they can.

The difference is that the machine doesn't guarantee a MSR to be migrated unless the corresponding feature bit reports the MSR as supported.

Quoted 2 lines

e.g. using the PV EOI MSR without having the PV EOI CPUID bit set is undefined behavior. The guest shouldn't do it.

So the problematic scenario involves a guest that sets PV EOI when CPUID bit is off on source. Why do we even care what happens on migration?

Quoted 7 lines

On the other hand, if the PV EOI CPUID bit is set, the MSR should be always kept, no matter what.

This means that we need to migrate the MSR only if the corresponding CPUID bit is enabled.

I do not follow. Why does it mean that?
It seems completely safe to migrate this MSR no matter what.

Quoted 43 lines

Quoted 9 lines

Quoted 9 lines

Quoted 8 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 3 lines

Quoted 3 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Quoted 16 lines

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

In this example, we should migrate CPUID (don't we?). Destination should validate that CPUID supplied by source matches what it can support (doesn't it?).

If we do and it does not, it's an unrelated bug:
CPUID changing under guest's feet.

Quoted 175 lines

Quoted 172 lines

Quoted 172 lines

Quoted 172 lines

Quoted 4 lines

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 54 lines · Michael S. Tsirkin

Quoted 53 lines · Eduardo Habkost

Quoted 39 lines · Michael S. Tsirkin

Quoted 38 lines · Michael S. Tsirkin

Quoted 35 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

Quoted 2 lines

Quoted 2 lines

We have a ton of code that reads in values then just ignores them, for compat with old qemu.

The difference is that you ignore only the values you _know_ to be safe to be ignored. You can't ignore a value simply because you don't know what it means, and if it's important or not.

Quoted 4 lines

Quoted 4 lines

This will be exactly such a case:
we don't drop anything - protocol does not
support this. We read and simply do not tell kvm
about it.

This fits the definition of "dropping", to me.

Quoted 8 lines

Quoted 5 lines

Quoted 2 lines

We also have tons of code that sends

Quoted 1 line

useless values again for compatibility.

But these values should be ignored only if the receiver knows exactly what it means, and knows exactly why/when it can be ignored.

Sorry I just do not understand these meta arguments. Do you mean the example below? If yes let us focus on it please.

We have to have these meta arguments, because the problem is about features/MSRs that will be introduced in the future. But okay, let's follow with the specific argument:

Quoted 25 lines

Quoted 24 lines

Quoted 4 lines

Quoted 4 lines

Quoted 4 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

It may be not directly visible, but if it's part of the machine state, it affects the guest in some way. If it didn't affect the guest in any way, we wouldn't even have to send it while migrating, in the first place.

Quoted 7 lines

Quoted 4 lines

Quoted 1 line

False, we often send internal device state which is not directly guest visible.

Besides, all MSRs in KVM's MSR list are actually guest visible, even if guests normally do not invoke these MSRs - they can.

The difference is that the machine doesn't guarantee a MSR to be migrated unless the corresponding feature bit reports the MSR as supported.

e.g. using the PV EOI MSR without having the PV EOI CPUID bit set is undefined behavior. The guest shouldn't do it.

Quoted 2 lines

So the problematic scenario involves a guest that sets PV EOI when CPUID bit is off on source. Why do we even care what happens on migration?

This is the case where there is _no_ problem. But the migration destination can't know that, how can it decide if it's safe to drop the MSR value or not?

Quoted 8 lines

Quoted 7 lines

On the other hand, if the PV EOI CPUID bit is set, the MSR should be always kept, no matter what.

This means that we need to migrate the MSR only if the corresponding CPUID bit is enabled.

Quoted 2 lines

I do not follow. Why does it mean that?
It seems completely safe to migrate this MSR no matter what.

It is completely safe to migrate the MSR, but it is completely unsafe to _ignore_ the incoming MSR value. That's the problem.

Quoted 64 lines

Quoted 43 lines

Quoted 9 lines

Quoted 9 lines

Quoted 8 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 3 lines

Quoted 3 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Quoted 16 lines

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

In this example, we should migrate CPUID (don't we?). Destination should validate that CPUID supplied by source matches what it can support (doesn't it?).

Quoted 2 lines

If we do and it does not, it's an unrelated bug:
CPUID changing under guest's feet.

CPUID changing under guest's feet is another problem, that we also have to solve.
But we also have the problem of migration compatibility between different host kernels.

Note that I am not saying that migrating all MSRs is wrong. It _is_ good, as long as:
- The destination never ignores any of the incoming MSR values.

- We don't do that by default on "pc-<version>", or we defeat the purpose of machine-types.

I'll try to enumerate the problems I am trying to address (that are related, but are separate problems):

- MSR not being migrated when it should:
- Possible solution: migrate all MSRs even if qemu doesn't know what they are.
- Constraint: migration destination must _never_ ignore any incoming MSR value, because it can't decide if it is important to the guest or not (with the current KVM interfaces).
- Problem with this solution: if we do that by default on "pc-<version>", we break migration compatibility between hosts with different kernel versions.

- Changing CPUID bits under guest's feet.
- Proposed solution: migrating CPUID bits, refuse migration if destination doesn't support the same bits.
- It solves the compatibility problem for migration to a newer kernel, but not to an older kernel. It helps to solve part of the problem, but not all.

- Alternative solution: simply make the resulting CPUID bits not be a function of the host kernel capabilities, but only of the qemu configuration (except on "-cpu host" and "-M pc-next").

