x86/tdx: Fix arch_safe_halt() execution for TDX VMs

Direct HLT instruction execution causes #VEs for TDX VMs which is routed
to hypervisor via TDCALL. If HLT is executed in STI-shadow, resulting #VE
handler will enable interrupts before TDCALL is routed to hypervisor
leading to missed wakeup events, as current TDX spec doesn't expose
interruptibility state information to allow #VE handler to selectively
enable interrupts.

Commit bfe6ed0c6727 ("x86/tdx: Add HLT support for TDX guests")
prevented the idle routines from executing HLT instruction in STI-shadow.
But it missed the paravirt routine which can be reached via this path
as an example:

	kvm_wait()       =>
        safe_halt()      =>
        raw_safe_halt()  =>
        arch_safe_halt() =>
        irq.safe_halt()  =>
        pv_native_safe_halt()

To reliably handle arch_safe_halt() for TDX VMs, introduce explicit
dependency on CONFIG_PARAVIRT and override paravirt halt()/safe_halt()
routines with TDX-safe versions that execute direct TDCALL and needed
interrupt flag updates. Executing direct TDCALL brings in additional
benefit of avoiding HLT related #VEs altogether.

As tested by Ryan Afranji:

  "Tested with the specjbb2015 benchmark. It has heavy lock contention which leads
   to many halt calls. TDX VMs suffered a poor score before this patchset.

   Verified the major performance improvement with this patchset applied."

Fixes: bfe6ed0c6727 ("x86/tdx: Add HLT support for TDX guests")
Signed-off-by: Vishal Annapurve <vannapurve@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Tested-by: Ryan Afranji <afranji@google.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20250228014416.3925664-3-vannapurve@google.com

4 files changed, 29 insertions, 4 deletions

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 05b4eca156cf..f614c0522a0b 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -878,6 +878,7 @@ config INTEL_TDX_GUEST
 	depends on X86_64 && CPU_SUP_INTEL
 	depends on X86_X2APIC
 	depends on EFI_STUB
+	depends on PARAVIRT
 	select ARCH_HAS_CC_PLATFORM
 	select X86_MEM_ENCRYPT
 	select X86_MCE
diff --git a/arch/x86/coco/tdx/tdx.c b/arch/x86/coco/tdx/tdx.c
index 7772b01ab738..aa0eb4057226 100644
--- a/arch/x86/coco/tdx/tdx.c
+++ b/arch/x86/coco/tdx/tdx.c
@@ -14,6 +14,7 @@
 #include <asm/ia32.h>
 #include <asm/insn.h>
 #include <asm/insn-eval.h>
+#include <asm/paravirt_types.h>
 #include <asm/pgtable.h>
 #include <asm/set_memory.h>
 #include <asm/traps.h>
@@ -398,7 +399,7 @@ static int handle_halt(struct ve_info *ve)
 	return ve_instr_len(ve);
 }
 
-void __cpuidle tdx_safe_halt(void)
+void __cpuidle tdx_halt(void)
 {
 	const bool irq_disabled = false;
 
@@ -409,6 +410,16 @@ void __cpuidle tdx_safe_halt(void)
 		WARN_ONCE(1, "HLT instruction emulation failed\n");
 }
 
+static void __cpuidle tdx_safe_halt(void)
+{
+	tdx_halt();
+	/*
+	 * "__cpuidle" section doesn't support instrumentation, so stick
+	 * with raw_* variant that avoids tracing hooks.
+	 */
+	raw_local_irq_enable();
+}
+
 static int read_msr(struct pt_regs *regs, struct ve_info *ve)
 {
 	struct tdx_module_args args = {
@@ -1110,6 +1121,19 @@ void __init tdx_early_init(void)
 	x86_platform.guest.enc_kexec_finish	     = tdx_kexec_finish;
 
 	/*
+	 * Avoid "sti;hlt" execution in TDX guests as HLT induces a #VE that
+	 * will enable interrupts before HLT TDCALL invocation if executed
+	 * in STI-shadow, possibly resulting in missed wakeup events.
+	 *
+	 * Modify all possible HLT execution paths to use TDX specific routines
+	 * that directly execute TDCALL and toggle the interrupt state as
+	 * needed after TDCALL completion. This also reduces HLT related #VEs
+	 * in addition to having a reliable halt logic execution.
+	 */
+	pv_ops.irq.safe_halt = tdx_safe_halt;
+	pv_ops.irq.halt = tdx_halt;
+
+	/*
 	 * TDX intercepts the RDMSR to read the X2APIC ID in the parallel
 	 * bringup low level code. That raises #VE which cannot be handled
 	 * there.
diff --git a/arch/x86/include/asm/tdx.h b/arch/x86/include/asm/tdx.h
index 65394aa9b49f..4a1922ec80cf 100644
--- a/arch/x86/include/asm/tdx.h
+++ b/arch/x86/include/asm/tdx.h
@@ -58,7 +58,7 @@ void tdx_get_ve_info(struct ve_info *ve);
 
 bool tdx_handle_virt_exception(struct pt_regs *regs, struct ve_info *ve);
 
-void tdx_safe_halt(void);
+void tdx_halt(void);
 
 bool tdx_early_handle_ve(struct pt_regs *regs);
 
@@ -72,7 +72,7 @@ void __init tdx_dump_td_ctls(u64 td_ctls);
 #else
 
 static inline void tdx_early_init(void) { };
-static inline void tdx_safe_halt(void) { };
+static inline void tdx_halt(void) { };
 
 static inline bool tdx_early_handle_ve(struct pt_regs *regs) { return false; }
 
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 91f6ff618852..962c3ce39323 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -939,7 +939,7 @@ void __init select_idle_routine(void)
 		static_call_update(x86_idle, mwait_idle);
 	} else if (cpu_feature_enabled(X86_FEATURE_TDX_GUEST)) {
 		pr_info("using TDX aware idle routine\n");
-		static_call_update(x86_idle, tdx_safe_halt);
+		static_call_update(x86_idle, tdx_halt);
 	} else {
 		static_call_update(x86_idle, default_idle);
 	}