patches/kernel/0013-kvm-vmx-Reinstate-support-for-CPUs-without-virtual-N.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Paolo Bonzini <pbonzini@redhat.com>
   3 Date: Mon, 6 Nov 2017 13:31:12 +0100
   4 Subject: [PATCH] kvm: vmx: Reinstate support for CPUs without virtual NMI
   5 MIME-Version: 1.0
   6 Content-Type: text/plain; charset=UTF-8
   7 Content-Transfer-Encoding: 8bit
   8
   9 commit 8a1b43922d0d1279e7936ba85c4c2a870403c95f upstream.
  10
  11 This is more or less a revert of commit 2c82878b0cb3 ("KVM: VMX: require
  12 virtual NMI support", 2017-03-27); it turns out that Core 2 Duo machines
  13 only had virtual NMIs in some SKUs.
  14
  15 The revert is not trivial because in the meanwhile there have been several
  16 fixes to nested NMI injection.  Therefore, the entire vNMI state is moved
  17 to struct loaded_vmcs.
  18
  19 Another change compared to before the patch is a simplification here:
  20
  21        if (unlikely(!cpu_has_virtual_nmis() && vmx->soft_vnmi_blocked &&
  22            !(is_guest_mode(vcpu) && nested_cpu_has_virtual_nmis(
  23                                        get_vmcs12(vcpu))))) {
  24
  25 The final condition here is always true (because nested_cpu_has_virtual_nmis
  26 is always false) and is removed.
  27
  28 Fixes: 2c82878b0cb38fd516fd612c67852a6bbf282003
  29 Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1490803
  30 Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
  31 Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
  32 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  33 Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
  34 ---
  35  arch/x86/kvm/vmx.c | 150 +++++++++++++++++++++++++++++++++++++----------------
  36  1 file changed, 106 insertions(+), 44 deletions(-)
  37
  38 diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
  39 index 118709e7597d..a2c95522ac99 100644
  40 --- a/arch/x86/kvm/vmx.c
  41 +++ b/arch/x86/kvm/vmx.c
  42 @@ -202,6 +202,10 @@ struct loaded_vmcs {
  43         bool nmi_known_unmasked;
  44         unsigned long vmcs_host_cr3;    /* May not match real cr3 */
  45         unsigned long vmcs_host_cr4;    /* May not match real cr4 */
  46 +       /* Support for vnmi-less CPUs */
  47 +       int soft_vnmi_blocked;
  48 +       ktime_t entry_time;
  49 +       s64 vnmi_blocked_time;
  50         struct list_head loaded_vmcss_on_cpu_link;
  51  };
  52
  53 @@ -1288,6 +1292,11 @@ static inline bool cpu_has_vmx_invpcid(void)
  54                 SECONDARY_EXEC_ENABLE_INVPCID;
  55  }
  56
  57 +static inline bool cpu_has_virtual_nmis(void)
  58 +{
  59 +       return vmcs_config.pin_based_exec_ctrl & PIN_BASED_VIRTUAL_NMIS;
  60 +}
  61 +
  62  static inline bool cpu_has_vmx_wbinvd_exit(void)
  63  {
  64         return vmcs_config.cpu_based_2nd_exec_ctrl &
  65 @@ -1339,11 +1348,6 @@ static inline bool nested_cpu_has2(struct vmcs12 *vmcs12, u32 bit)
  66                 (vmcs12->secondary_vm_exec_control & bit);
  67  }
  68
  69 -static inline bool nested_cpu_has_virtual_nmis(struct vmcs12 *vmcs12)
  70 -{
  71 -       return vmcs12->pin_based_vm_exec_control & PIN_BASED_VIRTUAL_NMIS;
  72 -}
  73 -
  74  static inline bool nested_cpu_has_preemption_timer(struct vmcs12 *vmcs12)
  75  {
  76         return vmcs12->pin_based_vm_exec_control &
  77 @@ -3676,9 +3680,9 @@ static __init int setup_vmcs_config(struct vmcs_config *vmcs_conf)
  78                                 &_vmexit_control) < 0)
  79                 return -EIO;
  80
  81 -       min = PIN_BASED_EXT_INTR_MASK | PIN_BASED_NMI_EXITING |
  82 -               PIN_BASED_VIRTUAL_NMIS;
  83 -       opt = PIN_BASED_POSTED_INTR | PIN_BASED_VMX_PREEMPTION_TIMER;
  84 +       min = PIN_BASED_EXT_INTR_MASK | PIN_BASED_NMI_EXITING;
  85 +       opt = PIN_BASED_VIRTUAL_NMIS | PIN_BASED_POSTED_INTR |
  86 +                PIN_BASED_VMX_PREEMPTION_TIMER;
  87         if (adjust_vmx_controls(min, opt, MSR_IA32_VMX_PINBASED_CTLS,
