Skip to contentSkip to conversation

Introduce support for confidential guest reset (x86)

PATCH Superseded Ani Sinha 6 participants 44 messages Feb 12 – Feb 13

Subsystems (3)

v4 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Ani Sinha

author
#0 lore raw

v4 cover Introduce support for confidential guest reset (x86)

This change introduces support for confidential guests
(SEV-ES, SEV-SNP and TDX) to reset/reboot just like other non-confidential guests. Currently, a reboot intiated from the confidential guest results in termination of the QEMU hypervisor as the CPUs are not resettable. As the initial state of the guest including private memory is locked and encrypted, the contents of that memory will not be accessible post reset. Hence a new KVM file descriptor must be opened to create a new confidential VM context closing the old one. All KVM VM specific ioctls must be called again. New VCPU file descriptors must be created against the new KVM fd and most VCPU ioctls must be called again as well.

This change perfoms closing of the old KVM fd and creating a new one. After the new KVM fd is opened, all generic and architecture specific ioctl calls are issued again. Notifiers are added to notify subsystems that:
- The KVM file fd is about to be changed to state sync-ing from KVM to QEMU should be done if required.
- The KVM file fd has changed, so ioctl calls to the new KVM fd has to be performed again.
- That new VCPU fds are created so that VCPU ioctl calls must be called again where required.

Specific subsystems use these notifiers to re-issue ioctl calls where required.

Changes are made to SEV and TDX modules to reinitialize the confidential guest state and seal it again. Along the way, some bug fixes are made so that some initialization functions can be called again. Some refactoring of existing code is done so that both init and reset paths can use them.

Tested on TDX, SEV-ES and SEV-SNP. Tested on non-coco hardware. Tested with Xen emulation enabled.

CI pipeline passes (one failure is unrelated):
https://gitlab.com/anisinha/qemu/-/pipelines/2320974514

Added functional test passed:

$ ./build/run tests/functional/x86_64/test_vmfd_change_reboot.py 
TAP version 13
ok 1 test_vmfd_change_reboot.KVMGuest.test_reset_console
ok 2 test_vmfd_change_reboot.KVMGuest.test_reset_hyperv_vmbus
ok 3 test_vmfd_change_reboot.KVMGuest.test_reset_kvmpit
ok 4 test_vmfd_change_reboot.KVMGuest.test_reset_qmp
ok 5 test_vmfd_change_reboot.KVMGuest.test_reset_xen_emulation
1..5

Please review and test.

Changelog:

v4:
- Fixed reset on non-coco with qmp "system_reset" command.
- Numerious misc fixes.
- addressed review comments from v3.
- dropped three patches that are not required.
- Added more functional tests including one vmbus test.
- added noop callbacks to stubs/kvm removing them from arch specific headers.
- Tags added.
- Rebased.

v3:
- Combined pre and post vmfd change notifier into one.
- rename kvm_arch_vmfd_change_ops() -> kvm_arch_on_vmfd_change()
- reuse kvm_arch_init() code in kvm_arch_on_vmfd_change()
- moved around migration blockers and notifers to more appropriate place.
- fixed Xen emulation.
- fixed SEV-ES reset.
- fixed/reorganized reset code in system/runstate.c
- can_rebuild_guest_state is now a boolean not a callback.
- misc fixes.
- added a functional test for Xen emulation with vmfd change.
- rebased.

v2:
- Bugfixes.
- Added a new machine option so that we can exercize most of the non-coco changes related to reboot on non-coco platforms.
- added a new functional test. Currently its skipped on CI pipeline as KVM is not enabled (no /dev/kvm on the container)for QEMU CI tests. It can be run manually and it passes on those systems where KVM is enabled.
- Addressed comments from v1 with regards to refactoring of code, code simplication by removal of redundant stuff, moved around code
so that notifiers, migration blockers are added only on one place.
- Added some tracepoints for future debugging on newly added functions.
- Rebased.

Ani Sinha (31):
i386/kvm: avoid installing duplicate msr entries in msr_handlers accel/kvm: add confidential class member to indicate guest rebuild
capability
hw/accel: add a per-accelerator callback to change VM accelerator
handle
system/physmem: add helper to reattach existing memory after KVM VM fd
change
accel/kvm: add changes required to support KVM VM file descriptor
change
accel/kvm: add a notifier to indicate KVM VM file descriptor has
changed
accel/kvm: notify when KVM VM file fd is about to be changed i386/kvm: unregister smram listeners prior to vm file descriptor
change
kvm/i386: implement architecture support for kvm file descriptor
change
hw/i386: refactor x86_bios_rom_init for reuse in confidential guest
reset
hw/i386: export a new function x86_bios_rom_reload kvm/i386: reload firmware for confidential guest reset accel/kvm: rebind current VCPUs to the new KVM VM file descriptor upon
reset
i386/tdx: refactor TDX firmware memory initialization code into a new
function
i386/tdx: finalize TDX guest state upon reset
i386/tdx: add a pre-vmfd change notifier to reset tdx state i386/sev: add migration blockers only once
i386/sev: add notifiers only once
i386/sev: free existing launch update data and kernel hashes data on
init
i386/sev: add support for confidential guest reset hw/vfio: generate new file fd for pseudo device and rebind existing
descriptors
kvm/i8254: refactor pit initialization into a helper kvm/i8254: add support for confidential guest reset kvm/hyperv: add synic feature to CPU only if its not enabled hw/hyperv/vmbus: add support for confidential guest reset kvm/xen-emu: re-initialize capabilities during confidential guest
reset
ppc/openpic: create a new openpic device and reattach mem region on
coco reset
kvm/vcpu: add notifiers to inform vcpu file descriptor change kvm/clock: add support for confidential guest reset hw/machine: introduce machine specific option 'x-change-vmfd-on-reset' tests/functional/x86_64: add functional test to exercise vm fd change
on reset

 MAINTAINERS                                   |   7 +
 accel/kvm/kvm-all.c                           | 360 +++++++++++++++---
 accel/kvm/trace-events                        |   2 +
 accel/stubs/kvm-stub.c                        |  18 +
 hw/core/machine.c                             |  22 ++
 hw/hyperv/trace-events                        |   1 +
 hw/hyperv/vmbus.c                             |  37 ++
 hw/i386/kvm/clock.c                           |  59 +++
 hw/i386/kvm/i8254.c                           |  91 +++--
 hw/i386/kvm/trace-events                      |   1 +
 hw/i386/x86-common.c                          |  71 +++-
 hw/intc/openpic_kvm.c                         | 112 ++++--
 hw/vfio/helpers.c                             |  94 +++++
 include/accel/accel-ops.h                     |   1 +
 include/hw/core/boards.h                      |   6 +
 include/hw/i386/x86.h                         |   1 +
 include/system/confidential-guest-support.h   |  20 +
 include/system/kvm.h                          |  43 +++
 include/system/physmem.h                      |   1 +
 stubs/kvm.c                                   |  22 ++
 stubs/meson.build                             |   1 +
 system/physmem.c                              |  28 ++
 system/runstate.c                             |  43 ++-
 target/i386/kvm/kvm.c                         | 139 +++++--
 target/i386/kvm/tdx.c                         | 141 +++++--
 target/i386/kvm/tdx.h                         |   1 +
 target/i386/kvm/trace-events                  |   4 +
 target/i386/kvm/xen-emu.c                     |  50 ++-
 target/i386/sev.c                             | 127 ++++--
 target/i386/trace-events                      |   1 +
 tests/functional/x86_64/meson.build           |   1 +
 .../x86_64/test_vmfd_change_reboot.py         | 135 +++++++
 32 files changed, 1422 insertions(+), 218 deletions(-)
 create mode 100644 stubs/kvm.c
 create mode 100755 tests/functional/x86_64/test_vmfd_change_reboot.py
Signature · 1 line
2.42.0
author
#1 lore raw

PATCH v4 1/31 i386/kvm: avoid installing duplicate msr entries in msr_handlers

kvm_filter_msr() does not check if an msr entry is already present in the msr_handlers table and installs a new handler unconditionally. If the function is called again with the same MSR, it will result in duplicate entries in the table and multiple such calls will fill up the table needlessly. Fix that.

---
 target/i386/kvm/kvm.c | 26 ++++++++++++++++----------
 1 file changed, 16 insertions(+), 10 deletions(-)
target/i386/kvm/kvm.c +16 -10
index 0c940d4b64..da1ed3b62a 100644
@@ -6042,27 +6042,33 @@ static int kvm_install_msr_filters(KVMState *s)
 static int kvm_filter_msr(KVMState *s, uint32_t msr, QEMURDMSRHandler *rdmsr,
                           QEMUWRMSRHandler *wrmsr)
 {
-    int i, ret;
+    int i, ret = 0;
 
     for (i = 0; i < ARRAY_SIZE(msr_handlers); i++) {
-        if (!msr_handlers[i].msr) {
+        if (msr_handlers[i].msr == msr) {
+            break;
+        } else if (!msr_handlers[i].msr) {
             msr_handlers[i] = (KVMMSRHandlers) {
                 .msr = msr,
                 .rdmsr = rdmsr,
                 .wrmsr = wrmsr,
             };
+            break;
+        }
+    }
 
-            ret = kvm_install_msr_filters(s);
-            if (ret) {
-                msr_handlers[i] = (KVMMSRHandlers) { };
-                return ret;
-            }
+    if (i == ARRAY_SIZE(msr_handlers)) {
+        ret = -EINVAL;
+        goto end;
+    }
 
-            return 0;
-        }
+    ret = kvm_install_msr_filters(s);
+    if (ret) {
+        msr_handlers[i] = (KVMMSRHandlers) { };
     }
 
-    return -EINVAL;
+ end:
+    return ret;
 }
 
 static int kvm_handle_rdmsr(X86CPU *cpu, struct kvm_run *run)
Signature · 1 line
2.42.0
author
#2 lore raw

PATCH v4 2/31 accel/kvm: add confidential class member to indicate guest rebuild capability

In reply to Ani Sinha

As a part of the confidential guest reset process, the existing encrypted guest state must be made mutable since it would be discarded after reset. A new encrypted and locked guest state must be established after the reset. To this end, a new boolean member per confidential guest support class (eg, tdx or sev-snp) is added that will indicate whether its possible to rebuild guest state:

bool can_rebuild_guest_state;

This is true if rebuilding guest state is possible, false otherwise. A KVM based confidential guest reset is only possible when the existing state is locked but its possible to rebuild guest state. Otherwise, the guest is not resettable.

---
 include/system/confidential-guest-support.h | 20 ++++++++++++++++++++
 system/runstate.c                           |  6 +++---
 target/i386/kvm/tdx.c                       |  1 +
 target/i386/sev.c                           |  1 +
 4 files changed, 25 insertions(+), 3 deletions(-)
include/system/confidential-guest-support.h +20 -0
index 0cc8b26e64..5dca717308 100644
@@ -152,6 +152,11 @@ typedef struct ConfidentialGuestSupportClass {
      */
     int (*get_mem_map_entry)(int index, ConfidentialGuestMemoryMapEntry *entry,
                              Error **errp);
+
+    /*
+     * is it possible to rebuild the guest state?
+     */
+    bool can_rebuild_guest_state;
 } ConfidentialGuestSupportClass;
 
 static inline int confidential_guest_kvm_init(ConfidentialGuestSupport *cgs,
@@ -167,6 +172,21 @@ static inline int confidential_guest_kvm_init(ConfidentialGuestSupport *cgs,
     return 0;
 }
 
+static inline bool
+confidential_guest_can_rebuild_state(ConfidentialGuestSupport *cgs)
+{
+    ConfidentialGuestSupportClass *klass;
+
+    if (!cgs) {
+        /* non-confidential guests */
+        return true;
+    }
+
+    klass = CONFIDENTIAL_GUEST_SUPPORT_GET_CLASS(cgs);
+    return klass->can_rebuild_guest_state;
+
+}
+
 static inline int confidential_guest_kvm_reset(ConfidentialGuestSupport *cgs,
                                                Error **errp)
 {
system/runstate.c +3 -3
index ed2db56480..5d58260ed5 100644
@@ -57,6 +57,7 @@
 #include "system/reset.h"
 #include "system/runstate.h"
 #include "system/runstate-action.h"
+#include "system/confidential-guest-support.h"
 #include "system/system.h"
 #include "system/tpm.h"
 #include "trace.h"
@@ -543,8 +544,6 @@ void qemu_system_reset(ShutdownCause reason)
      */
     if (cpus_are_resettable()) {
         cpu_synchronize_all_post_reset();
-    } else {
-        assert(runstate_check(RUN_STATE_PRELAUNCH));
     }
 
     vm_set_suspended(false);
@@ -693,7 +692,8 @@ void qemu_system_reset_request(ShutdownCause reason)
     if (reboot_action == REBOOT_ACTION_SHUTDOWN &&
         reason != SHUTDOWN_CAUSE_SUBSYSTEM_RESET) {
         shutdown_requested = reason;
-    } else if (!cpus_are_resettable()) {
+    } else if (!cpus_are_resettable() &&
+               !confidential_guest_can_rebuild_state(current_machine->cgs)) {
         error_report("cpus are not resettable, terminating");
         shutdown_requested = reason;
     } else {
target/i386/kvm/tdx.c +1 -0
index 0161985768..a3e81e1c0c 100644
@@ -1543,6 +1543,7 @@ static void tdx_guest_class_init(ObjectClass *oc, const void *data)
     X86ConfidentialGuestClass *x86_klass = X86_CONFIDENTIAL_GUEST_CLASS(oc);
 
     klass->kvm_init = tdx_kvm_init;
+    klass->can_rebuild_guest_state = true;
     x86_klass->kvm_type = tdx_kvm_type;
     x86_klass->cpu_instance_init = tdx_cpu_instance_init;
     x86_klass->adjust_cpuid_features = tdx_adjust_cpuid_features;
target/i386/sev.c +1 -0
index acdcb9c4e6..66e38ca32e 100644
@@ -2760,6 +2760,7 @@ sev_common_instance_init(Object *obj)
     cgs->set_guest_state = cgs_set_guest_state;
     cgs->get_mem_map_entry = cgs_get_mem_map_entry;
     cgs->set_guest_policy = cgs_set_guest_policy;
+    cgs->can_rebuild_guest_state = true;
 
     QTAILQ_INIT(&sev_common->launch_vmsa);
 }
Signature · 1 line
2.42.0
author
#3 lore raw

PATCH v4 3/31 hw/accel: add a per-accelerator callback to change VM accelerator handle

In reply to Ani Sinha

When a confidential virtual machine is reset, a new guest context in the accelerator must be generated post reset. Therefore, the old accelerator guest file handle must be closed and a new one created. To this end, a per-accelerator callback, "reset_vmfd" is introduced that would get called when a confidential guest is reset. Subsequent patches will introduce specific implementation of this callback for KVM accelerator.

---
 include/accel/accel-ops.h |  1 +
 system/runstate.c         | 37 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 37 insertions(+), 1 deletion(-)
include/accel/accel-ops.h +1 -0
index 23a8c246e1..998a95ca69 100644
@@ -23,6 +23,7 @@ struct AccelClass {
     AccelOpsClass *ops;
 
     int (*init_machine)(AccelState *as, MachineState *ms);
+    int (*reset_vmfd)(MachineState *ms);
     bool (*cpu_common_realize)(CPUState *cpu, Error **errp);
     void (*cpu_common_unrealize)(CPUState *cpu);
     /* get_stats: Append statistics to @buf */
system/runstate.c +36 -1
index 5d58260ed5..0a74e3ade5 100644
@@ -42,6 +42,7 @@
 #include "qapi/qapi-commands-run-state.h"
 #include "qapi/qapi-events-run-state.h"
 #include "qemu/accel.h"
+#include "accel/accel-ops.h"
 #include "qemu/error-report.h"
 #include "qemu/job.h"
 #include "qemu/log.h"
@@ -509,6 +510,9 @@ void qemu_system_reset(ShutdownCause reason)
 {
     MachineClass *mc;
     ResetType type;
+    AccelClass *ac = ACCEL_GET_CLASS(current_accel());
+    bool vmfd_reset = false;
+    int ret;
 
     mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL;
 
@@ -521,6 +525,29 @@ void qemu_system_reset(ShutdownCause reason)
     default:
         type = RESET_TYPE_COLD;
     }
+
+    /*
+     * different accelerators implement how to close the old file handle of
+     * the accelerator descriptor and create a new one here. Resetting
+     * file handle is necessary to create a new confidential VM context post
+     * VM reset.
+     */
+    if (!cpus_are_resettable() &&
+        (reason == SHUTDOWN_CAUSE_GUEST_RESET ||
+         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET)) {
+        if (ac->reset_vmfd) {
+            ret = ac->reset_vmfd(current_machine);
+            if (ret < 0) {
+                error_report("unable to reset vmfd: %s(%d)",
+                             strerror(-ret), ret);
+                vm_stop(RUN_STATE_INTERNAL_ERROR);
+            }
+            vmfd_reset = true;
+        } else {
+            error_report("accelerator does not support reset");
+        }
+    }
+
     if (mc && mc->reset) {
         mc->reset(current_machine, type);
     } else {
@@ -543,7 +570,15 @@ void qemu_system_reset(ShutdownCause reason)
      * it does _more_  than cpu_synchronize_all_post_reset().
      */
     if (cpus_are_resettable()) {
-        cpu_synchronize_all_post_reset();
+        if (vmfd_reset) {
+            /*
+             * If vmfd has changed, then vcpufds have also changed.
+             * Need to sync full cpu state for non confidential guests.
+             */
+            cpu_synchronize_all_post_init();
+        } else {
+            cpu_synchronize_all_post_reset();
+        }
     }
 
     vm_set_suspended(false);
Signature · 1 line
2.42.0
#4 lore raw

Re: [PATCH v4 03/31] hw/accel: add a per-accelerator callback to change VM accelerator handle

Hi Ani,

Quoted 18 lines · Ani Sinha

When a confidential virtual machine is reset, a new guest context in the accelerator must be generated post reset. Therefore, the old accelerator guest file handle must be closed and a new one created. To this end, a per-accelerator callback, "reset_vmfd" is introduced that would get called when a confidential guest is reset. Subsequent patches will introduce specific implementation of this callback for KVM accelerator.

---
  include/accel/accel-ops.h |  1 +
  system/runstate.c         | 37 ++++++++++++++++++++++++++++++++++++-
  2 files changed, 37 insertions(+), 1 deletion(-)
include/accel/accel-ops.h +0 -0
index 23a8c246e1..998a95ca69 100644

include/accel/accel-ops.h struct AccelClass { · line 26

Quoted 4 lines
include/accel/accel-ops.h +1 -0
@@ -23,6 +23,7 @@ struct AccelClass {
      AccelOpsClass *ops;
 
      int (*init_machine)(AccelState *as, MachineState *ms);
+    int (*reset_vmfd)(MachineState *ms);

So far 'vmfd' is a KVM concept. Can we use a more generic name?

Please add a @docstring description for this handler.

