Skip to contentSkip to conversation

kvm: add support for guest physical bits

PATCH Superseded Gerd Hoffmann 4 participants 7 messages Mar 1, 2024 – Mar 6, 2024 KVM

v1 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

cover kvm: add support for guest physical bits

The matching kernel bits are here:
https://lore.kernel.org/kvm/[email protected]/T/

Gerd Hoffmann (1):
kvm: add support for guest physical bits

 target/i386/cpu.h     | 1 +
 target/i386/cpu.c     | 1 +
 target/i386/kvm/kvm.c | 8 ++++++++
 3 files changed, 10 insertions(+)
Signature · 1 line
2.44.0
#1 lore raw

PATCH 1/1 kvm: add support for guest physical bits

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS. Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
 target/i386/cpu.h     | 1 +
 target/i386/cpu.c     | 1 +
 target/i386/kvm/kvm.c | 8 ++++++++
 3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
 
     /* Number of physical address bits supported */
     uint32_t phys_bits;
+    uint32_t guest_phys_bits;
 
     /* in order to simplify APIC support, we leave this pointer to the
        user */
target/i386/cpu.c +1 -0
index 2666ef380891..1a6cfc75951e 100644
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
         if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
             /* 64 bit processor */
              *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);
         }
         *ebx = env->features[FEAT_8000_0008_EBX];
         if (cs->nr_cores * cs->nr_threads > 1) {
target/i386/kvm/kvm.c +8 -0
index 42970ab046fa..e06c9d66bb01 100644
@@ -1716,6 +1716,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
     X86CPU *cpu = X86_CPU(cs);
     CPUX86State *env = &cpu->env;
     uint32_t limit, i, j, cpuid_i;
+    uint32_t guest_phys_bits;
     uint32_t unused;
     struct kvm_cpuid_entry2 *c;
     uint32_t signature[3];
@@ -1751,6 +1752,13 @@ int kvm_arch_init_vcpu(CPUState *cs)
 
     env->apic_bus_freq = KVM_APIC_BUS_FREQUENCY;
 
+    guest_phys_bits = kvm_check_extension(cs->kvm_state, KVM_CAP_VM_GPA_BITS);
+    if (guest_phys_bits &&
+        (cpu->guest_phys_bits == 0 ||
+         cpu->guest_phys_bits > guest_phys_bits)) {
+        cpu->guest_phys_bits = guest_phys_bits;
+    }
+
     /*
      * kvm_hyperv_expand_features() is called here for the second time in case
      * KVM_CAP_SYS_HYPERV_CPUID is not supported. While we can't possibly handle
Signature · 1 line
2.44.0
#2 lore raw
Quoted 28 lines · Gerd Hoffmann

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS. Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
  target/i386/cpu.h     | 1 +
  target/i386/cpu.c     | 1 +
  target/i386/kvm/kvm.c | 8 ++++++++
  3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
 
      /* Number of physical address bits supported */
      uint32_t phys_bits;
+    uint32_t guest_phys_bits;
 
      /* in order to simplify APIC support, we leave this pointer to the
         user */
target/i386/cpu.c +0 -0
index 2666ef380891..1a6cfc75951e 100644

target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6573

Quoted 4 lines
target/i386/cpu.c +1 -0
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
          if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
              /* 64 bit processor */
               *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);

I think you misunderstand this field.

If you expose this field to guest, it's the information for nested guest. i.e., the guest itself runs as a hypervisor will know its nested guest can have guest_phys_bits for physical addr.

Quoted 29 lines
target/i386/cpu.c +0 -0
          }
          *ebx = env->features[FEAT_8000_0008_EBX];
          if (cs->nr_cores * cs->nr_threads > 1) {
target/i386/kvm/kvm.c +8 -0
index 42970ab046fa..e06c9d66bb01 100644
@@ -1716,6 +1716,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
      X86CPU *cpu = X86_CPU(cs);
      CPUX86State *env = &cpu->env;
      uint32_t limit, i, j, cpuid_i;
+    uint32_t guest_phys_bits;
      uint32_t unused;
      struct kvm_cpuid_entry2 *c;
      uint32_t signature[3];
@@ -1751,6 +1752,13 @@ int kvm_arch_init_vcpu(CPUState *cs)
 
      env->apic_bus_freq = KVM_APIC_BUS_FREQUENCY;
 
