Skip to contentSkip to conversation

: Make 32bit legacy guests boot again

PATCH In review Christoph Egger 2 participants 7 messages Jun 26, 2009 – Jun 29, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 : Make 32bit legacy guests boot again

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb().
Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

Signature · 6 lines
---to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Thomas M. McCoy, Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632
xen/arch/x86/hvm/svm/svm.c +10 -4
diff -r 02003bee3e80 xen/arch/x86/hvm/svm/svm.c
@@ -111,10 +111,11 @@ static void svm_cpu_down(void)
     write_efer(read_efer() & ~EFER_SVME);
 }
 
-static enum handler_return long_mode_do_msr_write(struct cpu_user_regs *regs)
+static enum handler_return long_mode_do_msr_write(struct vcpu *v,
+                                                  struct cpu_user_regs *regs)
 {
-    u64 msr_content = (u32)regs->eax | ((u64)regs->edx << 32);
-    u32 ecx = regs->ecx;
+    uint64_t msr_content = (uint32_t)regs->eax | ((uint64_t)regs->edx << 32);
+    uint32_t ecx = regs->ecx;
 
     HVM_DBG_LOG(DBG_LEVEL_0, "msr %x msr_content %"PRIx64,
                 ecx, msr_content);
@@ -124,6 +125,11 @@ static enum handler_return long_mode_do_
     case MSR_EFER:
         if ( hvm_set_efer(msr_content) )
             return HNDL_exception_raised;
+        if (!( msr_content & EFER_LME ))
+            break;
+        svm_enable_intercept_for_msr(v, MSR_IA32_SYSENTER_CS);
+        svm_enable_intercept_for_msr(v, MSR_IA32_SYSENTER_ESP);
+        svm_enable_intercept_for_msr(v, MSR_IA32_SYSENTER_EIP);
         break;
 
     case MSR_IA32_MC4_MISC: /* Threshold register */
@@ -1139,7 +1145,7 @@ static int svm_msr_write_intercept(struc
         if ( wrmsr_viridian_regs(ecx, regs->eax, regs->edx) )
             break;
 
-        switch ( long_mode_do_msr_write(regs) )
+        switch ( long_mode_do_msr_write(v, regs) )
         {
         case HNDL_unhandled:
             wrmsr_hypervisor_regs(ecx, regs->eax, regs->edx);
xen/arch/x86/hvm/svm/vmcb.c +22 -7
diff -r 02003bee3e80 xen/arch/x86/hvm/svm/vmcb.c
@@ -78,29 +78,36 @@ struct host_save_area *alloc_host_save_a
     return hsa;
 }
 
-void svm_disable_intercept_for_msr(struct vcpu *v, u32 msr)
+void svm_intercept_msr(struct vcpu *v, uint32_t msr, int enable)
 {
     unsigned long *msr_bitmap = v->arch.hvm_svm.msrpm;
+    unsigned long *msr_bit = NULL;
 
     /*
      * See AMD64 Programmers Manual, Vol 2, Section 15.10 (MSR-Bitmap Address).
      */
     if ( msr <= 0x1fff )
     {
-        __clear_bit(msr*2, msr_bitmap + 0x000/BYTES_PER_LONG); 
-        __clear_bit(msr*2+1, msr_bitmap + 0x000/BYTES_PER_LONG); 
+        msr_bit = msr_bitmap + 0x0000 / BYTES_PER_LONG;
     }
     else if ( (msr >= 0xc0000000) && (msr <= 0xc0001fff) )
     {
         msr &= 0x1fff;
-        __clear_bit(msr*2, msr_bitmap + 0x800/BYTES_PER_LONG);
-        __clear_bit(msr*2+1, msr_bitmap + 0x800/BYTES_PER_LONG);
+        msr_bit = msr_bitmap + 0x0800 / BYTES_PER_LONG;
     } 
     else if ( (msr >= 0xc001000) && (msr <= 0xc0011fff) )
     {
         msr &= 0x1fff;
-        __clear_bit(msr*2, msr_bitmap + 0x1000/BYTES_PER_LONG);
-        __clear_bit(msr*2+1, msr_bitmap + 0x1000/BYTES_PER_LONG);
+        msr_bit = msr_bitmap + 0x1000 / BYTES_PER_LONG;
+    }
+
+    BUG_ON(msr_bit == NULL);
+    if (enable) {
+        __set_bit(msr * 2, msr_bit);
+        __set_bit(msr * 2 + 1, msr_bit);
+    } else {
+        __clear_bit(msr * 2, msr_bit);
+        __clear_bit(msr * 2 + 1, msr_bit);
     }
 }
 
@@ -151,6 +158,14 @@ static int construct_vmcb(struct vcpu *v
     svm_disable_intercept_for_msr(v, MSR_STAR);
     svm_disable_intercept_for_msr(v, MSR_SYSCALL_MASK);
 
