[PATCH v2 15/30] xen/x86: Improvements to in-hypervisor cpuid sanity checks

[Andrew Cooper <andrew.cooper3@citrix.com>]

* Use the boot-generated pv and hvm featureset to clamp the visible features,
  rather than picking and choosing at individual features.  This subsumes the
  static feature manipulation.
* More use of compiler-visible &'s and |'s, rather than clear,set bit.
* Remove logic which hides PSE36 out of PAE mode.  This is not how real
  hardware behaves.
* Improve logic to set OSXSAVE.  The bit is cleared by virtue of not being
  valid in a featureset, and should be a strict fast-forward from %cr4.
  Provide a very big health warning for OXSAVE for PV guests, which is
  non-architectural.

Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
---
CC: Jan Beulich <JBeulich@suse.com>

v2:
 * Reinstate some of the dynamic checks for now.  Future development work will
   instate a complete per-domain policy.
 * Fix OSXSAVE handling for PV guests.
---
 xen/arch/x86/hvm/hvm.c |  56 +++++++++---------
 xen/arch/x86/traps.c   | 151 ++++++++++++++++++++++++-------------------------
 2 files changed, 100 insertions(+), 107 deletions(-)

diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
index 35ec6c9..03b3868 100644
--- a/xen/arch/x86/hvm/hvm.c
+++ b/xen/arch/x86/hvm/hvm.c
@@ -71,6 +71,7 @@
 #include <public/memory.h>
 #include <public/vm_event.h>
 #include <public/arch-x86/cpuid.h>
+#include <asm/cpuid.h>
 
 bool_t __read_mostly hvm_enabled;
 
@@ -4617,50 +4618,39 @@ void hvm_cpuid(unsigned int input, unsigned int *eax, unsigned int *ebx,
         /* Fix up VLAPIC details. */
         *ebx &= 0x00FFFFFFu;
         *ebx |= (v->vcpu_id * 2) << 24;
+
+        *ecx &= hvm_featureset[FEATURESET_1c];
+        *edx &= hvm_featureset[FEATURESET_1d];
+
         if ( vlapic_hw_disabled(vcpu_vlapic(v)) )
-            __clear_bit(X86_FEATURE_APIC & 31, edx);
+            *edx &= ~cpufeat_bit(X86_FEATURE_APIC);
 
         /* Fix up OSXSAVE. */
-        if ( cpu_has_xsave )
-            *ecx |= (v->arch.hvm_vcpu.guest_cr[4] & X86_CR4_OSXSAVE) ?
-                     cpufeat_mask(X86_FEATURE_OSXSAVE) : 0;
+        if ( v->arch.hvm_vcpu.guest_cr[4] & X86_CR4_OSXSAVE )
+            *ecx |= cpufeat_mask(X86_FEATURE_OSXSAVE);
 
         /* Don't expose PCID to non-hap hvm. */
         if ( !hap_enabled(d) )
             *ecx &= ~cpufeat_mask(X86_FEATURE_PCID);
-
-        /* Only provide PSE36 when guest runs in 32bit PAE or in long mode */
-        if ( !(hvm_pae_enabled(v) || hvm_long_mode_enabled(v)) )
-            *edx &= ~cpufeat_mask(X86_FEATURE_PSE36);
         break;
+
     case 0x7:
         if ( count == 0 )
         {
-            if ( !cpu_has_smep )
-                *ebx &= ~cpufeat_mask(X86_FEATURE_SMEP);
-
-            if ( !cpu_has_smap )
-                *ebx &= ~cpufeat_mask(X86_FEATURE_SMAP);
-
-            /* Don't expose MPX to hvm when VMX support is not available */
-            if ( !(vmx_vmexit_control & VM_EXIT_CLEAR_BNDCFGS) ||
-                 !(vmx_vmentry_control & VM_ENTRY_LOAD_BNDCFGS) )
-                *ebx &= ~cpufeat_mask(X86_FEATURE_MPX);
+            *ebx &= hvm_featureset[FEATURESET_7b0];
+            *ecx &= hvm_featureset[FEATURESET_7c0];
 