Quoted 37 lines
include/accel/accel-ops.h +0 -0
      bool (*cpu_common_realize)(CPUState *cpu, Error **errp);
      void (*cpu_common_unrealize)(CPUState *cpu);
      /* get_stats: Append statistics to @buf */
system/runstate.c +12 -0
index 5d58260ed5..0a74e3ade5 100644
@@ -42,6 +42,7 @@
  #include "qapi/qapi-commands-run-state.h"
  #include "qapi/qapi-events-run-state.h"
  #include "qemu/accel.h"
+#include "accel/accel-ops.h"
  #include "qemu/error-report.h"
  #include "qemu/job.h"
  #include "qemu/log.h"
@@ -509,6 +510,9 @@ void qemu_system_reset(ShutdownCause reason)
  {
      MachineClass *mc;
      ResetType type;
+    AccelClass *ac = ACCEL_GET_CLASS(current_accel());
+    bool vmfd_reset = false;
+    int ret;
 
      mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL;
 
@@ -521,6 +525,29 @@ void qemu_system_reset(ShutdownCause reason)
      default:
          type = RESET_TYPE_COLD;
      }
+
+    /*
+     * different accelerators implement how to close the old file handle of
+     * the accelerator descriptor and create a new one here. Resetting
+     * file handle is necessary to create a new confidential VM context post
+     * VM reset.
+     */
+    if (!cpus_are_resettable() &&

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 545

Quoted 10 lines
system/runstate.c +10 -0
+        (reason == SHUTDOWN_CAUSE_GUEST_RESET ||
+         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET)) {
+        if (ac->reset_vmfd) {
+            ret = ac->reset_vmfd(current_machine);
+            if (ret < 0) {
+                error_report("unable to reset vmfd: %s(%d)",
+                             strerror(-ret), ret);
+                vm_stop(RUN_STATE_INTERNAL_ERROR);
+            }
+            vmfd_reset = true;

If we need a such flag, please rename it generically.

Quoted 9 lines
system/runstate.c +5 -0
+        } else {
+            error_report("accelerator does not support reset");
+        }
+    }
+
      if (mc && mc->reset) {
          mc->reset(current_machine, type);
      } else {

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 578

Quoted 10 lines
system/runstate.c +6 -1
@@ -543,7 +570,15 @@ void qemu_system_reset(ShutdownCause reason)
       * it does _more_  than cpu_synchronize_all_post_reset().
       */
      if (cpus_are_resettable()) {
-        cpu_synchronize_all_post_reset();
+        if (vmfd_reset) {
+            /*
+             * If vmfd has changed, then vcpufds have also changed.
+             * Need to sync full cpu state for non confidential guests.
+             */
+            cpu_synchronize_all_post_init();

Calling post_init() in reset() is dubious. It might work with KVM by chance (because only KVM implements .reset_vmfd), but by design I don't expect it to work on other accelerators. If you really think this is the only solution, then you'll need to document cpu_sync* methods very carefully.

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 584

Quoted 6 lines
system/runstate.c +3 -0
+        } else {
+            cpu_synchronize_all_post_reset();
+        }
      }
 
      vm_set_suspended(false);

Isn't the confidential_guest_kvm_reset() API more appropriate here?

author
#5 lore raw

Re: [PATCH v4 03/31] hw/accel: add a per-accelerator callback to change VM accelerator handle

Quoted 20 lines · Philippe Mathieu-Daudé

Hi Ani,

Quoted 13 lines · Ani Sinha

When a confidential virtual machine is reset, a new guest context in the accelerator must be generated post reset. Therefore, the old accelerator guest file handle must be closed and a new one created. To this end, a per-accelerator callback, "reset_vmfd" is introduced that would get called when a confidential guest is reset. Subsequent patches will introduce specific implementation of this callback for KVM accelerator.

---
 include/accel/accel-ops.h |  1 +
 system/runstate.c         | 37 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 37 insertions(+), 1 deletion(-)
include/accel/accel-ops.h +0 -0
index 23a8c246e1..998a95ca69 100644
Quoted 8 lines

include/accel/accel-ops.h struct AccelClass { · line 25

Quoted 6 lines
include/accel/accel-ops.h +1 -0
@@ -23,6 +23,7 @@ struct AccelClass {
     AccelOpsClass *ops;
       int (*init_machine)(AccelState *as, MachineState *ms);
+    int (*reset_vmfd)(MachineState *ms);

So far 'vmfd' is a KVM concept. Can we use a more generic name?

This is more like a hypervisor interface to manage the VM. For example, for Xen it would be what xc_interface_open() would return. So I can rename it to something like “reset_hypervisor_interface”.

Quoted 84 lines

Please add a @docstring description for this handler.

Quoted 35 lines
include/accel/accel-ops.h +0 -0
     bool (*cpu_common_realize)(CPUState *cpu, Error **errp);
     void (*cpu_common_unrealize)(CPUState *cpu);
     /* get_stats: Append statistics to @buf */
system/runstate.c +4 -0
index 5d58260ed5..0a74e3ade5 100644
@@ -42,6 +42,7 @@
 #include "qapi/qapi-commands-run-state.h"
 #include "qapi/qapi-events-run-state.h"
 #include "qemu/accel.h"
+#include "accel/accel-ops.h"
 #include "qemu/error-report.h"
 #include "qemu/job.h"
 #include "qemu/log.h"
@@ -509,6 +510,9 @@ void qemu_system_reset(ShutdownCause reason)
 {
     MachineClass *mc;
     ResetType type;
+    AccelClass *ac = ACCEL_GET_CLASS(current_accel());
+    bool vmfd_reset = false;
+    int ret;
       mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL;
 @@ -521,6 +525,29 @@ void qemu_system_reset(ShutdownCause reason)
     default:
         type = RESET_TYPE_COLD;
     }
system/runstate.c +8 -0
+
+    /*
+     * different accelerators implement how to close the old file handle of
+     * the accelerator descriptor and create a new one here. Resetting
+     * file handle is necessary to create a new confidential VM context post
+     * VM reset.
+     */
+    if (!cpus_are_resettable() &&

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 545

Quoted 10 lines
system/runstate.c +10 -0
+        (reason == SHUTDOWN_CAUSE_GUEST_RESET ||
+         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET)) {
+        if (ac->reset_vmfd) {
+            ret = ac->reset_vmfd(current_machine);
+            if (ret < 0) {
+                error_report("unable to reset vmfd: %s(%d)",
+                             strerror(-ret), ret);
+                vm_stop(RUN_STATE_INTERNAL_ERROR);
+            }
+            vmfd_reset = true;

If we need a such flag, please rename it generically.

Quoted 9 lines
system/runstate.c +5 -0
+        } else {
+            error_report("accelerator does not support reset");
+        }
+    }
+
     if (mc && mc->reset) {
         mc->reset(current_machine, type);
     } else {

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 578

Quoted 10 lines
system/runstate.c +6 -1
@@ -543,7 +570,15 @@ void qemu_system_reset(ShutdownCause reason)
      * it does _more_  than cpu_synchronize_all_post_reset().
      */
     if (cpus_are_resettable()) {
-        cpu_synchronize_all_post_reset();
+        if (vmfd_reset) {
+            /*
+             * If vmfd has changed, then vcpufds have also changed.
+             * Need to sync full cpu state for non confidential guests.
+             */
+            cpu_synchronize_all_post_init();

Calling post_init() in reset() is dubious. It might work with KVM by chance (because only KVM implements .reset_vmfd), but by design I don't expect it to work on other accelerators. If you really think this is the only solution, then you'll need to document cpu_sync* methods very carefully.

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 584

Quoted 5 lines
system/runstate.c +3 -0
+        } else {
+            cpu_synchronize_all_post_reset();
+        }
     }
       vm_set_suspended(false);

Isn't the confidential_guest_kvm_reset() API more appropriate here?

maintainer
#6 lore raw

Re: [PATCH v4 03/31] hw/accel: add a per-accelerator callback to change VM accelerator handle

Quoted 20 lines · Ani Sinha
Quoted 20 lines · Philippe Mathieu-Daudé

Hi Ani,

Quoted 13 lines · Ani Sinha

When a confidential virtual machine is reset, a new guest context in the accelerator must be generated post reset. Therefore, the old accelerator guest file handle must be closed and a new one created. To this end, a per-accelerator callback, "reset_vmfd" is introduced that would get called when a confidential guest is reset. Subsequent patches will introduce specific implementation of this callback for KVM accelerator.

---
  include/accel/accel-ops.h |  1 +
  system/runstate.c         | 37 ++++++++++++++++++++++++++++++++++++-
  2 files changed, 37 insertions(+), 1 deletion(-)
include/accel/accel-ops.h +0 -0
index 23a8c246e1..998a95ca69 100644
Quoted 8 lines
Quoted 8 lines

include/accel/accel-ops.h struct AccelClass { · line 25

Quoted 6 lines
include/accel/accel-ops.h +1 -0
@@ -23,6 +23,7 @@ struct AccelClass {
      AccelOpsClass *ops;
        int (*init_machine)(AccelState *as, MachineState *ms);
+    int (*reset_vmfd)(MachineState *ms);

So far 'vmfd' is a KVM concept. Can we use a more generic name?

For what it's worth, mshv and nvmm also have it (whpx has a handle because it's Windows, hvf probably has a file descriptor internally but only exposes a single VM per process).

Quoted 2 lines

This is more like a hypervisor interface to manage the VM. For example, for Xen it would be what xc_interface_open() would return. So I can rename it to something like “reset_hypervisor_interface”.

Or rebuild_vm perhaps? It can also be done on applying, since it's an easy mass change on the .patch files themselves.

Paolo

author
#7 lore raw

Re: [PATCH v4 03/31] hw/accel: add a per-accelerator callback to change VM accelerator handle

In reply to Philippe Mathieu-Daudé
Quoted 107 lines · Philippe Mathieu-Daudé

Hi Ani,

Quoted 18 lines · Ani Sinha

When a confidential virtual machine is reset, a new guest context in the accelerator must be generated post reset. Therefore, the old accelerator guest file handle must be closed and a new one created. To this end, a per-accelerator callback, "reset_vmfd" is introduced that would get called when a confidential guest is reset. Subsequent patches will introduce specific implementation of this callback for KVM accelerator.

---
  include/accel/accel-ops.h |  1 +
  system/runstate.c         | 37 ++++++++++++++++++++++++++++++++++++-
  2 files changed, 37 insertions(+), 1 deletion(-)
include/accel/accel-ops.h +0 -0
index 23a8c246e1..998a95ca69 100644

include/accel/accel-ops.h struct AccelClass { · line 26

Quoted 4 lines
include/accel/accel-ops.h +1 -0
@@ -23,6 +23,7 @@ struct AccelClass {
      AccelOpsClass *ops;
 
      int (*init_machine)(AccelState *as, MachineState *ms);
+    int (*reset_vmfd)(MachineState *ms);

So far 'vmfd' is a KVM concept. Can we use a more generic name?

Please add a @docstring description for this handler.

Quoted 37 lines
include/accel/accel-ops.h +0 -0
      bool (*cpu_common_realize)(CPUState *cpu, Error **errp);
      void (*cpu_common_unrealize)(CPUState *cpu);
      /* get_stats: Append statistics to @buf */
system/runstate.c +12 -0
index 5d58260ed5..0a74e3ade5 100644
@@ -42,6 +42,7 @@
  #include "qapi/qapi-commands-run-state.h"
  #include "qapi/qapi-events-run-state.h"
  #include "qemu/accel.h"
+#include "accel/accel-ops.h"
  #include "qemu/error-report.h"
  #include "qemu/job.h"
  #include "qemu/log.h"
@@ -509,6 +510,9 @@ void qemu_system_reset(ShutdownCause reason)
  {
      MachineClass *mc;
      ResetType type;
+    AccelClass *ac = ACCEL_GET_CLASS(current_accel());
+    bool vmfd_reset = false;
+    int ret;
 
      mc = current_machine ? MACHINE_GET_CLASS(current_machine) : NULL;
 
@@ -521,6 +525,29 @@ void qemu_system_reset(ShutdownCause reason)
      default:
          type = RESET_TYPE_COLD;
      }
+
+    /*
+     * different accelerators implement how to close the old file handle of
+     * the accelerator descriptor and create a new one here. Resetting
+     * file handle is necessary to create a new confidential VM context post
+     * VM reset.
+     */
+    if (!cpus_are_resettable() &&

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 545

Quoted 10 lines
system/runstate.c +10 -0
+        (reason == SHUTDOWN_CAUSE_GUEST_RESET ||
+         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET)) {
+        if (ac->reset_vmfd) {
+            ret = ac->reset_vmfd(current_machine);
+            if (ret < 0) {
+                error_report("unable to reset vmfd: %s(%d)",
+                             strerror(-ret), ret);
+                vm_stop(RUN_STATE_INTERNAL_ERROR);
+            }
+            vmfd_reset = true;

If we need a such flag, please rename it generically.

Quoted 9 lines
system/runstate.c +5 -0
+        } else {
+            error_report("accelerator does not support reset");
+        }
+    }
+
      if (mc && mc->reset) {
          mc->reset(current_machine, type);
      } else {

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 578

Quoted 10 lines
system/runstate.c +6 -1
@@ -543,7 +570,15 @@ void qemu_system_reset(ShutdownCause reason)
       * it does _more_  than cpu_synchronize_all_post_reset().
       */
      if (cpus_are_resettable()) {
-        cpu_synchronize_all_post_reset();
+        if (vmfd_reset) {
+            /*
+             * If vmfd has changed, then vcpufds have also changed.
+             * Need to sync full cpu state for non confidential guests.
+             */
+            cpu_synchronize_all_post_init();

Calling post_init() in reset() is dubious. It might work with KVM by chance (because only KVM implements .reset_vmfd), but by design I don't expect it to work on other accelerators. If you really think this is the only solution, then you'll need to document cpu_sync* methods very carefully.

Quoted 8 lines

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 584

Quoted 6 lines
system/runstate.c +3 -0
+        } else {
+            cpu_synchronize_all_post_reset();
+        }
      }
 
      vm_set_suspended(false);

Isn't the confidential_guest_kvm_reset() API more appropriate here?

This above hunk I did, calling post_init() to do a full put_registers() etc is required only for non-coco case in KVM after CPU reset when we are enabling vm file descriptor change. In the coco case, the call is a noop as the guest state is protected. This part I am not very familiar with and used my best judgement. To address your concern that other accelerators might implement reset_vmfd() but does not like to call cpu_synchronize_all_post_init(), we can wrap it inside another accelerator callback like ac->cpu_synchronize_all_post_vmfd_change() or some such.

Paolo has some context as I was debugging this area with his help. I am curious also for his inputs on this.

maintainer
#8 lore raw

Re: [PATCH v4 03/31] hw/accel: add a per-accelerator callback to change VM accelerator handle

Quoted 7 lines · Ani Sinha
Quoted 7 lines
Quoted 6 lines
system/runstate.c +6 -0
+        if (vmfd_reset) {
+            /*
+             * If vmfd has changed, then vcpufds have also changed.
+             * Need to sync full cpu state for non confidential guests.
+             */
+            cpu_synchronize_all_post_init();
Quoted 5 lines
Quoted 5 lines

Calling post_init() in reset() is dubious. It might work with KVM by chance (because only KVM implements .reset_vmfd), but by design I don't expect it to work on other accelerators. If you really think this is the only solution, then you'll need to document cpu_sync* methods very carefully.

post_init is called to do a first-time initialization of the VM file descriptor. This is exactly why you need to call it here, the VM file descriptor is new and has never been fully synchronized before. Otherwise, kvm_arch_set_tsc_khz() is never called (doing it in the reset_vmfd callback would be a hack).

Other accelerators don't implement separate post-init and post-reset callbacks at all:

mshv => same except for abort vs. RUN_STATE_INTERNAL_ERROR whpx => WHPX_SET_RESET_STATE vs. WHPX_SET_FULL_STATE, actually no change nvmm => same implementation
hvf => same implementation
tcg => does not implement either

so they're unaffected.

Quoted 9 lines
Quoted 8 lines

system/runstate.c void qemu_system_reset(ShutdownCause reason) · line 584

Quoted 6 lines
system/runstate.c +3 -0
+        } else {
+            cpu_synchronize_all_post_reset();
+        }
       }
 
       vm_set_suspended(false);

Isn't the confidential_guest_kvm_reset() API more appropriate here?

Quoted 10 lines

This above hunk I did, calling post_init() to do a full put_registers() etc is required only for non-coco case in KVM after CPU reset when we are enabling vm file descriptor change. In the coco case, the call is a noop as the guest state is protected. This part I am not very familiar with and used my best judgement. To address your concern that other accelerators might implement reset_vmfd() but does not like to call cpu_synchronize_all_post_init(), we can wrap it inside another accelerator callback like ac->cpu_synchronize_all_post_vmfd_change() or some such.

No, calling the post_init callback is the right thing to do.

I must say I am not 100% sure of the initialization flow in general; I think kvm_arch_set_tsc_khz() must be called for confidential guests too and, in order to do so, at initialization time QEMU could call a post_init call that only does kvm_arch_set_tsc_khz(), and then call post_reset normally. But this is independent of this series, which doesn't make the fix any easier or harder.

Paolo

author
#9 lore raw

PATCH v4 4/31 system/physmem: add helper to reattach existing memory after KVM VM fd change

In reply to Ani Sinha

After the guest KVM file descriptor has changed as a part of the process of confidential guest reset mechanism, existing memory needs to be reattached to the new file descriptor. This change adds a helper function ram_block_rebind() for this purpose. The next patch will make use of this function.

---
 include/system/physmem.h |  1 +
 system/physmem.c         | 28 ++++++++++++++++++++++++++++
 2 files changed, 29 insertions(+)
include/system/physmem.h +1 -0
index 7bb7d3e154..da91b77bd9 100644
@@ -51,5 +51,6 @@ physical_memory_snapshot_and_clear_dirty(MemoryRegion *mr, hwaddr offset,
 bool physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
                                         ram_addr_t start,
                                         ram_addr_t length);
+int ram_block_rebind(Error **errp);
 
 #endif
system/physmem.c +28 -0
index 2fb0c25c93..e5ff26acec 100644
@@ -2826,6 +2826,34 @@ found:
     return block;
 }
 
+/*
+ * Creates new guest memfd for the ramblocks and closes the
+ * existing memfd.
+ */
+int ram_block_rebind(Error **errp)
+{
+    RAMBlock *block;
+
+    qemu_mutex_lock_ramlist();
+
+    RAMBLOCK_FOREACH(block) {
+        if (block->flags & RAM_GUEST_MEMFD) {
+            if (block->guest_memfd >= 0) {
+                close(block->guest_memfd);
+            }
+            block->guest_memfd = kvm_create_guest_memfd(block->max_length,
+                                                        0, errp);
+            if (block->guest_memfd < 0) {
+                qemu_mutex_unlock_ramlist();
+                return -1;
+            }
+
+        }
+    }
+    qemu_mutex_unlock_ramlist();
+    return 0;
+}
+
 /*
  * Finds the named RAMBlock
  *
Signature · 1 line
2.42.0
author
#10 lore raw

PATCH v4 5/31 accel/kvm: add changes required to support KVM VM file descriptor change

In reply to Ani Sinha

This change adds common kvm specific support to handle KVM VM file descriptor change. KVM VM file descriptor can change as a part of confidential guest reset mechanism. A new function api kvm_arch_on_vmfd_change() per architecture platform is added in order to implement architecture specific changes required to support it. A subsequent patch will add x86 specific implementation for kvm_arch_on_vmfd_change() as currently only x86 supports confidential guest reset.