  88                                 &_pin_based_exec_control) < 0)
  89                 return -EIO;
  90 @@ -5538,7 +5542,8 @@ static void enable_irq_window(struct kvm_vcpu *vcpu)
  91
  92  static void enable_nmi_window(struct kvm_vcpu *vcpu)
  93  {
  94 -       if (vmcs_read32(GUEST_INTERRUPTIBILITY_INFO) & GUEST_INTR_STATE_STI) {
  95 +       if (!cpu_has_virtual_nmis() ||
  96 +           vmcs_read32(GUEST_INTERRUPTIBILITY_INFO) & GUEST_INTR_STATE_STI) {
  97                 enable_irq_window(vcpu);
  98                 return;
  99         }
 100 @@ -5578,6 +5583,19 @@ static void vmx_inject_nmi(struct kvm_vcpu *vcpu)
 101  {
 102         struct vcpu_vmx *vmx = to_vmx(vcpu);
 103
 104 +       if (!cpu_has_virtual_nmis()) {
 105 +               /*
 106 +                * Tracking the NMI-blocked state in software is built upon
 107 +                * finding the next open IRQ window. This, in turn, depends on
 108 +                * well-behaving guests: They have to keep IRQs disabled at
 109 +                * least as long as the NMI handler runs. Otherwise we may
 110 +                * cause NMI nesting, maybe breaking the guest. But as this is
 111 +                * highly unlikely, we can live with the residual risk.
 112 +                */
 113 +               vmx->loaded_vmcs->soft_vnmi_blocked = 1;
 114 +               vmx->loaded_vmcs->vnmi_blocked_time = 0;
 115 +       }
 116 +
 117         ++vcpu->stat.nmi_injections;
 118         vmx->loaded_vmcs->nmi_known_unmasked = false;
 119
 120 @@ -5596,6 +5614,8 @@ static bool vmx_get_nmi_mask(struct kvm_vcpu *vcpu)
 121         struct vcpu_vmx *vmx = to_vmx(vcpu);
 122         bool masked;
 123
 124 +       if (!cpu_has_virtual_nmis())
 125 +               return vmx->loaded_vmcs->soft_vnmi_blocked;
 126         if (vmx->loaded_vmcs->nmi_known_unmasked)
 127                 return false;
 128         masked = vmcs_read32(GUEST_INTERRUPTIBILITY_INFO) & GUEST_INTR_STATE_NMI;
 129 @@ -5607,13 +5627,20 @@ static void vmx_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
 130  {
 131         struct vcpu_vmx *vmx = to_vmx(vcpu);
 132
 133 -       vmx->loaded_vmcs->nmi_known_unmasked = !masked;
 134 -       if (masked)
 135 -               vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 136 -                             GUEST_INTR_STATE_NMI);
 137 -       else
 138 -               vmcs_clear_bits(GUEST_INTERRUPTIBILITY_INFO,
 139 -                               GUEST_INTR_STATE_NMI);
 140 +       if (!cpu_has_virtual_nmis()) {
 141 +               if (vmx->loaded_vmcs->soft_vnmi_blocked != masked) {
 142 +                       vmx->loaded_vmcs->soft_vnmi_blocked = masked;
 143 +                       vmx->loaded_vmcs->vnmi_blocked_time = 0;
 144 +               }
 145 +       } else {
 146 +               vmx->loaded_vmcs->nmi_known_unmasked = !masked;
 147 +               if (masked)
 148 +                       vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 149 +                                     GUEST_INTR_STATE_NMI);
 150 +               else
 151 +                       vmcs_clear_bits(GUEST_INTERRUPTIBILITY_INFO,
 152 +                                       GUEST_INTR_STATE_NMI);
 153 +       }
 154  }
 155
 156  static int vmx_nmi_allowed(struct kvm_vcpu *vcpu)
 157 @@ -5621,6 +5648,10 @@ static int vmx_nmi_allowed(struct kvm_vcpu *vcpu)
 158         if (to_vmx(vcpu)->nested.nested_run_pending)
 159                 return 0;
 160
 161 +       if (!cpu_has_virtual_nmis() &&
 162 +           to_vmx(vcpu)->loaded_vmcs->soft_vnmi_blocked)
 163 +               return 0;
 164 +
 165         return  !(vmcs_read32(GUEST_INTERRUPTIBILITY_INFO) &
 166                   (GUEST_INTR_STATE_MOV_SS | GUEST_INTR_STATE_STI
 167                    | GUEST_INTR_STATE_NMI));
 168 @@ -6348,6 +6379,7 @@ static int handle_ept_violation(struct kvm_vcpu *vcpu)
 169          * AAK134, BY25.
