Hi Gustavo,

On 25-09-2024 09:17 pm, Gustavo Romero wrote:
Hi Ganapatrao,

Sorry for the delay on replying it. I was attending KVM Forum and commuting.

On 9/20/24 09:37, Ganapatrao Kulkarni wrote:
Extend the 'mte' property for the virt machine to cover KVM as
well. For KVM, we don't allocate tag memory, but instead enable
the capability.

If MTE has been enabled, we need to disable migration, as we do not
yet have a way to migrate the tags as well. Therefore, MTE will stay
off with KVM unless requested explicitly.

This patch is rework of commit b320e21c48ce64853904bea6631c0158cc2ef227
which broke TCG since it made the TCG -cpu max
report the presence of MTE to the guest even if the board hadn't
enabled MTE by wiring up the tag RAM. This meant that if the guest
then tried to use MTE QEMU would segfault accessing the
non-existent tag RAM.

Signed-off-by: Cornelia Huck <coh...@redhat.com>
Signed-off-by: Ganapatrao Kulkarni <gankulka...@os.amperecomputing.com>
---

Changes since V2:
    Updated with review comments.

Changes since V1:
    Added code to enable MTE before reading register
id_aa64pfr1 (unmasked MTE bits).

This patch is boot tested on ARM64 with KVM and on X86 with TCG for mte=on
and default case(i.e, no mte).

  hw/arm/virt.c        | 72 ++++++++++++++++++++++++++------------------
  target/arm/cpu.c     | 11 +++++--
  target/arm/cpu.h     |  2 ++
  target/arm/kvm.c     | 57 +++++++++++++++++++++++++++++++++++
  target/arm/kvm_arm.h | 19 ++++++++++++
  5 files changed, 129 insertions(+), 32 deletions(-)