---
 MAINTAINERS            |  6 ++++
 accel/kvm/kvm-all.c    | 80 ++++++++++++++++++++++++++++++++++++++++--
 accel/kvm/trace-events |  1 +
 include/system/kvm.h   |  3 ++
 stubs/kvm.c            | 22 ++++++++++++
 stubs/meson.build      |  1 +
 target/i386/kvm/kvm.c  | 10 ++++++
 7 files changed, 120 insertions(+), 3 deletions(-)
 create mode 100644 stubs/kvm.c
MAINTAINERS +6 -0
index 65ac60b86b..a88901a7d7 100644
@@ -152,6 +152,12 @@ F: tools/i386/
 F: tests/functional/i386/
 F: tests/functional/x86_64/
 
+X86 VM file descriptor change on reset test
+M: Ani Sinha <[email protected]>
+M: Paolo Bonzini <[email protected]>
+S: Maintained
+F: stubs/kvm.c
+
 Guest CPU cores (TCG)
 ---------------------
 Overall TCG CPUs
accel/kvm/kvm-all.c +77 -3

107 lines not on this page · raw message

accel/kvm/trace-events +1 -0
index e43d18a869..e4beda0148 100644
@@ -14,6 +14,7 @@ kvm_destroy_vcpu(int cpu_index, unsigned long arch_cpu_id) "index: %d id: %lu"
 kvm_park_vcpu(int cpu_index, unsigned long arch_cpu_id) "index: %d id: %lu"
 kvm_unpark_vcpu(unsigned long arch_cpu_id, const char *msg) "id: %lu %s"
 kvm_irqchip_commit_routes(void) ""
+kvm_reset_vmfd(void) ""
 kvm_irqchip_add_msi_route(char *name, int vector, int virq) "dev %s vector %d virq %d"
 kvm_irqchip_update_msi_route(int virq) "Updating MSI route virq=%d"
 kvm_irqchip_release_virq(int virq) "virq %d"
include/system/kvm.h +3 -0
index 8f9eecf044..5fc7251fd9 100644
@@ -456,6 +456,9 @@ int kvm_physical_memory_addr_from_host(KVMState *s, void *ram_addr,
 
 #endif /* COMPILING_PER_TARGET */
 
+bool kvm_arch_supports_vmfd_change(void);
+int kvm_arch_on_vmfd_change(MachineState *ms, KVMState *s);
+
 void kvm_cpu_synchronize_state(CPUState *cpu);
 
 void kvm_init_cpu_signals(CPUState *cpu);
stubs/kvm.c new file +22 -0
new file mode 100644
index 0000000000..2db61d89a7
@@ -0,0 +1,22 @@
+/*
+ * kvm target arch specific stubs
+ *
+ * Copyright (c) 2026 Red Hat, Inc.
+ *
+ * Author:
+ *   Ani Sinha <[email protected]>
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ */
+#include "qemu/osdep.h"
+#include "system/kvm.h"
+
+int kvm_arch_on_vmfd_change(MachineState *ms, KVMState *s)
+{
+    abort();
+}
+
+bool kvm_arch_supports_vmfd_change(void)
+{
+    return false;
+}
stubs/meson.build +1 -0
index 8a07059500..6ae478bacc 100644
@@ -74,6 +74,7 @@ if have_system
   if igvm.found()
     stub_ss.add(files('igvm.c'))
   endif
+  stub_ss.add(files('kvm.c'))
   stub_ss.add(files('target-get-monitor-def.c'))
   stub_ss.add(files('target-monitor-defs.c'))
   stub_ss.add(files('win32-kbd-hook.c'))
target/i386/kvm/kvm.c +10 -0
index da1ed3b62a..83c15f098e 100644
@@ -3252,6 +3252,16 @@ static int kvm_vm_enable_energy_msrs(KVMState *s)
     return 0;
 }
 
+int kvm_arch_on_vmfd_change(MachineState *ms, KVMState *s)
+{
+    abort();
+}
+
+bool kvm_arch_supports_vmfd_change(void)
+{
+    return false;
+}
+
 int kvm_arch_init(MachineState *ms, KVMState *s)
 {
     int ret;
Signature · 1 line
2.42.0
author
#11 lore raw

PATCH v4 6/31 accel/kvm: add a notifier to indicate KVM VM file descriptor has changed

In reply to Ani Sinha

A notifier callback can be used by various subsystems to perform actions when KVM file descriptor for a virtual machine changes as a part of confidential guest reset process. This change adds this notifier mechanism. Subsequent patches will add specific implementations for various notifier callbacks corresponding to various subsystems that need to take action when KVM VM file descriptor changed.

---
 accel/kvm/kvm-all.c    | 30 ++++++++++++++++++++++++++++++
 accel/stubs/kvm-stub.c |  8 ++++++++
 include/system/kvm.h   | 21 +++++++++++++++++++++
 3 files changed, 59 insertions(+)
accel/kvm/kvm-all.c +30 -0
index 46243cfcf2..30d6295f53 100644
@@ -90,6 +90,7 @@ struct KVMParkedVcpu {
 };
 
 KVMState *kvm_state;
+VmfdChangeNotifier vmfd_notifier;
 bool kvm_kernel_irqchip;
 bool kvm_split_irqchip;
 bool kvm_async_interrupts_allowed;
@@ -123,6 +124,9 @@ static const KVMCapabilityInfo kvm_required_capabilites[] = {
 static NotifierList kvm_irqchip_change_notifiers =
     NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers);
 
+static NotifierWithReturnList register_vmfd_changed_notifiers =
+    NOTIFIER_WITH_RETURN_LIST_INITIALIZER(register_vmfd_changed_notifiers);
+
 struct KVMResampleFd {
     int gsi;
     EventNotifier *resample_event;
@@ -2173,6 +2177,22 @@ void kvm_irqchip_change_notify(void)
     notifier_list_notify(&kvm_irqchip_change_notifiers, NULL);
 }
 
+void kvm_vmfd_add_change_notifier(NotifierWithReturn *n)
+{
+    notifier_with_return_list_add(&register_vmfd_changed_notifiers, n);
+}
+
+void kvm_vmfd_remove_change_notifier(NotifierWithReturn *n)
+{
+    notifier_with_return_remove(n);
+}
+
+static int kvm_vmfd_change_notify(Error **errp)
+{
+    return notifier_with_return_list_notify(&register_vmfd_changed_notifiers,
+                                            &vmfd_notifier, errp);
+}
+
 int kvm_irqchip_get_virq(KVMState *s)
 {
     int next_virq;
@@ -2668,6 +2688,16 @@ static int kvm_reset_vmfd(MachineState *ms)
         do_kvm_irqchip_create(s);
     }
 
+    /*
+     * notify everyone that vmfd has changed.
+     */
+    vmfd_notifier.vmfd = s->vmfd;
+    ret = kvm_vmfd_change_notify(&err);
+    if (ret < 0) {
+        return ret;
+    }
+    assert(!err);
+
     /* these can be only called after ram_block_rebind() */
     memory_listener_register(&kml->listener, &address_space_memory);
     memory_listener_register(&kvm_io_listener, &address_space_io);
accel/stubs/kvm-stub.c +8 -0
index 68cd33ba97..a6e8a6e16c 100644
@@ -79,6 +79,14 @@ void kvm_irqchip_change_notify(void)
 {
 }
 
+void kvm_vmfd_add_change_notifier(NotifierWithReturn *n)
+{
+}
+
+void kvm_vmfd_remove_change_notifier(NotifierWithReturn *n)
+{
+}
+
 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
                                        EventNotifier *rn, int virq)
 {
include/system/kvm.h +21 -0
index 5fc7251fd9..f11729f432 100644
@@ -181,6 +181,7 @@ DECLARE_INSTANCE_CHECKER(KVMState, KVM_STATE,
 
 extern KVMState *kvm_state;
 typedef struct Notifier Notifier;
+typedef struct NotifierWithReturn NotifierWithReturn;
 
 typedef struct KVMRouteChange {
      KVMState *s;
@@ -567,4 +568,24 @@ int kvm_set_memory_attributes_shared(hwaddr start, uint64_t size);
 
 int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private);
 
+/* argument to vmfd change notifier */
+typedef struct VmfdChangeNotifier {
+    int vmfd;
+} VmfdChangeNotifier;
+
+/**
+ * kvm_vmfd_add_change_notifier - register a notifier to get notified when
+ * a KVM vm file descriptor changes as a part of the confidential guest "reset"
+ * process. Various subsystems should use this mechanism to take actions such
+ * as creating new fds against this new vm file descriptor.
+ * @n: notifier with return value.
+ */
+void kvm_vmfd_add_change_notifier(NotifierWithReturn *n);
+/**
+ * kvm_vmfd_remove_change_notifier - de-register a notifer previously
+ * registered with kvm_vmfd_add_change_notifier call.
+ * @n: notifier that was previously registered.
+ */
+void kvm_vmfd_remove_change_notifier(NotifierWithReturn *n);
+
 #endif
Signature · 1 line
2.42.0
author
#12 lore raw

PATCH v4 7/31 accel/kvm: notify when KVM VM file fd is about to be changed

In reply to Ani Sinha

Various subsystems might need to take some steps before the KVM file descriptor for a virtual machine is changed. So a new boolean attribute is added to the vmfd_notifier structure which is passed to the notifier callbacks. vmfd_notifer.pre is true for pre-notification of vmfd change and false for post notification. Notifier callback implementations can simply check the boolean value for (vmfd_notifer*)->pre and can take actions for pre or post vmfd change based on the value.

Subsequent patches will add callback implementations for specific components that need this pre-notification.

---
 accel/kvm/kvm-all.c  | 9 +++++++++
 include/system/kvm.h | 6 ++++--
 2 files changed, 13 insertions(+), 2 deletions(-)
accel/kvm/kvm-all.c +9 -0
index 30d6295f53..a6ff16a037 100644
@@ -2654,6 +2654,13 @@ static int kvm_reset_vmfd(MachineState *ms)
     memory_listener_unregister(&kml->listener);
     memory_listener_unregister(&kvm_io_listener);
 
+    vmfd_notifier.pre = true;
+    ret = kvm_vmfd_change_notify(&err);
+    if (ret < 0) {
+        return ret;
+    }
+    assert(!err);
+
     if (s->vmfd >= 0) {
         close(s->vmfd);
     }
@@ -2692,6 +2699,8 @@ static int kvm_reset_vmfd(MachineState *ms)
      * notify everyone that vmfd has changed.
      */
     vmfd_notifier.vmfd = s->vmfd;
+    vmfd_notifier.pre = false;
+
     ret = kvm_vmfd_change_notify(&err);
     if (ret < 0) {
         return ret;
include/system/kvm.h +4 -2
index f11729f432..fbe23608a1 100644
@@ -571,12 +571,14 @@ int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private);
 /* argument to vmfd change notifier */
 typedef struct VmfdChangeNotifier {
     int vmfd;
+    bool pre;
 } VmfdChangeNotifier;
 
 /**
  * kvm_vmfd_add_change_notifier - register a notifier to get notified when
- * a KVM vm file descriptor changes as a part of the confidential guest "reset"
- * process. Various subsystems should use this mechanism to take actions such
+ * a KVM vm file descriptor changes or about to be changed as a part of the
+ * confidential guest "reset" process.
+ * Various subsystems should use this mechanism to take actions such
  * as creating new fds against this new vm file descriptor.
  * @n: notifier with return value.
  */
Signature · 1 line
2.42.0
author
#13 lore raw

PATCH v4 8/31 i386/kvm: unregister smram listeners prior to vm file descriptor change

In reply to Ani Sinha

We will re-register smram listeners after the VM file descriptors has changed. We need to unregister them first to make sure addresses and reference counters work properly.

---
 target/i386/kvm/kvm.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
target/i386/kvm/kvm.c +18 -0
index 83c15f098e..5d9b79529f 100644
@@ -112,6 +112,11 @@ typedef struct {
 static void kvm_init_msrs(X86CPU *cpu);
 static int kvm_filter_msr(KVMState *s, uint32_t msr, QEMURDMSRHandler *rdmsr,
                           QEMUWRMSRHandler *wrmsr);
+static int unregister_smram_listener(NotifierWithReturn *notifier,
+                                     void *data, Error** errp);
+NotifierWithReturn kvm_vmfd_change_notifier = {
+    .notify = unregister_smram_listener,
+};
 
 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
     KVM_CAP_INFO(SET_TSS_ADDR),
@@ -2748,6 +2753,17 @@ static void register_smram_listener(Notifier *n, void *unused)
     }
 }
 
+static int unregister_smram_listener(NotifierWithReturn *notifier,
+                                     void *data, Error** errp)
+{
+    if (!((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    memory_listener_unregister(&smram_listener.listener);
+    return 0;
+}
+
 /* It should only be called in cpu's hotplug callback */
 void kvm_smm_cpu_address_space_init(X86CPU *cpu)
 {
@@ -3400,6 +3416,8 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
         }
     }
 
+    kvm_vmfd_add_change_notifier(&kvm_vmfd_change_notifier);
+
     return 0;
 }
 
Signature · 1 line
2.42.0
author
#14 lore raw

PATCH v4 9/31 kvm/i386: implement architecture support for kvm file descriptor change

In reply to Ani Sinha

When the kvm file descriptor changes as a part of confidential guest reset, some architecture specific setups including SEV/SEV-SNP/TDX specific setups needs to be redone. These changes are implemented as a part of the kvm_arch_on_vmfd_change() callback which was introduced previously.

---
 target/i386/kvm/kvm.c        | 82 ++++++++++++++++++++++++++----------
 target/i386/kvm/trace-events |  1 +
 2 files changed, 60 insertions(+), 23 deletions(-)
target/i386/kvm/kvm.c +59 -23
index 5d9b79529f..e82a6e9eda 100644
@@ -3268,14 +3268,52 @@ static int kvm_vm_enable_energy_msrs(KVMState *s)
     return 0;
 }
 
+static int xen_init_wrapper(MachineState *ms, KVMState *s);
+
 int kvm_arch_on_vmfd_change(MachineState *ms, KVMState *s)
 {
-    abort();
+    int ret;
+
+    ret = kvm_arch_init(ms, s);
+    if (ret < 0) {
+        return ret;
+    }
+
+    if (object_dynamic_cast(OBJECT(ms), TYPE_X86_MACHINE)) {
+        X86MachineState *x86ms = X86_MACHINE(ms);
+
+        if (x86_machine_is_smm_enabled(x86ms)) {
+            memory_listener_register(&smram_listener.listener,
+                                     &smram_address_space);
+        }
+        kvm_set_max_apic_id(x86ms->apic_id_limit);
+    }
+
+    trace_kvm_arch_on_vmfd_change();
+    return 0;
+}
+
+static int xen_init_wrapper(MachineState *ms, KVMState *s)
+{
+    int ret = 0;
+#ifdef CONFIG_XEN_EMU
+    if (!object_dynamic_cast(OBJECT(ms), TYPE_PC_MACHINE)) {
+        error_report("kvm: Xen support only available in PC machine");
+        return -ENOTSUP;
+    }
+    /* hyperv_enabled() doesn't work yet. */
+    uint32_t msr = XEN_HYPERCALL_MSR;
+    ret = kvm_xen_init(s, msr);
+#else
+    error_report("kvm: Xen support not enabled in qemu");
+    return -ENOTSUP;
+#endif
+    return ret;
 }
 
 bool kvm_arch_supports_vmfd_change(void)
 {
-    return false;
+    return true;
 }
 
 int kvm_arch_init(MachineState *ms, KVMState *s)
@@ -3283,6 +3321,7 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
     int ret;
     struct utsname utsname;
     Error *local_err = NULL;
+    static bool first = true;
 
     /*
      * Initialize confidential guest (SEV/TDX) context, if required
@@ -3311,21 +3350,10 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
     }
 
     if (s->xen_version) {
-#ifdef CONFIG_XEN_EMU
-        if (!object_dynamic_cast(OBJECT(ms), TYPE_PC_MACHINE)) {
-            error_report("kvm: Xen support only available in PC machine");
-            return -ENOTSUP;
-        }
-        /* hyperv_enabled() doesn't work yet. */
-        uint32_t msr = XEN_HYPERCALL_MSR;
-        ret = kvm_xen_init(s, msr);
+        ret = xen_init_wrapper(ms, s);
         if (ret < 0) {
             return ret;
         }
-#else
-        error_report("kvm: Xen support not enabled in qemu");
-        return -ENOTSUP;
-#endif
     }
 
     ret = kvm_get_supported_msrs(s);
@@ -3352,16 +3380,17 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
         return ret;
     }
 
-    /* Tell fw_cfg to notify the BIOS to reserve the range. */
-    e820_add_entry(KVM_IDENTITY_BASE, 0x4000, E820_RESERVED);
-
+    if (first) {
+        /* Tell fw_cfg to notify the BIOS to reserve the range. */
+        e820_add_entry(KVM_IDENTITY_BASE, 0x4000, E820_RESERVED);
+    }
     ret = kvm_vm_set_nr_mmu_pages(s);
     if (ret < 0) {
         return ret;
     }
 
     if (object_dynamic_cast(OBJECT(ms), TYPE_X86_MACHINE) &&
-        x86_machine_is_smm_enabled(X86_MACHINE(ms))) {
+        x86_machine_is_smm_enabled(X86_MACHINE(ms)) && first) {
         smram_machine_done.notify = register_smram_listener;
         qemu_add_machine_init_done_notifier(&smram_machine_done);
     }
@@ -3408,15 +3437,22 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
                 return ret;
             }
 
-            ret = kvm_msr_energy_thread_init(s, ms);
-            if (ret < 0) {
-                error_report("kvm : error RAPL feature requirement not met");
-                return ret;
+            if (first) {
+                ret = kvm_msr_energy_thread_init(s, ms);
+                if (ret < 0) {
+                    error_report("kvm : "
+                                 "error RAPL feature requirement not met");
+                    return ret;
+                }
             }
         }
     }
 
-    kvm_vmfd_add_change_notifier(&kvm_vmfd_change_notifier);
+    if (first) {
+        kvm_vmfd_add_change_notifier(&kvm_vmfd_change_notifier);
+    }
+
+    first = false;
 
     return 0;
 }
target/i386/kvm/trace-events +1 -0
index 74a6234ff7..2d213c9f9b 100644
@@ -6,6 +6,7 @@ kvm_x86_add_msi_route(int virq) "Adding route entry for virq %d"
 kvm_x86_remove_msi_route(int virq) "Removing route entry for virq %d"
 kvm_x86_update_msi_routes(int num) "Updated %d MSI routes"
 kvm_hc_map_gpa_range(uint64_t gpa, uint64_t size, uint64_t attributes, uint64_t flags) "gpa 0x%" PRIx64 " size 0x%" PRIx64 " attributes 0x%" PRIx64 " flags 0x%" PRIx64
+kvm_arch_on_vmfd_change(void) ""
 
 # xen-emu.c
 kvm_xen_hypercall(int cpu, uint8_t cpl, uint64_t input, uint64_t a0, uint64_t a1, uint64_t a2, uint64_t ret) "xen_hypercall: cpu %d cpl %d input %" PRIu64 " a0 0x%" PRIx64 " a1 0x%" PRIx64 " a2 0x%" PRIx64" ret 0x%" PRIx64
Signature · 1 line
2.42.0
author
#15 lore raw

PATCH v4 10/31 hw/i386: refactor x86_bios_rom_init for reuse in confidential guest reset

In reply to Ani Sinha

For confidential guests, bios image must be reinitialized upon reset. This is because bios memory is encrypted and hence once the old confidential kvm context is destroyed, it cannot be decrypted. It needs to be reinitilized. In order to do that, this change refactors x86_bios_rom_init() code so that parts of it can be called during confidential guest reset.