- Migration compatibility/predictability:
- See my example above: feature introduced in a newer kernel, migration to an older kernel.

- The only way I see to guarantee this is to never enable unknown CPUID bits or MSRs. People who don't care about predictable migration compatibility can use "-M pc-next", then.

Quoted 175 lines

Quoted 175 lines

Quoted 172 lines

Quoted 172 lines

Quoted 172 lines

Quoted 4 lines

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 96 lines · Eduardo Habkost

Quoted 54 lines · Michael S. Tsirkin

Quoted 53 lines · Eduardo Habkost

Quoted 39 lines · Michael S. Tsirkin

Quoted 38 lines · Michael S. Tsirkin

Quoted 35 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

Quoted 2 lines

Quoted 2 lines

We have a ton of code that reads in values then just ignores them, for compat with old qemu.

The difference is that you ignore only the values you _know_ to be safe to be ignored. You can't ignore a value simply because you don't know what it means, and if it's important or not.

Quoted 4 lines

Quoted 4 lines

This will be exactly such a case:
we don't drop anything - protocol does not
support this. We read and simply do not tell kvm
about it.

This fits the definition of "dropping", to me.

Quoted 8 lines

Quoted 5 lines

Quoted 2 lines

We also have tons of code that sends

Quoted 1 line

useless values again for compatibility.

But these values should be ignored only if the receiver knows exactly what it means, and knows exactly why/when it can be ignored.

Sorry I just do not understand these meta arguments. Do you mean the example below? If yes let us focus on it please.

We have to have these meta arguments, because the problem is about features/MSRs that will be introduced in the future. But okay, let's follow with the specific argument:

Quoted 27 lines

Quoted 18 lines

Quoted 4 lines

Quoted 4 lines

Quoted 4 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

It may be not directly visible, but if it's part of the machine state, it affects the guest in some way. If it didn't affect the guest in any way, we wouldn't even have to send it while migrating, in the first place.

Quoted 7 lines

Quoted 4 lines

Quoted 1 line

False, we often send internal device state which is not directly guest visible.

Besides, all MSRs in KVM's MSR list are actually guest visible, even if guests normally do not invoke these MSRs - they can.

Quoted 6 lines

The difference is that the machine doesn't guarantee a MSR to be migrated unless the corresponding feature bit reports the MSR as supported.

e.g. using the PV EOI MSR without having the PV EOI CPUID bit set is undefined behavior. The guest shouldn't do it.

So the problematic scenario involves a guest that sets PV EOI when CPUID bit is off on source. Why do we even care what happens on migration?

Quoted 3 lines

This is the case where there is _no_ problem. But the migration destination can't know that, how can it decide if it's safe to drop the MSR value or not?

I looked and it seems always safe to drop MSRs that we have. Any counter examples.

Quoted 9 lines

Quoted 9 lines

Quoted 8 lines

On the other hand, if the PV EOI CPUID bit is set, the MSR should be always kept, no matter what.

This means that we need to migrate the MSR only if the corresponding CPUID bit is enabled.

Quoted 5 lines

Quoted 2 lines

I do not follow. Why does it mean that?
It seems completely safe to migrate this MSR no matter what.

It is completely safe to migrate the MSR, but it is completely unsafe to _ignore_ the incoming MSR value. That's the problem.

Normally CPUID will tell you if such important MSR is available. So we can check that at destination.

Quoted 68 lines

Quoted 67 lines

Quoted 44 lines

Quoted 10 lines

Quoted 10 lines

Quoted 9 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 3 lines

Quoted 3 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Quoted 16 lines

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

In this example, we should migrate CPUID (don't we?). Destination should validate that CPUID supplied by source matches what it can support (doesn't it?).

If we do and it does not, it's an unrelated bug:
CPUID changing under guest's feet.

Quoted 4 lines

CPUID changing under guest's feet is another problem, that we also have to solve.
But we also have the problem of migration compatibility between different host kernels.

So here is the solution for both: on destination pass CPUID to kvm and it should come back unchanged. If it changed you fail migration.

Quoted 5 lines

Note that I am not saying that migrating all MSRs is wrong. It _is_ good, as long as:
- The destination never ignores any of the incoming MSR values.

What I am saying for MSRs added in last 2 years it is OK to ignore because CPUID check will tell you if it is supported and fail migration.

Quoted 7 lines

- We don't do that by default on "pc-<version>", or we defeat the purpose of machine-types.

I'll try to enumerate the problems I am trying to address (that are related, but are separate problems):

Quoted 9 lines

- MSR not being migrated when it should:
- Possible solution: migrate all MSRs even if qemu doesn't know what they are.
- Constraint: migration destination must _never_ ignore any incoming MSR value, because it can't decide if it is important to the guest or not (with the current KVM interfaces).
- Problem with this solution: if we do that by default on "pc-<version>", we break migration compatibility between hosts with different kernel versions.

Solution: add vcpu ioctl that tells you which MSRs to migrate (on source), depending on CPUID.

Quoted 7 lines

- Changing CPUID bits under guest's feet.
- Proposed solution: migrating CPUID bits, refuse migration if destination doesn't support the same bits.
- It solves the compatibility problem for migration to a newer kernel, but not to an older kernel. It helps to solve part of the problem, but not all.

How does not it save all of the problem? If destination kernel can support cpuid, then we are fine - it is new enough.