+    guest_phys_bits = kvm_check_extension(cs->kvm_state, KVM_CAP_VM_GPA_BITS);
+    if (guest_phys_bits &&
+        (cpu->guest_phys_bits == 0 ||
+         cpu->guest_phys_bits > guest_phys_bits)) {
+        cpu->guest_phys_bits = guest_phys_bits;
+    }
+
      /*
       * kvm_hyperv_expand_features() is called here for the second time in case
       * KVM_CAP_SYS_HYPERV_CPUID is not supported. While we can't possibly handle
#3 lore raw
Quoted 34 lines · Xiaoyao Li
Quoted 26 lines · Gerd Hoffmann

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS. Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
  target/i386/cpu.h     | 1 +
  target/i386/cpu.c     | 1 +
  target/i386/kvm/kvm.c | 8 ++++++++
  3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
      /* Number of physical address bits supported */
      uint32_t phys_bits;
+    uint32_t guest_phys_bits;
      /* in order to simplify APIC support, we leave this pointer to the
         user */
target/i386/cpu.c +0 -0
index 2666ef380891..1a6cfc75951e 100644

target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6573

Quoted 4 lines
target/i386/cpu.c +1 -0
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
          if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
              /* 64 bit processor */
               *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);

I think you misunderstand this field.

Quoted 3 lines

If you expose this field to guest, it's the information for nested guest. i.e., the guest itself runs as a hypervisor will know its nested guest can have guest_phys_bits for physical addr.

I'm also thinking about this issue...

Currently guest KVM doesn't use this field to advertise MAXPHYADDR because nested guest hasn't tdp. And this patch only affects KVM hypervisor now.

Thanks,
Tao

Quoted 30 lines
Quoted 28 lines
target/i386/cpu.c +0 -0
          }
          *ebx = env->features[FEAT_8000_0008_EBX];
          if (cs->nr_cores * cs->nr_threads > 1) {
target/i386/kvm/kvm.c +8 -0
index 42970ab046fa..e06c9d66bb01 100644
@@ -1716,6 +1716,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
      X86CPU *cpu = X86_CPU(cs);
      CPUX86State *env = &cpu->env;
      uint32_t limit, i, j, cpuid_i;
+    uint32_t guest_phys_bits;
      uint32_t unused;
      struct kvm_cpuid_entry2 *c;
      uint32_t signature[3];
@@ -1751,6 +1752,13 @@ int kvm_arch_init_vcpu(CPUState *cs)
      env->apic_bus_freq = KVM_APIC_BUS_FREQUENCY;
+    guest_phys_bits = kvm_check_extension(cs->kvm_state, KVM_CAP_VM_GPA_BITS);
+    if (guest_phys_bits &&
+        (cpu->guest_phys_bits == 0 ||
+         cpu->guest_phys_bits > guest_phys_bits)) {
+        cpu->guest_phys_bits = guest_phys_bits;
+    }
+
      /*
       * kvm_hyperv_expand_features() is called here for the second time in case
       * KVM_CAP_SYS_HYPERV_CPUID is not supported. While we can't possibly handle
#4 lore raw
In reply to Xiaoyao Li
Quoted 34 lines · Xiaoyao Li
Quoted 26 lines · Gerd Hoffmann

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS. Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
  target/i386/cpu.h     | 1 +
  target/i386/cpu.c     | 1 +
  target/i386/kvm/kvm.c | 8 ++++++++
  3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
      /* Number of physical address bits supported */
      uint32_t phys_bits;
+    uint32_t guest_phys_bits;
      /* in order to simplify APIC support, we leave this pointer to the
         user */
target/i386/cpu.c +0 -0
index 2666ef380891..1a6cfc75951e 100644

target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6573

Quoted 4 lines
target/i386/cpu.c +1 -0
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
          if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
              /* 64 bit processor */
               *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);

I think you misunderstand this field.

Quoted 3 lines

If you expose this field to guest, it's the information for nested guest. i.e., the guest itself runs as a hypervisor will know its nested guest can have guest_phys_bits for physical addr.

I think those limits (l1 + l2 guest phys-bits) are identical, no?

The problem this tries to solve is that making the guest phys-bits smaller than the host phys-bits is problematic (which why we have allow_smaller_maxphyaddr), but nevertheless there are cases where the usable guest physical address space is smaller than the host physical address space. One case is intel processors with phys-bits larger than 48 and 4-level EPT. Another case is amd processors with phys-bits larger than 48 and the l0 hypervisor using 4-level paging.