---
 hw/i386/x86-common.c | 50 ++++++++++++++++++++++++++++++++------------
 1 file changed, 37 insertions(+), 13 deletions(-)
hw/i386/x86-common.c +37 -13
index c1c9224039..18f2e2cfb9 100644
@@ -1024,17 +1024,11 @@ void x86_isa_bios_init(MemoryRegion *isa_bios, MemoryRegion *isa_memory,
     memory_region_set_readonly(isa_bios, read_only);
 }
 
-void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
-                       MemoryRegion *rom_memory, bool isapc_ram_fw)
+static int get_bios_size(X86MachineState *x86ms,
+                         const char *bios_name, char *filename)
 {
-    const char *bios_name;
-    char *filename;
     int bios_size;
-    ssize_t ret;
 
-    /* BIOS load */
-    bios_name = MACHINE(x86ms)->firmware ?: default_firmware;
-    filename = qemu_find_file(QEMU_FILE_TYPE_BIOS, bios_name);
     if (filename) {
         bios_size = get_image_size(filename, NULL);
     } else {
@@ -1044,6 +1038,21 @@ void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
         (bios_size % 65536) != 0) {
         goto bios_error;
     }
+
+    return bios_size;
+
+ bios_error:
+    fprintf(stderr, "qemu: could not load PC BIOS '%s'\n", bios_name);
+    exit(1);
+}
+
+static void load_bios_from_file(X86MachineState *x86ms, const char *bios_name,
+                                char *filename, int bios_size,
+                                bool isapc_ram_fw)
+{
+    ssize_t ret;
+
+    /* BIOS load */
     if (machine_require_guest_memfd(MACHINE(x86ms))) {
         memory_region_init_ram_guest_memfd(&x86ms->bios, NULL, "pc.bios",
                                            bios_size, &error_fatal);
@@ -1072,7 +1081,26 @@ void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
             goto bios_error;
         }
     }
-    g_free(filename);
+
+    return;
+
+ bios_error:
+    fprintf(stderr, "qemu: could not load PC BIOS '%s'\n", bios_name);
+    exit(1);
+}
+
+void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
+                       MemoryRegion *rom_memory, bool isapc_ram_fw)
+{
+    int bios_size;
+    const char *bios_name;
+    g_autofree char *filename;
+
+    bios_name = MACHINE(x86ms)->firmware ?: default_firmware;
+    filename = qemu_find_file(QEMU_FILE_TYPE_BIOS, bios_name);
+
+    bios_size = get_bios_size(x86ms, bios_name, filename);
+    load_bios_from_file(x86ms, bios_name, filename, bios_size, isapc_ram_fw);
 
     if (!machine_require_guest_memfd(MACHINE(x86ms))) {
         /* map the last 128KB of the BIOS in ISA space */
@@ -1085,8 +1113,4 @@ void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
                                 (uint32_t)(-bios_size),
                                 &x86ms->bios);
     return;
-
-bios_error:
-    fprintf(stderr, "qemu: could not load PC BIOS '%s'\n", bios_name);
-    exit(1);
 }
Signature · 1 line
2.42.0
author
#16 lore raw

PATCH v4 11/31 hw/i386: export a new function x86_bios_rom_reload

In reply to Ani Sinha

Confidential guest smust reload their bios rom upon reset. This is because bios memory is encrypted and upon reset, the contents of the old bios memory is lost and cannot be re-used. To this end, export a new x86 function x86_bios_rom_reload() to reload the bios again. This function will be used in the subsequent patches.

---
 hw/i386/x86-common.c  | 21 +++++++++++++++++++++
 include/hw/i386/x86.h |  1 +
 2 files changed, 22 insertions(+)
hw/i386/x86-common.c +21 -0
index 18f2e2cfb9..b5759f6993 100644
@@ -1089,6 +1089,27 @@ static void load_bios_from_file(X86MachineState *x86ms, const char *bios_name,
     exit(1);
 }
 
+void x86_bios_rom_reload(X86MachineState *x86ms)
+{
+    int bios_size;
+    const char *bios_name;
+    char *filename;
+
+    if (memory_region_size(&x86ms->bios) == 0) {
+        /* if -bios is not used */
+        return;
+    }
+
+    bios_name = MACHINE(x86ms)->firmware ?: "bios.bin";
+    filename = qemu_find_file(QEMU_FILE_TYPE_BIOS, bios_name);
+
+    bios_size = get_bios_size(x86ms, bios_name, filename);
+
+    void *ptr = memory_region_get_ram_ptr(&x86ms->bios);
+    load_image_size(filename, ptr, bios_size);
+    x86_firmware_configure(0x100000000ULL - bios_size, ptr, bios_size);
+}
+
 void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
                        MemoryRegion *rom_memory, bool isapc_ram_fw)
 {
include/hw/i386/x86.h +1 -0
index 0dffba95f9..66080da54a 100644
@@ -127,6 +127,7 @@ void x86_isa_bios_init(MemoryRegion *isa_bios, MemoryRegion *isa_memory,
                        MemoryRegion *bios, bool read_only);
 void x86_bios_rom_init(X86MachineState *x86ms, const char *default_firmware,
                        MemoryRegion *rom_memory, bool isapc_ram_fw);
+void x86_bios_rom_reload(X86MachineState *x86ms);
 
 void x86_load_linux(X86MachineState *x86ms,
                     FWCfgState *fw_cfg,
Signature · 1 line
2.42.0
author
#17 lore raw

PATCH v4 12/31 kvm/i386: reload firmware for confidential guest reset

In reply to Ani Sinha

When IGVM is not being used by the confidential guest, the guest firmware has to be reloaded explicitly again into memory. This is because, the memory into which the firmware was loaded before reset was encrypted and is thus lost upon reset. When IGVM is used, it is expected that the IGVM will contain the guest firmware and the execution of the IGVM directives will set up the guest firmware memory.

---
 target/i386/kvm/kvm.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
target/i386/kvm/kvm.c +8 -1
index e82a6e9eda..eec2f27a0f 100644
@@ -3281,7 +3281,14 @@ int kvm_arch_on_vmfd_change(MachineState *ms, KVMState *s)
 
     if (object_dynamic_cast(OBJECT(ms), TYPE_X86_MACHINE)) {
         X86MachineState *x86ms = X86_MACHINE(ms);
-
+        /*
+         * For confidential guests, reload bios ROM if IGVM is not specified.
+         * If an IGVM file is specified then the firmware must be provided
+         * in the IGVM file.
+         */
+        if (ms->cgs && !x86ms->igvm) {
+                x86_bios_rom_reload(x86ms);
+        }
         if (x86_machine_is_smm_enabled(x86ms)) {
             memory_listener_register(&smram_listener.listener,
                                      &smram_address_space);
Signature · 1 line
2.42.0
author
#18 lore raw

PATCH v4 13/31 accel/kvm: rebind current VCPUs to the new KVM VM file descriptor upon reset

In reply to Ani Sinha

Confidential guests needs to generate a new KVM file descriptor upon virtual machine reset. Existing VCPUs needs to be reattached to this new KVM VM file descriptor. As a part of this, new VCPU file descriptors against this new KVM VM file descriptor needs to be created and re-initialized. Resources allocated against the old VCPU fds needs to be released. This change makes this happen.

---
 accel/kvm/kvm-all.c    | 215 +++++++++++++++++++++++++++++++++--------
 accel/kvm/trace-events |   1 +
 2 files changed, 174 insertions(+), 42 deletions(-)
accel/kvm/kvm-all.c +173 -42
index a6ff16a037..5e81bf8ad2 100644
@@ -127,6 +127,10 @@ static NotifierList kvm_irqchip_change_notifiers =
 static NotifierWithReturnList register_vmfd_changed_notifiers =
     NOTIFIER_WITH_RETURN_LIST_INITIALIZER(register_vmfd_changed_notifiers);
 
+static int map_kvm_run(KVMState *s, CPUState *cpu, Error **errp);
+static int map_kvm_dirty_gfns(KVMState *s, CPUState *cpu, Error **errp);
+static int vcpu_unmap_regions(KVMState *s, CPUState *cpu);
+
 struct KVMResampleFd {
     int gsi;
     EventNotifier *resample_event;
@@ -420,6 +424,90 @@ err:
     return ret;
 }
 
+static void kvm_create_vcpu_internal(CPUState *cpu, KVMState *s, int kvm_fd)
+{
+    cpu->kvm_fd = kvm_fd;
+    cpu->kvm_state = s;
+    if (!s->guest_state_protected) {
+        cpu->vcpu_dirty = true;
+    }
+    cpu->dirty_pages = 0;
+    cpu->throttle_us_per_full = 0;
+
+    return;
+}
+
+static int kvm_rebind_vcpus(Error **errp)
+{
+    CPUState *cpu;
+    unsigned long vcpu_id;
+    KVMState *s = kvm_state;
+    int kvm_fd, ret = 0;
+
+    CPU_FOREACH(cpu) {
+        vcpu_id = kvm_arch_vcpu_id(cpu);
+
+        if (cpu->kvm_fd) {
+            close(cpu->kvm_fd);
+        }
+
+        ret = kvm_arch_destroy_vcpu(cpu);
+        if (ret < 0) {
+            goto err;
+        }
+
+        if (s->coalesced_mmio_ring == (void *)cpu->kvm_run + PAGE_SIZE) {
+            s->coalesced_mmio_ring = NULL;
+        }
+
+        ret = vcpu_unmap_regions(s, cpu);
+        if (ret < 0) {
+            goto err;
+        }
+
+        ret = kvm_arch_pre_create_vcpu(cpu, errp);
+        if (ret < 0) {
+            goto err;
+        }
+
+        kvm_fd = kvm_vm_ioctl(s, KVM_CREATE_VCPU, vcpu_id);
+        if (kvm_fd < 0) {
+            error_report("KVM_CREATE_VCPU IOCTL failed for vCPU %lu (%s)",
+                         vcpu_id, strerror(kvm_fd));
+            return kvm_fd;
+        }
+
+        kvm_create_vcpu_internal(cpu, s, kvm_fd);
+
+        ret = map_kvm_run(s, cpu, errp);
+        if (ret < 0) {
+            goto err;
+        }
+
+        if (s->kvm_dirty_ring_size) {
+            ret = map_kvm_dirty_gfns(s, cpu, errp);
+            if (ret < 0) {
+                goto err;
+            }
+        }
+
+        ret = kvm_arch_init_vcpu(cpu);
+        if (ret < 0) {
+            error_setg_errno(errp, -ret,
+                             "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
+                             vcpu_id);
+        }
+
+        close(cpu->kvm_vcpu_stats_fd);
+        cpu->kvm_vcpu_stats_fd = kvm_vcpu_ioctl(cpu, KVM_GET_STATS_FD, NULL);
+        kvm_init_cpu_signals(cpu);
+    }
+    trace_kvm_rebind_vcpus();
+
+ err:
+    return ret;
+}
+
 static void kvm_park_vcpu(CPUState *cpu)
 {
     struct KVMParkedVcpu *vcpu;
@@ -483,13 +571,7 @@ static int kvm_create_vcpu(CPUState *cpu)
         }
     }
 
-    cpu->kvm_fd = kvm_fd;
-    cpu->kvm_state = s;
-    if (!s->guest_state_protected) {
-        cpu->vcpu_dirty = true;
-    }
-    cpu->dirty_pages = 0;
-    cpu->throttle_us_per_full = 0;
+    kvm_create_vcpu_internal(cpu, s, kvm_fd);
 
     trace_kvm_create_vcpu(cpu->cpu_index, vcpu_id, kvm_fd);
 
@@ -508,19 +590,11 @@ int kvm_create_and_park_vcpu(CPUState *cpu)
     return ret;
 }
 
-static int do_kvm_destroy_vcpu(CPUState *cpu)
+static int vcpu_unmap_regions(KVMState *s, CPUState *cpu)
 {
-    KVMState *s = kvm_state;
     int mmap_size;
     int ret = 0;
 
-    trace_kvm_destroy_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
-
-    ret = kvm_arch_destroy_vcpu(cpu);
-    if (ret < 0) {
-        goto err;
-    }
-
     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
     if (mmap_size < 0) {
         ret = mmap_size;
@@ -548,39 +622,47 @@ static int do_kvm_destroy_vcpu(CPUState *cpu)
         cpu->kvm_dirty_gfns = NULL;
     }
 
-    kvm_park_vcpu(cpu);
-err:
+ err:
     return ret;
 }
 
-void kvm_destroy_vcpu(CPUState *cpu)
-{
-    if (do_kvm_destroy_vcpu(cpu) < 0) {
-        error_report("kvm_destroy_vcpu failed");
-        exit(EXIT_FAILURE);
-    }
-}
-
-int kvm_init_vcpu(CPUState *cpu, Error **errp)
+static int do_kvm_destroy_vcpu(CPUState *cpu)
 {
     KVMState *s = kvm_state;
-    int mmap_size;
-    int ret;
+    int ret = 0;
 
-    trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
+    trace_kvm_destroy_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
 
-    ret = kvm_arch_pre_create_vcpu(cpu, errp);
+    ret = kvm_arch_destroy_vcpu(cpu);
     if (ret < 0) {
         goto err;
     }
 
-    ret = kvm_create_vcpu(cpu);
+    /* If I am the CPU that created coalesced_mmio_ring, then discard it */
+    if (s->coalesced_mmio_ring == (void *)cpu->kvm_run + PAGE_SIZE) {
+        s->coalesced_mmio_ring = NULL;
+    }
+
+    ret = vcpu_unmap_regions(s, cpu);
     if (ret < 0) {
-        error_setg_errno(errp, -ret,
-                         "kvm_init_vcpu: kvm_create_vcpu failed (%lu)",
-                         kvm_arch_vcpu_id(cpu));
         goto err;
     }
+    kvm_park_vcpu(cpu);
+err:
+    return ret;
+}
+
+void kvm_destroy_vcpu(CPUState *cpu)
+{
+    if (do_kvm_destroy_vcpu(cpu) < 0) {
+        error_report("kvm_destroy_vcpu failed");
+        exit(EXIT_FAILURE);
+    }
+}
+
+static int map_kvm_run(KVMState *s, CPUState *cpu, Error **errp)
+{
+    int mmap_size, ret = 0;
 
     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
     if (mmap_size < 0) {
@@ -605,14 +687,53 @@ int kvm_init_vcpu(CPUState *cpu, Error **errp)
             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
     }
 
+ err:
+    return ret;
+}
+
+static int map_kvm_dirty_gfns(KVMState *s, CPUState *cpu, Error **errp)
+{
+    int ret = 0;
+    /* Use MAP_SHARED to share pages with the kernel */
+    cpu->kvm_dirty_gfns = mmap(NULL, s->kvm_dirty_ring_bytes,
+                               PROT_READ | PROT_WRITE, MAP_SHARED,
+                               cpu->kvm_fd,
+                               PAGE_SIZE * KVM_DIRTY_LOG_PAGE_OFFSET);
+    if (cpu->kvm_dirty_gfns == MAP_FAILED) {
+        ret = -errno;
+    }
+
+    return ret;
+}
+
+int kvm_init_vcpu(CPUState *cpu, Error **errp)
+{
+    KVMState *s = kvm_state;
+    int ret;
+
+    trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
+
+    ret = kvm_arch_pre_create_vcpu(cpu, errp);
+    if (ret < 0) {
+        goto err;
+    }
+
+    ret = kvm_create_vcpu(cpu);
+    if (ret < 0) {
+        error_setg_errno(errp, -ret,
+                         "kvm_init_vcpu: kvm_create_vcpu failed (%lu)",
+                         kvm_arch_vcpu_id(cpu));
+        goto err;
+    }
+
+    ret = map_kvm_run(s, cpu, errp);
+    if (ret < 0) {
+        goto err;
+    }
+
     if (s->kvm_dirty_ring_size) {
-        /* Use MAP_SHARED to share pages with the kernel */
-        cpu->kvm_dirty_gfns = mmap(NULL, s->kvm_dirty_ring_bytes,
-                                   PROT_READ | PROT_WRITE, MAP_SHARED,
-                                   cpu->kvm_fd,
-                                   PAGE_SIZE * KVM_DIRTY_LOG_PAGE_OFFSET);
-        if (cpu->kvm_dirty_gfns == MAP_FAILED) {
-            ret = -errno;
+        ret = map_kvm_dirty_gfns(s, cpu, errp);
+        if (ret < 0) {
             goto err;
         }
     }
@@ -2707,6 +2828,16 @@ static int kvm_reset_vmfd(MachineState *ms)
     }
     assert(!err);
 
+    /*
+     * rebind new vcpu fds with the new kvm fds
+     * These can only be called after kvm_arch_on_vmfd_change()
+     */
+    ret = kvm_rebind_vcpus(&err);
+    if (ret < 0) {
+        return ret;
+    }
+    assert(!err);
+
     /* these can be only called after ram_block_rebind() */
     memory_listener_register(&kml->listener, &address_space_memory);
     memory_listener_register(&kvm_io_listener, &address_space_io);
accel/kvm/trace-events +1 -0
index e4beda0148..4a8921c632 100644
@@ -15,6 +15,7 @@ kvm_park_vcpu(int cpu_index, unsigned long arch_cpu_id) "index: %d id: %lu"
 kvm_unpark_vcpu(unsigned long arch_cpu_id, const char *msg) "id: %lu %s"
 kvm_irqchip_commit_routes(void) ""
 kvm_reset_vmfd(void) ""
+kvm_rebind_vcpus(void) ""
 kvm_irqchip_add_msi_route(char *name, int vector, int virq) "dev %s vector %d virq %d"
 kvm_irqchip_update_msi_route(int virq) "Updating MSI route virq=%d"
 kvm_irqchip_release_virq(int virq) "virq %d"
Signature · 1 line
2.42.0
author
#19 lore raw

PATCH v4 14/31 i386/tdx: refactor TDX firmware memory initialization code into a new function

In reply to Ani Sinha

A new helper function is introduced that refactors all firmware memory initialization code into a separate function. No functional change.

---
 target/i386/kvm/tdx.c | 73 ++++++++++++++++++++++++-------------------
 1 file changed, 40 insertions(+), 33 deletions(-)
target/i386/kvm/tdx.c +40 -33
index a3e81e1c0c..fd8e3de969 100644
@@ -295,14 +295,51 @@ static void tdx_post_init_vcpus(void)
     }
 }
 
-static void tdx_finalize_vm(Notifier *notifier, void *unused)
+static void tdx_init_fw_mem_region(void)
 {
     TdxFirmware *tdvf = &tdx_guest->tdvf;
     TdxFirmwareEntry *entry;
-    RAMBlock *ram_block;
     Error *local_err = NULL;
     int r;
 
+    for_each_tdx_fw_entry(tdvf, entry) {
+        struct kvm_tdx_init_mem_region region;
+        uint32_t flags;
+
+        region = (struct kvm_tdx_init_mem_region) {
+            .source_addr = (uintptr_t)entry->mem_ptr,
+            .gpa = entry->address,
+            .nr_pages = entry->size >> 12,
+        };
+
+        flags = entry->attributes & TDVF_SECTION_ATTRIBUTES_MR_EXTEND ?
+                KVM_TDX_MEASURE_MEMORY_REGION : 0;
+
+        do {
+            error_free(local_err);
+            local_err = NULL;
+            r = tdx_vcpu_ioctl(first_cpu, KVM_TDX_INIT_MEM_REGION, flags,
+                               &region, &local_err);
+        } while (r == -EAGAIN || r == -EINTR);
+        if (r < 0) {
+            error_report_err(local_err);
+            exit(1);
+        }
+
+        if (entry->type == TDVF_SECTION_TYPE_TD_HOB ||
+            entry->type == TDVF_SECTION_TYPE_TEMP_MEM) {
+            qemu_ram_munmap(-1, entry->mem_ptr, entry->size);
+            entry->mem_ptr = NULL;
+        }
+    }
+}
+
+static void tdx_finalize_vm(Notifier *notifier, void *unused)
+{
+    TdxFirmware *tdvf = &tdx_guest->tdvf;
+    TdxFirmwareEntry *entry;
+    RAMBlock *ram_block;
+
     tdx_init_ram_entries();
 
     for_each_tdx_fw_entry(tdvf, entry) {
@@ -339,37 +376,7 @@ static void tdx_finalize_vm(Notifier *notifier, void *unused)
     tdvf_hob_create(tdx_guest, tdx_get_hob_entry(tdx_guest));
 
     tdx_post_init_vcpus();
-
-    for_each_tdx_fw_entry(tdvf, entry) {
-        struct kvm_tdx_init_mem_region region;
-        uint32_t flags;
-
-        region = (struct kvm_tdx_init_mem_region) {
-            .source_addr = (uintptr_t)entry->mem_ptr,
-            .gpa = entry->address,
-            .nr_pages = entry->size >> 12,
-        };
-
-        flags = entry->attributes & TDVF_SECTION_ATTRIBUTES_MR_EXTEND ?
-                KVM_TDX_MEASURE_MEMORY_REGION : 0;
-
-        do {
-            error_free(local_err);
-            local_err = NULL;
-            r = tdx_vcpu_ioctl(first_cpu, KVM_TDX_INIT_MEM_REGION, flags,
-                               &region, &local_err);
-        } while (r == -EAGAIN || r == -EINTR);
-        if (r < 0) {
-            error_report_err(local_err);
-            exit(1);
-        }
-
-        if (entry->type == TDVF_SECTION_TYPE_TD_HOB ||
-            entry->type == TDVF_SECTION_TYPE_TEMP_MEM) {
-            qemu_ram_munmap(-1, entry->mem_ptr, entry->size);
-            entry->mem_ptr = NULL;
-        }
-    }
+    tdx_init_fw_mem_region();
 
     /*
      * TDVF image has been copied into private region above via
Signature · 1 line
2.42.0
author
#20 lore raw

PATCH v4 15/31 i386/tdx: finalize TDX guest state upon reset

In reply to Ani Sinha

When the confidential virtual machine KVM file descriptor changes due to the guest reset, some TDX specific setup steps needs to be done again. This includes finalizing the initial guest launch state again. This change re-executes some parts of the TDX setup during the device reset phaze using a resettable interface. This finalizes the guest launch state again and locks it in. Machine done notifier which was previously used is no longer needed as the same code is now executed as a part of VM reset.