Quoted 4 lines

- Alternative solution: simply make the resulting CPUID bits not be a function of the host kernel capabilities, but only of the qemu configuration (except on "-cpu host" and "-M pc-next").

This perpetuates existing duplication of code between kvm and qemu. We are better off with logic in 1 place.

Quoted 4 lines

- Migration compatibility/predictability:
- See my example above: feature introduced in a newer kernel, migration to an older kernel.

If it is enabled then migration fails.

Quoted 4 lines

- The only way I see to guarantee this is to never enable unknown CPUID bits or MSRs. People who don't care about predictable migration compatibility can use "-M pc-next", then.

Guarantee what?

Just check dst can support all msrs and cpuid bits of src. Way to check is to ask kvm :) Not to add logic in qemu.

Quoted 181 lines

Quoted 178 lines

Quoted 178 lines

Quoted 175 lines

Quoted 175 lines

Quoted 175 lines

Quoted 7 lines

Too late for 1.2?

Absolutely (in my opinion).

Quoted 9 lines

Quoted 6 lines

While we don't make the KVM feature-bit handling sane (with defaults that are not blindly derived from the host kernel capabilities), maybe the safest bet is to expect users to not migrate between hosts running kernels with different KVM capabilities? (I am not sure which option is better)

Sorry not sure what you talk about here. What has KVM feature-bit handling to do with this patchset?

Everything? The whole point of this patch is to filter out the PV_EOI KVM feature bit.

This part of the current code, specifically, is wrong:

Quoted 1 line

Quoted 1 line

Quoted 1 line

plus_kvm_features = ~0; /* not supported bits will be filtered out later */

The QEMU-side list of KVM features should be whitelist-based, not blacklist-based (unless the user doesn't need migration, in that case he can use "-cpu host" and get every feature blindly enabled), because QEMU can't know if a new feature involves guest-visible state that has to be migrated.

Quoted 137 lines

Quoted 137 lines

Quoted 134 lines

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Signature · 1 line

Eduardo

Quoted 2 lines

Normally CPUID will tell you if such important MSR is available. So we can check that at destination.

How can qemu check it, if when the qemu code was written when the MSR didn't even exist yet?

(You could add an interface to check for that, but there's no KVM ioctl() to tell qemu "given these CPUID bits, can I safely drop this MSR that I don't even know about?")

Quoted 68 lines

Quoted 68 lines

Quoted 67 lines

Quoted 44 lines

Quoted 10 lines

Quoted 10 lines

Quoted 9 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 3 lines

Quoted 3 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Quoted 16 lines

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

In this example, we should migrate CPUID (don't we?). Destination should validate that CPUID supplied by source matches what it can support (doesn't it?).

If we do and it does not, it's an unrelated bug:
CPUID changing under guest's feet.

Quoted 7 lines

Quoted 4 lines

CPUID changing under guest's feet is another problem, that we also have to solve.
But we also have the problem of migration compatibility between different host kernels.

So here is the solution for both: on destination pass CPUID to kvm and it should come back unchanged. If it changed you fail migration.

This doesn't solve the problem of having predictable migration compatibility for "-M pc-<old-version>".

The whole point of machine-types is to expose the "same machine" to the guest, even if you change the hardware or host kernel. "qemu -M pc-1.x" must expose the same machine configuration to the guest, it doesn 't matter what's the host kernel version.

Quoted 9 lines

Quoted 5 lines

Note that I am not saying that migrating all MSRs is wrong. It _is_ good, as long as:
- The destination never ignores any of the incoming MSR values.

What I am saying for MSRs added in last 2 years it is OK to ignore because CPUID check will tell you if it is supported and fail migration.

Existing MSRs are easy to make work. The problem is about MSRs added to the "msrs_to_save" list in the future.

Also, the problem is not about being "safe" to ignore the MSR values, it's about being "correct" (part of the expected behavior of the virtual machine). The fact that most guests doen't crash when the virtual machine doesn't behave as it should doesn't mean we should do it.

Either the MSR is part of the machine state (and relevant to the guest), or not. If it is relevant, it must be _always_ migrated and never dropped by the destination. If it is not, it's useless to migrate it.

Quoted 17 lines

Quoted 7 lines

- We don't do that by default on "pc-<version>", or we defeat the purpose of machine-types.

I'll try to enumerate the problems I am trying to address (that are related, but are separate problems):

Quoted 9 lines

- MSR not being migrated when it should:
- Possible solution: migrate all MSRs even if qemu doesn't know what they are.
- Constraint: migration destination must _never_ ignore any incoming MSR value, because it can't decide if it is important to the guest or not (with the current KVM interfaces).
- Problem with this solution: if we do that by default on "pc-<version>", we break migration compatibility between hosts with different kernel versions.

Quoted 2 lines

Solution: add vcpu ioctl that tells you which MSRs to migrate (on source), depending on CPUID.

This may be a solution for old-kernel => new-kernel migration, yes. But this still doesn't solve the migration compatibility problem for new-kernel => old-kernel migration (see below).

Quoted 10 lines

Quoted 7 lines

- Changing CPUID bits under guest's feet.
- Proposed solution: migrating CPUID bits, refuse migration if destination doesn't support the same bits.
- It solves the compatibility problem for migration to a newer kernel, but not to an older kernel. It helps to solve part of the problem, but not all.

How does not it save all of the problem? If destination kernel can support cpuid, then we are fine - it is new enough.

See below. The problem is being able to migrate to an older host. That's the whole point of machine-types!