The guest needs to know that limit, specifically the guest firmware so it knows where it can map PCI bars.

take care,
Gerd

#5 lore raw
Quoted 35 lines · Gerd Hoffmann
Quoted 34 lines · Xiaoyao Li
Quoted 26 lines · Gerd Hoffmann

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS. Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
   target/i386/cpu.h     | 1 +
   target/i386/cpu.c     | 1 +
   target/i386/kvm/kvm.c | 8 ++++++++
   3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
       /* Number of physical address bits supported */
       uint32_t phys_bits;
+    uint32_t guest_phys_bits;
       /* in order to simplify APIC support, we leave this pointer to the
          user */
target/i386/cpu.c +0 -0
index 2666ef380891..1a6cfc75951e 100644

target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count, · line 6573

Quoted 4 lines
target/i386/cpu.c +1 -0
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
           if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
               /* 64 bit processor */
                *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);

I think you misunderstand this field.

Quoted 5 lines
Quoted 3 lines

If you expose this field to guest, it's the information for nested guest. i.e., the guest itself runs as a hypervisor will know its nested guest can have guest_phys_bits for physical addr.

I think those limits (l1 + l2 guest phys-bits) are identical, no?

Sorry, I didn't know this patch was based on the off-list proposal made by Paolo that changing the definition of CPUID.0x80000008:EAX[23:16] to advertise the "maximum usable physical address bits".

If you call out this in the change log, it can avoid the misunderstanding.

As I replied to KVM series, I think the info is better to setup by KVM and reported by GET_SUPPORTED_CPUID.

Quoted 14 lines

The problem this tries to solve is that making the guest phys-bits smaller than the host phys-bits is problematic (which why we have allow_smaller_maxphyaddr), but nevertheless there are cases where the usable guest physical address space is smaller than the host physical address space. One case is intel processors with phys-bits larger than 48 and 4-level EPT. Another case is amd processors with phys-bits larger than 48 and the l0 hypervisor using 4-level paging.

The guest needs to know that limit, specifically the guest firmware so it knows where it can map PCI bars.

take care,
Gerd

maintainer
#6 lore raw
In reply to Xiaoyao Li
Quoted 35 lines · Xiaoyao Li
Quoted 26 lines · Gerd Hoffmann

query kvm for supported guest physical address bits using KVM_CAP_VM_GPA_BITS.  Expose the value to the guest via cpuid (leaf 0x80000008, eax, bits 16-23).

---
  target/i386/cpu.h     | 1 +
  target/i386/cpu.c     | 1 +
  target/i386/kvm/kvm.c | 8 ++++++++

  3 files changed, 10 insertions(+)
target/i386/cpu.h +1 -0
index 952174bb6f52..d427218827f6 100644
@@ -2026,6 +2026,7 @@ struct ArchCPU {
       /* Number of physical address bits supported */
       uint32_t phys_bits;
+    uint32_t guest_phys_bits;
       /* in order to simplify APIC support, we leave this pointer to the

         user */

target/i386/cpu.c +0 -0
index 2666ef380891..1a6cfc75951e 100644

target/i386/cpu.c void cpu_x86_cpuid(CPUX86State *env, uint32_t · line 6574

Quoted 5 lines
target/i386/cpu.c +1 -0
@@ -6570,6 +6570,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t 
 index, uint32_t count,
           if (env->features[FEAT_8000_0001_EDX] & CPUID_EXT2_LM) {
               /* 64 bit processor */
                *eax |= (cpu_x86_virtual_addr_width(env) << 8);
+             *eax |= (cpu->guest_phys_bits << 16);

I think you misunderstand this field.

Quoted 1 line

If you expose this field to guest, it's the information for nested guest. i.e., the guest itself runs as a hypervisor will know its nested guest can have guest_phys_bits for physical addr.

It's one possible interpretation of AMD's definition. However there's no processor that has different MAXPHYADDR with/without nested paging, so there's no real benefit in adopting that interpretation.

The only architectural case in which you have two conflicting values for the guest MAXPHYADDR is hCR4.LA57=0 (and likewise for Intel 4-level EPT) with MAXPHYADDR=52, so it's useful to treat GuestPhysAddrSize as a way to communicate this situation to the guest.

Paolo