---
 target/i386/kvm/tdx.c        | 38 +++++++++++++++++++++++++++++++-----
 target/i386/kvm/tdx.h        |  1 +
 target/i386/kvm/trace-events |  3 +++
 3 files changed, 37 insertions(+), 5 deletions(-)
target/i386/kvm/tdx.c +33 -5
index fd8e3de969..37e91d95e1 100644
@@ -19,6 +19,7 @@
 #include "crypto/hash.h"
 #include "system/kvm_int.h"
 #include "system/runstate.h"
+#include "system/reset.h"
 #include "system/system.h"
 #include "system/ramblock.h"
 #include "system/address-spaces.h"
@@ -38,6 +39,7 @@
 #include "kvm_i386.h"
 #include "tdx.h"
 #include "tdx-quote-generator.h"
+#include "trace.h"
 
 #include "standard-headers/asm-x86/kvm_para.h"
 
@@ -389,9 +391,19 @@ static void tdx_finalize_vm(Notifier *notifier, void *unused)
     CONFIDENTIAL_GUEST_SUPPORT(tdx_guest)->ready = true;
 }
 
-static Notifier tdx_machine_done_notify = {
-    .notify = tdx_finalize_vm,
-};
+static void tdx_handle_reset(Object *obj, ResetType type)
+{
+    if (!runstate_is_running() && !phase_check(PHASE_MACHINE_READY)) {
+        return;
+    }
+
+    if (!kvm_enable_hypercall(BIT_ULL(KVM_HC_MAP_GPA_RANGE))) {
+        error_setg(&error_fatal, "KVM_HC_MAP_GPA_RANGE not enabled for guest");
+    }
+
+    tdx_finalize_vm(NULL, NULL);
+    trace_tdx_handle_reset();
+}
 
 /*
  * Some CPUID bits change from fixed1 to configurable bits when TDX module
@@ -738,8 +750,6 @@ static int tdx_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
      */
     kvm_readonly_mem_allowed = false;
 
-    qemu_add_machine_init_done_notifier(&tdx_machine_done_notify);
-
     tdx_guest = tdx;
     return 0;
 }
@@ -1505,6 +1515,7 @@ OBJECT_DEFINE_TYPE_WITH_INTERFACES(TdxGuest,
                                    TDX_GUEST,
                                    X86_CONFIDENTIAL_GUEST,
                                    { TYPE_USER_CREATABLE },
+                                   { TYPE_RESETTABLE_INTERFACE },
                                    { NULL })
 
 static void tdx_guest_init(Object *obj)
@@ -1538,16 +1549,24 @@ static void tdx_guest_init(Object *obj)
 
     tdx->event_notify_vector = -1;
     tdx->event_notify_apicid = -1;
+    qemu_register_resettable(obj);
 }
 
 static void tdx_guest_finalize(Object *obj)
 {
 }
 
+static ResettableState *tdx_reset_state(Object *obj)
+{
+    TdxGuest *tdx = TDX_GUEST(obj);
+    return &tdx->reset_state;
+}
+
 static void tdx_guest_class_init(ObjectClass *oc, const void *data)
 {
     ConfidentialGuestSupportClass *klass = CONFIDENTIAL_GUEST_SUPPORT_CLASS(oc);
     X86ConfidentialGuestClass *x86_klass = X86_CONFIDENTIAL_GUEST_CLASS(oc);
+    ResettableClass *rc = RESETTABLE_CLASS(oc);
 
     klass->kvm_init = tdx_kvm_init;
     klass->can_rebuild_guest_state = true;
@@ -1555,4 +1574,13 @@ static void tdx_guest_class_init(ObjectClass *oc, const void *data)
     x86_klass->cpu_instance_init = tdx_cpu_instance_init;
     x86_klass->adjust_cpuid_features = tdx_adjust_cpuid_features;
     x86_klass->check_features = tdx_check_features;
+
+    /*
+     * the exit phase makes sure sev handles reset after all legacy resets
+     * have taken place (in the hold phase) and IGVM has also properly
+     * set up the boot state.
+     */
+    rc->phases.exit = tdx_handle_reset;
+    rc->get_state = tdx_reset_state;
+
 }
target/i386/kvm/tdx.h +1 -0
index 1c38faf983..264fbe530c 100644
@@ -70,6 +70,7 @@ typedef struct TdxGuest {
 
     uint32_t event_notify_vector;
     uint32_t event_notify_apicid;
+    ResettableState reset_state;
 } TdxGuest;
 
 #ifdef CONFIG_TDX
target/i386/kvm/trace-events +3 -0
index 2d213c9f9b..a386234571 100644
@@ -14,3 +14,6 @@ kvm_xen_soft_reset(void) ""
 kvm_xen_set_shared_info(uint64_t gfn) "shared info at gfn 0x%" PRIx64
 kvm_xen_set_vcpu_attr(int cpu, int type, uint64_t gpa) "vcpu attr cpu %d type %d gpa 0x%" PRIx64
 kvm_xen_set_vcpu_callback(int cpu, int vector) "callback vcpu %d vector %d"
+
+# tdx.c
+tdx_handle_reset(void) ""
Signature · 1 line
2.42.0
author
#21 lore raw

PATCH v4 16/31 i386/tdx: add a pre-vmfd change notifier to reset tdx state

In reply to Ani Sinha

During reset, when the VM file descriptor is changed, the TDX state needs to be re-initialized. A notifier callback is implemented to reset the old state and free memory before the new state is initialized post VM file descriptor change.

---
 target/i386/kvm/tdx.c | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)
target/i386/kvm/tdx.c +31 -0
index 37e91d95e1..4cae99c281 100644
@@ -405,6 +405,36 @@ static void tdx_handle_reset(Object *obj, ResetType type)
     trace_tdx_handle_reset();
 }
 
+/* TDX guest reset will require us to reinitialize some of tdx guest state. */
+static int set_tdx_vm_uninitialized(NotifierWithReturn *notifier,
+                                    void *data, Error** errp)
+{
+    TdxFirmware *fw = &tdx_guest->tdvf;
+
+    if (!((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    if (tdx_guest->initialized) {
+        tdx_guest->initialized = false;
+    }
+
+    g_free(tdx_guest->ram_entries);
+
+    /*
+     * the firmware entries will be parsed again, see
+     * x86_firmware_configure() -> tdx_parse_tdvf()
+     */
+    fw->entries = 0;
+    g_free(fw->entries);
+
+    return 0;
+}
+
+static NotifierWithReturn tdx_vmfd_change_notifier = {
+    .notify = set_tdx_vm_uninitialized,
+};
+
 /*
  * Some CPUID bits change from fixed1 to configurable bits when TDX module
  * supports TDX_FEATURES0.VE_REDUCTION. e.g., MCA/MCE/MTRR/CORE_CAPABILITY.
@@ -1549,6 +1579,7 @@ static void tdx_guest_init(Object *obj)
 
     tdx->event_notify_vector = -1;
     tdx->event_notify_apicid = -1;
+    kvm_vmfd_add_change_notifier(&tdx_vmfd_change_notifier);
     qemu_register_resettable(obj);
 }
 
Signature · 1 line
2.42.0
author
#22 lore raw

PATCH v4 17/31 i386/sev: add migration blockers only once

In reply to Ani Sinha

sev_launch_finish() and sev_snp_launch_finish() could be called multiple times when the confidential guest is being reset/rebooted. The migration blockers should not be added multiple times, once per invocation. This change makes sure that the migration blockers are added only one time by adding the migration blockers to the vm state change handler when the vm transitions to the running state. Subsequent reboots do not change the state of the vm.

---
 target/i386/sev.c | 20 +++++---------------
 1 file changed, 5 insertions(+), 15 deletions(-)
target/i386/sev.c +5 -15
index 66e38ca32e..260d8ef88b 100644
@@ -1421,11 +1421,6 @@ sev_launch_finish(SevCommonState *sev_common)
     }
 
     sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV: Migration is not implemented");
-    migrate_add_blocker(&sev_mig_blocker, &error_fatal);
 }
 
 static int snp_launch_update_data(uint64_t gpa, void *hva, size_t len,
@@ -1608,7 +1603,6 @@ static void
 sev_snp_launch_finish(SevCommonState *sev_common)
 {
     int ret, error;
-    Error *local_err = NULL;
     OvmfSevMetadata *metadata;
     SevLaunchUpdateData *data;
     SevSnpGuestState *sev_snp = SEV_SNP_GUEST(sev_common);
@@ -1655,15 +1649,6 @@ sev_snp_launch_finish(SevCommonState *sev_common)
 
     kvm_mark_guest_state_protected();
     sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV-SNP: Migration is not implemented");
-    ret = migrate_add_blocker(&sev_mig_blocker, &local_err);
-    if (local_err) {
-        error_report_err(local_err);
-        exit(1);
-    }
 }
 
 
@@ -1676,6 +1661,11 @@ sev_vm_state_change(void *opaque, bool running, RunState state)
     if (running) {
         if (!sev_check_state(sev_common, SEV_STATE_RUNNING)) {
             klass->launch_finish(sev_common);
+
+            /* add migration blocker */
+            error_setg(&sev_mig_blocker,
+                       "SEV: Migration is not implemented");
+            migrate_add_blocker(&sev_mig_blocker, &error_fatal);
         }
     }
 }
Signature · 1 line
2.42.0
Reviewed-by on v4 17/31
#23 lore raw

Re: [PATCH v4 17/31] i386/sev: add migration blockers only once

Quoted 55 lines · Ani Sinha

sev_launch_finish() and sev_snp_launch_finish() could be called multiple times when the confidential guest is being reset/rebooted. The migration blockers should not be added multiple times, once per invocation. This change makes sure that the migration blockers are added only one time by adding the migration blockers to the vm state change handler when the vm transitions to the running state. Subsequent reboots do not change the state of the vm.

---
 target/i386/sev.c | 20 +++++---------------
 1 file changed, 5 insertions(+), 15 deletions(-)
target/i386/sev.c +0 -15
index 66e38ca32e..260d8ef88b 100644
@@ -1421,11 +1421,6 @@ sev_launch_finish(SevCommonState *sev_common)
     }
 
     sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV: Migration is not implemented");
-    migrate_add_blocker(&sev_mig_blocker, &error_fatal);
 }
 
 static int snp_launch_update_data(uint64_t gpa, void *hva, size_t len,
@@ -1608,7 +1603,6 @@ static void
 sev_snp_launch_finish(SevCommonState *sev_common)
 {
     int ret, error;
-    Error *local_err = NULL;
     OvmfSevMetadata *metadata;
     SevLaunchUpdateData *data;
     SevSnpGuestState *sev_snp = SEV_SNP_GUEST(sev_common);
@@ -1655,15 +1649,6 @@ sev_snp_launch_finish(SevCommonState *sev_common)
 
     kvm_mark_guest_state_protected();
     sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV-SNP: Migration is not implemented");
-    ret = migrate_add_blocker(&sev_mig_blocker, &local_err);
-    if (local_err) {
-        error_report_err(local_err);
-        exit(1);
-    }
 }
 
 
@@ -1676,6 +1661,11 @@ sev_vm_state_change(void *opaque, bool running, RunState state)
     if (running) {
         if (!sev_check_state(sev_common, SEV_STATE_RUNNING)) {

target/i386/sev.c sev_vm_state_change(void *opaque, bool running, RunState state)

Quoted 10 lines
target/i386/sev.c +5 -1
             klass->launch_finish(sev_common);
+
+            /* add migration blocker */
+            error_setg(&sev_mig_blocker,
+                       "SEV: Migration is not implemented");
+            migrate_add_blocker(&sev_mig_blocker, &error_fatal);
         }
     }
 }
--

* 'sev_mig_blocker' is a global static variable, so it's the same blocker (address) added each time, maybe add_blocker() should do a check to avoid duplicates.

* Otherwise it looks okay.

Thank you.
---
- Prasad

author
#24 lore raw

Re: [PATCH v4 17/31] i386/sev: add migration blockers only once

Quoted 66 lines · Prasad Pandit
Quoted 61 lines · Ani Sinha

sev_launch_finish() and sev_snp_launch_finish() could be called multiple times when the confidential guest is being reset/rebooted. The migration blockers should not be added multiple times, once per invocation. This change makes sure that the migration blockers are added only one time by adding the migration blockers to the vm state change handler when the vm transitions to the running state. Subsequent reboots do not change the state of the vm.

---
target/i386/sev.c | 20 +++++---------------

1 file changed, 5 insertions(+), 15 deletions(-)
target/i386/sev.c +5 -15
index 66e38ca32e..260d8ef88b 100644
@@ -1421,11 +1421,6 @@ sev_launch_finish(SevCommonState *sev_common)
    }
 
    sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV: Migration is not implemented");
-    migrate_add_blocker(&sev_mig_blocker, &error_fatal);
 }
 
 static int snp_launch_update_data(uint64_t gpa, void *hva, size_t len,
@@ -1608,7 +1603,6 @@ static void
 sev_snp_launch_finish(SevCommonState *sev_common)
 {
    int ret, error;
-    Error *local_err = NULL;
    OvmfSevMetadata *metadata;
    SevLaunchUpdateData *data;
    SevSnpGuestState *sev_snp = SEV_SNP_GUEST(sev_common);
@@ -1655,15 +1649,6 @@ sev_snp_launch_finish(SevCommonState *sev_common)
 
    kvm_mark_guest_state_protected();
    sev_set_guest_state(sev_common, SEV_STATE_RUNNING);
-
-    /* add migration blocker */
-    error_setg(&sev_mig_blocker,
-               "SEV-SNP: Migration is not implemented");
-    ret = migrate_add_blocker(&sev_mig_blocker, &local_err);
-    if (local_err) {
-        error_report_err(local_err);
-        exit(1);
-    }
 }
 
 
@@ -1676,6 +1661,11 @@ sev_vm_state_change(void *opaque, bool running, RunState state)
    if (running) {
        if (!sev_check_state(sev_common, SEV_STATE_RUNNING)) {
            klass->launch_finish(sev_common);
+
+            /* add migration blocker */
+            error_setg(&sev_mig_blocker,
+                       "SEV: Migration is not implemented");
+            migrate_add_blocker(&sev_mig_blocker, &error_fatal);
Quoted 8 lines

target/i386/sev.c sev_vm_state_change(void *opaque, bool running, RunState state)

Quoted 4 lines
target/i386/sev.c +0 -1
        }
    }
 }
--

* 'sev_mig_blocker' is a global static variable, so it's the same blocker (address) added each time, maybe add_blocker() should do a check to avoid duplicates.

Maybe we could do something like this (I have not tested it) but perhaps outside the scope of this change set:

migration/migration.c +5 -2
index b103a82fc0..7d85821315 100644
@@ -1696,8 +1696,11 @@ static int add_blockers(Error **reasonp, unsigned modes, Error **errp)
 {
     for (MigMode mode = 0; mode < MIG_MODE__MAX; mode++) {
         if (modes & BIT(mode)) {
-            migration_blockers[mode] = g_slist_prepend(migration_blockers[mode],
-                                                       *reasonp);
+            if (g_slist_index(migration_blockers[mode],
+                                 *reasonp) == -1) {
+                migration_blockers[mode] = g_slist_prepend(migration_blockers[mode],
+                                                           *reasonp);
+            }
         }
     }
     return 0;
Quoted 8 lines

* Otherwise it looks okay.

Thank you.
---
- Prasad

#25 lore raw

Re: [PATCH v4 17/31] i386/sev: add migration blockers only once

Quoted 10 lines · Ani Sinha

Maybe we could do something like this (I have not tested it) but perhaps outside the scope of this change set:

migration/migration.c +0 -0
index b103a82fc0..7d85821315 100644
@@ -1696,8 +1696,11 @@ static int add_blockers(Error **reasonp, unsigned modes, Error **errp)
 {
     for (MigMode mode = 0; mode < MIG_MODE__MAX; mode++) {
         if (modes & BIT(mode)) {

migration/migration.c static int add_blockers(Error **reasonp, unsigned modes, Error **errp) · line 1706

Quoted 10 lines
migration/migration.c +5 -2
-            migration_blockers[mode] = g_slist_prepend(migration_blockers[mode],
-                                                       *reasonp);
+            if (g_slist_index(migration_blockers[mode],
+                                 *reasonp) == -1) {
+                migration_blockers[mode] = g_slist_prepend(migration_blockers[mode],
+                                                           *reasonp);
+            }
         }
     }
     return 0;

* Yes, and when g_slist_index() returns >= 0, return -1 OR an indication that the blocker (address) already exists in the list. That way other such instances where the same blocker is added multiple times will be caught.

* But let's make it a separate patch. Restrict current patch to removing add_clocker calls from sev_*_launch_finish() calls.

Thank you.
---
- Prasad

author
#26 lore raw

PATCH v4 18/31 i386/sev: add notifiers only once

In reply to Ani Sinha

The various notifiers that are used needs to be installed only once not on every initialization. This includes the vm state change notifier and others. This change uses 'cgs->ready' flag to install the notifiers only one time, the first time.

---
 target/i386/sev.c | 36 +++++++++++++++++++-----------------
 1 file changed, 19 insertions(+), 17 deletions(-)
target/i386/sev.c +19 -17
index 260d8ef88b..647f4bf63d 100644
@@ -1920,8 +1920,9 @@ static int sev_common_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
         return -1;
     }
 
-    qemu_add_vm_change_state_handler(sev_vm_state_change, sev_common);
-
+    if (!cgs->ready) {
+        qemu_add_vm_change_state_handler(sev_vm_state_change, sev_common);
+    }
     cgs->ready = true;
 
     return 0;
@@ -1943,22 +1944,23 @@ static int sev_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
         return -1;
     }
 
-    /*
-     * SEV uses these notifiers to register/pin pages prior to guest use,
-     * but SNP relies on guest_memfd for private pages, which has its
-     * own internal mechanisms for registering/pinning private memory.
-     */
-    ram_block_notifier_add(&sev_ram_notifier);
-
-    /*
-     * The machine done notify event is used for SEV guests to get the
-     * measurement of the encrypted images. When SEV-SNP is enabled, the
-     * measurement is part of the guest attestation process where it can
-     * be collected without any reliance on the VMM. So skip registering
-     * the notifier for SNP in favor of using guest attestation instead.
-     */
-    qemu_add_machine_init_done_notifier(&sev_machine_done_notify);
+    if (!cgs->ready) {
+        /*
+         * SEV uses these notifiers to register/pin pages prior to guest use,
+         * but SNP relies on guest_memfd for private pages, which has its
+         * own internal mechanisms for registering/pinning private memory.
+         */
+        ram_block_notifier_add(&sev_ram_notifier);
 
+        /*
+         * The machine done notify event is used for SEV guests to get the
+         * measurement of the encrypted images. When SEV-SNP is enabled, the
+         * measurement is part of the guest attestation process where it can
+         * be collected without any reliance on the VMM. So skip registering
+         * the notifier for SNP in favor of using guest attestation instead.
+         */
+        qemu_add_machine_init_done_notifier(&sev_machine_done_notify);
+    }
     return 0;
 }
 
Signature · 1 line
2.42.0
author
#27 lore raw

PATCH v4 19/31 i386/sev: free existing launch update data and kernel hashes data on init

In reply to Ani Sinha

If there is existing launch update data and kernel hashes data, they need to be freed when initialization code is executed. This is important for resettable confidential guests where the initialization happens once every reset.