Quoted 5 lines

Quoted 4 lines

- Alternative solution: simply make the resulting CPUID bits not be a function of the host kernel capabilities, but only of the qemu configuration (except on "-cpu host" and "-M pc-next").

Quoted 2 lines

This perpetuates existing duplication of code between kvm and qemu. We are better off with logic in 1 place.

Yes, it does, and I would love to avoid having the list inside QEMU, too. But we can't avoid that because each machine-type defines a set of available features/MSRs, so we have to have a machine-type => list-of-features list somewhere, unfortunately.

Quoted 12 lines

Quoted 4 lines

- Migration compatibility/predictability:
- See my example above: feature introduced in a newer kernel, migration to an older kernel.

If it is enabled then migration fails.

Quoted 4 lines

- The only way I see to guarantee this is to never enable unknown CPUID bits or MSRs. People who don't care about predictable migration compatibility can use "-M pc-next", then.

Quoted 1 line

Guarantee what?

Guarantee that "-M pc-1.1" machines can be migrated to any host that is already capable of running "-M pc-1.1".

Quoted 2 lines

Just check dst can support all msrs and cpuid bits of src. Way to check is to ask kvm :) Not to add logic in qemu.

Checking and making migration fail when it has to fail is not the problem. The problem is that now "qemu -M pc-1.x" will result in a different machine, depending on the host kernel version. This causes two problems:

- Now you don't know if your existing machine can be migrated to a host running an older kernel (because now migration can fail even when you are using the same machine-type on both sides).
- Different VMs using the same machine-type will get different machines (with different sets of features), because they are running on different kernel versions.

This may be acceptable for "pc-next", but not for "pc-<version>".

Quoted 10 lines · Eduardo Habkost

Quoted 2 lines

Normally CPUID will tell you if such important MSR is available. So we can check that at destination.

How can qemu check it, if when the qemu code was written when the MSR didn't even exist yet?

(You could add an interface to check for that, but there's no KVM ioctl() to tell qemu "given these CPUID bits, can I safely drop this MSR that I don't even know about?")

So this is what I suggest exactly. Add a new ioctl like that.

Quoted 76 lines

Quoted 68 lines

Quoted 68 lines

Quoted 67 lines

Quoted 44 lines

Quoted 10 lines

Quoted 10 lines

Quoted 9 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

Quoted 3 lines

Quoted 3 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Yes:

Suppose kernel 3.7 introduces KVM_FOO_MSR and CPUID_KVM_FOO.

Also, suppose QEMU 1.2 doesn't know anything about KVM_FOO, because it was release before this feature was introduced.

Then you run "qemu-1.2 -M pc-1.2" on a 3.7 host kernel. qemu-1.2 can do two things here:

1) Not enable CPUID_KVM_FOO

In this case, the guest should not use KVM_FOO_MSR and the MSR does not need to be migrated (the guest may try to use it, but the behavior when trying to use it is undefined). Sending the MSR value when migrating would be useless.

2) Enable CPUID_KVM_FOO.

In this case, the guest may try to use the feature and write something into KVM_FOO_MSR. Sending the MSR value when migrating is absolutely necessary

---

Now assume you run "qemu-1.2 -M pc-1.2" in the destination host, running the 3.6 kernel (without KVM_FOO).

Quoted 16 lines

Then qemu-1.2 receives the KVM_FOO_MSR data in the migration stream. In this case, qemu-1.2 simply can't decide if it's safe to drop the data (and not tell KVM about it), or not.
If we simply send every MSR reported by the kernel, both the origin and destination qemu-1.2 processes can't ever know if the MSR value is important or not, because it doesn't know if it's part of the machine state that has to be kept consistent.
We could introduce a mode of operation where _every_ MSR reported by KVM is important and sent by the origin (and also where every MSR must be handled by the destination, otherwise migration would fail). But this new mode would break migration compatibility between two hosts running the same machine-type, only because they are running different kernel versions. But it may be useful for some use cases, so maybe it's appropriate for a future "pc-next" machine-type (and for "-cpu host"), but not for "pc-<version>".

In this example, we should migrate CPUID (don't we?). Destination should validate that CPUID supplied by source matches what it can support (doesn't it?).

If we do and it does not, it's an unrelated bug:
CPUID changing under guest's feet.

Quoted 7 lines

Quoted 4 lines

CPUID changing under guest's feet is another problem, that we also have to solve.
But we also have the problem of migration compatibility between different host kernels.

So here is the solution for both: on destination pass CPUID to kvm and it should come back unchanged. If it changed you fail migration.

Quoted 7 lines

This doesn't solve the problem of having predictable migration compatibility for "-M pc-<old-version>".

The whole point of machine-types is to expose the "same machine" to the guest, even if you change the hardware or host kernel. "qemu -M pc-1.x" must expose the same machine configuration to the guest, it doesn 't matter what's the host kernel version.

I'd tend to disagree. The point is to make migration work and avoid things like windows re-activation trigger. Let's not be purists - many internal changes in qemu introduce subtle guest visible changes, if even in timing.

Quoted 19 lines

Quoted 10 lines

Quoted 6 lines

Note that I am not saying that migrating all MSRs is wrong. It _is_ good, as long as:
- The destination never ignores any of the incoming MSR values.

What I am saying for MSRs added in last 2 years it is OK to ignore because CPUID check will tell you if it is supported and fail migration.

