Fix for AMD erratum 383 on Family 10h CPUs
PATCH 1/1 Fix for AMD erratum 383 on Family 10h CPUs
This patches implements the workaround of AMD erratum 383 on family 10h CPUs. It destroys the guest VM when a MC error with a special pattern is detected. Without this patch, a guest VM failure can potentially crash Xen hypervisor and the whole system. The erratum will be published in next version of guide.
xen/arch/x86/hvm/svm/svm.c +60 -0
@@ -72,6 +72,9 @@/* vmcb used for extended host state */ static void *root_vmcb[NR_CPUS ]__read_mostly ; + +/* for AMD erratum 383 */ +static bool_t amd_erratum383_found__read_mostly ;static void inline __update_guest_eip (struct cpu_user_regs *regs,unsigned int inst_len)
@@ -828,6 +831,20 @@return 0 ; } +static void svm_init_erratum_383 (struct cpuinfo_x86 *c) +{ +uint64_t msr_content; + +/* only family 10h is affected */ +if ( c->x86 !=0x10 ) +return ; + +rdmsrl (MSR_AMD64_DC_CFG , msr_content); +wrmsrl (MSR_AMD64_DC_CFG , msr_content | (1ULL <<47 )); + + amd_erratum383_found =1 ; +} +static int svm_cpu_up (struct cpuinfo_x86 *c) {u32 eax, edx, phys_hsa_lo, phys_hsa_hi;
@@ -852,6 +869,9 @@ phys_hsa_lo = (u32 )phys_hsa; phys_hsa_hi = (u32)(phys_hsa >>32 );wrmsr (MSR_K8_VM_HSAVE_PA , phys_hsa_lo, phys_hsa_hi); + +/* check for erratum 383 */ +svm_init_erratum_383 (c);/* Initialize core's ASID handling. */ svm_asid_init (c);
@@ -1275,6 +1295,45 @@ } } +extern unsigned int nr_mce_banks;/* from mce.h */ + +static int svm_is_erratum_383 (struct cpu_user_regs *regs) +{ +uint64_t msr_content; +uint32_t i; +struct vcpu *v = current; + +if ( !amd_erratum383_found ) +return 0 ; + +rdmsrl (MSR_IA32_MC0_STATUS , msr_content); +/* Bit 62 may or may not be set for this mce */ + msr_content &= ~(1ULL <<62 ); + +if ( msr_content !=0xb600000000010015ULL ) +return 0 ; + +/* Clear MCi_STATUS registers */ +for (i =0 ; i < nr_mce_banks; i++) +wrmsrl (MSR_IA32_MCx_STATUS (i),0ULL ); + +rdmsrl (MSR_IA32_MCG_STATUS , msr_content); +wrmsrl (MSR_IA32_MCG_STATUS , msr_content & ~(1ULL <<2 )); + +/* flush TLB */ +flush_tlb_mask (&v->domain->domain_dirty_cpumask); + +return 1 ; +} + +static void svm_vmexit_mce_intercept (struct vcpu *v,struct cpu_user_regs *regs) +{ +if (svm_is_erratum_383 (regs) ) { +gdprintk (XENLOG_ERR ,"SVM hits AMD erratum 383\n" ); +domain_crash (v->domain); + } +} +static void wbinvd_ipi (void *info) {wbinvd ();
@@ -1480,6 +1539,7 @@/* Asynchronous event, handled when we STGI'd after the VMEXIT. */ case VMEXIT_EXCEPTION_MC :HVMTRACE_0D (MCE ); +svm_vmexit_mce_intercept (v, regs);break ;case VMEXIT_VINTR :
xen/include/asm-x86/msr-index.h +6 -0
@@ -146,6 +146,11 @@#define MSR_IA32_MC8_ADDR 0x00000422#define MSR_IA32_MC8_MISC 0x00000423 +#define MSR_IA32_MCx_CTL (x) (MSR_IA32_MC0_CTL + 4*(x)) +#define MSR_IA32_MCx_STATUS (x) (MSR_IA32_MC0_STATUS + 4*(x)) +#define MSR_IA32_MCx_ADDR (x) (MSR_IA32_MC0_ADDR + 4*(x)) +#define MSR_IA32_MCx_MISC (x) (MSR_IA32_MC0_MISC + 4*(x)) +#define MSR_P6_PERFCTR0 0x000000c1#define MSR_P6_PERFCTR1 0x000000c2#define MSR_P6_EVNTSEL0 0x00000186
@@ -225,6 +230,7 @@/* AMD64 MSRs */ #define MSR_AMD64_NB_CFG 0xc001001f +#define MSR_AMD64_DC_CFG 0xc0011022#define AMD64_NB_CFG_CF8_EXT_ENABLE_BIT 46/* AMD Family10h machine check MSRs */
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Quoted 6 lines
Quoted 1 line
Quoted 1 line
Wei Huang <[email protected]> 17.05.10 20:57 >>>
This patches implements the workaround of AMD erratum 383 on family 10h CPUs. It destroys the guest VM when a MC error with a special pattern is detected. Without this patch, a guest VM failure can potentially crash Xen hypervisor and the whole system. The erratum will be published in next version of guide.
While I realize that the patch has already been applied, I still have some reservations:
- It fiddles with bit 47 of MSR_AMD64_DC_CFG *and* marks that the erratum is present. Normally I would expect that MSR modifications are done to work around an erratum, but here it is completely unclear what the modification is good for.
- It clears all MCi_STATUS registers, despite (as I understand it) the problem only involving bank 0.
- It flushes the TLBs for the affected domain - shouldn't this be done globally, if it is necessary at all?
- Is it guaranteed that the #MC cannot happen when a #VMEXIT is already being processed by the CPU?
Jan