---
 target/i386/sev.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
target/i386/sev.c +12 -0
index 647f4bf63d..b3893e431c 100644
@@ -1773,6 +1773,7 @@ static int sev_common_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
     uint32_t ebx;
     uint32_t host_cbitpos;
     struct sev_user_data_status status = {};
+    SevLaunchUpdateData *data, *next_elm;
     SevCommonState *sev_common = SEV_COMMON(cgs);
     SevCommonStateClass *klass = SEV_COMMON_GET_CLASS(cgs);
     X86ConfidentialGuestClass *x86_klass =
@@ -1780,6 +1781,11 @@ static int sev_common_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
 
     sev_common->state = SEV_STATE_UNINIT;
 
+    /* free existing launch update data if any */
+    QTAILQ_FOREACH_SAFE(data, &launch_update, next, next_elm) {
+        g_free(data);
+    }
+
     host_cpuid(0x8000001F, 0, NULL, &ebx, NULL, NULL);
     host_cbitpos = ebx & 0x3f;
 
@@ -1968,6 +1974,8 @@ static int sev_snp_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
 {
     MachineState *ms = MACHINE(qdev_get_machine());
     X86MachineState *x86ms = X86_MACHINE(ms);
+    SevCommonState *sev_common = SEV_COMMON(cgs);
+    SevSnpGuestState *sev_snp_guest = SEV_SNP_GUEST(sev_common);
 
     if (x86ms->smm == ON_OFF_AUTO_AUTO) {
         x86ms->smm = ON_OFF_AUTO_OFF;
@@ -1976,6 +1984,10 @@ static int sev_snp_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
         return -1;
     }
 
+    /* free existing kernel hashes data if any */
+    g_free(sev_snp_guest->kernel_hashes_data);
+    sev_snp_guest->kernel_hashes_data = NULL;
+
     return 0;
 }
 
Signature · 1 line
2.42.0
author
#28 lore raw

PATCH v4 20/31 i386/sev: add support for confidential guest reset

In reply to Ani Sinha

When the KVM VM file descriptor changes as a part of the confidential guest reset mechanism, it necessary to create a new confidential guest context and re-encrypt the VM memory. This happens for SEV-ES and SEV-SNP virtual machines as a part of SEV_LAUNCH_FINISH, SEV_SNP_LAUNCH_FINISH operations.

A new resettable interface for SEV module has been added. A new reset callback for the reset 'exit' state has been implemented to perform the above operations when the VM file descriptor has changed during VM reset.

Tracepoints has been added also for tracing purpose.

---
 target/i386/sev.c        | 58 ++++++++++++++++++++++++++++++++++++++++
 target/i386/trace-events |  1 +
 2 files changed, 59 insertions(+)
target/i386/sev.c +58 -0
index b3893e431c..549e624176 100644
@@ -30,8 +30,10 @@
 #include "system/kvm.h"
 #include "kvm/kvm_i386.h"
 #include "sev.h"
+#include "system/cpus.h"
 #include "system/system.h"
 #include "system/runstate.h"
+#include "system/reset.h"
 #include "trace.h"
 #include "migration/blocker.h"
 #include "qom/object.h"
@@ -86,6 +88,10 @@ typedef struct QEMU_PACKED PaddedSevHashTable {
     uint8_t padding[ROUND_UP(sizeof(SevHashTable), 16) - sizeof(SevHashTable)];
 } PaddedSevHashTable;
 
+static void sev_handle_reset(Object *obj, ResetType type);
+
+SevKernelLoaderContext sev_load_ctx = {};
+
 QEMU_BUILD_BUG_ON(sizeof(PaddedSevHashTable) % 16 != 0);
 
 #define SEV_INFO_BLOCK_GUID     "00f771de-1a7e-4fcb-890e-68c77e2fb44e"
@@ -129,6 +135,7 @@ struct SevCommonState {
     uint8_t build_id;
     int sev_fd;
     SevState state;
+    ResettableState reset_state;
 
     QTAILQ_HEAD(, SevLaunchVmsa) launch_vmsa;
 };
@@ -1666,6 +1673,11 @@ sev_vm_state_change(void *opaque, bool running, RunState state)
             error_setg(&sev_mig_blocker,
                        "SEV: Migration is not implemented");
             migrate_add_blocker(&sev_mig_blocker, &error_fatal);
+            /*
+             * mark SEV guest as resettable so that we can reinitialize
+             * SEV upon reset.
+             */
+            qemu_register_resettable(OBJECT(sev_common));
         }
     }
 }
@@ -1991,6 +2003,41 @@ static int sev_snp_kvm_init(ConfidentialGuestSupport *cgs, Error **errp)
     return 0;
 }
 
+/*
+ * handle sev vm reset
+ */
+static void sev_handle_reset(Object *obj, ResetType type)
+{
+    SevCommonState *sev_common = SEV_COMMON(MACHINE(qdev_get_machine())->cgs);
+    SevCommonStateClass *klass = SEV_COMMON_GET_CLASS(sev_common);
+
+    if (!sev_common) {
+        return;
+    }
+
+    if (!runstate_is_running()) {
+        return;
+    }
+
+    sev_add_kernel_loader_hashes(&sev_load_ctx, &error_fatal);
+    if (sev_es_enabled() && !sev_snp_enabled()) {
+        sev_launch_get_measure(NULL, NULL);
+    }
+    if (!sev_check_state(sev_common, SEV_STATE_RUNNING)) {
+        /* this calls sev_snp_launch_finish() etc */
+        klass->launch_finish(sev_common);
+    }
+
+    trace_sev_handle_reset();
+    return;
+}
+
+static ResettableState *sev_reset_state(Object *obj)
+{
+    SevCommonState *sev_common = SEV_COMMON(obj);
+    return &sev_common->reset_state;
+}
+
 int
 sev_encrypt_flash(hwaddr gpa, uint8_t *ptr, uint64_t len, Error **errp)
 {
@@ -2469,6 +2516,8 @@ bool sev_add_kernel_loader_hashes(SevKernelLoaderContext *ctx, Error **errp)
         return false;
     }
 
+    /* save the context here so that it can be re-used when vm is reset */
+    memcpy(&sev_load_ctx, ctx, sizeof(*ctx));
     return klass->build_kernel_loader_hashes(sev_common, area, ctx, errp);
 }
 
@@ -2729,8 +2778,16 @@ static void
 sev_common_class_init(ObjectClass *oc, const void *data)
 {
     ConfidentialGuestSupportClass *klass = CONFIDENTIAL_GUEST_SUPPORT_CLASS(oc);
+    ResettableClass *rc = RESETTABLE_CLASS(oc);
 
     klass->kvm_init = sev_common_kvm_init;
+    /*
+     * the exit phase makes sure sev handles reset after all legacy resets
+     * have taken place (in the hold phase) and IGVM has also properly
+     * set up the boot state.
+     */
+    rc->phases.exit = sev_handle_reset;
+    rc->get_state = sev_reset_state;
 
     object_class_property_add_str(oc, "sev-device",
                                   sev_common_get_sev_device,
@@ -2780,6 +2837,7 @@ static const TypeInfo sev_common_info = {
     .abstract = true,
     .interfaces = (const InterfaceInfo[]) {
         { TYPE_USER_CREATABLE },
+        { TYPE_RESETTABLE_INTERFACE },
         { }
     }
 };
target/i386/trace-events +1 -0
index 51301673f0..b320f655ee 100644
@@ -14,3 +14,4 @@ kvm_sev_attestation_report(const char *mnonce, const char *data) "mnonce %s data
 kvm_sev_snp_launch_start(uint64_t policy, char *gosvw) "policy 0x%" PRIx64 " gosvw %s"
 kvm_sev_snp_launch_update(uint64_t src, uint64_t gpa, uint64_t len, const char *type) "src 0x%" PRIx64 " gpa 0x%" PRIx64 " len 0x%" PRIx64 " (%s page)"
 kvm_sev_snp_launch_finish(char *id_block, char *id_auth, char *host_data) "id_block %s id_auth %s host_data %s"
+sev_handle_reset(void) ""
Signature · 1 line
2.42.0
author
#29 lore raw

PATCH v4 21/31 hw/vfio: generate new file fd for pseudo device and rebind existing descriptors

In reply to Ani Sinha

Normally the vfio pseudo device file descriptor lives for the life of the VM. However, when the kvm VM file descriptor changes, a new file descriptor for the pseudo device needs to be generated against the new kvm VM descriptor. Other existing vfio descriptors needs to be reattached to the new pseudo device descriptor. This change performs the above steps.

This chnage is untested.

---
 hw/vfio/helpers.c | 94 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 94 insertions(+)
hw/vfio/helpers.c +94 -0
index f68f8165d0..2ff189a53b 100644
@@ -116,6 +116,91 @@ bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
  * we'll re-use it should another vfio device be attached before then.
  */
 int vfio_kvm_device_fd = -1;
+
+/*
+ * Confidential virtual machines:
+ * During reset of confidential vms, the kvm vm file descriptor changes.
+ * In that case, the old vfio kvm file descriptor is
+ * closed and a new descriptor is created agaist the new kvm vm file
+ * descriptor.
+ */
+
+typedef struct KVMVfioFileFd {
+    int fd;
+    QLIST_ENTRY(KVMVfioFileFd) node;
+} KVMVfioFileFd;
+
+static QLIST_HEAD(, KVMVfioFileFd) kvm_vfio_file_fds =
+    QLIST_HEAD_INITIALIZER(kvm_vfio_file_fds);
+
+static void insert_fd_to_list(int fd)
+{
+    KVMVfioFileFd *file_fd;
+    file_fd = g_malloc0(sizeof(*file_fd));
+    file_fd->fd = fd;
+    QLIST_INSERT_HEAD(&kvm_vfio_file_fds, file_fd, node);
+    return;
+}
+
+static void remove_fd_from_list(int fd)
+{
+    KVMVfioFileFd *file_fd, *next;
+
+    QLIST_FOREACH_SAFE(file_fd, &kvm_vfio_file_fds, node, next) {
+        if (file_fd->fd == fd) {
+            QLIST_REMOVE(file_fd, node);
+            g_free(file_fd);
+            break;
+        }
+    }
+    return;
+}
+
+static int kvm_vfio_filefd_rebind(NotifierWithReturn *notifier, void *data,
+                                  Error **errp)
+{
+    KVMVfioFileFd *file_fd;
+    int ret = 0;
+    struct kvm_device_attr attr = {
+        .group = KVM_DEV_VFIO_FILE,
+        .attr = KVM_DEV_VFIO_FILE_ADD,
+    };
+    struct kvm_create_device cd = {
+        .type = KVM_DEV_TYPE_VFIO,
+    };
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
+        error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
+        return -errno;
+    }
+
+    if (vfio_kvm_device_fd != -1) {
+        close(vfio_kvm_device_fd);
+    }
+
+    vfio_kvm_device_fd = cd.fd;
+
+    QLIST_FOREACH(file_fd, &kvm_vfio_file_fds, node) {
+        attr.addr = (uint64_t)(unsigned long)&file_fd->fd;
+        if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
+            error_setg_errno(errp, errno,
+                             "Failed to add fd %d to KVM VFIO device",
+                             file_fd->fd);
+            ret = -errno;
+        }
+    }
+    return ret;
+}
+
+static struct NotifierWithReturn kvm_vfio_vmfd_change_notifier = {
+    .notify = kvm_vfio_filefd_rebind,
+};
+
 #endif
 
 void vfio_kvm_device_close(void)
@@ -153,6 +238,11 @@ int vfio_kvm_device_add_fd(int fd, Error **errp)
         }
 
         vfio_kvm_device_fd = cd.fd;
+        /*
+         * If the vm file descriptor changes, add a notifier so that we can
+         * re-create the vfio_kvm_device_fd.
+         */
+        kvm_vmfd_add_change_notifier(&kvm_vfio_vmfd_change_notifier);
     }
 
     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
@@ -160,6 +250,8 @@ int vfio_kvm_device_add_fd(int fd, Error **errp)
                          fd);
         return -errno;
     }
+
+    insert_fd_to_list(fd);
 #endif
     return 0;
 }
@@ -183,6 +275,8 @@ int vfio_kvm_device_del_fd(int fd, Error **errp)
                          "Failed to remove fd %d from KVM VFIO device", fd);
         return -errno;
     }
+
+    remove_fd_from_list(fd);
 #endif
     return 0;
 }
Signature · 1 line
2.42.0
#30 lore raw

Re: [PATCH v4 21/31] hw/vfio: generate new file fd for pseudo device and rebind existing descriptors

Quoted 7 lines · Ani Sinha

Normally the vfio pseudo device file descriptor lives for the life of the VM. However, when the kvm VM file descriptor changes, a new file descriptor for the pseudo device needs to be generated against the new kvm VM descriptor. Other existing vfio descriptors needs to be reattached to the new pseudo device descriptor. This change performs the above steps.

This chnage is untested.

It is now.

Quoted 14 lines
---
  hw/vfio/helpers.c | 94 +++++++++++++++++++++++++++++++++++++++++++++++
  1 file changed, 94 insertions(+)
hw/vfio/helpers.c +0 -0
index f68f8165d0..2ff189a53b 100644
@@ -116,6 +116,91 @@ bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
   * we'll re-use it should another vfio device be attached before then.
   */
  int vfio_kvm_device_fd = -1;

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 128

Quoted 10 lines
hw/vfio/helpers.c +10 -0
+
+/*
+ * Confidential virtual machines:
+ * During reset of confidential vms, the kvm vm file descriptor changes.
+ * In that case, the old vfio kvm file descriptor is
+ * closed and a new descriptor is created agaist the new kvm vm file
+ * descriptor.
+ */
+
+typedef struct KVMVfioFileFd {

The KVM prefix should be reserved to the KVM subsystem. Let's use VFIO (Yeah, it's not CalmelCase)

So, VFIODeviceFd should be correct enough for the type and 'vfio_device_fd_' prefix for the routines below?

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 133

Quoted 5 lines
hw/vfio/helpers.c +5 -0
+    int fd;
+    QLIST_ENTRY(KVMVfioFileFd) node;
+} KVMVfioFileFd;
+
+static QLIST_HEAD(, KVMVfioFileFd) kvm_vfio_file_fds =

vfio_device_fds.

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 136

Quoted 3 lines
hw/vfio/helpers.c +3 -0
+    QLIST_HEAD_INITIALIZER(kvm_vfio_file_fds);
+
+static void insert_fd_to_list(int fd)

This function name is too generic. How about vfio_device_fd_list_add()

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 142

Quoted 6 lines
hw/vfio/helpers.c +6 -0
+{
+    KVMVfioFileFd *file_fd;
+    file_fd = g_malloc0(sizeof(*file_fd));
+    file_fd->fd = fd;
+    QLIST_INSERT_HEAD(&kvm_vfio_file_fds, file_fd, node);
+    return;

Please drop the return.

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 145

Quoted 3 lines
hw/vfio/helpers.c +3 -0
+}
+
+static void remove_fd_from_list(int fd)

and vfio_device_fd_list_remove() ?

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 156

Quoted 11 lines
hw/vfio/helpers.c +11 -0
+{
+    KVMVfioFileFd *file_fd, *next;
+
+    QLIST_FOREACH_SAFE(file_fd, &kvm_vfio_file_fds, node, next) {
+        if (file_fd->fd == fd) {
+            QLIST_REMOVE(file_fd, node);
+            g_free(file_fd);
+            break;
+        }
+    }
+    return;

Please remove.

hw/vfio/helpers.c bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, · line 159

Quoted 3 lines
hw/vfio/helpers.c +3 -0
+}
+
+static int kvm_vfio_filefd_rebind(NotifierWithReturn *notifier, void *data,

vfio_device_fd_rebind() for this one and there are some more below.

Thanks,

C.

Quoted 77 lines
hw/vfio/helpers.c +53 -0
+                                  Error **errp)
+{
+    KVMVfioFileFd *file_fd;
+    int ret = 0;
+    struct kvm_device_attr attr = {
+        .group = KVM_DEV_VFIO_FILE,
+        .attr = KVM_DEV_VFIO_FILE_ADD,
+    };
+    struct kvm_create_device cd = {
+        .type = KVM_DEV_TYPE_VFIO,
+    };
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
+        error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
+        return -errno;
+    }
+
+    if (vfio_kvm_device_fd != -1) {
+        close(vfio_kvm_device_fd);
+    }
+
+    vfio_kvm_device_fd = cd.fd;
+
+    QLIST_FOREACH(file_fd, &kvm_vfio_file_fds, node) {
+        attr.addr = (uint64_t)(unsigned long)&file_fd->fd;
+        if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
+            error_setg_errno(errp, errno,
+                             "Failed to add fd %d to KVM VFIO device",
+                             file_fd->fd);
+            ret = -errno;
+        }
+    }
+    return ret;
+}
+
+static struct NotifierWithReturn kvm_vfio_vmfd_change_notifier = {
+    .notify = kvm_vfio_filefd_rebind,
+};
+
  #endif
 
  void vfio_kvm_device_close(void)
@@ -153,6 +238,11 @@ int vfio_kvm_device_add_fd(int fd, Error **errp)
          }
 
          vfio_kvm_device_fd = cd.fd;
+        /*
+         * If the vm file descriptor changes, add a notifier so that we can
+         * re-create the vfio_kvm_device_fd.
+         */
+        kvm_vmfd_add_change_notifier(&kvm_vfio_vmfd_change_notifier);
      }
 
      if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
@@ -160,6 +250,8 @@ int vfio_kvm_device_add_fd(int fd, Error **errp)
                           fd);
          return -errno;
      }
+
+    insert_fd_to_list(fd);
  #endif
      return 0;
  }
@@ -183,6 +275,8 @@ int vfio_kvm_device_del_fd(int fd, Error **errp)
                           "Failed to remove fd %d from KVM VFIO device", fd);
          return -errno;
      }
+
+    remove_fd_from_list(fd);
  #endif
      return 0;
  }
author
#31 lore raw

PATCH v4 22/31 kvm/i8254: refactor pit initialization into a helper

In reply to Ani Sinha

The initialization code will be used again by VM file descriptor change notifier callback in a subsequent change. So refactor common code into a new helper function.