Existing MSRs are easy to make work. The problem is about MSRs added to the "msrs_to_save" list in the future.

Also, the problem is not about being "safe" to ignore the MSR values, it's about being "correct" (part of the expected behavior of the virtual machine). The fact that most guests doen't crash when the virtual machine doesn't behave as it should doesn't mean we should do it.

Quoted 4 lines

Either the MSR is part of the machine state (and relevant to the guest), or not. If it is relevant, it must be _always_ migrated and never dropped by the destination. If it is not, it's useless to migrate it.

Yes. But it is better to keep all knowledge which is which in one place which is in kvm.

Quoted 48 lines

Quoted 18 lines

Quoted 8 lines

- We don't do that by default on "pc-<version>", or we defeat the purpose of machine-types.

I'll try to enumerate the problems I am trying to address (that are related, but are separate problems):

Quoted 9 lines

- MSR not being migrated when it should:
- Possible solution: migrate all MSRs even if qemu doesn't know what they are.
- Constraint: migration destination must _never_ ignore any incoming MSR value, because it can't decide if it is important to the guest or not (with the current KVM interfaces).
- Problem with this solution: if we do that by default on "pc-<version>", we break migration compatibility between hosts with different kernel versions.

Quoted 2 lines

Solution: add vcpu ioctl that tells you which MSRs to migrate (on source), depending on CPUID.

This may be a solution for old-kernel => new-kernel migration, yes. But this still doesn't solve the migration compatibility problem for new-kernel => old-kernel migration (see below).

Quoted 10 lines

Quoted 7 lines

- Changing CPUID bits under guest's feet.
- Proposed solution: migrating CPUID bits, refuse migration if destination doesn't support the same bits.
- It solves the compatibility problem for migration to a newer kernel, but not to an older kernel. It helps to solve part of the problem, but not all.

How does not it save all of the problem? If destination kernel can support cpuid, then we are fine - it is new enough.

See below. The problem is being able to migrate to an older host. That's the whole point of machine-types!

Quoted 7 lines

Quoted 4 lines

- Alternative solution: simply make the resulting CPUID bits not be a function of the host kernel capabilities, but only of the qemu configuration (except on "-cpu host" and "-M pc-next").

This perpetuates existing duplication of code between kvm and qemu. We are better off with logic in 1 place.

Quoted 5 lines

Yes, it does, and I would love to avoid having the list inside QEMU, too. But we can't avoid that because each machine-type defines a set of available features/MSRs, so we have to have a machine-type => list-of-features list somewhere, unfortunately.

Yes. But let us have machine type->cpuid list in qemu. kvm will have the cpuid->MSR logic.

Quoted 16 lines

Quoted 15 lines

Quoted 6 lines

- Migration compatibility/predictability:
- See my example above: feature introduced in a newer kernel, migration to an older kernel.

If it is enabled then migration fails.

Quoted 4 lines

- The only way I see to guarantee this is to never enable unknown CPUID bits or MSRs. People who don't care about predictable migration compatibility can use "-M pc-next", then.

Guarantee what?

Quoted 3 lines

Guarantee that "-M pc-1.1" machines can be migrated to any host that is already capable of running "-M pc-1.1".

I can't change the past. I am suggesting forward compatible approach so we'll be able to guarantee this for
-M pc-1.2 and on.

Quoted 16 lines

Quoted 3 lines

Just check dst can support all msrs and cpuid bits of src. Way to check is to ask kvm :) Not to add logic in qemu.

Checking and making migration fail when it has to fail is not the problem. The problem is that now "qemu -M pc-1.x" will result in a different machine, depending on the host kernel version. This causes two problems:

- Now you don't know if your existing machine can be migrated to a host running an older kernel (because now migration can fail even when you are using the same machine-type on both sides).
- Different VMs using the same machine-type will get different machines (with different sets of features), because they are running on different kernel versions.

Quoted 1 line

This may be acceptable for "pc-next", but not for "pc-<version>".

So you can whitelist CPUID bits. But leave MSRs alone, it is nasty enough with CPUID.

Quoted 30 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

Quoted 5 lines

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

We have a ton of code that reads in values then just ignores them, for compat with old qemu.
This will be exactly such a case:
we don't drop anything - protocol does not support this. We read and simply do not tell kvm about it. We also have tons of code that sends useless values again for compatibility.

Quoted 3 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

False, we often send internal device state which is not directly guest visible.

Quoted 6 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 3 lines

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

Quoted 8 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 3 lines

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

My concern here is that you are introducing infrastructure by using _one example_, claiming it to solve a general problem, without the appropriate amount of energy spent in this solution.

Using old machines is supposed to be supported on older guests.

Michael, not migrating the PVEOI MSR actually crashes the guest?

Quoted 13 lines · Marcelo Tosatti

Quoted 8 lines

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 3 lines

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

My concern here is that you are introducing infrastructure by using _one example_, claiming it to solve a general problem, without the appropriate amount of energy spent in this solution.

Quoted 3 lines

Using old machines is supposed to be supported on older guests.

Michael, not migrating the PVEOI MSR actually crashes the guest?

Which is irrelevant since it won't be able to migrate to an older qemu.

Quoted 65 lines · Michael S. Tsirkin

Quoted 30 lines · Eduardo Habkost

Quoted 20 lines · Michael S. Tsirkin

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Quoted 8 lines

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all
- send all MSRS in key/value format
- destination gets list of MSRs to be migrated from KVM and
only restores the supported ones

Quoted 5 lines

As far as I understand the migration code requirements/expectations, if the origin is sending some data, it is because it is part of the guest-visible machine state that must be kept while migrating. Because of that, the destination is not allowed to drop anything it doesn't know about.