             /* Don't expose INVPCID to non-hap hvm. */
             if ( !hap_enabled(d) )
                 *ebx &= ~cpufeat_mask(X86_FEATURE_INVPCID);
-
-            /* Don't expose PCOMMIT to hvm when VMX support is not available */
-            if ( !cpu_has_vmx_pcommit )
-                *ebx &= ~cpufeat_mask(X86_FEATURE_PCOMMIT);
         }
-
         break;
+
     case 0xb:
         /* Fix the x2APIC identifier. */
         *edx = v->vcpu_id * 2;
         break;
+
     case 0xd:
         /* EBX value of main leaf 0 depends on enabled xsave features */
         if ( count == 0 && v->arch.xcr0 ) 
@@ -4677,9 +4667,12 @@ void hvm_cpuid(unsigned int input, unsigned int *eax, unsigned int *ebx,
                     *ebx = _eax + _ebx;
             }
         }
+
         if ( count == 1 )
         {
-            if ( cpu_has_xsaves && cpu_has_vmx_xsaves )
+            *eax &= hvm_featureset[FEATURESET_Da1];
+
+            if ( *eax & cpufeat_mask(X86_FEATURE_XSAVES) )
             {
                 *ebx = XSTATE_AREA_MIN_SIZE;
                 if ( v->arch.xcr0 | v->arch.hvm_vcpu.msr_xss )
@@ -4694,6 +4687,9 @@ void hvm_cpuid(unsigned int input, unsigned int *eax, unsigned int *ebx,
         break;
 
     case 0x80000001:
+        *ecx &= hvm_featureset[FEATURESET_e1c];
+        *edx &= hvm_featureset[FEATURESET_e1d];
+
         /* We expose RDTSCP feature to guest only when
            tsc_mode == TSC_MODE_DEFAULT and host_tsc_is_safe() returns 1 */
         if ( d->arch.tsc_mode != TSC_MODE_DEFAULT ||
@@ -4702,12 +4698,10 @@ void hvm_cpuid(unsigned int input, unsigned int *eax, unsigned int *ebx,
         /* Hide 1GB-superpage feature if we can't emulate it. */
         if (!hvm_pse1gb_supported(d))
             *edx &= ~cpufeat_mask(X86_FEATURE_PAGE1GB);
-        /* Only provide PSE36 when guest runs in 32bit PAE or in long mode */
-        if ( !(hvm_pae_enabled(v) || hvm_long_mode_enabled(v)) )
-            *edx &= ~cpufeat_mask(X86_FEATURE_PSE36);
-        /* Hide data breakpoint extensions if the hardware has no support. */
-        if ( !boot_cpu_has(X86_FEATURE_DBEXT) )
-            *ecx &= ~cpufeat_mask(X86_FEATURE_DBEXT);
+        break;
+
+    case 0x80000007:
+        *edx &= hvm_featureset[FEATURESET_e7d];
         break;
 
     case 0x80000008:
@@ -4725,6 +4719,8 @@ void hvm_cpuid(unsigned int input, unsigned int *eax, unsigned int *ebx,
         hvm_cpuid(0x80000001, NULL, NULL, NULL, &_edx);
         *eax = (*eax & ~0xffff00) | (_edx & cpufeat_mask(X86_FEATURE_LM)
                                      ? 0x3000 : 0x2000);
+
+        *ebx &= hvm_featureset[FEATURESET_e8b];
         break;
     }
 }
diff --git a/xen/arch/x86/traps.c b/xen/arch/x86/traps.c
index 6a181bb..d0f836c 100644
--- a/xen/arch/x86/traps.c
+++ b/xen/arch/x86/traps.c
@@ -73,6 +73,7 @@
 #include <asm/hpet.h>
 #include <asm/vpmu.h>
 #include <public/arch-x86/cpuid.h>
+#include <asm/cpuid.h>
 #include <xsm/xsm.h>
 
 /*
@@ -841,69 +842,70 @@ void pv_cpuid(struct cpu_user_regs *regs)
     else
         cpuid_count(leaf, subleaf, &a, &b, &c, &d);
 
