Message ID | 20200320131707.GF5122@8bytes.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [RFC,v2.1] x86/sev-es: Handle NMI State | expand |
On 3/20/20 6:17 AM, Joerg Roedel wrote: > On Thu, Mar 19, 2020 at 08:35:59AM -0700, Andy Lutomirski wrote: >> 1. Just put the NMI unmask in do_nmi(). The kernel *already* knows >> how to handle running do_nmi() with NMIs unmasked. This is much, much >> simpler than your code. > Okay, attached is the updated patch which implements this approach. I > tested it in an SEV-ES guest with 'perf top' running for a little more > than 30 minutes and all looked good. I also removed the dead code from > the patch. FWIW, perf plus the x86 selftests run in a big loop was my best way of stressing the NMI path when we mucked with it for PTI. The selftests make sure to hit some of the more rare entry/exit paths.
On Fri, Mar 20, 2020 at 07:42:09AM -0700, Dave Hansen wrote: > FWIW, perf plus the x86 selftests run in a big loop was my best way of > stressing the NMI path when we mucked with it for PTI. The selftests > make sure to hit some of the more rare entry/exit paths. Yeah, I ran the x86 selftests in an SEV-ES guest on-top of these patches, that works. But doing this together with 'perf top' is also on the list of tests to do. Regards, Joerg
diff --git a/arch/x86/include/asm/sev-es.h b/arch/x86/include/asm/sev-es.h index 63acf50e6280..441ec1ba2cc7 100644 --- a/arch/x86/include/asm/sev-es.h +++ b/arch/x86/include/asm/sev-es.h @@ -82,11 +82,13 @@ struct real_mode_header; #ifdef CONFIG_AMD_MEM_ENCRYPT int sev_es_setup_ap_jump_table(struct real_mode_header *rmh); +void sev_es_nmi_complete(void); #else /* CONFIG_AMD_MEM_ENCRYPT */ static inline int sev_es_setup_ap_jump_table(struct real_mode_header *rmh) { return 0; } +static inline void sev_es_nmi_complete(void) { } #endif /* CONFIG_AMD_MEM_ENCRYPT*/ #endif diff --git a/arch/x86/include/uapi/asm/svm.h b/arch/x86/include/uapi/asm/svm.h index 20a05839dd9a..0f837339db66 100644 --- a/arch/x86/include/uapi/asm/svm.h +++ b/arch/x86/include/uapi/asm/svm.h @@ -84,6 +84,7 @@ /* SEV-ES software-defined VMGEXIT events */ #define SVM_VMGEXIT_MMIO_READ 0x80000001 #define SVM_VMGEXIT_MMIO_WRITE 0x80000002 +#define SVM_VMGEXIT_NMI_COMPLETE 0x80000003 #define SVM_VMGEXIT_AP_HLT_LOOP 0x80000004 #define SVM_VMGEXIT_AP_JUMP_TABLE 0x80000005 #define SVM_VMGEXIT_SET_AP_JUMP_TABLE 0 diff --git a/arch/x86/kernel/nmi.c b/arch/x86/kernel/nmi.c index 54c21d6abd5a..fc872a7e0ed1 100644 --- a/arch/x86/kernel/nmi.c +++ b/arch/x86/kernel/nmi.c @@ -37,6 +37,7 @@ #include <asm/reboot.h> #include <asm/cache.h> #include <asm/nospec-branch.h> +#include <asm/sev-es.h> #define CREATE_TRACE_POINTS #include <trace/events/nmi.h> @@ -510,6 +511,13 @@ NOKPROBE_SYMBOL(is_debug_stack); dotraplinkage notrace void do_nmi(struct pt_regs *regs, long error_code) { + /* + * Re-enable NMIs right here when running as an SEV-ES guest. This might + * cause nested NMIs, but those can be handled safely. + */ + if (sev_es_active()) + sev_es_nmi_complete(); + if (IS_ENABLED(CONFIG_SMP) && cpu_is_offline(smp_processor_id())) return; diff --git a/arch/x86/kernel/sev-es.c b/arch/x86/kernel/sev-es.c index 3c22f256645e..a7e2739771e7 100644 --- a/arch/x86/kernel/sev-es.c +++ b/arch/x86/kernel/sev-es.c @@ -270,6 +270,24 @@ static phys_addr_t vc_slow_virt_to_phys(struct ghcb *ghcb, long vaddr) /* Include code shared with pre-decompression boot stage */ #include "sev-es-shared.c" +void sev_es_nmi_complete(void) +{ + struct ghcb_state state; + struct ghcb *ghcb; + + ghcb = sev_es_get_ghcb(&state); + + vc_ghcb_invalidate(ghcb); + ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_NMI_COMPLETE); + ghcb_set_sw_exit_info_1(ghcb, 0); + ghcb_set_sw_exit_info_2(ghcb, 0); + + sev_es_wr_ghcb_msr(__pa(ghcb)); + VMGEXIT(); + + sev_es_put_ghcb(&state); +} + static u64 sev_es_get_jump_table_addr(void) { struct ghcb_state state;