+    /* Assume 32bit legacy guest. We intercept MSR_EFER and when guest enables
+     * longmode, we enable intercept for the SYSENTER MSRs below
+     * (needed for sysenter/sysexit emulation).
+     */
+    svm_disable_intercept_for_msr(v, MSR_IA32_SYSENTER_CS);
+    svm_disable_intercept_for_msr(v, MSR_IA32_SYSENTER_ESP);
+    svm_disable_intercept_for_msr(v, MSR_IA32_SYSENTER_EIP);
+
     vmcb->msrpm_base_pa = (u64)virt_to_maddr(arch_svm->msrpm);
     vmcb->iopm_base_pa  = (u64)virt_to_maddr(hvm_io_bitmap);
 
xen/include/asm-x86/hvm/svm/vmcb.h +3 -1
diff -r 02003bee3e80 xen/include/asm-x86/hvm/svm/vmcb.h
@@ -481,7 +481,9 @@ void svm_destroy_vmcb(struct vcpu *v);
 
 void setup_vmcb_dump(void);
 
-void svm_disable_intercept_for_msr(struct vcpu *v, u32 msr);
+#define svm_disable_intercept_for_msr(v, msr) svm_intercept_msr((v), (msr), 0)
+#define svm_enable_intercept_for_msr(v, msr) svm_intercept_msr((v), (msr), 1)
+void svm_intercept_msr(struct vcpu *v, uint32_t msr, int enable);
 
 #endif /* ASM_X86_HVM_SVM_VMCS_H__ */
 

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical issues:
* Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy mode).

-- Keir

Quoted 20 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb().
Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

#2 lore raw

Tnx. I will check as soon as it appears in the public staging tree.

Christoph

Quoted 26 lines · On Saturday 27 June 2009 11:35:20 Keir Fraser wrote:

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical issues:
* Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy mode).

-- Keir

Quoted 18 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb(). Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

Signature · 6 lines
---to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Thomas M. McCoy, Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632
#3 lore raw

I forgot to push. Should be there now.

-- Keir

Quoted 35 lines · Christoph Egger

Tnx. I will check as soon as it appears in the public staging tree.

Christoph

Quoted 26 lines · On Saturday 27 June 2009 11:35:20 Keir Fraser wrote:

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical issues:
* Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy mode).

-- Keir

Quoted 18 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb(). Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

#4 lore raw

Your changes broke the boot of 64bit guests. I see a triple fault now.

Christoph

Quoted 36 lines · On Monday 29 June 2009 11:17:39 Keir Fraser wrote:

I forgot to push. Should be there now.

-- Keir

Quoted 31 lines · Christoph Egger

Tnx. I will check as soon as it appears in the public staging tree.

Christoph

Quoted 26 lines · On Saturday 27 June 2009 11:35:20 Keir Fraser wrote:

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical
issues: * Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy mode).

-- Keir

Quoted 18 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb(). Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

Signature · 6 lines
---to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Thomas M. McCoy, Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632
#5 lore raw

I'm flying blind since I don't have an AMD box to hand to test on, but I think changeset 19869 will probably fix this.

-- Keir

Quoted 45 lines · Christoph Egger

Your changes broke the boot of 64bit guests. I see a triple fault now.

Christoph

Quoted 36 lines · On Monday 29 June 2009 11:17:39 Keir Fraser wrote:

I forgot to push. Should be there now.

-- Keir

Quoted 31 lines · Christoph Egger

Tnx. I will check as soon as it appears in the public staging tree.

Christoph

Quoted 26 lines · On Saturday 27 June 2009 11:35:20 Keir Fraser wrote:

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical
issues: * Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy mode).

-- Keir

Quoted 18 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb(). Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

#6 lore raw

Confirmed. Now both 32bit and 64bit guests boot.

Christoph

Quoted 48 lines · On Monday 29 June 2009 16:51:19 Keir Fraser wrote:

I'm flying blind since I don't have an AMD box to hand to test on, but I think changeset 19869 will probably fix this.

-- Keir

Quoted 42 lines · Christoph Egger

Your changes broke the boot of 64bit guests. I see a triple fault now.

Christoph

Quoted 37 lines · On Monday 29 June 2009 11:17:39 Keir Fraser wrote:

I forgot to push. Should be there now.

-- Keir

Quoted 32 lines · Christoph Egger

Tnx. I will check as soon as it appears in the public staging tree.

Christoph

Quoted 27 lines · On Saturday 27 June 2009 11:35:20 Keir Fraser wrote:

Changeset 19856 is a cleaned up and streamlined version of this patch. Please take a look. It also fixes a couple of largely theoretical
issues: * Should depend on EFER.LMA not EFER.LME
* Should handle the LMA 1->0 transition (i.e., return to legacy
mode).

-- Keir

Quoted 18 lines · Christoph Egger

Hi!

Attached patch fixes a bug introduced in c/s 19648.

32bit legacy guests have the sysenter/sysexit instructions available. Therefore, we have to disable intercepts for the sysenter MSRs or the guest stucks in an infinite loop of #GPs, otherwise.

For guests in 64bit mode and 32bit compat mode, sysenter/sysexit instructions aren't available. The sysenter MSRs have to be intercepted to make the instruction emulation working.

Attach patch first assumes the guest is in 32bit legacy mode and therefore disables the sysenter MSRs in construct_vmcb(). Access to the MSR_EFER is intercepted. When the guest enables longmode, then enable interception of the sysenter MSRs.

Signature · 6 lines
---to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Thomas M. McCoy, Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632