-    if ( (leaf & 0x7fffffff) == 0x00000001 )
-    {
-        /* Modify Feature Information. */
-        if ( !cpu_has_apic )
-            __clear_bit(X86_FEATURE_APIC, &d);
-
-        if ( !is_pvh_domain(currd) )
-        {
-            __clear_bit(X86_FEATURE_PSE, &d);
-            __clear_bit(X86_FEATURE_PGE, &d);
-            __clear_bit(X86_FEATURE_PSE36, &d);
-            __clear_bit(X86_FEATURE_VME, &d);
-        }
-    }
-
     switch ( leaf )
     {
     case 0x00000001:
-        /* Modify Feature Information. */
-        if ( !cpu_has_sep )
-            __clear_bit(X86_FEATURE_SEP, &d);
-        __clear_bit(X86_FEATURE_DS, &d);
-        __clear_bit(X86_FEATURE_ACC, &d);
-        __clear_bit(X86_FEATURE_PBE, &d);
-        if ( is_pvh_domain(currd) )
-            __clear_bit(X86_FEATURE_MTRR, &d);
-
-        __clear_bit(X86_FEATURE_DTES64 % 32, &c);
-        __clear_bit(X86_FEATURE_MWAIT % 32, &c);
-        __clear_bit(X86_FEATURE_DSCPL % 32, &c);
-        __clear_bit(X86_FEATURE_VMXE % 32, &c);
-        __clear_bit(X86_FEATURE_SMXE % 32, &c);
-        __clear_bit(X86_FEATURE_TM2 % 32, &c);
+        c &= pv_featureset[FEATURESET_1c];
+        d &= pv_featureset[FEATURESET_1d];
+
         if ( is_pv_32bit_domain(currd) )
-            __clear_bit(X86_FEATURE_CX16 % 32, &c);
-        __clear_bit(X86_FEATURE_XTPR % 32, &c);
-        __clear_bit(X86_FEATURE_PDCM % 32, &c);
-        __clear_bit(X86_FEATURE_PCID % 32, &c);
-        __clear_bit(X86_FEATURE_DCA % 32, &c);
-        if ( !cpu_has_xsave )
-        {
-            __clear_bit(X86_FEATURE_XSAVE % 32, &c);
-            __clear_bit(X86_FEATURE_AVX % 32, &c);
-        }
-        if ( !cpu_has_apic )
-           __clear_bit(X86_FEATURE_X2APIC % 32, &c);
-        __set_bit(X86_FEATURE_HYPERVISOR % 32, &c);
+            c &= ~cpufeat_mask(X86_FEATURE_CX16);
+
+        /*
+         * !!! Warning - OSXSAVE handling for PV guests is non-architectural !!!
+         *
+         * Architecturally, the correct code here is simply:
+         *
+         *   if ( curr->arch.pv_vcpu.ctrlreg[4] & X86_CR4_OSXSAVE )
+         *       c |= cpufeat_mask(X86_FEATURE_OSXSAVE);
+         *
+         * However because of bugs in Xen (before c/s bd19080b, Nov 2010, the
+         * XSAVE cpuid flag leaked into guests despite the feature not being
+         * avilable for use), buggy workarounds where introduced to Linux (c/s
+         * 947ccf9c, also Nov 2010) which relied on the fact that Xen also
+         * incorrectly leaked OSXSAVE into the guest.
+         *
+         * Furthermore, providing architectural OSXSAVE behaviour to a many
+         * Linux PV guests triggered a further kernel bug when the fpu code
+         * observes that XSAVEOPT is available, assumes that xsave state had
+         * been set up for the task, and follows a wild pointer.
+         *
+         * Therefore, the leaking of Xen's OSXSAVE setting has become a
+         * defacto part of the PV ABI and can't reasonably be corrected.
+         *
+         * The following situations and logic now applies:
+         *
+         * - Hardware without CPUID faulting support and native CPUID:
+         *    There is nothing Xen can do here.  The hosts XSAVE flag will
+         *    leak through and Xen's OSXSAVE choice will leak through.
+         *
+         *    In the case that the guest kernel has not set up OSXSAVE, only
+         *    SSE will be set in xcr0, and guest userspace can't do too much
+         *    damage itself.
+         *
+         * - Enlightened CPUID or CPUID faulting available:
+         *    Xen can fully control what is seen here.  Guest kernels need to
+         *    see the leaked OSXSAVE, but guest userspace is given
+         *    architectural behaviour, to reflect the guest kernels
+         *    intentions.
+         */
+        if ( (is_pv_domain(currd) && guest_kernel_mode(curr, regs) &&
+              (this_cpu(cr4) & X86_CR4_OSXSAVE)) ||
+             (curr->arch.pv_vcpu.ctrlreg[4] & X86_CR4_OSXSAVE) )
+            c |= cpufeat_mask(X86_FEATURE_OSXSAVE);
+
+        c |= cpufeat_mask(X86_FEATURE_HYPERVISOR);
         break;
 