---
 hw/i386/kvm/i8254.c | 68 +++++++++++++++++++++++++--------------------
 1 file changed, 38 insertions(+), 30 deletions(-)
hw/i386/kvm/i8254.c +38 -30
index 81e742f866..255047458a 100644
@@ -60,6 +60,43 @@ struct KVMPITClass {
     DeviceRealize parent_realize;
 };
 
+static void do_pit_initialize(KVMPITState *s, Error **errp)
+{
+    struct kvm_pit_config config = {
+        .flags = 0,
+    };
+    int ret;
+
+    ret = kvm_vm_ioctl(kvm_state, KVM_CREATE_PIT2, &config);
+    if (ret < 0) {
+        error_setg(errp, "Create kernel PIC irqchip failed: %s",
+                   strerror(-ret));
+        return;
+    }
+    switch (s->lost_tick_policy) {
+    case LOST_TICK_POLICY_DELAY:
+        break; /* enabled by default */
+    case LOST_TICK_POLICY_DISCARD:
+        if (kvm_check_extension(kvm_state, KVM_CAP_REINJECT_CONTROL)) {
+            struct kvm_reinject_control control = { .pit_reinject = 0 };
+
+            ret = kvm_vm_ioctl(kvm_state, KVM_REINJECT_CONTROL, &control);
+            if (ret < 0) {
+                error_setg(errp,
+                           "Can't disable in-kernel PIT reinjection: %s",
+                           strerror(-ret));
+                return;
+            }
+        }
+        break;
+    default:
+        error_setg(errp, "Lost tick policy not supported.");
+        return;
+    }
+
+    return;
+}
+
 static void kvm_pit_update_clock_offset(KVMPITState *s)
 {
     int64_t offset, clock_offset;
@@ -241,42 +278,13 @@ static void kvm_pit_realizefn(DeviceState *dev, Error **errp)
     PITCommonState *pit = PIT_COMMON(dev);
     KVMPITClass *kpc = KVM_PIT_GET_CLASS(dev);
     KVMPITState *s = KVM_PIT(pit);
-    struct kvm_pit_config config = {
-        .flags = 0,
-    };
-    int ret;
 
     if (!kvm_check_extension(kvm_state, KVM_CAP_PIT_STATE2) ||
         !kvm_check_extension(kvm_state, KVM_CAP_PIT2)) {
         error_setg(errp, "In-kernel PIT not available");
     }
 
-    ret = kvm_vm_ioctl(kvm_state, KVM_CREATE_PIT2, &config);
-    if (ret < 0) {
-        error_setg(errp, "Create kernel PIC irqchip failed: %s",
-                   strerror(-ret));
-        return;
-    }
-    switch (s->lost_tick_policy) {
-    case LOST_TICK_POLICY_DELAY:
-        break; /* enabled by default */
-    case LOST_TICK_POLICY_DISCARD:
-        if (kvm_check_extension(kvm_state, KVM_CAP_REINJECT_CONTROL)) {
-            struct kvm_reinject_control control = { .pit_reinject = 0 };
-
-            ret = kvm_vm_ioctl(kvm_state, KVM_REINJECT_CONTROL, &control);
-            if (ret < 0) {
-                error_setg(errp,
-                           "Can't disable in-kernel PIT reinjection: %s",
-                           strerror(-ret));
-                return;
-            }
-        }
-        break;
-    default:
-        error_setg(errp, "Lost tick policy not supported.");
-        return;
-    }
+    do_pit_initialize(s, errp);
 
     memory_region_init_io(&pit->ioports, OBJECT(dev), NULL, NULL, "kvm-pit", 4);
 
Signature · 1 line
2.42.0
author
#32 lore raw

PATCH v4 23/31 kvm/i8254: add support for confidential guest reset

In reply to Ani Sinha

A confidential guest reset involves closing the old virtual machine KVM file descriptor and opening a new one. Since its a new KVM fd, PIT needs to be re-initialized again. This is done with the help of a notifier which is invoked upon KVM vm file descriptor change during the confidential guest reset process.

---
 hw/i386/kvm/i8254.c      | 23 +++++++++++++++++++++++
 hw/i386/kvm/trace-events |  1 +
 2 files changed, 24 insertions(+)
hw/i386/kvm/i8254.c +23 -0
index 255047458a..70e8fd83cd 100644
@@ -35,6 +35,7 @@
 #include "hw/core/qdev-properties-system.h"
 #include "system/kvm.h"
 #include "target/i386/kvm/kvm_i386.h"
+#include "trace.h"
 #include "qom/object.h"
 
 #define KVM_PIT_REINJECT_BIT 0
@@ -52,6 +53,8 @@ struct KVMPITState {
     LostTickPolicy lost_tick_policy;
     bool vm_stopped;
     int64_t kernel_clock_offset;
+
+    NotifierWithReturn kvmpit_vmfd_change_notifier;
 };
 
 struct KVMPITClass {
@@ -203,6 +206,23 @@ static void kvm_pit_put(PITCommonState *pit)
     }
 }
 
+static int kvmpit_post_vmfd_change(NotifierWithReturn *notifier,
+                                   void *data, Error** errp)
+{
+    KVMPITState *s = container_of(notifier, KVMPITState,
+                                  kvmpit_vmfd_change_notifier);
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    do_pit_initialize(s, errp);
+
+    trace_kvmpit_post_vmfd_change();
+    return 0;
+}
+
 static void kvm_pit_set_gate(PITCommonState *s, PITChannelState *sc, int val)
 {
     kvm_pit_get(s);
@@ -292,6 +312,9 @@ static void kvm_pit_realizefn(DeviceState *dev, Error **errp)
 
     qemu_add_vm_change_state_handler(kvm_pit_vm_state_change, s);
 
+    s->kvmpit_vmfd_change_notifier.notify = kvmpit_post_vmfd_change;
+    kvm_vmfd_add_change_notifier(&s->kvmpit_vmfd_change_notifier);
+
     kpc->parent_realize(dev, errp);
 }
 
hw/i386/kvm/trace-events +1 -0
index 67bf7f174e..33680ff82b 100644
@@ -20,3 +20,4 @@ xenstore_reset_watches(void) ""
 xenstore_watch_event(const char *path, const char *token) "path %s token %s"
 xen_primary_console_create(void) ""
 xen_primary_console_reset(int port) "port %u"
+kvmpit_post_vmfd_change(void) ""
Signature · 1 line
2.42.0
author
#33 lore raw

PATCH v4 24/31 kvm/hyperv: add synic feature to CPU only if its not enabled

In reply to Ani Sinha

We need to make sure that synic CPU feature is not already enabled. If it is, trying to enable it again will result in the following assertion:

Unexpected error in object_property_try_add() at ../qom/object.c:1268:
qemu-system-x86_64: attempt to add duplicate property 'synic' to object (type 'host-x86_64-cpu')

So enable synic only if its not enabled already.

---
 target/i386/kvm/kvm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
target/i386/kvm/kvm.c +1 -1
index eec2f27a0f..a6b773db34 100644
@@ -1752,7 +1752,7 @@ static int hyperv_init_vcpu(X86CPU *cpu)
             return ret;
         }
 
-        if (!cpu->hyperv_synic_kvm_only) {
+        if (!cpu->hyperv_synic_kvm_only && !hyperv_is_synic_enabled()) {
             ret = hyperv_x86_synic_add(cpu);
             if (ret < 0) {
                 error_report("failed to create HyperV SynIC: %s",
Signature · 1 line
2.42.0
author
#34 lore raw

PATCH v4 25/31 hw/hyperv/vmbus: add support for confidential guest reset

In reply to Ani Sinha

On confidential guests when the KVM virtual machine file descriptor changes as a part of the reset process, event file descriptors needs to be reassociated with the new KVM VM file descriptor. This is achieved with the help of a callback handler that gets called when KVM VM file descriptor changes during the confidential guest reset process.

This patch is tested on non-confidential platform only.

---
 hw/hyperv/trace-events |  1 +
 hw/hyperv/vmbus.c      | 37 +++++++++++++++++++++++++++++++++++++
 2 files changed, 38 insertions(+)
hw/hyperv/trace-events +1 -0
index 7963c215b1..d8c96f18e9 100644
@@ -16,6 +16,7 @@ vmbus_gpadl_torndown(uint32_t gpadl_id) "gpadl #%d"
 vmbus_open_channel(uint32_t chan_id, uint32_t gpadl_id, uint32_t target_vp) "channel #%d gpadl #%d target vp %d"
 vmbus_channel_open(uint32_t chan_id, uint32_t status) "channel #%d status %d"
 vmbus_close_channel(uint32_t chan_id) "channel #%d"
+vmbus_handle_vmfd_change(void) ""
 
 # hv-balloon
 hv_balloon_state_change(const char *tostr) "-> %s"
hw/hyperv/vmbus.c +37 -0
index c5bab5d245..64abe4c4c1 100644
@@ -20,6 +20,7 @@
 #include "hw/hyperv/vmbus-bridge.h"
 #include "hw/core/sysbus.h"
 #include "exec/cpu-common.h"
+#include "system/kvm.h"
 #include "exec/target_page.h"
 #include "trace.h"
 
@@ -248,6 +249,12 @@ struct VMBus {
      * interrupt page
      */
     EventNotifier notifier;
+
+    /*
+     * Notifier to inform when vmfd is changed as a part of confidential guest
+     * reset mechanism.
+     */
+    NotifierWithReturn vmbus_vmfd_change_notifier;
 };
 
 static bool gpadl_full(VMBusGpadl *gpadl)
@@ -2347,6 +2354,33 @@ static void vmbus_dev_unrealize(DeviceState *dev)
     free_channels(vdev);
 }
 
+/*
+ * If the KVM fd changes because of VM reset in confidential guests,
+ * reassociate event fd with the new KVM fd.
+ */
+static int vmbus_handle_vmfd_change(NotifierWithReturn *notifier,
+                                    void *data, Error** errp)
+{
+    VMBus *vmbus = container_of(notifier, VMBus,
+                                vmbus_vmfd_change_notifier);
+    int ret = 0;
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    ret = hyperv_set_event_flag_handler(VMBUS_EVENT_CONNECTION_ID,
+                                            &vmbus->notifier);
+    /* if we are only using userland event handler, it may already exist */
+    if (ret != 0 && ret != -EEXIST) {
+        error_setg(errp, "hyperv set event handler failed with %d", ret);
+    }
+
+    trace_vmbus_handle_vmfd_change();
+    return ret;
+}
+
 static const Property vmbus_dev_props[] = {
     DEFINE_PROP_UUID("instanceid", VMBusDevice, instanceid),
 };
@@ -2429,6 +2463,9 @@ static void vmbus_realize(BusState *bus, Error **errp)
         goto clear_event_notifier;
     }
 
+    vmbus->vmbus_vmfd_change_notifier.notify = vmbus_handle_vmfd_change;
+    kvm_vmfd_add_change_notifier(&vmbus->vmbus_vmfd_change_notifier);
+
     return;
 
 clear_event_notifier:
Signature · 1 line
2.42.0
author
#35 lore raw

PATCH v4 26/31 kvm/xen-emu: re-initialize capabilities during confidential guest reset

In reply to Ani Sinha

On confidential guests KVM virtual machine file descriptor changes as a part of the guest reset process. Xen capabilities needs to be re-initialized in KVM against the new file descriptor.

---
 target/i386/kvm/xen-emu.c | 50 +++++++++++++++++++++++++++++++++++++--
 1 file changed, 48 insertions(+), 2 deletions(-)
target/i386/kvm/xen-emu.c +48 -2
index 52de019834..69527145eb 100644
@@ -44,9 +44,12 @@
 
 #include "xen-compat.h"
 
+NotifierWithReturn xen_vmfd_change_notifier;
+static bool hyperv_enabled;
 static void xen_vcpu_singleshot_timer_event(void *opaque);
 static void xen_vcpu_periodic_timer_event(void *opaque);
 static int vcpuop_stop_singleshot_timer(CPUState *cs);
+static int do_initialize_xen_caps(KVMState *s, uint32_t hypercall_msr);
 
 #ifdef TARGET_X86_64
 #define hypercall_compat32(longmode) (!(longmode))
@@ -54,6 +57,30 @@ static int vcpuop_stop_singleshot_timer(CPUState *cs);
 #define hypercall_compat32(longmode) (false)
 #endif
 
+static int xen_handle_vmfd_change(NotifierWithReturn *n,
+                                  void *data, Error** errp)
+{
+    int ret;
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    ret = do_initialize_xen_caps(kvm_state, XEN_HYPERCALL_MSR);
+    if (ret < 0) {
+        return ret;
+    }
+
+    if (hyperv_enabled) {
+        ret = do_initialize_xen_caps(kvm_state, XEN_HYPERCALL_MSR_HYPERV);
+        if (ret < 0) {
+            return ret;
+        }
+    }
+    return 0;
+}
+
 static bool kvm_gva_to_gpa(CPUState *cs, uint64_t gva, uint64_t *gpa,
                            size_t *len, bool is_write)
 {
@@ -111,15 +138,16 @@ static inline int kvm_copy_to_gva(CPUState *cs, uint64_t gva, void *buf,
     return kvm_gva_rw(cs, gva, buf, sz, true);
 }
 
-int kvm_xen_init(KVMState *s, uint32_t hypercall_msr)
+static int do_initialize_xen_caps(KVMState *s, uint32_t hypercall_msr)
 {
+    int xen_caps, ret;
     const int required_caps = KVM_XEN_HVM_CONFIG_HYPERCALL_MSR |
         KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL | KVM_XEN_HVM_CONFIG_SHARED_INFO;
+
     struct kvm_xen_hvm_config cfg = {
         .msr = hypercall_msr,
         .flags = KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL,
     };
-    int xen_caps, ret;
 
     xen_caps = kvm_check_extension(s, KVM_CAP_XEN_HVM);
     if (required_caps & ~xen_caps) {
@@ -143,6 +171,21 @@ int kvm_xen_init(KVMState *s, uint32_t hypercall_msr)
                      strerror(-ret));
         return ret;
     }
+    return xen_caps;
+}
+
+int kvm_xen_init(KVMState *s, uint32_t hypercall_msr)
+{
+    int xen_caps;
+
+    xen_caps = do_initialize_xen_caps(s, hypercall_msr);
+    if (xen_caps < 0) {
+        return xen_caps;
+    }
+
+    if (!hyperv_enabled && (hypercall_msr == XEN_HYPERCALL_MSR_HYPERV)) {
+        hyperv_enabled = true;
+    }
 
     /* If called a second time, don't repeat the rest of the setup. */
     if (s->xen_caps) {
@@ -185,6 +228,9 @@ int kvm_xen_init(KVMState *s, uint32_t hypercall_msr)
     xen_primary_console_reset();
     xen_xenstore_reset();
 
+    xen_vmfd_change_notifier.notify = xen_handle_vmfd_change;
+    kvm_vmfd_add_change_notifier(&xen_vmfd_change_notifier);
+
     return 0;
 }
 
Signature · 1 line
2.42.0
author
#36 lore raw

PATCH v4 27/31 ppc/openpic: create a new openpic device and reattach mem region on coco reset

In reply to Ani Sinha

For confidential guests during the reset process, the old KVM VM file descriptor is closed and a new one is created. When a new file descriptor is created, a new openpic device needs to be created against this new KVM VM file descriptor as well. Additionally, existing memory region needs to be reattached to this new openpic device and proper CPU attributes set associating new file descriptor. This change makes this happen with the help of a callback handler that gets called when the KVM VM file descriptor changes as a part of the confidential guest reset process.

---
 hw/intc/openpic_kvm.c | 112 +++++++++++++++++++++++++++++++++---------
 1 file changed, 88 insertions(+), 24 deletions(-)
hw/intc/openpic_kvm.c +88 -24
index fbf0bdbe07..b099da20eb 100644
@@ -49,6 +49,7 @@ struct KVMOpenPICState {
     uint32_t fd;
     uint32_t model;
     hwaddr mapped;
+    NotifierWithReturn vmfd_change_notifier;
 };
 
 static void kvm_openpic_set_irq(void *opaque, int n_IRQ, int level)
@@ -114,6 +115,88 @@ static const MemoryRegionOps kvm_openpic_mem_ops = {
     },
 };
 
+static int kvm_openpic_setup(KVMOpenPICState *opp, Error **errp)
+{
+    int kvm_openpic_model;
+    struct kvm_create_device cd = {0};
+    KVMState *s = kvm_state;
+    int ret;
+
+    switch (opp->model) {
+    case OPENPIC_MODEL_FSL_MPIC_20:
+        kvm_openpic_model = KVM_DEV_TYPE_FSL_MPIC_20;
+        break;
+
+    case OPENPIC_MODEL_FSL_MPIC_42:
+        kvm_openpic_model = KVM_DEV_TYPE_FSL_MPIC_42;
+        break;
+
+    default:
+        error_setg(errp, "Unsupported OpenPIC model %" PRIu32, opp->model);
+        return -1;
+    }
+
+    cd.type = kvm_openpic_model;
+    ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &cd);
+    if (ret < 0) {
+        error_setg(errp, "Can't create device %d: %s",
+                   cd.type, strerror(errno));
+        return -1;
+    }
+    opp->fd = cd.fd;
+
+    return 0;
+}
+
+static int kvm_openpic_handle_vmfd_change(NotifierWithReturn *notifier,
+                                          void *data, Error **errp)
+{
+    KVMOpenPICState *opp = container_of(notifier, KVMOpenPICState,
+                                        vmfd_change_notifier);
+    uint64_t reg_base;
+    struct kvm_device_attr attr;
+    CPUState *cs;
+    int ret;
+
+    /* we are not interested in pre vmfd change notification */
+    if (((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+
+    /* close the old descriptor */
+    close(opp->fd);
+
+    if (kvm_openpic_setup(opp, errp) < 0) {
+        return -1;
+    }
+
+    if (!opp->mapped) {
+        return 0;
+    }
+
+    reg_base = opp->mapped;
+    attr.group = KVM_DEV_MPIC_GRP_MISC;
+    attr.attr = KVM_DEV_MPIC_BASE_ADDR;
+    attr.addr = (uint64_t)(unsigned long)&reg_base;
+
+    ret = ioctl(opp->fd, KVM_SET_DEVICE_ATTR, &attr);
+    if (ret < 0) {
+        error_setg(errp, "%s: %s %" PRIx64, __func__,
+                   strerror(errno), reg_base);
+        return -1;
+    }
+
+    CPU_FOREACH(cs) {
+        ret = kvm_vcpu_enable_cap(cs, KVM_CAP_IRQ_MPIC, 0, opp->fd,
+                                   kvm_arch_vcpu_id(cs));
+        if (ret < 0) {
+            return ret;
+        }
+    }
+
+    return 0;
+}
+
 static void kvm_openpic_region_add(MemoryListener *listener,
                                    MemoryRegionSection *section)
 {
@@ -197,36 +280,14 @@ static void kvm_openpic_realize(DeviceState *dev, Error **errp)
     SysBusDevice *d = SYS_BUS_DEVICE(dev);
     KVMOpenPICState *opp = KVM_OPENPIC(dev);
     KVMState *s = kvm_state;
-    int kvm_openpic_model;
-    struct kvm_create_device cd = {0};
-    int ret, i;
+    int i;
 
     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
         error_setg(errp, "Kernel is lacking Device Control API");
         return;
     }
 
-    switch (opp->model) {
-    case OPENPIC_MODEL_FSL_MPIC_20:
-        kvm_openpic_model = KVM_DEV_TYPE_FSL_MPIC_20;
-        break;
-
-    case OPENPIC_MODEL_FSL_MPIC_42:
-        kvm_openpic_model = KVM_DEV_TYPE_FSL_MPIC_42;
-        break;
-
-    default:
-        error_setg(errp, "Unsupported OpenPIC model %" PRIu32, opp->model);
-        return;
-    }
-
-    cd.type = kvm_openpic_model;
-    ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &cd);
-    if (ret < 0) {
-        error_setg_errno(errp, errno, "Can't create device %d", cd.type);
-        return;
-    }
-    opp->fd = cd.fd;
+    kvm_openpic_setup(opp, errp);
 
     sysbus_init_mmio(d, &opp->mem);
     qdev_init_gpio_in(dev, kvm_openpic_set_irq, OPENPIC_MAX_IRQ);
@@ -235,6 +296,9 @@ static void kvm_openpic_realize(DeviceState *dev, Error **errp)
     opp->mem_listener.region_del = kvm_openpic_region_del;
     opp->mem_listener.name = "openpic-kvm";
     memory_listener_register(&opp->mem_listener, &address_space_memory);
+    opp->vmfd_change_notifier.notify =
+        kvm_openpic_handle_vmfd_change;
+    kvm_vmfd_add_change_notifier(&opp->vmfd_change_notifier);
 
     /* indicate pic capabilities */
     msi_nonbroken = true;
Signature · 1 line
2.42.0
author
#37 lore raw

PATCH v4 28/31 kvm/vcpu: add notifiers to inform vcpu file descriptor change

In reply to Ani Sinha

When new vcpu file descriptors are created and bound to the new kvm file descriptor as a part of the confidential guest reset mechanism, various subsystems needs to know about it. This change adds notifiers so that various subsystems can take appropriate actions when vcpu fds change by registering their handlers to this notifier.
Subsequent changes will register specific handlers to this notifier.