We have a ton of code that reads in values then just ignores them, for compat with old qemu.
This will be exactly such a case:
we don't drop anything - protocol does not support this. We read and simply do not tell kvm about it. We also have tons of code that sends useless values again for compatibility.

Quoted 3 lines

At the same time, if it's not part of guest-visible machine state, it doesn't have to be sent by the migration origin.

False, we often send internal device state which is not directly guest visible.

Quoted 6 lines

On the other hand, a mode of operation that doesn't require updating QEMU every time there's a new bit of guest-visible state to be migrated would be nice (just like the "-cpu host" mode, that doesn't require updating QEMU for every new CPU feature, is nice for some use cases). I just don't know how to make work with the current migration protocol.

I don't understand. What is the problem with the proposal? What will not work with our protocol?
Can you give an example please?

Quoted 3 lines

Quoted 3 lines

Quoted 1 line

Too late for 1.2?

Absolutely (in my opinion).

My concern here is that you are introducing infrastructure by using _one example_, claiming it to solve a general problem, without the appropriate amount of energy spent in this solution.

Using old machines is supposed to be supported on older guests.

Michael, not migrating the PVEOI MSR actually crashes the guest?

Quoted 1 line

Michael, not migrating the PVEOI MSR actually crashes the guest?

Not at all. migration currently simply disables PV EOI, so each EOI triggers exits after migration.

Signature · 1 line

MST

Quoted 11 lines · Eduardo Habkost

Quoted 9 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Quoted 7 lines

What about users that are already running "qemu-1.1 -M pc-1.1" on a host kernel that supports PV EOI already? They would get PV EOI disabled when migrating to a destination running "qemu-1.2 -M pc-1.1".

(On the other hand, people running "qemu-1.1 -M pc-1.1" on a host kernel supporting PV EOI already have migration broken, so there's not much we can do for them)

Exactly.

Talked to Gleb, long term I think we should rework code to make it forward-compatible wrt adding new MSRs:
- source gets list of MSRs to be migrated from KVM and simply sends them all

- send all MSRS in key/value format

- destination gets list of MSRs to be migrated from KVM and only restores the supported ones

Too late for 1.2?

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST

Quoted 10 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

As best I can tell, we're papering over an ABI breakage in KVM.

If an old QEMU attempts to do a live migration on a new kernel, migrating to an QEMU on a different box with an older kernel, it will fail because there is state that isn't being migrated.

This ought to be fixed in the kernel by making these features whitelisted by userspace.

Quoted 19 lines · Anthony Liguori

Quoted 10 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

As best I can tell, we're papering over an ABI breakage in KVM.

If an old QEMU attempts to do a live migration on a new kernel, migrating to an QEMU on a different box with an older kernel, it will fail because there is state that isn't being migrated.

Quoted 3 lines

This ought to be fixed in the kernel by making these features whitelisted by userspace.

What do you mean. Userspace and only userspace decides what cpuid bits will be seen by a guest. Currently userspace enables all PV cpuid bits it finds.

--
Gleb.

Quoted 23 lines · Gleb Natapov

Quoted 22 lines · Anthony Liguori

Quoted 10 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

As best I can tell, we're papering over an ABI breakage in KVM.

If an old QEMU attempts to do a live migration on a new kernel, migrating to an QEMU on a different box with an older kernel, it will fail because there is state that isn't being migrated.

This ought to be fixed in the kernel by making these features whitelisted by userspace.

Quoted 3 lines

What do you mean. Userspace and only userspace decides what cpuid bits will be seen by a guest. Currently userspace enables all PV cpuid bits it finds.

Right, I misunderstood from the commit message. I see now that the problem is that bfee7546df51c08e395dc8a7676a5c7f20186fee unconditionally enabled kvm_pv_eoi without taking into account migration support.

I think for 1.2 we should simply revert the above commit and then we can restore it for 1.3 with proper support for migration.

Regards,

Anthony Liguori

Quoted 3 lines

--
Gleb.

Quoted 23 lines · Gleb Natapov

Quoted 22 lines · Anthony Liguori

Quoted 10 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

As best I can tell, we're papering over an ABI breakage in KVM.

If an old QEMU attempts to do a live migration on a new kernel, migrating to an QEMU on a different box with an older kernel, it will fail because there is state that isn't being migrated.

This ought to be fixed in the kernel by making these features whitelisted by userspace.

Quoted 3 lines

What do you mean. Userspace and only userspace decides what cpuid bits will be seen by a guest. Currently userspace enables all PV cpuid bits it finds.

Right, I misunderstood from the commit message. I see now that the problem is that bfee7546df51c08e395dc8a7676a5c7f20186fee unconditionally enabled kvm_pv_eoi without taking into account migration support.

I think for 1.2 we should simply revert the above commit and then we can restore it for 1.3 with proper support for migration.

We can't simply expose the GET_SUPPORTED_CPUID results directly to the guest, or the resulting guest-visible CPUID bits may change under the guest's feet when we live-migrate.

We still have to implement proper per-machine-type migration compatibility bits, but this at least is a workaround for cases where PV EOI would be silently enabled after a host kernel upgrade.

We also have to implement proper check/warnings about unsupported bits when using "-cpu check" or "-cpu enforce", because the default bits may be silently disabled if the host kernel doesn't support them.