 170          */
 171         if (!(to_vmx(vcpu)->idt_vectoring_info & VECTORING_INFO_VALID_MASK) &&
 172 +                       cpu_has_virtual_nmis() &&
 173                         (exit_qualification & INTR_INFO_UNBLOCK_NMI))
 174                 vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO, GUEST_INTR_STATE_NMI);
 175
 176 @@ -6820,7 +6852,7 @@ static struct loaded_vmcs *nested_get_current_vmcs02(struct vcpu_vmx *vmx)
 177         }
 178
 179         /* Create a new VMCS */
 180 -       item = kmalloc(sizeof(struct vmcs02_list), GFP_KERNEL);
 181 +       item = kzalloc(sizeof(struct vmcs02_list), GFP_KERNEL);
 182         if (!item)
 183                 return NULL;
 184         item->vmcs02.vmcs = alloc_vmcs();
 185 @@ -7837,6 +7869,7 @@ static int handle_pml_full(struct kvm_vcpu *vcpu)
 186          * "blocked by NMI" bit has to be set before next VM entry.
 187          */
 188         if (!(to_vmx(vcpu)->idt_vectoring_info & VECTORING_INFO_VALID_MASK) &&
 189 +                       cpu_has_virtual_nmis() &&
 190                         (exit_qualification & INTR_INFO_UNBLOCK_NMI))
 191                 vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 192                                 GUEST_INTR_STATE_NMI);
 193 @@ -8554,6 +8587,25 @@ static int vmx_handle_exit(struct kvm_vcpu *vcpu)
 194                 return 0;
 195         }
 196
 197 +       if (unlikely(!cpu_has_virtual_nmis() &&
 198 +                    vmx->loaded_vmcs->soft_vnmi_blocked)) {
 199 +               if (vmx_interrupt_allowed(vcpu)) {
 200 +                       vmx->loaded_vmcs->soft_vnmi_blocked = 0;
 201 +               } else if (vmx->loaded_vmcs->vnmi_blocked_time > 1000000000LL &&
 202 +                          vcpu->arch.nmi_pending) {
 203 +                       /*
 204 +                        * This CPU don't support us in finding the end of an
 205 +                        * NMI-blocked window if the guest runs with IRQs
 206 +                        * disabled. So we pull the trigger after 1 s of
 207 +                        * futile waiting, but inform the user about this.
 208 +                        */
 209 +                       printk(KERN_WARNING "%s: Breaking out of NMI-blocked "
 210 +                              "state on VCPU %d after 1 s timeout\n",
 211 +                              __func__, vcpu->vcpu_id);
 212 +                       vmx->loaded_vmcs->soft_vnmi_blocked = 0;
 213 +               }
 214 +       }
 215 +
 216         if (exit_reason < kvm_vmx_max_exit_handlers
 217             && kvm_vmx_exit_handlers[exit_reason])
 218                 return kvm_vmx_exit_handlers[exit_reason](vcpu);
 219 @@ -8837,33 +8889,38 @@ static void vmx_recover_nmi_blocking(struct vcpu_vmx *vmx)
 220
 221         idtv_info_valid = vmx->idt_vectoring_info & VECTORING_INFO_VALID_MASK;
 222
 223 -       if (vmx->loaded_vmcs->nmi_known_unmasked)
 224 -               return;
 225 -       /*
 226 -        * Can't use vmx->exit_intr_info since we're not sure what
 227 -        * the exit reason is.