---
 accel/kvm/kvm-all.c    | 26 ++++++++++++++++++++++++++
 accel/stubs/kvm-stub.c | 10 ++++++++++
 include/system/kvm.h   | 17 +++++++++++++++++
 3 files changed, 53 insertions(+)
accel/kvm/kvm-all.c +26 -0
index 5e81bf8ad2..4d4d361708 100644
@@ -127,6 +127,9 @@ static NotifierList kvm_irqchip_change_notifiers =
 static NotifierWithReturnList register_vmfd_changed_notifiers =
     NOTIFIER_WITH_RETURN_LIST_INITIALIZER(register_vmfd_changed_notifiers);
 
+static NotifierWithReturnList register_vcpufd_changed_notifiers =
+    NOTIFIER_WITH_RETURN_LIST_INITIALIZER(register_vcpufd_changed_notifiers);
+
 static int map_kvm_run(KVMState *s, CPUState *cpu, Error **errp);
 static int map_kvm_dirty_gfns(KVMState *s, CPUState *cpu, Error **errp);
 static int vcpu_unmap_regions(KVMState *s, CPUState *cpu);
@@ -2314,6 +2317,22 @@ static int kvm_vmfd_change_notify(Error **errp)
                                             &vmfd_notifier, errp);
 }
 
+void kvm_vcpufd_add_change_notifier(NotifierWithReturn *n)
+{
+    notifier_with_return_list_add(&register_vcpufd_changed_notifiers, n);
+}
+
+void kvm_vcpufd_remove_change_notifier(NotifierWithReturn *n)
+{
+    notifier_with_return_remove(n);
+}
+
+static int kvm_vcpufd_change_notify(Error **errp)
+{
+    return notifier_with_return_list_notify(&register_vcpufd_changed_notifiers,
+                                            &vmfd_notifier, errp);
+}
+
 int kvm_irqchip_get_virq(KVMState *s)
 {
     int next_virq;
@@ -2838,6 +2857,13 @@ static int kvm_reset_vmfd(MachineState *ms)
     }
     assert(!err);
 
+    /* notify everyone that vcpu fd has changed. */
+    ret = kvm_vcpufd_change_notify(&err);
+    if (ret < 0) {
+        return ret;
+    }
+    assert(!err);
+
     /* these can be only called after ram_block_rebind() */
     memory_listener_register(&kml->listener, &address_space_memory);
     memory_listener_register(&kvm_io_listener, &address_space_io);
accel/stubs/kvm-stub.c +10 -0
index a6e8a6e16c..c4617caac6 100644
@@ -87,6 +87,16 @@ void kvm_vmfd_remove_change_notifier(NotifierWithReturn *n)
 {
 }
 
+void kvm_vcpufd_add_change_notifier(NotifierWithReturn *n)
+{
+    return;
+}
+
+void kvm_vcpufd_remove_change_notifier(NotifierWithReturn *n)
+{
+    return;
+}
+
 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
                                        EventNotifier *rn, int virq)
 {
include/system/kvm.h +17 -0
index fbe23608a1..4b0e1b4ab1 100644
@@ -590,4 +590,21 @@ void kvm_vmfd_add_change_notifier(NotifierWithReturn *n);
  */
 void kvm_vmfd_remove_change_notifier(NotifierWithReturn *n);
 
+/**
+ * kvm_vcpufd_add_change_notifier - register a notifier to get notified when
+ * a KVM vcpu file descriptors changes as a part of the confidential guest
+ * "reset" process. Various subsystems should use this mechanism to take
+ * actions such as re-issuing vcpu ioctls as a part of setting up vcpu
+ * features.
+ * @n: notifier with return value.
+ */
+void kvm_vcpufd_add_change_notifier(NotifierWithReturn *n);
+
+/**
+ * kvm_vcpufd_remove_change_notifier - de-register a notifer previously
+ * registered with kvm_vcpufd_add_change_notifier call.
+ * @n: notifier that was previously registered.
+ */
+void kvm_vcpufd_remove_change_notifier(NotifierWithReturn *n);
+
 #endif
Signature · 1 line
2.42.0
author
#38 lore raw

PATCH v4 29/31 kvm/clock: add support for confidential guest reset

In reply to Ani Sinha

Confidential guests change the KVM VM file descriptor upon reset and also create new VCPU file descriptors against the new KVM VM file descriptor. We need to save the clock state from kvm before KVM VM file descriptor changes and restore it after. Also after VCPU file descriptors changed, we must call KVM_KVMCLOCK_CTRL on the VCPU file descriptor to inform KVM that the VCPU is in paused state.

---
 hw/i386/kvm/clock.c | 59 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 59 insertions(+)
hw/i386/kvm/clock.c +59 -0
index aba6842a22..10d34254f0 100644
@@ -50,6 +50,9 @@ struct KVMClockState {
     /* whether the 'clock' value was obtained in a host with
      * reliable KVM_GET_CLOCK */
     bool clock_is_reliable;
+
+    NotifierWithReturn kvmclock_vcpufd_change_notifier;
+    NotifierWithReturn kvmclock_vmfd_change_notifier;
 };
 
 struct pvclock_vcpu_time_info {
@@ -63,6 +66,9 @@ struct pvclock_vcpu_time_info {
     uint8_t    pad[2];
 } __attribute__((__packed__)); /* 32 bytes */
 
+static int kvmclock_set_clock(NotifierWithReturn *notifier,
+                              void *data, Error** errp);
+
 static uint64_t kvmclock_current_nsec(KVMClockState *s)
 {
     CPUState *cpu = first_cpu;
@@ -219,6 +225,54 @@ static void kvmclock_vm_state_change(void *opaque, bool running,
     }
 }
 
+static int kvmclock_save_clock(NotifierWithReturn *notifier,
+                               void *data, Error** errp)
+{
+    if (!((VmfdChangeNotifier *)data)->pre) {
+        return 0;
+    }
+    KVMClockState *s = container_of(notifier, KVMClockState,
+                                    kvmclock_vmfd_change_notifier);
+    kvm_update_clock(s);
+    return 0;
+}
+
+static int kvmclock_set_clock(NotifierWithReturn *notifier,
+                              void *data, Error** errp)
+{
+    struct kvm_clock_data clock_data = {};
+    CPUState *cpu;
+    int ret;
+    KVMClockState *s = container_of(notifier, KVMClockState,
+                                    kvmclock_vcpufd_change_notifier);
+    int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
+
+    if (!s->clock_is_reliable) {
+        uint64_t pvclock_via_mem = kvmclock_current_nsec(s);
+        /* saved clock value before vmfd change is not reliable */
+        if (pvclock_via_mem) {
+            s->clock = pvclock_via_mem;
+        }
+    }
+
+    clock_data.clock = s->clock;
+    ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &clock_data);
+    if (ret < 0) {
+        fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(-ret));
+        abort();
+    }
+
+    if (!cap_clock_ctrl) {
+        return 0;
+    }
+    CPU_FOREACH(cpu) {
+        run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL);
+    }
+
+    return 0;
+}
+
+
 static void kvmclock_realize(DeviceState *dev, Error **errp)
 {
     KVMClockState *s = KVM_CLOCK(dev);
@@ -230,7 +284,12 @@ static void kvmclock_realize(DeviceState *dev, Error **errp)
 
     kvm_update_clock(s);
 
+    s->kvmclock_vcpufd_change_notifier.notify = kvmclock_set_clock;
+    s->kvmclock_vmfd_change_notifier.notify = kvmclock_save_clock;
+
     qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
+    kvm_vcpufd_add_change_notifier(&s->kvmclock_vcpufd_change_notifier);
+    kvm_vmfd_add_change_notifier(&s->kvmclock_vmfd_change_notifier);
 }
 
 static bool kvmclock_clock_is_reliable_needed(void *opaque)
Signature · 1 line
2.42.0
author
#39 lore raw

PATCH v4 30/31 hw/machine: introduce machine specific option 'x-change-vmfd-on-reset'

In reply to Ani Sinha

A new machine specific option 'x-change-vmfd-on-reset' is introduced for debugging and testing only (hence the 'x-' prefix). This option when enabled will force KVM VM file descriptor to be changed upon guest reset like in the case of confidential guests. This can be used to exercise the code changes that are specific for confidential guests on non-confidential guests as well (except changes that require hardware support for confidential guests).
A new functional test has been added in the next patch that uses this new parameter to test the VM file descriptor changes.

---
 hw/core/machine.c        | 22 ++++++++++++++++++++++
 include/hw/core/boards.h |  6 ++++++
 system/runstate.c        | 20 ++++++++++----------
 3 files changed, 38 insertions(+), 10 deletions(-)
hw/core/machine.c +22 -0
index 6411e68856..95d7650db9 100644
@@ -450,6 +450,21 @@ static void machine_set_dump_guest_core(Object *obj, bool value, Error **errp)
     ms->dump_guest_core = value;
 }
 
+static bool machine_get_new_accel_vmfd_on_reset(Object *obj, Error **errp)
+{
+    MachineState *ms = MACHINE(obj);
+
+    return ms->new_accel_vmfd_on_reset;
+}
+
+static void machine_set_new_accel_vmfd_on_reset(Object *obj,
+                                                bool value, Error **errp)
+{
+    MachineState *ms = MACHINE(obj);
+
+    ms->new_accel_vmfd_on_reset = value;
+}
+
 static bool machine_get_mem_merge(Object *obj, Error **errp)
 {
     MachineState *ms = MACHINE(obj);
@@ -1198,6 +1213,13 @@ static void machine_class_init(ObjectClass *oc, const void *data)
     object_class_property_set_description(oc, "dump-guest-core",
         "Include guest memory in a core dump");
 
+    object_class_property_add_bool(oc, "x-change-vmfd-on-reset",
+        machine_get_new_accel_vmfd_on_reset,
+        machine_set_new_accel_vmfd_on_reset);
+    object_class_property_set_description(oc, "x-change-vmfd-on-reset",
+        "Generate new accelerator fd on reset, "
+        "to be used only for testing and debugging.");
+
     object_class_property_add_bool(oc, "mem-merge",
         machine_get_mem_merge, machine_set_mem_merge);
     object_class_property_set_description(oc, "mem-merge",
include/hw/core/boards.h +6 -0
index 07f8938752..ee3cc9130e 100644
@@ -447,6 +447,12 @@ struct MachineState {
     struct NVDIMMState *nvdimms_state;
     struct NumaState *numa_state;
     bool acpi_spcr_enabled;
+    /*
+     * whether to change virtual machine accelerator file descriptor upon
+     * reset or not. used only for debugging and testing purpose.
+     * It should be set to false for all regular use.
+     */
+    bool new_accel_vmfd_on_reset;
 };
 
 /*
system/runstate.c +10 -10
index 0a74e3ade5..29953fb088 100644
@@ -526,25 +526,25 @@ void qemu_system_reset(ShutdownCause reason)
         type = RESET_TYPE_COLD;
     }
 
-    /*
-     * different accelerators implement how to close the old file handle of
-     * the accelerator descriptor and create a new one here. Resetting
-     * file handle is necessary to create a new confidential VM context post
-     * VM reset.
-     */
-    if (!cpus_are_resettable() &&
-        (reason == SHUTDOWN_CAUSE_GUEST_RESET ||
-         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET)) {
+    if ((reason == SHUTDOWN_CAUSE_GUEST_RESET ||
+         reason == SHUTDOWN_CAUSE_HOST_QMP_SYSTEM_RESET) &&
+        (current_machine->new_accel_vmfd_on_reset || !cpus_are_resettable())) {
         if (ac->reset_vmfd) {
             ret = ac->reset_vmfd(current_machine);
             if (ret < 0) {
                 error_report("unable to reset vmfd: %s(%d)",
                              strerror(-ret), ret);
                 vm_stop(RUN_STATE_INTERNAL_ERROR);
+            } else if (current_machine->new_accel_vmfd_on_reset) {
+                info_report("virtual machine accel file descriptor "
+                            "has changed.");
             }
             vmfd_reset = true;
+        } else if (!cpus_are_resettable())  {
+            error_report("accelerator does not support reset!");
         } else {
-            error_report("accelerator does not support reset");
+            error_report("accelerator does not support vmfd change reset, "
+                         "proceeding with normal reset!");
         }
     }
 
Signature · 1 line
2.42.0
author
#40 lore raw

PATCH v4 31/31 tests/functional/x86_64: add functional test to exercise vm fd change on reset

In reply to Ani Sinha

A new functional test is added that exercises the code changes related to closing of the old KVM VM file descriptor and opening a new one upon VM reset. This normally happens when confidential guests are reset but for non-confidential guests, we use a special machine specific debug/test parameter 'x-change-vmfd-on-reset' to enable this behavior.
Only specific code changes related to re-initialisation of SEV-ES, SEV-SNP and TDX platforms are not exercised in this test as they require hardware that supports running confidential guests.

---
 MAINTAINERS                                   |   1 +
 tests/functional/x86_64/meson.build           |   1 +
 .../x86_64/test_vmfd_change_reboot.py         | 135 ++++++++++++++++++
 3 files changed, 137 insertions(+)
 create mode 100755 tests/functional/x86_64/test_vmfd_change_reboot.py
MAINTAINERS +1 -0
index a88901a7d7..cd62cbd57a 100644
@@ -157,6 +157,7 @@ M: Ani Sinha <[email protected]>
 M: Paolo Bonzini <[email protected]>
 S: Maintained
 F: stubs/kvm.c
+F: tests/functional/x86_64/test_vmfd_change_reboot.py
 
 Guest CPU cores (TCG)
 ---------------------
tests/functional/x86_64/meson.build +1 -0
index f78eec5e6c..784d9791cb 100644
@@ -36,4 +36,5 @@ tests_x86_64_system_thorough = [
   'vfio_user_client',
   'virtio_balloon',
   'virtio_gpu',
+  'vmfd_change_reboot',
 ]
tests/functional/x86_64/test_vmfd_change_reboot.py new file +135 -0

135 lines not on this page · raw message

Signature · 1 line
2.42.0
#41 lore raw
In reply to Ani Sinha
Quoted 23 lines · Ani Sinha

This change introduces support for confidential guests
(SEV-ES, SEV-SNP and TDX) to reset/reboot just like other non-confidential guests. Currently, a reboot intiated from the confidential guest results in termination of the QEMU hypervisor as the CPUs are not resettable. As the initial state of the guest including private memory is locked and encrypted, the contents of that memory will not be accessible post reset. Hence a new KVM file descriptor must be opened to create a new confidential VM context closing the old one. All KVM VM specific ioctls must be called again. New VCPU file descriptors must be created against the new KVM fd and most VCPU ioctls must be called again as well.

This change perfoms closing of the old KVM fd and creating a new one. After the new KVM fd is opened, all generic and architecture specific ioctl calls are issued again. Notifiers are added to notify subsystems that:
- The KVM file fd is about to be changed to state sync-ing from KVM to QEMU should be done if required.
- The KVM file fd has changed, so ioctl calls to the new KVM fd has to be performed again.
- That new VCPU fds are created so that VCPU ioctl calls must be called again where required.

Specific subsystems use these notifiers to re-issue ioctl calls where required.

Quoted 7 lines

Changes are made to SEV and TDX modules to reinitialize the confidential guest state and seal it again. Along the way, some bug fixes are made so that some initialization functions can be called again. Some refactoring of existing code is done so that both init and reset paths can use them.

Tested on TDX, SEV-ES and SEV-SNP. Tested on non-coco hardware. Tested with Xen emulation enabled.

I see no QOM changes here, so is there any way for a mgmt app to detect that QEMU is now capable of guest reset with TDX/SEV/SNP, compared to the current behaviour ?

With regards,
Daniel

Signature · 3 lines
|: https://berrange.com       ~~        https://hachyderm.io/@berrange :|
|: https://libvirt.org          ~~          https://entangle-photo.org :|
|: https://pixelfed.art/berrange   ~~    https://fstop138.berrange.com :|
author
#42 lore raw
Quoted 33 lines · Daniel P. Berrangé
Quoted 28 lines · Ani Sinha

This change introduces support for confidential guests
(SEV-ES, SEV-SNP and TDX) to reset/reboot just like other non-confidential guests. Currently, a reboot intiated from the confidential guest results in termination of the QEMU hypervisor as the CPUs are not resettable. As the initial state of the guest including private memory is locked and encrypted, the contents of that memory will not be accessible post reset. Hence a new KVM file descriptor must be opened to create a new confidential VM context closing the old one. All KVM VM specific ioctls must be called again. New VCPU file descriptors must be created against the new KVM fd and most VCPU ioctls must be called again as well.

This change perfoms closing of the old KVM fd and creating a new one. After the new KVM fd is opened, all generic and architecture specific ioctl calls are issued again. Notifiers are added to notify subsystems that:
- The KVM file fd is about to be changed to state sync-ing from KVM to QEMU should be done if required.
- The KVM file fd has changed, so ioctl calls to the new KVM fd has to be performed again.
- That new VCPU fds are created so that VCPU ioctl calls must be called again where required.

Specific subsystems use these notifiers to re-issue ioctl calls where required.

Changes are made to SEV and TDX modules to reinitialize the confidential guest state and seal it again. Along the way, some bug fixes are made so that some initialization functions can be called again. Some refactoring of existing code is done so that both init and reset paths can use them.

Quoted 6 lines
Quoted 2 lines

Tested on TDX, SEV-ES and SEV-SNP. Tested on non-coco hardware. Tested with Xen emulation enabled.

I see no QOM changes here, so is there any way for a mgmt app to detect that QEMU is now capable of guest reset with TDX/SEV/SNP, compared to the current behaviour ?

The thought did not occur to me. So other than the availability of 'x-change-vmfd-on-reset’ machine property there is nothing else.

author
#43 lore raw
Quoted 37 lines · Ani Sinha
Quoted 34 lines · Daniel P. Berrangé
Quoted 28 lines · Ani Sinha

This change introduces support for confidential guests
(SEV-ES, SEV-SNP and TDX) to reset/reboot just like other non-confidential guests. Currently, a reboot intiated from the confidential guest results in termination of the QEMU hypervisor as the CPUs are not resettable. As the initial state of the guest including private memory is locked and encrypted, the contents of that memory will not be accessible post reset. Hence a new KVM file descriptor must be opened to create a new confidential VM context closing the old one. All KVM VM specific ioctls must be called again. New VCPU file descriptors must be created against the new KVM fd and most VCPU ioctls must be called again as well.

This change perfoms closing of the old KVM fd and creating a new one. After the new KVM fd is opened, all generic and architecture specific ioctl calls are issued again. Notifiers are added to notify subsystems that:
- The KVM file fd is about to be changed to state sync-ing from KVM to QEMU should be done if required.
- The KVM file fd has changed, so ioctl calls to the new KVM fd has to be performed again.
- That new VCPU fds are created so that VCPU ioctl calls must be called again where required.

Specific subsystems use these notifiers to re-issue ioctl calls where required.

Changes are made to SEV and TDX modules to reinitialize the confidential guest state and seal it again. Along the way, some bug fixes are made so that some initialization functions can be called again. Some refactoring of existing code is done so that both init and reset paths can use them.

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Tested on TDX, SEV-ES and SEV-SNP. Tested on non-coco hardware. Tested with Xen emulation enabled.

I see no QOM changes here, so is there any way for a mgmt app to detect that QEMU is now capable of guest reset with TDX/SEV/SNP, compared to the current behaviour ?

The thought did not occur to me. So other than the availability of 'x-change-vmfd-on-reset’ machine property there is nothing else.

Looking at the code, actually ‘can_rebuild_guest_state’ is the one we want to use. It’s a property of ConfidentialGuestSupportClass . Maybe we can expose this somehow.