Signed-off-by Eduardo Habkost <[email protected]>

---
 target-i386/cpu.c | 18 +++++++++++++++++-
 target-i386/cpu.h | 10 ++++++++++
 2 files changed, 27 insertions(+), 1 deletion(-)

target-i386/cpu.c +17 -1

index 120a2e3..52042f0 100644
@@ -283,6 +283,21 @@ typedef struct x86_def_t {
           CPUID_EXT3_CR8LEG | CPUID_EXT3_ABM | CPUID_EXT3_SSE4A)
 #define TCG_SVM_FEATURES 0
 
+
+/* List of KVM features we (try to) enable by default
+ * 
+ * CPUID_KVM_PV_EOI is not enabled until we have proper per-machine-type
+ * migration compatibility support
+ * 
+ * CPUID_KVM_MMU is not enabled because it is deprecated.
+ */
+#define DEFAULT_KVM_FEATURES (CPUID_KVM_CLOCKSOURCE | \
+                              CPUID_KVM_KVM_NOP_IO_DELAY | \
+                              CPUID_KVM_CLOCKSOURCE2 | \
+                              CPUID_KVM_KVM_ASYNC_PF | \
+                              CPUID_KVM_STEAL_TIME | \
+                              CPUID_KVM_CLOCKSOURCE_STABLE_BIT)
+
 /* maintains list of cpu model definitions
  */
 static x86_def_t *x86_defs = {NULL};

@@ -887,7 +902,8 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
         memcpy(x86_cpu_def, def, sizeof(*def));
     }
 
-    plus_kvm_features = ~0; /* not supported bits will be filtered out later */
+    /* not supported bits will be filtered out later */
+    plus_kvm_features = DEFAULT_KVM_FEATURES;
 
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,

target-i386/cpu.h +3 -0

index 60f9e97..5ce4baf 100644
@@ -440,6 +440,16 @@
 #define CPUID_SVM_PAUSEFILTER  (1 << 10)
 #define CPUID_SVM_PFTHRESHOLD  (1 << 12)
 
+/* CPUID[4000_0001h].EAX bits: */
+#define CPUID_KVM_CLOCKSOURCE             (1 << 0)
+#define CPUID_KVM_KVM_NOP_IO_DELAY        (1 << 1)

Signature · 1 line

1.7.11.4

Quoted 62 lines · Eduardo Habkost

We can't simply expose the GET_SUPPORTED_CPUID results directly to the guest, or the resulting guest-visible CPUID bits may change under the guest's feet when we live-migrate.

We still have to implement proper per-machine-type migration compatibility bits, but this at least is a workaround for cases where PV EOI would be silently enabled after a host kernel upgrade.

We also have to implement proper check/warnings about unsupported bits when using "-cpu check" or "-cpu enforce", because the default bits may be silently disabled if the host kernel doesn't support them.

Signed-off-by Eduardo Habkost <[email protected]>

---
 target-i386/cpu.c | 18 +++++++++++++++++-
 target-i386/cpu.h | 10 ++++++++++
 2 files changed, 27 insertions(+), 1 deletion(-)

target-i386/cpu.c +17 -1

index 120a2e3..52042f0 100644
@@ -283,6 +283,21 @@ typedef struct x86_def_t {
           CPUID_EXT3_CR8LEG | CPUID_EXT3_ABM | CPUID_EXT3_SSE4A)
 #define TCG_SVM_FEATURES 0
 
+
+/* List of KVM features we (try to) enable by default
+ * 
+ * CPUID_KVM_PV_EOI is not enabled until we have proper per-machine-type
+ * migration compatibility support
+ * 
+ * CPUID_KVM_MMU is not enabled because it is deprecated.
+ */
+#define DEFAULT_KVM_FEATURES (CPUID_KVM_CLOCKSOURCE | \
+                              CPUID_KVM_KVM_NOP_IO_DELAY | \
+                              CPUID_KVM_CLOCKSOURCE2 | \
+                              CPUID_KVM_KVM_ASYNC_PF | \
+                              CPUID_KVM_STEAL_TIME | \
+                              CPUID_KVM_CLOCKSOURCE_STABLE_BIT)
+
 /* maintains list of cpu model definitions
  */
 static x86_def_t *x86_defs = {NULL};

@@ -887,7 +902,8 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
         memcpy(x86_cpu_def, def, sizeof(*def));
     }
 
-    plus_kvm_features = ~0; /* not supported bits will be filtered out later */
+    /* not supported bits will be filtered out later */
+    plus_kvm_features = DEFAULT_KVM_FEATURES;
 
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,

target-i386/cpu.h +0 -0

index 60f9e97..5ce4baf 100644
@@ -440,6 +440,16 @@
 #define CPUID_SVM_PAUSEFILTER  (1 << 10)
 #define CPUID_SVM_PFTHRESHOLD  (1 << 12)
 

target-i386/cpu.h line 455

Quoted 13 lines

target-i386/cpu.h +10 -0

+/* CPUID[4000_0001h].EAX bits: */
+#define CPUID_KVM_CLOCKSOURCE             (1 << 0)
+#define CPUID_KVM_KVM_NOP_IO_DELAY        (1 << 1)
+#define CPUID_KVM_MMU                     (1 << 2)
+#define CPUID_KVM_CLOCKSOURCE2            (1 << 3)
+#define CPUID_KVM_KVM_ASYNC_PF            (1 << 4)
+#define CPUID_KVM_STEAL_TIME              (1 << 5)
+#define CPUID_KVM_KVM_PV_EOI              (1 << 6)
+#define CPUID_KVM_CLOCKSOURCE_STABLE_BIT  (1 << 24)
+
 #define CPUID_VENDOR_INTEL_1 0x756e6547 /* "Genu" */
 #define CPUID_VENDOR_INTEL_2 0x49656e69 /* "ineI" */
 #define CPUID_VENDOR_INTEL_3 0x6c65746e /* "ntel" */

Why duplicate macros that kvm linux header already has?