diff --git a/hw/arm/virt.c b/hw/arm/virt.c
index 7934b23651..a33af7d996 100644
--- a/hw/arm/virt.c
+++ b/hw/arm/virt.c
@@ -2211,7 +2211,7 @@ static void machvirt_init(MachineState *machine)
          exit(1);
      }
-    if (vms->mte && (kvm_enabled() || hvf_enabled())) {
+    if (vms->mte && hvf_enabled()) {
          error_report("mach-virt: %s does not support providing "
                       "MTE to the guest CPU",
                       current_accel_name());
@@ -2281,39 +2281,51 @@ static void machvirt_init(MachineState *machine)
          }
          if (vms->mte) {
-            /* Create the memory region only once, but link to all cpus. */
-            if (!tag_sysmem) {
-                /*
-                 * The property exists only if MemTag is supported.
-                 * If it is, we must allocate the ram to back that up.
-                 */
-                if (!object_property_find(cpuobj, "tag-memory")) {
-                    error_report("MTE requested, but not supported "
-                                 "by the guest CPU");
-                    exit(1);
+            if (tcg_enabled()) {
+                /* Create the memory region only once, but link to all cpus. */
+                if (!tag_sysmem) {
+                    /*
+                     * The property exists only if MemTag is supported.
+                     * If it is, we must allocate the ram to back that up.
+                     */
+                    if (!object_property_find(cpuobj, "tag-memory")) {
+                        error_report("MTE requested, but not supported "
+                                     "by the guest CPU");
+                        exit(1);
+                    }
+
+                    tag_sysmem = g_new(MemoryRegion, 1);
+                    memory_region_init(tag_sysmem, OBJECT(machine),
+                                       "tag-memory", UINT64_MAX / 32);
+
+                    if (vms->secure) {
+                        secure_tag_sysmem = g_new(MemoryRegion, 1);
+                        memory_region_init(secure_tag_sysmem, OBJECT(machine),
+                                           "secure-tag-memory",
+                                           UINT64_MAX / 32);
+
+                        /* As with ram, secure-tag takes precedence over tag. */ + memory_region_add_subregion_overlap(secure_tag_sysmem, +                                                            0, tag_sysmem, -1);
+                    }
                  }
-                tag_sysmem = g_new(MemoryRegion, 1);
-                memory_region_init(tag_sysmem, OBJECT(machine),
-                                   "tag-memory", UINT64_MAX / 32);
-
+                object_property_set_link(cpuobj, "tag-memory",
+                                         OBJECT(tag_sysmem), &error_abort);
                  if (vms->secure) {
-                    secure_tag_sysmem = g_new(MemoryRegion, 1);
-                    memory_region_init(secure_tag_sysmem, OBJECT(machine), -                                       "secure-tag-memory", UINT64_MAX / 32);
-
-                    /* As with ram, secure-tag takes precedence over tag.  */ - memory_region_add_subregion_overlap(secure_tag_sysmem, 0,
-                                                        tag_sysmem, -1);
+                    object_property_set_link(cpuobj, "secure-tag-memory",
+                                             OBJECT(secure_tag_sysmem),
+                                             &error_abort);
                  }
-            }
-
-            object_property_set_link(cpuobj, "tag-memory", OBJECT(tag_sysmem),
-                                     &error_abort);
-            if (vms->secure) {
-                object_property_set_link(cpuobj, "secure-tag-memory",
-                                         OBJECT(secure_tag_sysmem),
-                                         &error_abort);
+            } else if (kvm_enabled()) {
+                if (!kvm_arm_mte_supported()) {
+                    error_report("MTE requested, but not supported by KVM");
+                    exit(1);
+                }
+                kvm_arm_enable_mte(cpuobj, &error_abort);
+            } else {
+                    error_report("MTE requested, but not supported ");
+                    exit(1);
              }
          }
diff --git a/target/arm/cpu.c b/target/arm/cpu.c
index 19191c2391..8a2fc471ce 100644
--- a/target/arm/cpu.c
+++ b/target/arm/cpu.c
@@ -2390,14 +2390,21 @@ static void arm_cpu_realizefn(DeviceState *dev, Error **errp)
  #ifndef CONFIG_USER_ONLY
          /*
-         * If we do not have tag-memory provided by the machine,
+         * If we do not have tag-memory provided by the TCG,
           * reduce MTE support to instructions enabled at EL0.
           * This matches Cortex-A710 BROADCASTMTE input being LOW.
           */
-        if (cpu->tag_memory == NULL) {
+        if (tcg_enabled() && cpu->tag_memory == NULL) {
              cpu->isar.id_aa64pfr1 =
                  FIELD_DP64(cpu->isar.id_aa64pfr1, ID_AA64PFR1, MTE, 1);
          }
+
+        /*
+         * Clear MTE bits, if not enabled in KVM mode.
+         */
+        if (kvm_enabled() && !cpu->kvm_mte) {
+                FIELD_DP64(cpu->isar.id_aa64pfr1, ID_AA64PFR1, MTE, 0);
+        }

I 've discussed a bit with Richard about the need of setting the MTE field here

to 0. This is already a reduction since it's inside the condition block:

if (cpu_isar_feature(aa64_mte, cpu)) { ... }, which is only taken if we already

have MTE field >= 1. At this point the MTE bits in cpu->isar.id_aa64pfr1 should

already be set correctly accordingly to the host bits since kvm_arm_get_host_cpu_features()

was called.


The check for TCG (cpu->tag_memory == NULL) exists because even if MTE instructions

are supported by the CPU it's possible that the machine's memory does not support tags,

but we don't check for that in KVM afaics.


For KVM, isn't the cpu->isar.id_aa64pfr1 MTE bits already correct here so we don't need to touch them?

id_aa64pfr1.MTE bits are masked/zero when we read without enabling the MTE(KVM_CAP_ARM_MTE). Hence to read the real value we are enabling MTE while reading the register.

Later, If user is not passed argument "mte=on", we want to revert the value that is read with MTE enabled since we are enabling MTE on actual VM only if "mte=on".

Are you saying, let us have the MTE enabled for KVM by default and disable only if user passes mte=off?.


cpu->kvm_mte is false here only if kvm_check_extension(kvm_state, KVM_CAP_ARM_MTE) is true but

kvm_vm_enable_cap(kvm_state, KVM_CAP_ARM_MTE, 0) is false -- can that happen ever happen? _or_ when

This can be, if hardware is not supporting it.
Since user is trying to enable with mte=on but hardware is not capable, in that case qemu aborts with error message.


we can't block migration? So basically we are telling to the guest that MTE is not supported if we have

MTE supported in the host _but_ we can't block migration?



Cheers,

Gustavo


--
Thanks,
Ganapat/GK

Reply via email to