 228 -        */
 229 -       exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
 230 -       unblock_nmi = (exit_intr_info & INTR_INFO_UNBLOCK_NMI) != 0;
 231 -       vector = exit_intr_info & INTR_INFO_VECTOR_MASK;
 232 -       /*
 233 -        * SDM 3: 27.7.1.2 (September 2008)
 234 -        * Re-set bit "block by NMI" before VM entry if vmexit caused by
 235 -        * a guest IRET fault.
 236 -        * SDM 3: 23.2.2 (September 2008)
 237 -        * Bit 12 is undefined in any of the following cases:
 238 -        *  If the VM exit sets the valid bit in the IDT-vectoring
 239 -        *   information field.
 240 -        *  If the VM exit is due to a double fault.
 241 -        */
 242 -       if ((exit_intr_info & INTR_INFO_VALID_MASK) && unblock_nmi &&
 243 -           vector != DF_VECTOR && !idtv_info_valid)
 244 -               vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 245 -                             GUEST_INTR_STATE_NMI);
 246 -       else
 247 -               vmx->loaded_vmcs->nmi_known_unmasked =
 248 -                       !(vmcs_read32(GUEST_INTERRUPTIBILITY_INFO)
 249 -                         & GUEST_INTR_STATE_NMI);
 250 +       if (cpu_has_virtual_nmis()) {
 251 +               if (vmx->loaded_vmcs->nmi_known_unmasked)
 252 +                       return;
 253 +               /*
 254 +                * Can't use vmx->exit_intr_info since we're not sure what
 255 +                * the exit reason is.
 256 +                */
 257 +               exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
 258 +               unblock_nmi = (exit_intr_info & INTR_INFO_UNBLOCK_NMI) != 0;
 259 +               vector = exit_intr_info & INTR_INFO_VECTOR_MASK;
 260 +               /*
 261 +                * SDM 3: 27.7.1.2 (September 2008)
 262 +                * Re-set bit "block by NMI" before VM entry if vmexit caused by
 263 +                * a guest IRET fault.
 264 +                * SDM 3: 23.2.2 (September 2008)
 265 +                * Bit 12 is undefined in any of the following cases:
 266 +                *  If the VM exit sets the valid bit in the IDT-vectoring
 267 +                *   information field.
 268 +                *  If the VM exit is due to a double fault.
 269 +                */
 270 +               if ((exit_intr_info & INTR_INFO_VALID_MASK) && unblock_nmi &&
 271 +                   vector != DF_VECTOR && !idtv_info_valid)
 272 +                       vmcs_set_bits(GUEST_INTERRUPTIBILITY_INFO,
 273 +                                     GUEST_INTR_STATE_NMI);
 274 +               else
 275 +                       vmx->loaded_vmcs->nmi_known_unmasked =
 276 +                               !(vmcs_read32(GUEST_INTERRUPTIBILITY_INFO)
 277 +                                 & GUEST_INTR_STATE_NMI);
 278 +       } else if (unlikely(vmx->loaded_vmcs->soft_vnmi_blocked))
 279 +               vmx->loaded_vmcs->vnmi_blocked_time +=
 280 +                       ktime_to_ns(ktime_sub(ktime_get(),
 281 +                                             vmx->loaded_vmcs->entry_time));
 282  }
 283
 284  static void __vmx_complete_interrupts(struct kvm_vcpu *vcpu,
 285 @@ -8980,6 +9037,11 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
 286         struct vcpu_vmx *vmx = to_vmx(vcpu);
 287         unsigned long debugctlmsr, cr3, cr4;
 288
 289 +       /* Record the guest's net vcpu time for enforced NMI injections. */
 290 +       if (unlikely(!cpu_has_virtual_nmis() &&
 291 +                    vmx->loaded_vmcs->soft_vnmi_blocked))
 292 +               vmx->loaded_vmcs->entry_time = ktime_get();
 293 +
 294         /* Don't enter VMX if guest state is invalid, let the exit handler
 295            start emulation until we arrive back to a valid state */
 296         if (vmx->emulation_required)
 297 --
 298 2.14.2
 299