Quoted 10 lines · Michael S. Tsirkin

In preparation for adding PV EOI support, disable PV EOI by default for 1.1 and older machine types, to avoid CPUID changing during migration.

PV EOI can still be enabled/disabled by specifying it explicitly.
Enable for 1.1
-M pc-1.1 -cpu kvm64,+kvm_pv_eoi
Disable for 1.2
-M pc-1.2 -cpu kvm64,-kvm_pv_eoi

Signed-off-by Michael S. Tsirkin <[email protected]>

As best I can tell, we're papering over an ABI breakage in KVM.

If an old QEMU attempts to do a live migration on a new kernel, migrating to an QEMU on a different box with an older kernel, it will fail because there is state that isn't being migrated.

This ought to be fixed in the kernel by making these features whitelisted by userspace.

Regards,

Anthony Liguori

Quoted 129 lines

---
 hw/Makefile.objs  |  2 +-
 hw/cpu_flags.c    | 32 ++++++++++++++++++++++++++++++++
 hw/cpu_flags.h    |  9 +++++++++
 hw/pc_piix.c      |  2 ++
 target-i386/cpu.c |  8 ++++++++
 5 files changed, 52 insertions(+), 1 deletion(-)
 create mode 100644 hw/cpu_flags.c
 create mode 100644 hw/cpu_flags.h

hw/Makefile.objs +1 -1

index 850b87b..3f2532a 100644
@@ -1,5 +1,5 @@
 hw-obj-y = usb/ ide/
-hw-obj-y += loader.o
+hw-obj-y += loader.o cpu_flags.o
 hw-obj-$(CONFIG_VIRTIO) += virtio-console.o
 hw-obj-$(CONFIG_VIRTIO_PCI) += virtio-pci.o
 hw-obj-y += fw_cfg.o

hw/cpu_flags.c new file +32 -0

new file mode 100644
index 0000000..7a633c0
@@ -0,0 +1,32 @@
+/*
+ * CPU compatibility flags.
+ *
+ * Copyright (c) 2012 Red Hat Inc.
+ * Author: Michael S. Tsirkin.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, see <http://www.gnu.org/licenses/>.
+ */
+#include "hw/cpu_flags.h"
+
+static bool kvm_pv_eoi_disabled_state;
+
+void disable_kvm_pv_eoi(void)
+{
+       kvm_pv_eoi_disabled_state = true;
+}
+
+bool kvm_pv_eoi_disabled(void)
+{
+       return kvm_pv_eoi_disabled_state;
+}

hw/cpu_flags.h new file +9 -0

new file mode 100644
index 0000000..05777b6
@@ -0,0 +1,9 @@
+#ifndef HW_CPU_FLAGS_H
+#define HW_CPU_FLAGS_H
+
+#include <stdbool.h>
+
+void disable_kvm_pv_eoi(void);
+bool kvm_pv_eoi_disabled(void);
+
+#endif

hw/pc_piix.c +2 -0

index 008d42f..bdbceda 100644
@@ -46,6 +46,7 @@
 #ifdef CONFIG_XEN
 #  include <xen/hvm/hvm_info_table.h>
 #endif
+#include "cpu_flags.h"
 
 #define MAX_IDE_BUS 2
 

@@ -371,6 +372,7 @@ static QEMUMachine pc_machine_v1_2 = {
 
 static void pc_machine_v1_1_compat(void)
 {
+    disable_kvm_pv_eoi();
 }
 
 static void pc_init_pci_v1_1(ram_addr_t ram_size,

target-i386/cpu.c +8 -0

index 120a2e3..0d02fd1 100644
@@ -23,6 +23,7 @@
 
 #include "cpu.h"
 #include "kvm.h"
+#include "asm/kvm_para.h"
 
 #include "qemu-option.h"
 #include "qemu-config.h"

@@ -33,6 +34,7 @@
 #include "hyperv.h"
 
 #include "hw/hw.h"
+#include "hw/cpu_flags.h"
 
 /* feature flags taken from "Intel Processor Identification and the CPUID
  * Instruction" and AMD's "CPUID Specification".  In cases of disagreement

@@ -889,6 +891,12 @@ static int cpu_x86_find_by_name(x86_def_t *x86_cpu_def, const char *cpu_model)
 
     plus_kvm_features = ~0; /* not supported bits will be filtered out later */
 
+    /* Disable PV EOI for old machine types.
+     * Feature flags can still override. */
+    if (kvm_pv_eoi_disabled()) {
+        plus_kvm_features &= ~(0x1 << KVM_FEATURE_PV_EOI);
+    }
+
     add_flagname_to_bitmaps("hypervisor", &plus_features,
         &plus_ext_features, &plus_ext2_features, &plus_ext3_features,
         &plus_kvm_features, &plus_svm_features);

Signature · 1 line

MST