     case 0x00000007:
         if ( subleaf == 0 )
-            b &= (cpufeat_mask(X86_FEATURE_BMI1) |
-                  cpufeat_mask(X86_FEATURE_HLE)  |
-                  cpufeat_mask(X86_FEATURE_AVX2) |
-                  cpufeat_mask(X86_FEATURE_BMI2) |
-                  cpufeat_mask(X86_FEATURE_ERMS) |
-                  cpufeat_mask(X86_FEATURE_RTM)  |
-                  cpufeat_mask(X86_FEATURE_RDSEED)  |
-                  cpufeat_mask(X86_FEATURE_ADX)  |
-                  cpufeat_mask(X86_FEATURE_FSGSBASE));
+        {
+            b &= pv_featureset[FEATURESET_7b0];
+            c &= pv_featureset[FEATURESET_7c0];
+        }
         else
-            b = 0;
-        a = c = d = 0;
+            b = c = 0;
+        a = d = 0;
         break;
 
     case XSTATE_CPUID:
@@ -926,37 +928,32 @@ void pv_cpuid(struct cpu_user_regs *regs)
         }
 
         case 1:
-            a &= (boot_cpu_data.x86_capability[cpufeat_word(X86_FEATURE_XSAVEOPT)] &
-                  ~cpufeat_mask(X86_FEATURE_XSAVES));
+            a &= pv_featureset[FEATURESET_Da1];
             b = c = d = 0;
             break;
         }
         break;
 
     case 0x80000001:
-        /* Modify Feature Information. */
+        c &= pv_featureset[FEATURESET_e1c];
+        d &= pv_featureset[FEATURESET_e1d];
+
         if ( is_pv_32bit_domain(currd) )
         {
-            __clear_bit(X86_FEATURE_LM % 32, &d);
-            __clear_bit(X86_FEATURE_LAHF_LM % 32, &c);
-        }
-        if ( is_pv_32bit_domain(currd) &&
-             boot_cpu_data.x86_vendor != X86_VENDOR_AMD )
-            __clear_bit(X86_FEATURE_SYSCALL % 32, &d);
-        __clear_bit(X86_FEATURE_PAGE1GB % 32, &d);
-        __clear_bit(X86_FEATURE_RDTSCP % 32, &d);
-
-        __clear_bit(X86_FEATURE_SVM % 32, &c);
-        if ( !cpu_has_apic )
-           __clear_bit(X86_FEATURE_EXTAPIC % 32, &c);
-        __clear_bit(X86_FEATURE_OSVW % 32, &c);
-        __clear_bit(X86_FEATURE_IBS % 32, &c);
-        __clear_bit(X86_FEATURE_SKINIT % 32, &c);
-        __clear_bit(X86_FEATURE_WDT % 32, &c);
-        __clear_bit(X86_FEATURE_LWP % 32, &c);
-        __clear_bit(X86_FEATURE_NODEID_MSR % 32, &c);
-        __clear_bit(X86_FEATURE_TOPOEXT % 32, &c);
-        __clear_bit(X86_FEATURE_MWAITX % 32, &c);
+            d &= ~cpufeat_mask(X86_FEATURE_LM);
+            c &= ~cpufeat_mask(X86_FEATURE_LAHF_LM);
+
+            if ( boot_cpu_data.x86_vendor != X86_VENDOR_AMD )
+                d &= ~cpufeat_mask(X86_FEATURE_SYSCALL);
+        }
+        break;
+
+    case 0x80000007:
+        d &= pv_featureset[FEATURESET_e7d];
+        break;
+
+    case 0x80000008:
+        b &= pv_featureset[FEATURESET_e8b];
         break;
 
     case 0x0000000a: /* Architectural Performance Monitor Features (Intel) */
-- 
2.1.4